Citation
Remote sensing applications for environmental management

Material Information

Title:
Remote sensing applications for environmental management
Creator:
Hsu, Chengmin
Publication Date:
Language:
English
Physical Description:
vii, 143 leaves : illustrations ; 28 cm

Subjects

Subjects / Keywords:
Environmental management -- Remote sensing ( lcsh )
Environmental management -- Remote sensing ( fast )
Genre:
bibliography ( marcgt )
theses ( marcgt )
non-fiction ( marcgt )

Notes

Bibliography:
Includes bibliographical references (leaves 133-143).
General Note:
Department of Civil Engineering
Statement of Responsibility:
by Chengmin Hsu.

Record Information

Source Institution:
|University of Colorado Denver
Holding Location:
|Auraria Library
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
672300479 ( OCLC )
ocn672300479
Classification:
LD1193.E53 2010d H88 ( lcc )

Full Text
REMOTE SENSING APPLICATIONS FOR ENVIRONMENTAL MANAGEMENT
by
Chengmin Hsu
Bachelor of Architecture, Chung-Yuan Christian University, 1986
Master of Architecture, University of Colorado Denver, 2001
Master of Engineering, University of Colorado Denver, 2006
A thesis submitted to the
University of Colorado Denver
in partial fulfillment
of the requirements for the degree of
Doctor of Philosophy
Civil Engineering Department
2010


This thesis for the Doctor of Philosophy
degree by
Chengmin Hsu
has been approved
Silj_ I Zo/o
Date


Hsu, Chengmin (Ph.D., Civil Engineering)
Remote Sensing Applications for Environmental Management
Thesis directed by Professor Lynn E. Johnson
ABSTRACT
After decades of development of sensor technology and analytical theories, remote
sensing technologies have now become a very powerful tool in many disciplines, ranging
from reconnaissance to agriculture. Remote sensing is providing us an unprecedented
platform with which to better understand our Earth and more wisely manage resources, by
capturing data in every comer of the world with resolutions ranging from centimeters to
hundreds of kilometers. However, there are still many deficiencies that exist in remote
sensing technology. Without fully understanding the extent of the capabilities of remote
sensing, people will never truly benefit from the tremendous power of this new technology
to meet their specific needs.
Four research topics are used to illustrate the optimum uses of remotely sensed data and
to uncover potential pitfalls in their application. These four topics are:
1. Multi-Criteria Wetlands Mapping Using an Integrated Pixel-Based and Object-Based
Classification Approach
2. Route Formation and Land Disturbance: Applications of an Object-Based Road
Detection Algorithm
3. Evaluating Route Formation and Disturbance of Off-Highway Travel Using Logistic
Regression Analysis
4. Downscaling of Advanced Microwave Scanning Radiometer (AMSR-E) Soil
Moisture Using a Thermal Sensor and the Physically-Based Models
When applied in sophisticated situations, remote sensing alone cannot simulate natural
phenomena and resolve problems. It needs to be incorporated with ancillary data and
simulation techniques. By incorporating the applications of image preprocessing, GIS
theories, knowledge-based fuzzy logic, statistical theories, and physical models, these four
topics explore the tremendous potential of remote sensing when integrated with models
and theories from other disciplines.
This abstract accurately represents the content of the candidates thesis. I recommend its
publication.


DEDICATION
This thesis is dedicated to Dengdi and Shung.


ACKNOWLEDGMENTS
I would like to thank Dr. Lynn E. Johnson, my graduate thesis advisor, for recognizing my
ability and giving me flexibility to explore the applications of remote sensing. Thanks are
also due to Dr. Johnson for his understanding and acceptance of my various
circumstances throughout my journey in the PhD program. His help in many different areas
facilitates my persistence toward reaching the end of the line. I would also like to thank
Drs. Rajagopalan Balaji, Jacek Grodecki, Brian Muller, and Robert Zamora for serving on
my committee and providing constructive criticism.
I would like to thank Roland Wostl, the ex-director of the CDOT environmental department,
for providing an opportunity for me to explore the topic of wetland mapping. His remarks
from institutions point of view of using data helped the wetland mapping study result in
more useful data generation. Also, I thank Rebecca Pierce, CDOT wetland program
manager, for helping to identify wetlands in the field. I would also like to thank Dr. Muller for
guidance on ATV disturbance studies as well as Dr. Zamora and Dr. Timothy Schneider for
advice and support on the soil moisture downscaling project.
Financial support for these studies was provided by CDOT, BLM, College of Architecture
and Planning at University of Colorado Denver, Colorado Water Institute at Colorado State
University, and NOAA. The College of Engineering and Applied Science at University of
Colorado Denver granted scholarship money. This financial support is gratefully
acknowledged.
I am most grateful to my family, for their enduring patience and support.


TABLE OF CONTENTS
Figures.............................................................................iv
Tables.............................................................................vii
Chapter
1. Introduction ..................................................................1
1.1 Challenges Faced by Humans.....................................................1
1.2 Remote Sensing for Environmental Management....................................3
1.3 Study Outline..................................................................4
1.4 Research Topics and Approaches.................................................6
1.4.1 Multi-Criteria Wetlands Mapping Using an Integrated
Pixel-Based and Object-Based Classification Approach ..........................6
1.4.2 Recreational Travel, Route Formation and Disturbance:
Applications of an Object-Based Road Detection Algorithm......................7
1.4.3 Evaluating Route Formation and Disturbance of Off-Highway
Travel Using Logistic Regression Method .......................................8
1.4.4 Downscaling of Advanced Microwave Scanning Radiometer
(AMSR-E) Soil Moisture Using a Thermal Sensor and the
Physically-Based Models........................................................9
2. Multi-Criteria Wetlands Mapping Using an Integrated
Pixel-Based and Object-Based Classification Approach..........................11
2.1 The Necessity of Automation of Wetland Mapping................................11
2.1.1 The Importance of Wetland Information........................................11
2.1.2 Regulatory Background........................................................11
2.1.3 Objectives ..................................................................12
2.2 Wetland Definition and Parameters.............................................13
2.2.1 Parameters of Wetland Delineation of USACE...................................13
2.2.1.1 Vegetation.................................................................13
2.2.1.2 Soil ......................................................................14
2.2.1.3 Hydrology..................................................................14
2.2.2 Wetland Types Adopted for Classification.....................................14
2.3 Methods ...................................................................18
2.3.1 Study Area and Image Acquisition.............................................18
2.3.2 Field Survey ................................................................20
2.3.3 Data Preparation.............................................................21
2.3.3.1 Geo-referencing............................................................22
2.3.3.2 Gram-Schmidt Spectral Sharpening...........................................22
2.3.3.3 Transferring Digital Number to Reflectance.................................24
2.3.4 Vegetation Indices...........................................................28
2.3.4.1 Kauth-Thomas Tasseled Cap Transformation...................................28
2.3.4.2 NDVI (Normalized Difference Vegetation Index)..............................30
2.3.5 Hydrological Analysis........................................................32
2.3.6 Unsupervised Pixel-Based Classification......................................35
i


2.3.6.1 Minimum Noise Fraction Transformation..................................35
2.3.6.2 ISODATA Classification...................................................36
2.3.7 Object-Based Classification................................................39
2.3.7.1 Project Creation.........................................................39
2.3.7.2 Create Image Object......................................................40
2.3.7.3 Classification...........................................................42
2.4 Results and Discussion......................................................44
2.4.1 Areas of the Wetlands Identified............................................44
2.4.2 KAPPA Analysis.............................................................44
2.4.3 Findings and Conclusions...................................................45
2.5 Future Efforts..............................................................46
3. Recreational Travel, Route Formation and Land
Disturbance: Applications of an Object-Based Road
Detection Algorithm.........................................................48
3.1 Complexity of Recreational Travel...........................................48
3.1.1 Routes. The Source of Unmanaged Recreation..................................48
3.1.2 Evolution of Automatic Road Detection Method...............................49
3.1.3 Research Questions.........................................................51
3.2 Study Area .................................................................52
3.3 Data Used and Data Preprocessing............................................54
3.3.1 Data Used ..................................................................54
3.3.2 Image Preprocessing........................................................55
3.4 Methodology ............................................................... 57
3.4.1 Segmentation Method.........................................................57
3.4.1.1 Level 5 Segmentation.....................................................58
3.4.1.2 Level 4 Segmentation.....................................................59
3.4.1.3 Third to First Level Segmentation........................................60
3.4.2 Multi-Level Hierarchical Classification....................................61
3.4.3 Feature Extraction.........................................................64
3.4.4 Classification Rules.......................................................67
3.5 Results ....................................................................68
3.5.1 Validation .................................................................68
3.5.2 Sign of ATV Activity.......................................................69
3.5.3 Accuracy Assessment........................................................73
3.6 Conclusion and Discussion...................................................73
3.6.1 Conclusions ................................................................73
3.6.2 Recommendations for Future.................................................75
4. Evaluating Route Formation and Disturbance of
Off-Highway Travel Using Logistic Regression Method.........................77
4.1 The Phenomenon of OHVs......................................................77
4.2 Background .................................................................78
4.3 Approach ...................................................................80
4.3.1 Generalized Linear Model....................................................80
4.3.2 Study Site and Data........................................................81
4.4 Hypothesis .................................................................88
4.4.1 Radiation ..................................................................88
4.4.2 Geomorphology..............................................................89
4.4.3 Route Configuration........................................................89
4.4.4 Level of OHV Use...........................................................90
ii


4.5 Logistic Regression.........................................................90
4.6 Results ....................................................................96
4.6.1 Radiation Variables.........................................................96
4.6.2 Geomorphology Variables....................................................97
4.6.3 Route Configuration Variables..............................................97
4.6.4 Level of OHV Use Variable..................................................98
4.6.5 Modeling Outcome...........................................................99
4.7 Conclusions ...............................................................100
5. Downscaling of Advanced Microwave Scanning
Radiometer (AMSR-E) Soil Moisture Using a Thermal
Sensor and the Physically-Based Models.....................................103
5.1 Background and Research Objectives.........................................103
5.1.1 The Importance of High Resolution Soil Moisture............................103
5.1.2 The Quality of AMSR-E Soil Moisture.......................................105
5.1.3 Methods Used for Multi-Resolution Soil Moisture Retrieval.................107
5.1.4 Multiphase High Resolution Soil Moisture Estimation.......................108
5.2 Study Site and Data........................................................108
5.2.1 Study Site ................................................................108
5.2.2 Data 110
5.2.3 MODIS Data ...............................................................110
5.2.4 LANDS AT Data.............................................................110
5.2.5 Wind Speed ...............................................................111
5.2.6 Precipitation ............................................................111
5.2.7 Digital Elevation Model...................................................111
5.3 Approach ..................................................................112
5.3.1 MODIS-Derived Soil Evaporative Efficiency..................................113
5.3.2 Infiltration Estimation...................................................113
5.3.3 Geomorphologic Adjustment.................................................116
5.4 Results ...................................................................117
5.4.1 Validation ................................................................118
5.5 Discussion ................................................................119
5.6 Summary ...................................................................120
6. Epilogue ..................................................................123
6.1 Integration of Numerical Models, Statistical Analysis, and
the Remote Sensing Approach................................................123
6.2 The Semantic Expression of the Earth.......................................125
6.3 Planned Mission of Earth Observation Sensors...............................125
Appendix
A. Logistic Regression Analysis R Source Code....................................128
Bibliography .............................................................133


FIGURES
Figure
1.1 Schematic Diagram of Applying Remote Sensing and GIS for Environmental
Management..................................................................5
2.1 Aquatic Bed................................................................15
2.2 Marshes Commonly Seen Species Growing in Shallow Water in Colorado.......16
2.3 Wet Meadows in Fort Collins................................................17
2.4 Scrub/Shrub along the Cache La Poudre River................................17
2.5 Seasonal Floodplain Forest in Winter with Grass Cover Underneath Trees.....18
2.6 The Study Area with a Trapezoidal Boundary.................................20
2.7 GPS Data Collection of the Wetland Samples.................................21
2.8 ETM+ Imagery after Pan-sharpening..........................................23
2.9 Raw ETM+ Imagery (30-m Resolution).........................................24
2.10 Band 4 Reflectance of the LANDSAT ETM+ 04/16/2003 Imagery Covering Fort
Collins and Adjacencies....................................................26
2.11 Generated At-Satellite Surface Temperature from ETM+ 04/16/2003 Data.......27
2.12 Correspondence between the Greenness Layer (Right) and the Existence of
Wetland Locations (Left)...................................................29
2.13 Wetness (Right) Correspondence to the Wetlands and Ponds Displayed on the
NAIP Imagery (Left)........................................................29
2.14 Correspondence between the Brightness Layer (Right) and the Existence of
Wetland Locations (Left)...................................................29
2.15 Fluctuation of Reflectance of Different Materials Across Wavelengths from
0.4 pm to 1.1pm............................................................30
2.16 NDVI Generated from ASTER 08/13/2003 Imagery...............................31
2.17 NDVI Generated from E01 ALI October Imagery with Crops in the Irrigated
Farms Still Shown a High NDVI Value........................................31
2.18 Photo at the Same Location as in Figure 2.16 and 2.17......................32
2.19 Flow Direction Number Assignment...........................................32
2.20 Flow Direction (Left) to Flow Accumulation (Right).........................33
2.21 Generated Stream Network...................................................34
2.22 The Stream Passing through Downtown Fort Collins...........................34
2.23 Detected Wetlands along the Stream Comparison with the Figure 2.22.......35
2.24 MNF Band 1 of E01 ALI Imagery..............................................36
2.25 ISODATA Classification Result of the E01 ALI 10/26/2001 Imagery............37
2.26 ISODATA Classification Result of the LANDSAT 7 04/16/2003 Imagery..........37
2.27 ISODATA Classification of ETM+ 06/16/2002 Imagery with 3*3 Majority
Analysis ..................................................................38
2.28 Weight Overlay of ISODATA Classifications..................................38
2.29 Level 1 Image Objects......................................................40
2.30 Level 2 Image object segmentation..........................................41
2.31 Level 1 Classification Result..............................................42
2.32 Comparison of the Classification of a Marsh and its Presence in NAIP.......43
IV


2.33 Photograph of the Real World Marsh at the Same Location Shown in
Figure 2.32 ...................................................................44
3.1 The Study Site Located at the Edge of Las Cruces City in New Mexico............53
3.2 Magnified Study Area with LANDSAT GeoCover 2000 Imagery as Background ...54
3.3 Eigenvalue Spectrum of the Principal Components...............................56
3.4 Comparison of the Original Image (Right) with the Result after Executing Local
Sigma Filter on the First Component of the RGB 2005 Data (Left)...............57
3.5 A Portion of the Segmented Image Objects at Level 2...........................58
3.6 Image Segmentation Result in Level 5..........................................59
3.7 Image Segmentation Result in Level 4...........................................60
3.8 Image Segmentation Result of Level 1...........................................61
3.9 Diagram of the Flierarchical Classification....................................62
3.10 Ten Strategically Selected Validation Sites...................................69
3.11 Ruts 70
3.12 Major Routes...................................................................70
3.13 Uprooted Vegetation............................................................71
3.14 Connective ....................................................................71
3.15 Straight Vegetation Edges......................................................72
3.16 GPS Validation Points Collected in 2008 with CIR 2005 Photo as Background.....72
4.1 The Study Site in Las Cruces, New Mexico.......................................81
4.2 The Study Site (Hatched Purple Rectangle Area) on CIR Imagery..................82
4.3 Landscape Overlaid with 5m Cells Aggregation of the Attributes of 1 -m
Cells into 5-m Grids...........................................................83
4.4 The Scatter Plot of FPR/TPR with the Discriminating Cut-off Point of OHV 2005
Occurrence Probability Set at 0.02............................................92
4.5 The Scatter Plot of FPR/TPR with the Discriminating Cut-off Point of OHV 2005
Occurrence Probability Set at 0.05............................................92
4.6 The Scatter Plot of FPR/TPR with the Discriminating Cut-off Point of OHV 2005
Occurrence Probability Set at 0.09............................................93
4.7 The Scatter Plot of FPR/TPR with the Discriminating Cut-off Point of OHV 2005
Occurrence Probability Set at 0.14............................................93
4.8 The Scatter Plot of FPR/TPR with the Discriminating Cut-off Point of OHV 2005
Occurrence Probability Set at 0.2.............................................93
4.9 The Scatter Plot of FPR/TPR with the Discriminating Cut-off Point of OHV 2005
Occurrence Probability Set at 0.27............................................93
4.10 The Scatter Plot of FPR/TPR with the Discriminating Cut-off Point of OHV 2005
Occurrence Probability Set at 0.36............................................94
4.11 The Scatter Plot of FPR/TPR with the Discriminating Cut-off Point of OHV 2005
Occurrence Probability Set at 0.48............................................94
4.12 The Scatter Plot of FPR/TPR with the Discriminating Cut-off Point of OHV 2005
Occurrence Probability Set at 0.5.............................................94
4.13 The Scatter Plot of FPR/TPR with the Discriminating Cut-off Point of OHV 2005
Occurrence Probability Set at 0.57............................................94
4.14 The Scatter Plot of FPR/TPR with the Discriminating Cut-off Point of OHV 2005
Occurrence Probability Set at 0.75............................................95
4.15 The Scatter Plot of FPR/TPR with the Discriminating Cut-off Point of OHV 2005
Occurrence Probability Set at 0.92............................................95
4.16 The Box Plot of the 84 Root Mean Squared Errors (RMSEs) 0.34 of the Overall
RMSE ..........................................................................95
v


4.17 FPR-TPR Scatter Plot........................................................99
4.18 Predicted and Observed Routes of 2005 .....................................100
5.1 Three NOAA Soil Moisture Network Stations in the Babocomari River
Watershed Located within the AMSR-E Grid Canelo, Elgin, and
Freeman Spring Stations....................................................106
5.2 The Time-Series Comparison of the Average Soil Moisture of the
Three NOAA HMT Soil Moisture Stations with the Soil Moisture Values of the
AMSR-E Cell Which Contains the Three Stations from July 1, 2008 to
August 20, 2008.............................................................106
5.3 Babocomari River and Walnut Gulch Watershed with LANDSAT TM 5
06/22/2008 Imagery as Background............................................109
5.4 Fractional Vegetation Cover Calculated from MODIS 16-Day EVI Product of
the Period from 07/11/2008 to 07/26/2008....................................110
5.5 Ten Meter Resolution DEM with Hillshade Format.............................111
5.6 Three Stages of Downscaling.................................................112
5.7 Comparison of Precipitation Fluctuation and Soil Moisture Variation at the
Whetstone Station ..........................................................114
5.8 K Estimates of the Study Site for Calculating Infiltration Accumulation at the
1-km Resolution.............................................................116
5.9 Accumulated Infiltration on July 5, 2008....................................116
5.10 The 2-km Aggregation of the Natural Logarithm of the Flow Accumulation of
the Study Site..............................................................117
5.11 Derived 1-km Resolution Soil Moisture of 07/25/2008........................118
5.12 Scatter Plot of Observation against Estimation............................. 119
VI


TABLES
Table
2.1 ETM+ solar spectral irradiances...................................................25
2.2 Earth-Sun distance in astronomical units..........................................26
2.3 ETM+ thermal band calibration constants...........................................27
2.4 Wetlands area distribution according to its type..................................44
2.5 Error matrix .....................................................................45
3.1 Spectral vegetation indices generated for testing.................................56
3.2 Selected features in object-based classification..................................64
3.3 Rules used for the Roads_01 class...............................................67
3.4 Error matrix .....................................................................73
4.1 Variable description..............................................................84
4.2 Z value of the involved variables under the various model compositions............91
4.3 A logistic model without NDVI.....................................................96
6.1 Planned mission by NASA in the coming decade.....................................125
vii


1. Introduction
1.1 Challenges Faced by Humans
Humankind has placed the natural resources of the earth in severe jeopardy while pursuing
prosperity and convenience. In view of soaring oil prices and demand for water occurring
around the world, the rate of degradation and depletion of resources has become the
biggest challenge that we need to face in this generation and the future. Deforestation,
desertification, soil erosion, water pollution, and salinity problems have degraded the
environment, threatening the food security and economic development of most countries.
Although the United States has an abundance of natural resources, is a developed
country, and is considered as one of the greatest biodiversity pools genetically,
economically and ecologically, it is one of the countries that occupies a very high
percentage of the carbon emissions of the world. According to data from the United
Nations, carbon dioxide emission from United States occupies 20.2% of the global total
and is the second in the world in that respect, just behind China.
In the western U.S., the competition for limited water resources from various sectors is
always a severe problem. The situation has been made even worse given ineffective
decision-making systems used in all of the levels of municipal systems, such as
inappropriate cropping systems, imprudent urban development policies, reductions in crop
diversity, and the losing of important habitats. Also, the U.S. has already experienced a
wave of deforestation. According to the statistics of Global Deforestation of the University
of Michigan, 90% of virgin forests that once covered much of the lower 48 states in 1600
have now been cleared away. In the Pacific Northwest, about 80% of this forestland is
slated for logging. Today, forests are continuing to be cleared, degraded and fragmented
by timber harvesting, agricultural conversions, recreational activities, road-building, man-
made fires, and myriad other ways. Our ignorance of Earths system makes the situation
complicated; eventually, the well-being of humans around the world will be devastated if
we continue in this pattern. The extreme weather now occurring around the world is
changing water distribution patterns seasonally and spatially. If we dont understand the
interactions between vegetation and hydrology, we will not be able to manage all resources
wisely to face challenges from the changing of Earth systems. The albedo of mountain
snow cover and, in turn, global warming are sensitive to absorptive impurities such as dust
and soot. Hansen and Nazarenko (2004) have asserted that about a quarter of the
observed increase in global temperature over the last century owes to absorption by black
carbon in snow. But how do forests respond to the absorption of carbon in the snow and to
increased temperatures? Besides, montane forest dynamics rely on the spatial and
temporal variability of hydrologic variables, especially the variability in snow water
equivalent and snowmelt (Nagler and Rott, 2000). What all this does this mean to
agriculture and to an increasing population?
The per capita availability of fresh water in the western U .S. has decreased dramatically
and could further decrease within 30 years due to population pressure and poor
agricultural policies. Increasing population and industrialization along coastal areas are
adding pressure on coastal wetlands, seagrass, and coral reefs at an alarming rate.
1


Moreover, climate change is showing its impact by causing destruction of infrastructures
and vanishing of life in nature as the scale of hurricanes increase. The United States
Climate Change Science Program Synthesis and Assessment Product 4.3 (SAP 4.3): The
Effects of Climate Change on Agriculture, Land Resources, Water Resources, and
Biodiversity in the United States (CCSP, 2008,
http:/Av\v\v.sap43.ucar.edu/documents/SAP 4.3 6.18.pdf). describes aspects of
climate change that appear to be already affecting U.S. water resources, agriculture, land
resources, and biodiversity. They include:
1) More rapid maturation of grain and oilseed crops
2) Greater risk of crop failure, particularly if precipitation decreases or becomes more
variable
3) Rapid growth of weeds, northward migration of weed species, and greater resistance
of weeds to herbicides
4) Higher precipitation and streamflow with decreased drought severity and duration,
over the 20th century (the West and Southwest, however, are notable exceptions,
with increased drought conditions)
5) Reduced mountain snowpack and earlier spring snowmelt runoff in the Western
United States
6) Invasion by exotic grass species into arid lands causing increased fire frequency,
with rivers and riparian systems being negatively impacted.
7) Increased growing season length by 10 to 14 days as compared to the last 19 years
across temperate latitudes
The other factor strongly related to agricultural and water resource management is soil
moisture. Although soil only holds a small percentage of the total amount of water in the
world, estimation of soil moisture in root zones is vital for improving short- and medium-
term meteorological forecasting, hydrological modeling, monitoring of photosynthesis and
plant growth, and estimating and monitoring the terrestrial carbon cycle (ESA
Communications, 2009). Timely estimates of soil moisture are also important for
contributing to the forecasting of hazardous events such as floods, droughts and heat
waves. Currently, there are relatively few precise in situ measurements of soil moisture, but
if we are to better understand the water cycle so that the forecasting of weather and
climate can be improved, global datasets are urgently required. For this reason, accurate
specification and prediction of the soil moisture fields at different spatial scales (ranging
from local to continental) is a practical need for hydro-meteorological and climate modeling
applications. Though very precise in nature for soil moisture in situ data, the spatial density
of available in situ soil moisture measurements is not practical for deriving reliable area-
averaged soil moisture estimates in the range of 1-10 km grid resolution, which is
mandatory for initialization of meso scale atmospheric models at regional or continental
scales (Mostovoy & Anantharaj, 2008). To overcome this problem, scientists use land
surface models (LSMs) to simulate initial soil moisture states (Crawford et al. 2000; Chen
and Dudhia 2001). But still, varieties of land surface parameters are very difficult to obtain
and thus make the model simulation suffer obvious uncertainties.
To find solutions to these challenging problems, mapping and monitoring existing natural
resources and forecasting future scenarios are highly important. Since a variety of
satellites with various data capture abilities have been launched by NASA and by other
countries, remote sensing (RS) has played a significant role in providing geo-information in
a spatial format and also in determining, enhancing and monitoring the overall capacity of
2


the earth (Navalgund et. al., 2007). Satellite observations of terrain, oceans, land use,
atmosphere, hydrology, and specifically, natural and human-induced hazards have
become crucial for creating important databases and providing prompt information for the
purposes of environmental protection, life loss reduction, and decision-making
enhancements.
1.2 Remote Sensing for Environmental Management
Remote sensing was formally defined by the American Society for Photogrammetry and
Remote Sensing as:
The measurement of acquisition of information of some property of an object or
phenomenon, by a recording device that is not in physical or intimate contact with
the object or phenomenon under study (Colwell, 1983).
By using sensors onboard airborne (aircraft, balloons) or space-borne (satellites, space
shuttles) platforms, remote sensing technology can acquire information about the earths
surface (land and ocean) and atmosphere in an extraordinary manner. With calibrated
sensors, remote sensing observing system allows data to be collected with a synoptic
view, repetitive coverage, observations at different resolutions, and within a very wide
region. These observation systems, furthermore, possess the ability of detecting reflective
radiance from natural phenomena in a variety of wavelengths, making remote sensing a
better alternative for natural resource management, as compared with traditional methods.
Due to these strengths, remote sensing data enables timely generation, reliable and cost-
effective information on various natural resources, namely surface water, ground water,
land use/cover, soil, forest cover and environmental hazards such as water-logging.
Remote sensing can generally be classified as an optical or microwave system. In optical
remote sensing, sensors detect solar radiation in the visible, near-, middle- and thermal-
infrared wavelength regions, reflected, scattered or emitted from the earth, producing
images similar to photographs taken by a camera or a sensor located high in space.
Microwaves, on the other hand, are electromagnetic waves with frequencies between 109
and 1012 Hz. Radar is an active microwave system that fits into this category. The system
illuminates the terrain with electromagnetic energy, detects the scattered energy returning
from the terrain (called radar return) and then records it as an image. Unlike the signature
of the objects in the optical system, the microwave signatures are governed by sensor
parameters (frequency, polarization, incidence angle) and physical (surface roughness,
feature orientation) and electrical (dielectric constant) properties of the target (Jensen,
2007). For example, the radar backscattering coefficient of terrain is strongly dependent on
angle of incidence. The angular dependency of the backscattering coefficient is primarily
due to surface roughness. This interactive nature of radar systems with target physical
properties makes this system ideal for many environmental and hydrological applications.
One of the chapters of this thesis will explore the application of the integration of physical
models and passive microwave systems in generating high-resolution soil moisture
imagery.
The powerful capability of remote sensing in its ability to observe the Earth, covering a very
wide spatial expansion simultaneously, makes studies in many of the related issues
3


mentioned in the previous section possible. Many studies in the following areas
incorporating remote sensing and GIS technology have been already underway.
1) Monitoring climate change:
a. Development of and utilization of measurement methods to evaluate greenhouse
gas emissions resulting from changes in agricultural cropping systems
b. Incorporation of land use/land cover mapping and monitoring methods based on
remote sensing and geographic information systems (GIS)
c. Integration of in situ, airborne, and satellite based monitoring systems for
measuring long-term responses of agricultural lands to regional and global
change
d. Research in the development of easy, reliable means to accurately
ascertain mineral and carbon state of agricultural lands, particularly over
large areas
e. Developing satellite data assimilation techniques to monitor crop systems and
hydrological states over agricultural and urban landscapes
2) Predicting climate change
a. Research in the areas of understanding how climate change (long-term
trends) interacts with land use land cover change at regional and sub-regional
scales
b. Understanding the role of non-radiative forcing, including
agricultural intensification and green infrastructure and their relationship with
climate change
c. Forecasting associated land surface change and resulting changes in the
weather, regional climate, water availability and quality using climate models
d. Developing soil and water resource status prediction model validations through
multi-scale measurement scheme in suitably complex terrains
3) Evaluating impacts of climate change:
a. Research in the area of predicting impacts of changes in management
practices on soil quality and water availability and quality based on in situ and
remote sensing measurements and models
b. Research on the effects of droughts on hydrologic and chemical/pesticide flow
due to changes in soil structure
c. Research in the area of climate change on agricultural productivity in the
Midwest and in Asia and Africa
All of these climate change-related studies are actually based on applications of remote
sensing in various areas, such as sustainable agriculture, ocean color and fishery, water
security, environmental assessment and monitoring, disaster monitoring and mitigation,
weather and climate studies, community-centric applications, and tsunami relief and
rehabilitation. The list above includes just some of the possible applications of remote
sensing in enhancing our understanding of climate change. In fact, after decades of
development of sensor technologies and analytical theories, remote sensing technologies
have become a very powerful tool in many disciplines, including in reconnaissance and
homeland security.
1.3 Study Outline
4


To explore capabilities and deficiencies of remote sensing technologies, four topics were
studied for this thesis. These topics encompass directly and indirectly the use of remote
sensing data, including the applications of image-preprocessing, multivariate statistical
theories, rule-based knowledge systems, GIS analysis, non-linear regression modeling,
and physical models. This thesis is focused on the exploration of the data collected from
various satellite platforms and their applications for environmental management purposes.
In the process, imagery and the aforementioned theories and algorithms are interwoven to
create suitable models for specific purposes. This thesis covers the following independent
topics:
1) Multi-criteria wetland mapping
2) Trail Inventory and object-based classification algorithm
3) The application of generalized linear model (GLM) in predicting Off-High Way
Vehicle (OHV) trail formation
4) The application of the thermal sensor and the physical models in retrieving high
resolution soil moisture
Data Analysis
Figure 1.1 Schematic Diagram of Applying Remote Sensing and GIS for Environmental
Management
The applications of remote sensing technology in these four topics actually cover the
factors of vegetation, atmosphere, hydrology, energy emission, soil, and human behavior.
This wide range of applications illustrates the great capability of remote sensing and the
methods which can be used to overcome pitfalls when using remotely sensed data. From
an electromagnetic radiation spectrum point of view, this thesis applies for various satellite
5


sensors encompassing optical range to passive microwave channels. These four projects
incorporate the applications of remote sensing, GIS, and physical models. The process of
incorporation can be generalized as seen in Figure 1.1.
1.4 Research Topics and Approaches
1.4.1 Multi-Criteria Wetlands Mapping Using an Integrated
Pixel-Based and Object-Based Classification Approach
Wetlands are vanishing rapidly. Over the past two centuries, approximately 40% to 60%
(0.4-1.2 million ha; 1-3 million ac) of the original wetland areas of Colorado have been lost
(Dahl 1990, Wilen 1995). In the Great Plains of the United States, wetlands have been
drained and leveled for crop production. In one major watershed of this region, the Platte
River watershed, the results are devastating. It is estimated that 74 to 80% of wet meadow
systems have been drained for agriculture (Sidle et. al., 1989). The watershed on the
Nebraska side now comprises less than 5% of the total land area (U.S. Fish and Wildlife
Service 1997). The extensive degradation of wetlands in this region is of particular concern
because the Platte River is used extensively by migratory birds on the central flyway
(including the federally endangered interior least tern, Sterna antillarum, and the whooping
crane, Grus americana).
Wetland inventory is very valuable information for decision making for many organizations.
For example, the Colorado Department of Transportation (CDOT) has the challenging task
of protecting the environment while developing and maintaining the best transportation
systems and services possible for the citizens of Colorado. A wetland inventory database
is a key component required to meet the environmental protection mandate. The subject of
the research project is directed towards developing a semi-automated method to identify
and classify inland wetlands in the northern Front Range area of Colorado. A goal of the
project is to produce a database that accurately records wetland locations based on the
classification system that is commonly used by many organization and institutions. The
methodology is based on satellite imagery, high resolution aerial photos and digital
elevation model data, in conjunction with field global positioning system data collections.
Satellite imagery that is being used includes moderate resolution LANDSAT 7 ETM+, Terra
ASTER, and EO-1 Hyperion/ALI. The aerial photography comes from the National
Agriculture Imagery Program and is mainly used for validation and sample collection
purposes. The EO-1 imagery has a high spectral resolution and is used to develop a
wetlands spectrum signature library, which is then used to observe the correlations
between EO-1 and LANDSAT 7 and ASTER image bands. The image processing
approach that is being applied uses both pixel-based and object-based classification
techniques; the object-based technique accounts for the pattern of neighboring pixels (i.e.
context) and wetland boundary shapes. The variables generated for the object-based
classification algorithm are extracted from multi-spectral imagery and include image
texture, wetland shapes, greenness, wetness, brightness, normalized difference vegetation
index, principal components, stream networks, biological soil crust index, and land thermal
fluctuation. In the final stage, these variables are incorporated into hierarchical rule
creation for facilitating wetland classification operations. To complete these tasks, the
software used includes ArcGIS, ENVI, and DEFINIENS Professional. Results of research
indicate a high correspondence with wetlands mapped by field biologists and identification
of additional wetlands that were not previously recognized. As shown in Figure 1.1, many
6


important image processing steps will be executed, including geometric corrections,
atmospheric corrections, image classification, image transformations, image sharpening,
and others. These pre-processing and post-processing steps are mandatory because
wetlands, as seen in the real world, are very complex.
1.4.2 Recreational Travel, Route Formation and Land
Disturbance: Applications of an Object-Based Road
Detection Algorithm
Off-highway vehicles are popularly defined as 1) 4-wheel drive jeeps, automobiles, or sport
utility vehicles; 2) motorcycles designed for off-highway use; 3) all-terrain vehicles, better
known as ATVs and other specially designed off road motor vehicles used in a variety of
ways (Cordell et al., 2005). While people enjoy exploring terrain, the increasing number of
recreational activities in public lands is degrading the environment by affecting vegetation,
soil, and water. Douglass et al. (1999) points out that these components of the
environment would benefit from limiting of all Off-Highway Vehicle (OHV) use to official
routes, as well as devising timing and usage standards for moderate to high impact non-
motorized recreational areas. However, with a lack of a trail inventory of OHVs covering a
large amount of terrain, devising a management plan for monitoring and conserving
purposes would be impossible.
In this project we demonstrate a method for trail inventory and temporal measurement of
disturbances from color images supplemented with preexisting trail vector information.
Because the input data only includes the digital numbers of the three bands (NIR, red, and
green) of an aerial color photo of high resolution, effective application of the shape and
texture information from the images is required. The system includes four different
modules: data preprocessing, image segmentation based on five levels of spatial
resolution of the data, rule set creation for hierarchical classification, and a module for
system evaluation. In the first module, data preprocessing operations include the
application of principal component analysis, sigma local filters to aerial photos, and the
calculation of vegetation indices from multi-spectral images. The second module uses the
context-based Definiens Cognition Network Technology (TCNT) to identify objects at
multiple resolutions The TCNT technique consists of fusing information streaming from
various different sources for an image. In the third module, the obtained image objects are
classified by synthetically evaluating their spectral, texture, and shape criteria created in
the rule sets. The result is extracted ATV activity networks in various levels, including trails
and bulbs, which proved to be a structural set of elements that is geometrically and
topologically correct. The fourth module consists of an evaluation of the detection results
using field GPS maneuvers and qualitative interpretation. Validation results show that this
method is efficient in extracting and defining road networks from high-resolution satellites
or aerial imagery. The object-based trail inventory model comprises a class hierarchy
representing the features under study and their likely relationships. Configured with and,
or", and not operators, each object class in each level within this model contains criteria
that needs to be satisfied in order to strengthen the fact that an image of that object type
has been recognized. The system described in this study enables the extraction of ATV
features, linear or clustered, to be separated or merged from the classification process,
which gives the system the flexibility to build up evidence of class membership from a
variety of information extracted from different scales. In the process, many significant
shape and texture variables can be derived for subsequent statistical analysis. Statistical
7


data exploration serves to create a model that can be used to predict the locations of new
ATV disturbances that will likely occur in the future.
1.4.3 Evaluating Route Formation and Disturbance of
Off-Highway Travel Using Logistic Regression Method
To public land managers, the policy would be optimum if fitting to users' needs while
reducing the environmental impacts of OHV activities by foreseeing future recreational
route formation. After executing object-based classification, much contextual information
can be generated from images. In this study, two years of contextual information data sets
(1996 and 2005) were generated. Through field validation, the trail inventory for 2005 was
shown to have around 90% accuracy. The 2005 data sets high accuracy implies that it can
be transformed as a binary variable in terms of OHV disturbance or non-disturbance (1 or
0). With abundant semantic information accompanying the 1996 trail inventory having been
created, a non-linear regression model, logistic regression analysis, is thus deemed to be
suitable to explore the predictive capability of these semantic variables in evaluating OHV
route formation of 2005. The independent variables are categorized into four groups:
radiation, geomorphology, route configuration, and level of OHV use.
Nonlinear regression models are of a basic form:
Vi = /(Xi, Y) + e\ (1)
Observation Yi is the sum of a mean response f(Xj, y) given by the nonlinear response
function f(X, y) and the error term ei. The error terms are assumed to have expectation
zero, constant variance, and be uncorrelated (Kunter, Nachtsheim, & Neter, 2004). One of
the nonlinear regression models that will be used to predict OHV disturbances is logistic
regression analysis. The equation (1) can be more specifically written in the form in
equation (2) to fit the exponential function.
__________Vo____________
*x. Y) - 1 + Yi* exp(Y2 X) (2)
The nonlinear response function, f(X, y), in parameters yO, y1, and y2 creates an ideal
model to replicate the nonlinear phenomenon found in OHV zone formation. Generally,
logistic regression is well suited for describing and testing hypotheses about relationships
between a categorical outcome variable and one or more categorical or continuous
predictor variables. The central mathematical concept that underlies logistic regression is
the logisticthe natural logarithm of an odds ratio. This logistic transformation is applied to
the dependent variable, making the complicated S-shaped relationship of a dichotomous
dependent variable with continuous predictor variables describable with a linear equation.
Though the study was mainly performed on the R program using statistical theory, the data
generated using GIS and remote sensing functions is still the guarantee for being able to
model the non-linear pattern of recreational traveling. The relationship between GIS,
remote sensing, and statistical analysis is shown in Figure 1.1. GIS itself is not equipped
with the capability of executing multivariate statistical analysis on the attributes of
geographic entities. In this study, the means of integration is illustrated.
8


1.4.4 Downscaling of Advanced Microwave Scanning
Radiometer (AMSR-E) Soil Moisture Using a Thermal
Sensor and the Physically-Based Models
Currently, there are only two sources of operational global soil moisture data from
satellites. One is the data generated from the observations of the Advanced Microwave
Scanning Radiometer (AMSR-E) aboard NASA's Aqua satellite. The other is the data
produced from the recently launched Soil Moisture and Ocean Salinity (SMOS) satellite by
the European Space Agency. AMSR-E and SMOS soil moisture are extracted from passive
microwave. Although they show high accuracy, they have a coarse resolution due to
passive microwave radiometers having inherent limitations. It needs to accumulate
electromagnetic radiation from a large enough area to be of reasonable importance. In
practice, soil moisture at such a coarse resolution has only limited value. To retrieve higher
resolution soil moisture products, active microwave has the potential of satisfying
requirements because its resolution reaches less than 10 meters. However, active
microwave is easily impacted by interferences caused by the terrain and by radar waves.
Foreshortening, layovers, shadows, and speckles are the four common disturbances that
accompany active microwave imagery. Two types of active microwave sensors have been
used to retrieve surface soil moisture: SAR and scatterometry. Most SAR satellites operate
for scientific purposes at C-band. C-band returns are sensitive to vegetation, surface
roughness and the soil dielectric constant. A few studies have derived the surface soil
moisture from scatterometers (e.g. Wagner et al., 2003). Nonetheless, operational surface
soil moisture products from active microwave remote sensing are not yet available
(Wagner et al., 2007). Surface soil moisture retrieval is inevitably limited by the platforms
orbiting and sensing characteristics in terms of geo-biophysical parameters,
spatial/temporal resolutions, and the accuracy level of the data.
Given the limitations of active microwaves in producing daily operational soil moisture
products, passive microwave data was chosen to execute this study. A deterministic
approach for downscaling the 25 km resolution AMSR-E daily soil moisture data was
developed based on the 1 km resolution Moderate Resolution Imaging Spectroradiometer
(MODIS) data and rain gage precipitation data. A two-phase downscaling scheme was
developed in this study. The downscaling relationship is built on creating an association
between soil evaporative efficiency and near surface soil moisture through a physically-
based scaling function. This function is created by bringing together soil properties, the
Von Karman wind turbulence model, and aerodynamic resistance to form a semi-empirical
parameter. As for soil properties, we use percent clay and bulk density extracted from the
Soil Survey Geographic (SSURGO) database to infer published lab findings for application.
Aerodynamic resistance is calculated from wind speed measurements at wind gage
heights given soil roughness. Vegetation index and surface temperature data derived from
MODIS are used to estimate soil temperature and, subsequently, to calculate soil
evaporative efficiency. In this process, MODIS is used to downscale low resolution
microwave-derived soil moisture to the 5-km resolution. The second phase of downscaling
is disaggregating 5-km resolution soil moisture to 1-km resolution soil moisture. In the
second phase, precipitation data and Hortons infiltration equation are used because
precipitation is the major enforcing factor for soil moisture dynamics. The downscaling in
the second phase accounts for the lower soil moisture sensitivity of the MODIS surface
9


temperature and the poor capability of AMSR-E in differentiating soil and vegetational
signals.
The research site encompasses the Babocomari River watershed and Walnut Gulch River
watershed in Arizona. After two phases of downscaling operations, the geomorphology
factor is used to adjust the soil moisture distribution across terrain. In the monsoon season,
water runoff is another important factor that may considerably influence soil moisture
distribution. Flow accumulation, calculated by using the GIS hydrology analysis function, is
used to involve this matter into the downscaling process. Again, this numerical model-
oriented soil moisture downscaling study underlines the fact that GIS has its role in
studying geophysical phenomena and the impacts of climate change.
10


2. Multi-Criteria Wetlands Mapping Using an Integrated
Pixel-Based and Object-Based Classification Approach
2.1 The Necessity of Automation of Wetland Mapping
2.1.1 The Importance of Wetland Information
In Colorado, wetlands are recognized as one of the most productive ecosystems by virtue
of their abundant moisture in otherwise arid environments. In recognition of the extent of
ecological and human value provided by wetlands, government agencies are seeking to
establish wetland protection programs. The Colorado Department of Transportation
(CDOT), Colorado Division of Wildlife (CDOW), US Army Corps of Engineers (USACE),
Colorado Division of Water Resources (CDWR), local governments, and various
conservation organizations are among the institutions which may use a wetland inventory
database so that land use decision quality can be maintained and enhanced.
2.1.2 Regulatory Background
In recognizing wetlands' role in wildlife habitat and flood control, the federal government
has gradually enacted voluntary programs to conserve wetlands. The Water Bank program
(1970) and the Small Wetland Acquisition Program paid landowners to preserve wetlands
for public benefit. Later, additional efforts to inhibit further wetland losses, especially those
occurring at public expense were enacted (Heimlich et al., 1997). The Federal Water
Pollution Control Act Amendments of 1972, amended in 1977, was established to maintain
and restore the biological, chemical, and physical integrity of the waters of the United
States. In 1985, the Food Security Act began to discourage landowners from converting
wetlands by denying farm program benefits to producers who converted wetlands after 1985. But
such subjective tactics can cause dispute. Later in 1993, under the goal of no net loss,
the wetland management plan moved toward restructuring Section 404 permitting. With the
inclusion of drainage activities, the Section 404 process made wetland permitting
restrictions more comprehensive. But to make legislation consistent with a natural system
that is configured with hydrological and vegetative patterns, there are still some points of
debate in the legislation process. For example, the regulation on the number of the
consecutive days during the growing season that areas are inundated as the criteria for
wetlands delineation would dramatically influence the level of conservation of wetlands.
Wetlands have a very complex formation and have various functions toward environment
and human well being. The setup of any benchmark for wetland conservation would
inevitably fail without thorough scientific trials and validation processes.
To settle possible topics of debate, the Secretary of the Army, acting through the Chief of
Engineers, was authorized by Section 404 of the Act to issue permits for the discharge of
dredged or fill material into the waters of the United States, including wetlands. However,
to accomplish wetland regulatory functions requires knowledge of where the wetlands
themselves are located. Creating a wetlands inventory database has thus become an
urgent mission for Colorado. In response to the Clean Water Act, US Army Corps of
Engineers established the Wetlands Delineation Manual (USACE 1987) to provide
11


guidance for identification and delineation of wetlands potentially subject to regulation
under Section 404. This manual is viewed as a mandatory means by which public and
private sectors can legitimately identify wetlands. The technical guidelines for wetlands
delineation in the Corps manual do not specify a strict wetland classification system.
Rather it provides guidelines for determining whether a given area is a wetland or not for
legal regulatory purposes; such determination typically requires field delineation of the
wetland boundary by a professional trained in wetland delineation techniques.
The use of remote sensing for wetlands mapping has been established by a number of
state agencies as a guide for planning purposes. A notable example in this regard is the
Wisconsin wetland inventory program developed to augment the USACE procedures. For
our work, the wetland classification system of the Wisconsin Wetlands Association (WDNR
1992) was used for wetland categorization.
2.1.3 Objectives
The wetlands mapping research project has been motivated by requirements of the Safe,
Accountable, Flexible, and Efficient Transportation Equity Act: A Legacy for Users
(SAFETEA-LU). SAFETEA-LU authorizes federal surface transportation programs for
highway safety and transit for the 5-year period 2005-2009. SAFETEA-LU addresses the
many challenges involved in the development of transportation systems, such as improving
safety, reducing traffic congestion, improving efficiency in freight movement, increasing
intermodal connectivity, and protecting the environment. The provisions include a new
environmental review process for highways, transit, and multimodal projects, with
increased authority for transportation agencies such as CDOT.
The wetland identification methods developed in this research study are considered highly
accurate and appropriate for transportation facilities planning purposes. The generated
wetland inventory will therefore become a valuable component for developed cumulative
impact assessments, informed transportation planning, stewardship of natural resources,
and land use allocation decisions.
The objectives of this research are to:
1) Establish a highly reliable database of wetland occurrences and distributions to
assist planning level activities.
2) Develop procedures for wetland identification using inexpensive data.
3) Create algorithms that accurately identify wetlands in a wide area.
4) Simulate the delineation method established in the USACE Wetlands Delineation
Manual by applying recent advances of geospatial technology.
Traditional pixel-based classification algorithms, such as parallelepiped and ISODATA
(Iterative Self-Organizing Data Analysis Technique), are carried out based on the spectrum
information of each individual pixel. With a pixel-based algorithm, pixels are classified
according to the composition pattern of the radiation in various wavelengths emitted from
that pixel location. This type of clustering, however, doesnt involve the evaluation of the
geographic aspects of the pixel and its contextual relationship with the surroundings. Thus,
the pixel-based classification method is likely to generate many salt-and-pepper type
classifications that do not represent contiguous wetland areas. Especially when there are
not enough spectral bands in the imagery, the separation capability between classes would
12


be dramatically reduced by using traditional pixel-based approaches. To overcome this
deficiency, other techniques have been proposed, mainly belonging to one of three
categories: (a) image pre-processing, (b) contextual classification, (c) post-classification
processing, such as rule-based processing and morphological filtering. These techniques
increase classification accuracy, but their disadvantages are evident when applied to high
spatial resolution images, such as EO-1 Advanced Land Imager imagery, IKONOS, and
NAIP (National Agricultural Imagery Program) data. These methods either require intensive
computation or produce inaccurate results at the boundaries of distinctive land cover units.
In consideration of the complexities involved in mapping wetland occurrences, the object-
based classification algorithm was developed in this project to supplement the capability of
traditional pixel-based classification algorithms. Instead of analyzing the pixel spectrum
information for classification purposes, the mechanism of the object-based classification
algorithm is able to categorize the imagery by evaluating the geometric and textural
characteristics of wetland areas as objects. Simply stated, objects are formed by grouping
contiguous pixels with homogeneous aspects of ancillary conditions, such as smoothness,
spectrum, or shape, as guided by the imagery analyst. Within objects the local spectral
variation caused by gaps, shadows, and crown textures is mitigated by the creation of the
objects at various scales. Furthermore, with objects as the minimum map units, the object-
based algorithm is able to appraise many spatial properties of wetlands, such as the
shape, size, direction, density, distance, compactness, and texture of wetlands. Thus, the
object-based approach employed in this study is not limited to the evaluation of spectral
characteristics of the hydrophytic vegetation and hydric soils, but makes maximum use of
the spatial contextual variables of wetlands, such as the wetland distribution pattern and
distances to the stream courses. This greatly enhances mapping accuracy and
completeness. The object-based classification procedures developed in this study were
executed with DEFINIENS image processing software.
2.2 Wetland Definition and Parameters
Successful identification of wetlands starts with a clear definition of wetlands. According to
USACE (1987) wetlands are defined as
Those areas that are inundated or saturated by surface or ground water at a frequency
and duration sufficient to support, and that under normal circumstances do support, a
prevalence of vegetation typically adapted for life in saturated soil conditions. Wetlands
generally include swamps, marshes, bogs, and similar areas.
2.2.1 Parameters of Wetland Delineation of USACE
The USACE (1987) wetland definition introduces three mandatory environmental
characteristics for wetland detection: vegetation, soil, and hydrology. Within the manual,
these three diagnostic environmental characteristics are further characterized to provide
guidelines for this study.
2.2.1.1 Vegetation
The widespread vegetation in wetlands normally consists of macrophytes that are adapted
to areas with saturated hydrologic and soil environmental conditions (USACE 1987).
13


Hydrophytic species are vegetation which has the ability to grow, compete, reproduce,
and/or persist in anaerobic soil conditions. The delineation in the manual places emphasis
on the collection of plant species that have a controlling influence on the character of the
plant community rather than on indicator species. For this project characterization of the
hydrophytic plant communities in Colorado were based on data from the National Wetland
Inventory website and associated field observations.
2.2.1.2 Soil
The USACE Wetlands Delineation Manual (1987) set up criteria for determining the
presence of hydric soils. Normally, these soils possess characteristics that are connected
with anoxic soil conditions. They can be organic soil, histic epipedons, sulfidic, or soils of
the aquic/peraquic group. These soils are the products of prolonged anaerobic soil
conditions, which exist when soil is inundated or saturated for a sufficient duration, and will
result in the chemical reduction of some soil components (e.g., iron and manganese
oxides). Soil colors and other physical characteristics thus become the indicators of hydric
soils. For the remote sensing approach, biological soil crust index accompanied by the
generated emissivity or Thermal Response number was used to emulate the soil
identification process specified in the USACE delineation manual.
2.2.1.3 Hydrology
According to the USACE manual (1987), wetland identifications should include hydrologic
characteristics that result in inundation either permanently or periodically at mean water
depths < 6.6 ft. Otherwise, at some time during the growing season of the prevalent
vegetation, the soil is saturated to the surface. Topographically, the areas of lower
elevation in a floodplain or marsh have more frequent periods and/or a greater duration of
inundation than most areas at higher elevations. As for plant cover factors, areas of
abundant plant cover make additional water drain more slowly and thus increases the
frequency and duration of inundation or soil saturation. Conversely, the transpiration rates
of these sites may give the investigated field the entirely opposite effect. The evapo-
transpiration rates may be higher in areas with abundant plant cover and thereby reduce
the duration of soil saturation. This plant canopy factor, in association with other indicators
such as drainage patterns, drift lines, sediment deposition, watermarks, stream gage data
and flood predictions, historic records, visual observations of saturated soils and
inundation, form a rigorous criteria for the wetland hydrological process evaluation. In this
research study, the use of tasseled cap transformation techniques, hydrological analysis
on topographic data, and the concept of surface temperature allows for the incorporation of
some of the field hydrologic indicators mentioned above using remote sensing techniques.
2.2.2 Wetland Types Adopted for Classification
As summarized in the previous section, wetlands are characterized by vegetation, soil
type, and degree of saturation or water cover. According to the USACE (1987) wetland
definition, all of the three diagnostic environmental characteristics need to be satisfied to
classify an area as a wetland. On the other hand, the Classification of Wetlands and
Deepwater Habitats of the United States (Cowardin et al., 1979) from the US Fish and
Wildlife Service (USFWS) classifies a wetland by requiring only one attribute fulfillment
among the three diagnostic attributes. Considering that such a contradiction exists
14


between these two systems and that the USACE wetland definition has been adopted for
the purposes of this research, a more general wetland classification system was used to
categorize wetlands found in the field. However, the corresponding wetland classes of the
Wetlands and Deepwater Habitats Classification hierarchy from the USFWS were also
contained in the texts for comparison. Some of the prominent wetland types are listed
below; these form the basis for the wetland classification system for this research.
Aquatic Bed. Plants grow entirely on or in a water body for most of the growing season in
most years. Aquatic beds in the NFRMPO area can be found in the sheltered areas of
reservoirs or ponds that have little water movement. They generally occur in water no
deeper than 6 feet. Plants include pondweed, duckweed, lotus and water-lilies. They can
be classified into the Aquatic Bed class in the Lacustrine, Palustrine, or Riverine system if
the hierarchy of the Wetlands and Deepwater Habitats Classification of the USFWS is
used.
Marshes. Marshes are characterized by emergent aquatic plants growing in permanent
and seasonal shallow water with water depths of less than 6.6 feet (2 meters). The
counterparts of Marshes in the Wetlands and Deepwater Habitats Classification hierarchy
of the USFWS is Palustrine / Emergent or Lacustrine / Littoral / Emergent. In the NFRMPO
area, the marsh size can vary from a one-quarter acre pond to a long oxbow of a river or
shallow bay of a lake. The species are dominated by cattails, bulrushes, pickerelweed, lake
sedges and/or giant bur-reed (Figure 2.2).
Figure 2.1 Aquatic Bed
15


Figure 2.2 Marshes Commonly Seen Species Growing in Shallow Water in Colorado
Wet Meadows. Wet meadows normally exhibit saturated soil rather than standing water.
Meadows are essentially closed wetland communities (nearly 100% vegetative cover) and
often act as a transition zone between aquatic communities and the uplands. Sedges,
grasses and reeds are the governing species with the possible presence of sneezeweed,
marsh milkweed, mint and several species of aster and goldenrod. Plants occurring in
meadows include species found in other communities, such as the annuals of seasonally
flooded basins and emergent aquatic plants of marshes (Figure 2.3). With the USFWS
Wetlands and Deepwater Habitats Classification hierarchy, wet meadows can be
categorized as Lacustrine / Littoral / Emergent class or Palustrine / Emergent class. In the
Platte River region, wet meadows often are present as mesic prairies (grassland) with
dendritic linear slough complexes in low-lying areas (Meyer et al., 2008). Water levels in
the sloughs are regulated by groundwater connections to river channels and are influenced
by precipitation and evapo-transpiration (Whiles and Goldowitz 1998).
Scrub/Shrub. These areas, which include bogs and alder thickets, are characterized by
woody shrubs and small trees less than 20 feet tall, such as bog birches, willows, and
dogwoods (Figure 2.4). This type of wetland in the NFRMPO area mainly exists along
rivers, but does not belong to a riverine system. Instead, Scrub/Shrub is categorized as
Palustrine / Scrub-Shrub class according to the wetland hierarchy of the USFWS. Physico-
chemical properties of wetlands provide many positive attributes for remediating
contaminants. The expansive rhizosphere of wetland herbaceous shrubs and tree species
provide an enriched culture zone for microbes involved in decomposition (Williams, 2002).
By developing a wetland function value index, Wolfson et al. (2002) found that scrub/shrub
wetland type has significant value in flood control and in the recharging of ground water. It
also enhances fishes habitats and aquatic diversity.
16


Figure 2.3 Wet Meadows in Fort Collins
Figure 2.4 Scrub/Shrub along the Cache La Poudre River
17


Floodplain Forest. The floodplain forest is categorized in the USFWS wetland hierarchy as
Palustrine / Forested. These forested floodplain complexes are characterized by trees 20
feet or more in height such as cottonwood, elm, salix spp., water birch, and big-tooth
maple. Grass cover underneath trees is an important part of this system and consists of a
mix of tall grass species, including Panicum virgatum and Andropogon gerardii (Figure
2.5). There are not many flood plain forests in the NFRMPO area; most of the forested
floodplain complexes exist along the Cache La Poudre River. Floodplain forests are very
important places where groundwater recharging and the infiltration of dissolved nutrients
occur. Holland et al. (2006) points out that groundwater recharging by seasonal wetlands is
a major hydrologic mechanism in several arid wetland areas and that the infiltration of
dissolved constituents such as dissolved organic matter (DOM) has implications for human
and/or ecosystem health. For example, the seasonal floodplains of the Hadejia-Nguru
Wetlands of Nigeria collect groundwater in that area that is recharged by the floodplains
and is then relied upon heavily for potable water supply and irrigation (Goes, 1999).
Mladenov et al. (2007) find that during flood periods not influenced by fire, the organic
carbon derived from vegetation/litter sources within the floodplain contributes greater DOC
to the seasonal floodplain than organic carbon from microbial sources within the floodplain.
Through the input of organic matter, floodplain forests provide sources of energy for
aquatic organisms. Shade from streamside vegetation moderates temperature variations in
aquatic systems, preventing excessive warming of the water during summer months.
Woody debris from floodplain vegetation influences the development of channel
morphology and provides a necessary habitat for many aquatic organisms (Tepley, 2008).
Figure 2.5 Seasonal Floodplain Forest in Winter with Grass Cover Underneath Trees
2.3 Methods
2.3.1 Study Area and Image Acquisition
This research study was conducted in the northern Front Range of Colorado covering an
area from north of Fort Collins to north of Loveland, with I-25 crossing through the middle
of the site, north to south (Figure 2.6). The area is delineated by a trapezoid with corners at
1056'6.275'W, 4039'56.001" N; 10453'51.055"W, 4(T39'56.001"N; 10453'51,055'W,
4026'45.949"N; 10510'27.332"W, 4026'45.949"N. The study area is about 60 miles
away from Denver, Colorado. The area of the project research boundary is approximately
18


500 km2. This area consists of agriculture, urban, and suburban zones. Most of the study
area is located within the watershed of the Cache La Poudre River. The Cache la Poudre
in northern Colorado is renowned for its canal development heritage in importing, storing
and conveying water. Originating from above the tree-line in Rocky Mountain National
Park, the Cache la Poudre has become Colorado's first and only national Wild and Scenic
River. As the river and its tributaries wanders through the terrain, numerous fens, marshes,
potholes, and wet meadows are dotted throughout the area. Agricultural irrigation projects
have also created incidental wetlands.
To cover this area, several satellite images were used:
1) ASTER (Advanced Space-borne Thermal Emission Radiometer)
ASTER obtains high-resolution (15 to 90 square meters per pixel) images of the
Earth in 14 different wavelengths of the electromagnetic spectrum, ranging from
visible to thermal infrared light. The ASTER imagery used for this study was
captured during the dates of 08-13-2003 and 03-15-2003. For the date of 08-13-
2003, the L1B data, high-level kinetic temperature data, and surface radiance data
all had been acquired. Data purchased for the date of 03-15-2003 consists mainly of
L1B data and kinetic temperature data for thermal infrared bands. The data from
these two dates include all of the 14 bands of ASTER.
2) LANDSAT7ETM+
The purchased images are scenes captured on the dates of 06-16-2002 and 04-16-
2003. The LANDSAT 7 imagery is composed of four bands of 30 m 30 m visible
and near infrared data, two middle infrared bands, one band of thermal-infrared data
with 60 m resolution, and a 15m* 15 meter panchromatic image. The ETM+ data
was used to calculate various vegetation indices as well as the soil crust index. The
scene of 04-16-2003 was mainly used as supplementary analysis for the spring
season, when leaves had not yet totally turned green and there were fewer leaves
present. Thus, the impacts of image signaling from soil on vegetation radiation can
be observed. Comparison of wetland hydrology, vegetation, and soil conditions in
the spring and summer increases the accuracy of the mapping.
3) EO-1 Hyperion Hyper-spectral / Advance Land Imager
To further observe hydrophytic vegetation to the species level, hyper-spectral data
was employed to classify the vegetation species. With only a 7.7 kilometer swath,
the EO-1 Hyperion data was not used directly for delineating wetlands. Instead, it
was used for the creation of the spectrum library of various wetland classes and for
a comparison with other overlapping images. The obtained Hyperion hyper-spectral
image has more than 220 spectral bands (from 0.4 to 2.5 urn). The data was
captured on 10-26-2001.
4) NAIP (National Agriculture Imagery Program) Aerial Photographs
The one-meter resolution NAIP for Larimer and Weld Counties were downloaded
from the USGS (US Geological Survey) FTP site
(ftp://rockvftp.cr.usqs.qov/nqtoc/colo naip/). This data was mainly used for GPS
data checks and sample site selection.
19


Figure 2.6 The Study Area with a Trapezoidal Boundary
2.3.2 Field Survey
Field data collection is mandatory for successful wetland mapping. The wetland sample
sites can be either used as training sites to collect spectrum characteristics or used as
samples for validation purposes.
20


The sample sites of various wetlands were located using a Global Positioning System
(GPS) unit. The possible wetlands were first identified on the NAIP imagery. Field visits
were then planned accordingly. In the field, pictures of wetland samples were taken and
their location information was collected by using the handheld Trimble Geo-Explorer GPS
unit. In order to ensure that the accuracy of wetland locations was within 2m, the
recorded GPS data was post-processed using differential corrections. In the U.S., a
number of governmental and private agencies have made base files for differential
correction purposes freely available online. The data collected at the base station is used
to calculate generated differences with the GPS satellite signals (by finding the difference
between the positions calculated from the satellite signals and the known reference
position). In the Fort Collins area there are two base stations that continuously serve
ready-to-download data every hour. Two of the collected wetland sample sites are shown
in Figure 2.7.
2.3.3 Data Preparation
21


There are some operations that need to be performed on the raw images acquired before
obtaining environmental indices and implementing classification processes. These steps
include geo-referencing, sharpening, and converting to reflectance.
2.3.3.1 Geo-referencing
Since all of the images are collected from multiple satellite platforms, they need to be
registered in the same projection and geographic coordination system so that subsequent
operations can be accurately executed across layers generated from various satellite
platforms. This is done in order to make sure that the pixel discrepancy between images is
constrained to a minimum. The registration accuracy is targeted to smaller than 0.35 pixels
of root mean square error. All geo-referencing operations were set up to the third order of
the polynomial transformation. More than 80 ground control points were collected for each
registration to achieve a high RMSE standard. In some cases, there were actually more
than 100 ground control points collected for image registration.
NAIP, with its high geographic accuracy, was chosen as the reference map for geo-
referencing other satellite images. NAIP data uses the UTM North America Datum 1983 as
its coordination system; thus, NAD83, UTM zone 13N is used as a common spatial
reference system for all of the images in this research. The registration processes were
executed using ENVI software. Most of the images were registered by using the cubic
convolution algorithm given that wetlands are normally are located on the edge between
water bodies and uplands.
2.3.3.2 Gram-Schmidt Spectral Sharpening
Image fusion of remote sensing techniques aims to assimilate information conveyed by
data acquired with different spatial and spectral resolutions from satellite or aerial
platforms. Because the number and kind of satellite platforms are increasing, image fusion
techniques are increasing in importance in remote sensing. The theory of image fusion
says that an optimal quality for a fused image has Minimum Color Distortion (containing all
the spectral property of Multi\Hyper spectral images), Maximum Spatial Resolution
(containing all the spatial property of high resolution image) and Maximum Neutrality (the
best integration of spectral and spatial quality of input data). But this ideal situation only
can be obtained theoretically (Zhou, 1998). For example, the most straightforward fusion
method, the Intensity Hue Saturation (IHS) transformation, features a large spectral
distortion when displaying the fusion product in color composition.
To achieve optimal quality in the image sharpening process, Gram-Schmidt spectral
sharpening, which is one of the most sophisticated methods for executing fusion on multi-
spectral and panchromatic bands, was applied in this project. This algorithm is based on
the component substitution strategy devised by Laben and Braver in 1998 and patented by
Eastman Kodak. This algorithm is the method adopted by the ENVI package and is thus
used by this research.
Basically, Gram-Schmidt spectral sharpening extracts the high frequency variation of a
high resolution image and then inserts it into the multi-spectral framework of a
corresponding low resolution image. In this algorithm, algebraic procedures operate on
images at the level of the individual pixel to proportion spectral information among the
22


bands of the multi-spectral image. The replacing (high resolution) image substitutes one of
the bands of the original image and can then be assigned the correct spectral brightness.
A comparison of raw images and post sharpening of LANDSAT 7 ETM+ 06/16/2002 data
by the Gram-Schmidt sharpening method is shown in Figure 2.8 and 2.9. The comparison
exhibits an outstanding improvement of spatial quality after sharpening, though it shows a
little spectral distortion.
Figure 2.8 ETM+ Imagery after Pan-sharpening
23


Figure 2.9 Raw ETM+ Imagery (30-m Resolution)
2.3.3.3 Transferring Digital Number to Reflectance
In order to successfully generate vegetation and soil indices, the digital numbers of
LANDSAT 7 ETM+ need to be transferred to the reflectance. This process actually
includes two steps. The first step is to transfer the digital number of each band to its
radiation, as detected by the sensor. This process involves the calibration process that
brings 8-bit integer data to the 32-bit floating point. The equation used is shown below.
LA = "gain" QCAL + "offset" (1)
Equation (1) can be expressed as:
24


(LMAXA LMINA)
LA = -------------------- (QCAL- QCALMIN) + LMINA
(QCALMAX QCALMIN)
(2)
Where:
LA = Spectral radiance at the sensor aperture in watts/(meter squared ster pm)
"gain" = Rescaled gain (the data product "gain" contained in the Level 1 product header or
ancillary data record) in watts/(meter squared ster pm)
"offset = Rescaled bias (the data product "offset" contained in the Level 1 product header
or ancillary data record ) in watts/(meter squared ster pm)
QCAL = Quantized calibrated pixel value in DN
LMINA = Spectral radiance that is scaled to QCALMIN in watts/(meter squared ster pm)
LMAXA = Spectral radiance that is scaled to QCALMAX in watts/(meter squared ster *
pm)
QCALMIN = Minimum quantized calibrated pixel value (corresponding to LMINA) in DN
= 1 (LPGS Products)
= 0 (NLAPS Products)
QCALMAX = Maximum quantized calibrated pixel value (corresponding to LMAXA) in DN
= 255
All of the above variables can be obtained from the metadata accompanying the file. The
second step is to translate from surface radiation to reflectance at the sensor. The
LANDSAT images in this step were actually normalized for solar irradiance by converting
spectral radiance, as calculated above, to planetary reflectance or albedo. This is a
combination of the surface and atmospheric reflectance of the Earth. It was computed with
the following formula:
TT U d2
Pp= --------------
ENSUa cosGs
(3)
Where:
Pp = Unitless planetary reflectance
/.a = Spectral radiance at the sensor's aperture
d = Earth-Sun distance in astronomical units from nautical handbook or interpolated from
values listed in Table 2. In order to get a correct distance in astronomical units, the
Julian Day of the image capture date needs to be figured out.
ESUNk = Mean solar exo-atmospheric irradiances from Table 1
6S = Solar zenith angle in degrees
Table 2.1 ETM+ solar spectral irradiances
Band watts/(meter squared pm)
1 1969.00
2 1840.00
3 1551.00
25


Table 2.1 (Cont)
4 1044.00
5 225.70
7 82.07
8 1368.00
The Julian Day / Calendar Day Conversion information can be found from the following
webpage provided by NASA Goddard Space Flight Center.
http://rapidfire.sci.qsfc.nasa.gov/faq/calendar.html
Table 2.2 Earth-Sun distance in astronomical units
Julian Day Distance Julian Day Distance Julian Day Distance Julian Day Distance Julian Day Distance
1 0.9832 74 0.9945 152 1.0140 227 1.0128 305 0.9925
15 0.9836 91 0.9993 166 1.0158 242 1.0092 319 0.9892
32 0.9853 106 1.0033 182 1.0167 258 1.0057 335 0.9860
46 0.9878 121 1.0076 196 1.0165 274 1.0011 349 0.9843
60 0.9909 135 1.0109 213 1.0149 288 0.9972 365 0.9833
By using the equations (1 to 3) and the tabulated values listed above, the reflectance of
band 4 of the LANDSAT imagery for a portion of the study site captured on 04/16/2003 is
displayed with the Figure 2.10.
Figure 2.10 Band 4 Reflectance of the LANDSAT ETM+ 04/16/2003 Imagery Covering
Fort Collins and Adjacencies
26


After transferring the digital number of ETM+ Band 6 digital number to radiance with the
process described above, the ETM+ Band 6 imagery can also be converted from spectral
radiance to a more physically useful state. This is done with an assumption of unity in
emissivity and with using the pre-launch calibration constants listed in the Table below.
The effective at-satellite temperatures of the viewed earth-atmosphere system in the Fort
Collins area on the date of 04/16/2003 are displayed below. The conversion formula used
in research is:
T =
K2
m( u +1)
(4)
Where:
T = Effective at-satellite temperature in Kelvin
K2 = Calibration constant 2 from Table below
K1 = Calibration constant 1 from Table below
L = Spectral radiance in watts/ (meter squared ster pm)
Table 2.3 ETM+ thermal band calibration constants
Constant 1 K1 Watts / (meter squared ster pm) Constant 2 K2 Kelvin
LANDSAT 7 666.09 1282.71
LANDSAT 5 607.76 1260.56
Figure 2.11 Generated At-Satellite Surface Temperature from ETM+ 04/16/2003 Data
27


From the at-satellite surface temperature image shown above, the differences in surface
temperature for urban/suburban, agricultural, and wetter areas are clear. This data is
helpful for identifying wetlands because they are normally cooler in comparison with bare
soil or artificial structures.
2.3.4 Vegetation Indices
The acquired hyper-spectral data (EO-1 Hyperion) only covers a portion of the study area;
this makes thorough differentiation of wetland vegetation species across the study area
impossible. Therefore, instead of relying completely on EO-1 imagery, the vegetation and
soil biophysical variables extracted from ASTER and LANDSAT imagery were used in a
supplemental manner to increase the accuracy of wetland identification. In this study, these
vegetation indices, complemented with the spatial attributes of the image objects
generated in the object-based classification process, will create the thresholds for various
classification categories. The contributions of the vegetation indices in this study are
(Jensen, 2007):
1) Maximizing sensitivity to plant biophysical parameters,
2) Minimizing external effects, such as atmospheric effects, Sun angle, and viewing
angle,
3) Validating the classification results, and
4) Normalizing internal effects such as canopy background variations, including soil
interference, differences in senesced or woody vegetation, and topography (slope
and aspect).
Several indices were extracted from the images. They are NDVI, wetness, brightness, and
greenness. Some of the generated indices are displayed below. They are accompanied
with NAIP photographs for comparison.
2.3.4.1 Kauth-Thomas Tasseled Cap Transformation
Kauth and Thomas (1976) produced an orthogonal transformation of the original LANDSAT
MSS data space to a new four-dimensional feature space. This is the inauguration of the
application of the Kauth-Thomas tasseled cap transformation. Through the years, the
Kauth-Thomas tasseled cap transformation continues to be widely used and has been
reformed for ETM+ image application. The derived brightness, greenness, wetness indices
can provide subtle information concerning the occurrence status of the wetland
environment (Figure 2.12 to Figure 2.14).
28


Figure 2.12 Correspondence between the Greenness Layer (Right) and the Existence of
Wetland Locations (Left)
Figure 2.13 Wetness (Right) Correspondence to the Wetlands and Ponds Displayed on
the NAIP Imagery (Left).
Figure 2.14 Correspondence between the Brightness Layer (Right) and the Existence of
Wetland Locations (Left)
29


The coefficients developed by Huang et al. (2002) are used for executing the Kauth-
Thomas tasseled cap transformation in this study and are listed below. This tasseled cap
transformation process was performed on the LANDSAT ETM+ data by using the Band
Math function in ENVI. The character B following the coefficients in the listings below is
the band number.
Brightness
0.3561*B1+ 0.3972*B2+ 0.3904*B3+0.6966*B4+0.2286*B5+0.1596*B7
Greenness
-0.334*B1-0.354*B2-0.456*B3+0.6966*B4-0.024*B5-0.263*B7
Wetness
0.2626*B1+0.2141*B2+0.0926*B3+0.0656*B4 0.763*B5 0.539*B7
2.3.4.2 NDVI (Normalized Difference Vegetation Index)
NDVI can be used to discriminate between herbaceous and hard wood vegetations and
other non-vegetation land covers. This discrimination is based on differences in reflectance
in the NIR and in the red bands for vegetation and other land covers. The equation is listed
below.
pNIR PRed
NDVI = --------------------
PNIR + pRed
Where pN|R is the reflectance of the Near Infrared band and pRed is the reflectance of the
red band. This reflectance can be obtained through the same process depicted in the
section 2.3.3.3.
Figure 2.15 Fluctuation of Reflectance of Different Materials Across Wavelengths from 0.4
pm to 1.1pm
30


In this study, the NDVI was generated from the ASTER imagery of August and from the
E01 ALI imagery of October. The ASTER data is high-level surface radiance data
corrected for atmospheric effects, having a higher radiometric resolution with 16 bits. The
generated NDVI layers are shown in Figures 2.16 and 2.17. Compared with the NAIP
photograph in Figure 2.18, the generated NDVI from ASTER and E01 ALI images was
found to have potential to differentiate between herbaceous plants, woody plants, and non-
vegetative area. These land covers have different NDVI values, making them easily
distinguishable.
Figure 2.16 NDVI Generated from ASTER 08/13/2003 Imagery
Figure 2.17 NDVI Generated from E01 ALI October Imagery with Crops in the Irrigated
Farms Still Shown a High NDVI Value
31


Figure 2.18 NAIP Photo at the Same Location as in Figure 2.16 and 2.17
2.3.5 Hydrological Analysis
Wetlands and riparian often can be found in surface depressions and in areas along
streams. To discover depression storage areas and stream networks, a sequence of
operations on DEM data were executed using hydrological analysis functions. The
functions of hydrology analysis can be found in the Spatial Analyst Tools in ArcGIS. The
hydrology analysis starts from the flow direction function. Drainage flow directions are
determined by the prevalent D8 algorithm, which assigns the drainage value from one
point on the DEM grid to one of its eight bordering neighbors. The possible assigned
values are 1, 2, 4, 8, 16, 32", 64, and 128 as shown in Figure 2.19. This
FLOW_DIR raster is then used to execute flow accumulation analysis (Figure 2.20). In flow
accumulation analysis, the numbers of cells that will flow into each cell in the FLOW_DIR
grid along all possible directions are calculated and accumulated to produce a new raster
of FLOW_ACC. After the flow accumulation operation, the Map Algebra function is
employed as the final step to generate the stream network.
32 64 128
1 6 1
8 4 2
Figure 2.19 Flow Direction Number Assignment
32


2 2 2 4 4 8
2 2 2 4 4 8
1 1 2 4 8 4
128 128 1 2 4 8
2 2 1 4 4 4
1 1 1 1 4 16
0 0 0 0 0 0
0 1 1 2 2 0
0 3 7 5 4 0
0 0 0 20 1 1
0 0 0 1 24 0
0 2 4 7 35 2
Figure 2.20 Flow Direction (Left) to Flow Accumulation (Right)
The sequence of hydrologic operations is quite clear. In the path from flow direction to flow
accumulation, some confining hydrologic depressions will be generated, interrupting the
network at these depression spots. In the real world, these spots are areas where water
stops flowing. Though some of these depressions exist in the real world, most of them are
data assimilation errors when converting floating point values in the DEM to integer values.
This error may cause problems in establishing stream networks because in the real world,
terrain water flow fills small depressions and additional water will continue to flow along its
course. Therefore, depressions in the 10M DEM need to be filled to insure drainage
continuity throughout flat spots and out of depressions. However, some of the sinks are, in
fact, real depressions in the terrain. To avoid erasing these real depressions, these sinks
were filled back to the elevation of the outpour point of that specific drainage area. The
retained depression areas are prone to becoming flood detention sites and potential sites
for wet soil where wetland vegetation can build up.
In this study, the flow accumulation threshold for stream network generation was set at 45
pixels. In other words, only grids which possess more than 45 pixels of progressive
accumulation after the flow accumulation operation were counted as members of the
network. The CON tool syntax used to set up this condition is listed below:
streamnet = con (flowacc > 45, 1)
The results of the hydrologic analysis were quite accurate (Figure 2.21). The generated
synthetic stream network was overlaid on NAIP data and was found to be very close to the
real world network configuration. Meanwhile, the locations of depressions also proved to
be accurately mapped except for the fact that the true range of the depressions may not be
precisely as depicted, particularly in flat areas. The differences may be caused by the
coarse resolution of elevation data; a 10-meter DEM grid was employed in this study.
The stream network is a very important layer for later object-based classification. As land
developments encroach into stream buffers, identification of wetlands in neighborhoods
has become more difficult. This is due to hydrophytic vegetation being confused with plants
in residents backyards or vice versa. In this research study, the creation of the stream
network helps to resolve these mapping challenges. The buffering of the stream network
increases capability in identifying areas where water will stagnate and wetlands occur
(Figure 2.22 and 2.23).
33


Figure 2.21 Generated Stream Network
Figure 2.22 The Stream Passing through Downtown Fort Collins
34


Figure 2.23 Detected Wetlands along the Stream Comparison with the Figure 2.22
2.3.6 Unsupervised Pixel-Based Classification
Rapid assessment of the land cover distribution pattern for the study area was
accomplished using an ISODATA (Iterative Self-Organizing Data Analysis) technique. This
approach was chosen as a classification technique to classify the imagery of E01 ALI
10/26/2001, LANDSAT 7 ETM+ 04/16/2003, and LANDSAT 7 ETM+ 06/16/2002.
ISODATA is actually an unsupervised classification method and consists of three steps: (a)
classification into spectrally distinct clusters, (b) post-clustering treatment, and (c)
assignment of labels to the clusters. Since unsupervised classification clusters pixels into
spectral clusters, it is possible that classes yet unknown, a priori, can be discovered. This
is an iterative practice; the cluster properties are defined from pixels belonging to that
cluster at any iteration; then, all pixels are appointed to the "closest" cluster.
One of the properties of unsupervised classification algorithms is that they always implicitly
assume that the initial assignment of the clusters does not influence the outcome of the
classification. This is not always true. In this project, the ISODATA operations set with the
same thresholds had been tested upon E01 ALI 10/26/2001 imagery several times. The
classification result is slightly different in every operation. In other words, the classification
results cannot be exactly reproduced. If working in a relatively wide area, this classification
uncertainty problem can become quite noticeable. However, the ISODATA technique still
provides a preliminary land cover classification which can greatly enhance the accuracy of
the object-based classification operation in later steps.
2.3.6.1 Minimum Noise Fraction Transformation
35


When imagery is captured by the sensors there can be considerable variability (i.e. noise)
that becomes implanted into data due to problems of band overlapping and irradiance from
adjacent pixels. In this project, a minimum noise fraction (MNF) transformation algorithm
was used to segregate noise within the data and to determine the inherent dimensionality
of image data. MNF consists of two steps of separate principal components analysis
rotation: 1) by using the principal component analysis on the noise variance/covariance
matrix, the noise in the data is whitened. Thus the noise in the transformed data only has
unit variance. 2) The derived principal components with large eigenvalues are used for
further spectral processing. Figure 2.24 shows the MNF-transformed band 1 of E01 ALI
data. After checking the MNF images and eigenvalues spectrum, the first 6 MNF bands
were found to contain coherent variability. The MNF operation was performed by using
ENVI software.
Figure 2.24 MNF Band 1 of E01 ALI Imagery
2.3.6.2 ISODATA Classification
After the MNF transformation had been performed on the three image sets, ISODATA
classification was executed. In this step, 35 classes were set for ISODATA classification on
E01 ALI 10/26/2001, 20 classes for ISODATA operation on ETM+ 06/16/2002, and 18
classes for ISODATA classification on ETM+ 04/16/2003 imagery. The categorization
results from these images were then reclassified using a 1 to10 scale, depending on the
potential these generated classes had of becoming wetlands. Figure 2.25, 26 and 27 are
part of the reclassification results of the ISODATA operations based on the three data sets;
in these figures, the bluer the color, the higher the wetland potential. The classification
36


result in June (Figure 2.27) shows a more generalized extension of possible wetlands
because crops and hydrophytic species display a similar color in the growing season. This
classification provides clues that can supplement the classification of the April and October
Images in mapping wetlands. Figure 2.25 to 2.27 show that areas with heterogeneous
surface and elongated irregular shapes are possible wetlands.
E01 ALI 10/26/2001 Unsupervised Classification Results
Figure 2.25 ISODATA Classification Result of the E01 ALI 10/26/2001 Imagery
LANDSAT 1ETM+ 04/16/2003 Unsupervised Classification Results
Figure 2.26 ISODATA Classification Result of the LANDSAT 7 04/16/2003 Imagery
37


LANDSAT 7 ETM+ 06/16/2002 Unsupervised Classification Results
Figure 2.27 ISODATA Classification of ETM+ 06/16/2002 Imagery with 3*3 Majority
Analysis
Weight: 1.25
June, LANDSAT 7
Figure 2.28 Weight Overlay of ISODATA Classifications
38


Generally, the individual ISODATA classifications on images of various seasons proved
productive in locating possible wetlands. Still, some amount of misclassification and
inconsistency was found for these three unsupervised classifications. Specifically, many
irrigated agricultural lands were included in the wetland category. This misclassification
problem is largely due to the generic limitation of the multi-spectral data and pixel-based
classification approach. On the other hand, in comparing classification results for the three
different seasonal images, the classification of EOI ALI 10/26/2001 was found to have the
best quality. To overcome the inconsistencies of classifying different seasonal images and
to make the best use of multi-temporal observations, the products from ISODATA
classification were overlaid with different weights to produce a final pixel-based potential
wetland map (Figure 2.28). This map was later inserted into the DEFINIENS package as
a layer for object-based classification
2.3.7 Object-Based Classification
The basic idea of object-based classification is to cluster spatially adjacent pixels into
homogeneous objects, and then perform classification on these objects. Hay et al. (2001)
defined the objects as basic entities situated within an image; these objects possess an
inherent size, texture, shape, and geographic relationship with the real-world scene
component it represents. Essentially, object-based classification emulates human cognitive
processes that deductively extract conclusions from images. The workflow of object-based
classification in DEFIENS Professional consists of the following sequence of operations:
2.3.7.1 Project Creation
The data layers inserted into DEFINIENS for object-based classification and
segmentation are the layers listed below:
1) AST 09 Atmospheric Corrected Surface Radiance data. There are 9 bands, including
visible, near infrared, and short wave infrared, in this dataset.
2) LANDSAT 7 ETM+ panchromatic band
3) Brightness, greenness, and wetness layers of 06/16/2002 and 04/16/2003 ETM+
imagery generated from Kauth-Thomas tasseled cap transformation.
4) Principal components 1,2,3, and 4 of LANDSAT 7 imagery.
5) Convoluted thermal infrared band of LANDSAT 7 ETM+ 06/16/2002
6) At-satellite surface temperature of LANDSAT 7 ETM+ 04/16/2003 generated
according to the algorithm described in the previous section.
7) Five bands of AST09T 08/13/2003 Atmospheric Corrected Surface Radiance of
thermal infrared data
8) AST08 of 08/13/2003 surface kinetic temperature. This is the high level ASTER data
acquired from NASA. The data is obtained by applying temperature-emissivity
separation algorithm to atmospherically corrected surface radiance data.
9) NDVI of 08/13/2003 generated from AST09 surface radiance data
10) NDVI generated from EOI ALI 10/26/2001
11) Nine bands of E01 ALI of 10/26/2001 data. This dataset is a 16 bit data.
12) Stream buffer of 165 meters raster layer. The raw stream layer was downloaded
from the CDOT website. The purpose of the buffering of 165 meters is to identify the
floodplain forest. These forests in the Front Range normally are present along wider
rivers, such as Cache la Poudre or South Platte River.
39


13) Stream buffer of 32 meters raster layer. The purpose of the buffering of 32 meters is
to consolidate the capability of identifying marshes. As the stream network normally
indicates the presence of inundated water, the addition of stream buffer data into
object-based classification operation enhances the segregation of marshes and wet
meadows.
14) Generated wetland raster layer using overlay and the ISODATA classification
method.
2.3.7.2 Create Image Object
Unlike the ISODATA technique applied in the previous step, the segmentation technique
used in this step is a local behavior-based method which analyzes data variation in a
relatively small neighborhood. In essence, ISODATA produces clusters based on
similarities in the data space, whereas the segmentation technique used by DEFINIENS
not only lessens the degree of variable heterogeneity of pixels within an object but also
addresses the concern of spatial heterogeneity of the image space. The Fractal Net
Evolution approach is thus employed by DEFINIENS. This approach begins with 1-pixel
image objects and grows regionally. Currently DEFINIEN Professional provides four
different image object segmentation algorithms, including 1) chessboard segmentation, 2)
quad tree-based, 3) multi-resolution, and 4) spectral difference. Though the calculations
may be time consuming, multi-resolution segmentation meticulously generates objects
closely resembling ground features. Considering wetlands are clusters of vegetation and
water with a specific shape, the multi-resolution segmentation algorithm is assumed in this
study.
Figure 2.29 Level 1 Image Objects
In the level 1 (the most basic level) image segmentation, the scale parameter was set at
15. The composition of the homogeneity criterion was set as 0.7/0.3 for color/shape and at
0.4/0.6 for compactness/smoothness. The level 1 segmentation result is displayed in
Figure 2.29. Close examination of the results show that objects corresponded well to that
40


in the real situation. If an even smaller scale parameter is set, the segmentation results
might be even better; but this could result in excessive computer processing time. The
scale parameter for level 1 segmentation is therefore recommended to be set between 12
and 15.
For the level 2 image segmentation process, the scale parameter was set at 60. The
composition of the homogeneity criterion was set as 0.9/0.1 for color/shape and 0.4/0.6 for
compactness/smoothness. The color parameter in the segmentation operation of the level
2 segmentation process was set much higher than the shape parameter. This results in
spectral and data variables from the input layers making the greatest contribution to the
formation of image objects. The level 2 objects were used to support the correct
assignment of classes in level 1 classification; the involved layers for the creation of
objects in the level 2 were less than the layers used for level 1 image segmentation. These
layers include ASTER green and near infrared, ASTER band 7 and 9, ETM+ 06/16/2002
brightness and wetness layers, ETM+ 04/16/2003 at-satellite surface temperature, NDVI of
ASTER 08/13/2003 layer, and 32 meters stream buffering. The level 2 segmentation result
is displayed in Figure 2.30. From this display, the object outlines can be found very close to
communities, farm units, or water body boundaries. The classification executed on these
larger objects will furnish information in enhancing the classification accuracy in level 1
(child classes).
Figure 2.30 Level 2 Image object segmentation
41


2.3.7.3 Classification
Though intended for wetland identification, the classes created in this object-based
classification process are not limited to wetland-related classes. The classes created are
aquatic bed, commercial/industrial zone, farm land, floodplain forest, forest, golf course,
grassland, marsh, residential area, rock, scrub/shrub, water body, and wet meadow.
DEFINIENS employs a nearest neighbor function as its main classification algorithm. This
is a supervised classification process. The training site selection is very similar to the
traditional pixel-based supervised classification process, but the objects created previously
are used as the medium in place of pixels.
Figure 2.31 Level 1 Classification Result
42


The most common features that can be applied for classification in object-based
classification are layer mean and standard deviation. In addition to the layer mean and
standard deviation, the features that can be employed for wetland identification in the
object-based classification process include area of wetlands, length/width, density,
compactness, distance to streams, relationship to super-objects, gray level co-occurrence
matrix, and shape index of the wetland features. Use of these features dramatically
increases the wetland classification accuracy. However, due to computer capacity
concerns and to the priority of exploring procedures for integrated pixel-based and object-
based classification, only the features of layer mean value and distances to other classes
were adopted for classification in this pilot project.
Part of the classification results are displayed in Figures 2.31, 2.32, and 2.33, showing a
satisfactory result. The agricultural zone (peach-colored area) and residential zone
(magenta-colored area) in the Fort Collins area are clearly segregated. The various
wetlands are found to be present either along the water course or close to water bodies. In
addition, the misclassification problem with wetlands and irrigated farms has been
resolved. This result demonstrates the effectiveness of the object-based classification
approach.
Figure 2.32 Comparison of the Classification of a Marsh and its Presence in NAIP
43


Figure 2.33 Photograph of the Real World Marsh at the Same Location Shown in Figure
2.32
2.4 Results and Discussion
2.4.1 Areas of the Wetlands Identified
The areas of the wetlands as classified in the study area are listed in the Table 2.4.
Table 2.4 Wetlands area distribution according to its type
Category Area (Km2)
Marshes 37.6
Scrub/Shrub 10.3
Floodplain Forest 5.2
Aquatic Bed 5.5
Wet Meadows 17.6
Total 76.2
The above statistics show that mapped wetlands occupy around 15.3% of the research
area (500 km2). These figures may be slightly higher than in the real world due to a small
portion of irrigated farms that were misclassified as marshes or scrub/shrub. These minor
misclassification problems can easily be resolved if more object features and another level
of classification are be executed with object-based classification methods.
2.4.2 KAPPA Analysis
44


The final accuracy assessment of classification results was based on KAPPA analysis and
the error matrix. Results are shown below (Table 2.5). Classification is based on the 42
samples collected during the field work and from 79 samples directly extracted from NAIP
data. The samples from NAIP are believed to be highly reliable, given field visits to areas
with similar wetland formation characteristics.
Table 2.5 Error matrix
Classified Data
Reference Data Marshes Wet Meadows Aquatic Bed Scrub / Shrub Floodplain Forest Water Row Total
Marshes 29 2 0 2 0 0 33
Wet Meadows 2 11 0 2 1 0 16
Aquatic Bed 1 0 5 2 0 0 8
Scrub/Shrub 1 1 0 12 3 0 17
Floodplain Forest 1 0 0 2 19 0 22
Water 0 0 0 0 0 25 25
Column Total 34 14 5 20 23 25 121
Overall Accuracy 0.83
1) Producer's Accuracy (Omission Error)
Marshes
Wet Meadows
Aquatic Bed
Scrub/Shrub
Floodplain Forest
Water
Overall Accuracy
29/33 = 87.88%
11/16 = 68.8%
5/8 = 62.5%
12/17 = 70.6%
19/21 =86.4%
25/25 = 100.0%
(29 + 11 +5+12 + 19 + 25)/121 =0.83
2) Khat Coefficient
Khat=(N £Xii £(Xi+ X+i)) / (N2 £(Xi+*X+i))
N = 121
Xxii = 29+11+5+12+19+25 = 101
X(Xi+ X+i) = 33*34 + 16*14 + 8*5 + 17*19 + 22*23 + 25*25 =2857
Khat= 79.5%
2.4.3 Findings and Conclusions
This study examined the effectiveness of an integrated pixel-based and object-based
classification method for wetland mapping. Many variables were generated to enhance the
wetland identification process. Some of the variables, such as stream networks,
greenness, wetness, NDVI, surface temperature and the leading two principal components
45


were more influential than others in wetland detection. These influential variables represent
real world wetland factors; vegetation, soil, and hydrology. By incorporating geometric
features extracted from segmented objects, these influential variables contributed greatly
to the accuracy of wetland mapping using inexpensive imagery.
The integrated classification method described began with pre-processing the image to
obtain spectrally and spatially adequate data. The preprocessing steps included geo-
referencing, Gram-Schmidt spectral sharpening, and transferring the digital number to
reflectance. The final phase of the wetland classification process involved synthetic
analysis using the DEFINIENS software. The integrated approach has proved to be an
effective and efficient method for high accuracy wetland mapping.
In natural communities, small changes to one or more local conditions of altitude,
hydrology and climate can result in an entirely different makeup of soils, plants or animals.
This kind of complexity is one of the factors that make wetlands so difficult to classify into
distinct categories. Wetlands have not only variability of natural communities but also exist
on a gradual continuum in the field. Considering that the USACE (1987) wetland definition
was adopted for this research instead of the classification system of the U.S. Fish and
Wildlife Service, a general wetland classification system was applied in classifying wetland
types discovered in the study area. The classes of wetlands are aquatic bed, floodplain
forest, marshes, scrub/shrub, and wet meadows. A similar categorization system was used
in Wisconsin and can be found at the following link:
http://www.wisconsinwetlands.org/wetlofwisc.htm
As can be seen from the accuracy assessment, the developed classification approach
performed especially well at locating inland marshes throughout the study area. To achieve
a high accuracy of wetland mapping for all wetland types, more field observations would
need to be done so that the factors relevant for wetland identification can be collected and
transferred into variables for data analysis. Nevertheless, the results shown in this study
indicate that high quality wetland mapping can be achieved using inexpensive multi-
spectral LANDSAT ETM+, ASTER, and E01 Advanced Land Imager images.
Methodologically, regardless of computer capacity, the developed image processing
procedures are suitable for wetland mapping work for an even wider area. But to identify
wetlands in an extremely large area, some of the processes proclaimed in this research
need to be automated and customized as tools. In addition, this study demonstrated the
possibilities for classifying wetlands by setting up classification rules. For example, the
segregation of sub-emerged vegetated wetlands (aquatic bed) and low marshes depends
on rules that single out areas where vegetation appears in dry seasons but disappears in
wet seasons.
2.5 Future Efforts
Overall, wetland mapping accuracy is quite good. Though the mapping accuracy for wet
meadows and aquatic bed categories are only 68.8% and 62.5% respectively, accuracy
can be improved with additional research. For aquatic bed classification, accuracy can be
improved by doing more field work and observing the beds geographic relationship with
rivers, the seasonal water inundation, and the NDVI standard deviation. For wet meadows,
46


achieving higher accuracy could be difficult because this category is often confused with
general grassland. The accuracy of wet meadow identification can be improved by
employing multi-temporal satellite images. In addition, some ancillary data such as
SSURGO soil data and distances to other wetland types and water bodies could be useful
for enhancing mapping accuracy. These are all activities which deserve future effort.
The integrated pixel-based and object-based classification approach developed in this
study can be improved by developing a decision tree model to simulate the wetland
occurrence logic in the real world and applying such logic as a mathematical model in the
classification process. Successful application of these mathematical models within a large
geographical area requires a computer system with exceptional calculation capacity. A
parallel computer processing system is a means of future improvement.
To achieve an accurate and smooth vector format of the wetland boundaries layer covering
a large area is always demanding. The completed research study provides a solid
foundation for future work pertaining to this purpose. Since most of the images used here
are commonly used data and cover the whole state (except E01 ALI data), the idea of
mapping wetlands across the whole state is quite feasible. The methods and parameters
developed here are repeatable and can be written as processing functions by using the IDL
scripting language. DEFINIENS software allows users to create a process so that
repetitive tasks can be automated. Creation of such a wetland identification process would
lessen the burden for wetland mapping in a wide geographic area.
There are some cautionary notes to be considered. First, a large area should be divided
into numerous operation units with less than 1000 km2 to overcome possible limitations of
computer calculation capabilities. Second, for mountainous areas, more time for data
collection and validation may be required for each sampling site. Thirdly, additional remote
sensing software license seats should be purchased for production-level operations.
Though there may be challenges, we are optimistic about the capability of the developed
methodology of mapping wetlands for the whole state and are hopeful that this task can
indeed be carried out in the near future.
47


3. Recreational Travel, Route Formation and Land
Disturbance: Applications of an Object-Based Road
Detection Algorithm
3.1 Complexity of Recreational Travel
3.1.1 Routes: The Source of Unmanaged Recreation
Researchers have identified unmanaged recreational activities, including Off-Highway
Vehicle (OHV) use, as one of the top four threats to the health of public lands and
communities. As recreational activities in public areas become more popular, the
management of Off-Highway Vehicles in the many open spaces has become a very
significant issue, particularly in those lands managed by the United States Forest Service
(USFS) and the Bureau of Land Management (BLM). Public lands frequently have unique
topography, unique vegetation types and other unique natural features that make them
popular for a variety of activities. At the same time, they typically have sensitive
environmental and cultural resources that are degraded by recreational activities through
processes such as soil compaction, rutting, vegetation trampling, and water quality
dilapidation. Studies by Cole and Landres (1995) observed that vehicles off-road
movement generated impacted the soil in multiple ways including altered nutrient content,
water and oxygen characteristics, and pH and infiltration rate. Some of these negative
effects have been indirectly compounded by more natural forces, such as wind and rainfall.
Indeed, direct damage from the crushing, uprooting, and disruption of root systems of
larger plants by shear stresses to the soil has increased the degree of damage that would
normally be caused by these natural effects. These could include the undercutting of root
systems as vehicle paths are enlarged by erosion, or the creation of new erosion channels
caused by accelerated runoff or wind erosion on land adjacent to OHV activity, or interment
by debris eroded from areas used by vehicles, or the reduction of the biological capability
of the soil with the physical modification and stripping of the fertile upper soil layers
(Douglass et al., 1999). The impact on vegetation and the ecosystem is extensive. Many
studies sponsored by governmental organizations have found that OHVs have damaged a
variety of ecosystems in the United States. The various areas that have been damaged
include the beach grass-covered sand dunes on Cape Cod, the pine and Cyprus
woodlands in Florida, the hardwood forests in Indiana, the prairie grasslands in Montana,
the chaparral and sagebrush hills in Arizona, the alpine meadows in Colorado, the conifer
forests in Washington, and the arctic tundra in Alaska (White House Council on
Environmental Quality, 1979). These impacts contribute to habitat alteration or ecosystem
degradation. Studies using various landscape matrices have confirmed that the alteration
of native plants, animals, and natural ecosystem processes (e.g. nutrient cycling,
pollination, predator-prey interactions, and predator-prey interactions, and natural
disturbance regimes) are related to the fragmentation process (Heilman JR., Strittholt,
Slosser, & Dellasala, 2002). Heilman JR. et al. integrated road density index into the
landscape matrices to explore the effects of forest fragmentation and of disturbance by
using USGS road data (1:100,000 scale). However, they found that high-resolution road
data is desirable when aiming to explore ecosystem degradation at the local scale. High
accuracy road inventory provides a solid data infrastructure for non-linear ecosystem
48


modeling to evaluate the impacts of human activities. For example, the developed models
relating elk use to road densities show that 50% reductions in elk abundance will occur
when road densities are between 1.2-1.9 km/km2 (Lyon 1983).
Road inventory is often applied for one of two purposes. First, it is used for environmental
impact assessment. The importance of the road data in this aspect is immense. Many
studies have demonstrated the important role road data plays in exploring degrees of
disturbance and in creating a variety of indices for hydrological and geomorphologic
analyses. For instance, Tull & Brussard (2007) used Fluctuating Asymmetry, based on the
location of roads, as an indicator to measure environmental stress on populations of
western fence lizards at OHV and non-OHV sites. Also, by illustrating the various effects of
road networks on hydrology, geomorphology, and disturbance patches in stream networks,
Jones et al. (2000) asserted that roads increase the frequency and intensity of flood peaks.
Their study found that the proliferation and change of hydrological patterns may reduce the
availability of refuges in flooding scenarios and thus influence the rates of survival and
recovery of disturbed patches in stream networks. In this way, ecosystem resilience will be
inevitably affected. By studying road type and tree structure for harvest timber from African
tropical forests using Principal Component Analysis, Malcolm & Ray (2000) concluded that
canopy damage rather than stem damage during road formation was the main reason for
overall forest damage, impacting the composition and diversity of rodent and tree
communities.
Second, road data is often used for decision-making and law enforcement operations at
the management level. The BLM and USFS have both introduced planning and
management processes designed to reduce the negative effects of travel-based recreation
through the designation of routes and trails for specific types and levels of use. Land
management agencies are also adapting to the expanded use of roads and trails through
the development of new enforcement systems. The National OHV Implementation Team of
USFS (2005) listed various criteria for evaluating roads and areas for travel planning,
designation, and analysis purposes. All of these operations of planning, regulation and
enforcement of law on public lands require a database that supports analysis of travel
routes, travel patterns and change in patterns. Unfortunately, both BLM and USFS have
limited types of travel data available to them, which limits their capacity to effectively
manage travel on lands within their jurisdiction.
3.1.2 Evolution of Automatic Road Detection Method
One of the biggest challenges of generating a reliable road database is creating a trail
inventory that covers a large area of land with high resolution and accuracy. Due to the
progression of sensor machinery and the development of sophisticated analysis
algorithms, remote sensing technology provides the potential to overcome this challenge
with automatic/semi-automatic strength. A remote sensing observation system allows for
data to be collected at various spectrum channels with a synoptic view and for repeated
observation at different resolutions. When captured at the optimum position, remote
sensing data always presents scenes covering an extensive amount of terrain. Thus, travel
routes can be identified without conducting extensive ground surveys. Although there is
little published research on trail detection using remote sensing methods, feature extraction
methods have been widely applied to geologic lineament, road, and building detection
(Witztum & Stow, 2004).
49


Automatic road extraction methods have been extensively explored since the 1980s. Mena
(2003) categorized these methods according to extraction techniques; applied for seeding
methods, morphological methods, contour models and dynamic programming,
segmentation classification, multi-resolution analysis and Wavelet transform, stereoscopic
operations, multi-temporal analysis, artificial intellectual descriptions with fuzzy modeling,
and methods of spatial reasoning. These methods continue to be applied synthetically
today. Applications of theories such as mathematic morphology, contour models, Wavelet
transform, rule-based system, fuzzy logic, and semantic networks have especially attracted
researchers eyes and have proved that these systems efficiently extract major sections of
the road network, junctions and curved roads. As very high resolution imagery, such as
IKONOS, QuickBird, and many aerial photos, become more widely available for high
accuracy feature extraction work, many of these new methods are now being aimed at
resolving the problems of high intra-class and low interclass variability (Bruzzone and
Carlin 2006) occurring in the classification process. By integrating fuzzy theory and
mathematical morphology, Mohammadzadeh et al. (2006) developed a two-stage method
of detecting roads from high-resolution images. Their method involves the application of
the spectrum information (grey level) of the pixels, of information of size and shape pattern
(granulometry), and of fuzzy if-then rules. The application of the fuzzy logic system allows
them to generate segmented images ready for use in the subsequent granulometry,
remove, trivial opening, filling, and closing procedures. A complete theory using multi-scale
wavelet transform techniques was developed (Chen, Wang, & Zhang, 2002) to separate
roads with an evident difference in width. This technique proposed involving reprocessing
images, deleting undesirable interferences and thin lanes using wavelet transforms, and
performing post-processing to achieve a final result. The strength of the wavelet theory is
that it can successfully characterize images with their frequencies at different resolutions.
This means that wavelet transforms can provide information in both spatial and spatial-
frequency domains. However, the wavelet transformation coefficients originate from an
original pixel, indicating a very high dimensional feature space if spectral and texture
features of multiple bands and involved at multiple scales. To resolve the drawback of high
dimensionality, Huang et al. (2008) incorporated a support vector machine (SVM) classifier
to categorize integrated feature sets. They also executed a multiple windows pyramid
algorithm and an adaptive window algorithm respectively to fuse multi-scale information in
order to mimic human perception in identifying objects of different shapes and structures
on different scales and to increase classification accuracy in homogeneous areas.
Meanwhile, by skillfully organizing high pass and low pass filters in this process, Shankar
et al. (2007) effectively integrated the wavelet transformation-based feature extraction
algorithm with the multi-layer neural classifier for classification of multispectral images. The
performance of the proposed scheme was much improved.
As for the fuzzy logic knowledge-based system, Priestnall et al. (2004) introduced a
framework for automated extraction of linear networks in the Automatic Linear Feature
Identification and Extraction (ALFIE) project. They created an object-oriented data model
as a framework to store contextual knowledge and used this database in the control
module for feature extraction routines. The information collected can be used for many
classification requirements. Liu, Guo, & Kelly (2008) successfully applied the knowledge-
based system to extract roads and move vehicles from an aerial photo. Specifically, their
method involves the development of a framework of spatial relations between segmented
objects, so that shape information, such as topological relations, can benefit classification
operations. Knowledge-based classification has a different emphasis. Bruzzone and Carlin
50


(2006) use the spatial context of each pixel to perform classification according to a
hierarchical multilevel representation of the scene. Su et al. (2008) applied texture and
local statistics information to execute an object-oriented classification method in classifying
urban areas using QuickBird imagery. Their methods also inspired the development of an
object-based trail detection algorithm in our research. The method developed by Segl. &
Kaufmann (2001), which combines supervised shape classification with iterative
unsupervised image segmentation, can be considered as a prototype of the knowledge-
based classification algorithm applied in our research.
The other important advancements of feature extraction techniques include the application
of multi-sensor fusion, local-oriented energy function, and scale invariant feature transform
(SIFT). The study by McKeown et al. (1999), for example, applied data fusion techniques
by combining hyperspectral data and stereo panchromatic imagery to extract cartographic
features. In the 2000s, this multi-sensor fusion technique was further applied to a single
feature and object level. Based on the multi-sensor technique and the application of
processes such as smoothing, topological adjustment, and skeleton extraction, a semi-
automated road centerline extraction and attribution software, RoadTracker, was
developed by GeoEye Inc. Commonly used data fusion techniques include RGB color
composites, layer stacking, multi-resolution analysis, and so on. Walker et al. (2007)
explored a data fusion approach by integrating interferometric synthetic aperture radar
(InSAR), passive optical remote sensing, and the National Land Cover Database (NLCD)
from 2001 derived from LANDSAT imagery to estimate vegetation canopy height for the
contiguous United States. GIS spatial join technique, decision tree regression analysis, and
the image segmentation technique of eCognition were employed in their data fusion
process. Katartzis et al. (2001) modeled spatial information to extract linear features from
airborne images. Markov random field theory, which is defined on the hierarchy of a multi-
scale region adjacency graph, was used to segment images into regions and to classify
those regions. Sirmacek et al. (2009), on the other hand, used scale invariant feature
transform (SIFT) and graph theoretical tools in detecting urban areas and buildings. The
invariance properties of SIFT in this case was used to identify the locations that are
invariant to the scale change of the image. This was achieved by transforming the images
to the scale space. In order to implement graph matching between building templates and
test images, they created key-point descriptors as matching media. This key-point
descriptor was composed of the key-points and the magnitudes and orientations of the
image gradient in the surrounding area. The success of this study lies in that the multiple
sub-graphs matching algorithm can be used to detect buildings and urban areas within
various environmental (background) situations. However, window selection is very tricky in
SIFT because different window sizes will extract entirely different information from images
and hence generate different results. However, the methods proposed by Huang et al.
(2008) can overcome this problem.
3.1.3 Research Questions
All of the research described above displayed the enhancement of classification accuracy
in three aspects: 1) progressive application of information embedded within various scales
(resolutions); 2) resolution of the interference problem inherited in high-resolution imagery,
and 3) involvement of contextual information in mimicking human perception. However,
most of these studies were executed on small and homogeneous study sites. Most
approaches executed in the aforementioned studies were either restricted to the analysis
51


of spatial interactions over relatively small neighborhoods on a single scale (Shankar,
Meher, & Ghosh, 2007) or executed for paved roads in the urban areas. Large study areas
normally encompass more than one type of land cover, featuring a sizeable variation of
terrain formations. This land cover variation challenges the applicability of these
classification techniques in detecting ATV trails for a large area. A second challenge
comes from the fact that the behavior of OHV riders in the field is nonlinear. One of the
common assumptions in map-matching is that the vehicle is essentially constrained to a
finite network of roads. While this assumption is valid for most vehicles under most
operating conditions, it does not hold for off-road situations in areas such as car parks or
on private land. A third difficulty is that desert environments contain a great amount of bare
ground, so projected shadows can be easy to be confused with the ATV trails.
A multi-scale, object-based classification method is thus proposed for this study. An
object is defined here as a group of spectrally similar contiguous pixels. With a careful
segmentation scheme to divide imagery into objects, object-based derived classes should
represent a physically or ecologically homogeneous land class. One reason that object-
based methods perform well in classifying high-resolution images is because once the
object is created with a segmentation approach, more features, such as geometrical (e.g.
shape and width) and topological (e.g. relationship between objects) properties, can be
extracted from segmented images. This information is particularly useful in classifying high
spatial resolution images because these images often contain relatively fewer spectral
bands, (e.g. CIR, NAIP, QIUCKBIRD) as compared to coarser images (e.g. MOIDS,
Landsat ETM+). Indeed, object-oriented classification has been widely discussed in the
field of remote sensing.
3.2 Study Area
The study site, located in the Mesilla Valley in southern New Mexico, just borders the city
of Las Cruces (-10645 Longitude / 3220 Latitude). In contrast to the small sites chosen
in prior research, our research was executed on quite a large area; around 42 square
kilometers. The site is composed of lands managed by BLM and the state and can be
accessed from major city streets or community roads. This site is a transition zone where
an urban area and nature meet. In fact, commercial and residential land development has
been steadily encroaching into the study area, taking over areas that were originally
rangeland. Towards the east and towards the south of this site, a wild zone running south-
and-north by highway I-25 can be clearly identified from satellite images (Figure 3.2). This
site is characterized by its easy accessibility in terms of location, smooth connectivity to
adjacent recreational zones, and dramatic variability in geomorphologic alteration. It has
thus has attracted many recreational transportation-related activities. In addition its
uniqueness of terrain, the Las Cruces site was selected because it reflects
demographically affected patterns of recreational travel in arid regions of the western
United States. In this study site, recreational travelling was observed either parallel to or
cutting through desert washes, generating a wandering feel or a more up-and-down,
exhilarating style, respectively. In addition to these two major types of ATV routes,
squiggly" trails of infrequent use were found among the bushes. One of the purposes of
detecting social routes in this study site is to explore possible challenges being faced in
creating a route inventory in the semi-arid area in the western US. Indeed, significant
controversy in the development of travel plans has been faced by the organizations such
as BLM. The creation of a route inventory will be able to provide data infrastructure for
52


recreational transportation model simulation and benefit organizations adapting their
management framework to an environment of conflicting travel demands. Such demands
include OHV riding, mountain biking, hiking, hunting and fishing, oil, gas and mineral
developing, gathering of forest and grassland products, and a variety of passive uses such
as bird-watching and picnicking. Projections of future growth suggest that rapid population
growth trends in Las Cruces are likely to continue. The region has a diversity of resources
that may be affected by social roads and travel. These include a variety of biological
resources, encompassing both terrestrial and aquatic ecosystems, and significant cultural
resources.
53


3.3 Data Used and Data Preprocessing
3.3.1 Data Used
The dataset includes CIR DOQQ (Digital Orthophoto Quarter Quadrangles) and LANDSAT
7 ETM+. As the standard DOQQ products of USGS, CIR is a one-meter aerial photograph
in which image displacement caused by terrain relief and camera tilts have been removed.
CIR DOQQs are the products of the National Aerial Photography Program (NAPP). They
are acquired at 20,000 feet above mean terrain with a 6-inch focal length lens. The flight
lines are quarter quad-centered on the 1:24,000-scale USGS maps. The CIR DOQQs were
taken in 2005 and have three spectral bands (green, red, and near-infrared). In this project,
the accompanying RGB (Red, Green, and Blue) was also downloaded from the New
Mexico Information Technology Commission website. A mosaic sequence and subset on
the downloaded DOQQs was performed so that the size of the imagery was restricted to
the size of the study area. This size of the image almost reached the maximum dimension
that object-oriented remote sensing software (Definiens Professional 5.0) can process.
The Landsat ETM+ imagery used in this study is an USGS Global Land Survey (GLS)
2005 data set. The main portions were captured in June 8, 2005. The GLS 2005 is the data
set that has been developed with the collaboration of NASA and USGS. However, ETM+
images in GLS 2005 have 22 percent data loss due to the scan-line corrector failure in
54


2003. Therefore, the GLS 2005 data of ETM+ actually includes the combination of two of
the ETM+ images for each path/row, so that complete or at least near complete coverage
can be obtained. Furthermore, to fill areas with low image quality or excessive cloud cover,
data recorded in 2004 and 2007 may be used as needed. Nevertheless, these combined
products will not seriously impact trail detection work, because the Landsat imagery, with
its medium spatial resolution, was merely assigned an auxiliary role in this research.
Field data was collected in the summer of 2008 to provide training for calibrating ATV route
detection. The CIR hardcopy images were brought to the field, with the assistance of GPS,
to identify the objects corresponding to the trails. For the initial field data collection,
observations were limited to areas within 70 m of major trails, which were more accessible
and contained more disturbance-related trail features. Though these field-based
observations were conducted three years after CIR 2005 was captured, high consistency
between the bright linear features in the images and the true trail features across the
terrain can still be found, regardless of the small geomorphologic changes from human
activities.
3.3.2 Image Preprocessing
Some important information was derived by preprocessing the original images. These
derived information layers were then inserted into the Definiens software for the synthetic
classification operations. Image derived variables were used (1) to serve as a vegetation
index in detecting trails, (2) to minimize the contrast between the trail edges and adjacent
features so that connectivity of trails can be more observable, and (3) to remove the
interference existing in the imagery and perform de-correlation contrast stretching among
the bands. This clarified the trail features scattered among the other features in the
landscape. To meet these ends, Normalized Difference Vegetation Index (NDVI)
calculations (1), image enhancing using the convolution filters (2), and principal component
analysis were used (3). These image preprocessing operations were implemented with
ENVI remote sensing software.
There are many spectral vegetation indices which can be used to determine the
percentage of vegetation cover and can thus be applied as an indirect reference for
exploring trail formations. In a semi-arid area like the study area in Las Cruces, where
vegetation is sparsely distributed, compensation for both soil background effects and
senescent grasses using the soil line concept is crucial (Purevdorj et al. 1998). A variety of
vegetation indices were generated to test bare ground impacts and efficacy in identifying
trail features. The vegetation indices generated in this study include a simple ratio based
on the CIR DOQQ, a Normalized Difference Index (NDI) based on the Landsat ETM+, a
NDVI obtained from CIR 2005, a NDVI obtained from Landsat ETM+, and a Soil and
Atmosphere Resistant Vegetation Index (SARVI) obtained from Landsat ETM+. Table 3.1
lists the image-derived vegetation indices that were generated for the purpose of testing
their efficacy in identifying trail features. After reviewing vegetation indices of apparent
route features and non-disturbance features, we found that none of these indices can be
entirely depended on to detect trails. The NDI and SARVI were especially incapable of
generating practical vegetation information for the study in this semi-arid environment. This
discovery reflects the findings by Bork et al. (1999) that the Soil Adjusted Vegetation Index
did not increase understanding of the variance between spectral and ground
measurements in complex natural environments. However, the Simple Ratio and NDVI,
55


obtained from CIR or Landsat ETM+, still can provide supplemental information in
determining if objects are trail features.
Table 3.1 Spectral vegetation indices generated for testing
Simple Ration (CIR) NIR / Red
NDVI (Landsat ETM+) (NIR Red) / (NIR + Red)
NDVI (CIR) (NIR-Red)/(NIR + Red)
Normalized Difference Index (Landsat ETM+) (NIR-Blue)/(NIR + Blue)
Soil and Atmosphere Resistant Vegetation Index (Landsat ETM+) 2.5 (NIR Red) / (1 + NIR + 6 Red -7.5* Blue)
CIR DOQQ is an aerial photo, showing only blue, green, and red bands in the visible
range. Executing classification on a 3-band image with such a high spatial resolution will
result in severe speckle effects and confusion between classes (Cleve et al., 2008). To
reduce the interference and generate more interpretable data, principal component
analysis (PCA) was implemented on the RGB 2005 data. In the least squares sense, the
principal components obtained after using PCA supply the optimal representation of the
variance of the original image. Figure 3.3 shows that the resulting first component has
captured more than 95% of the variability of the original data.
Figure 3.3 Eigenvalue Spectrum of the Principal Components
To increase the tractability of trail features, a low pass filter and a local sigma filter were
executed on the principal components of RGB 2005 to decrease contrasts and highlight
homogeneous information within the class. By using the local standard deviation calculated
for the filter window to determine valid pixels, the local sigma filter replaces the pixel being
filtered with the mean calculated, using only valid pixels within the filter box. This
56


manipulation will dramatically enhance the results of image segmentation so that it is more
corresponding to the true existence of the ATV routes. The comparison of the results,
between image enhancement and the raw image, is shown in the Figure 3.4. Random
interference is significantly reduced.
Figure 3.4 Comparison of the Original Image (Right) with the Result after Executing
Local Sigma Filter on the First Component of the RGB 2005 Data (Left)
3.4 Methodology
3.4.1 Segmentation Method
Image objects are the basis for object-based classification and trail retrieval. An object is
defined as a group of spectrally similar contiguous pixels. Numerous algorithms are now
available for segmenting an image into image objects. In this study, we used the
segmentation method from the Definiens software, which is based on a multi-resolution
segmentation algorithm. In the multi-resolution segmentation method, the segmentation
results are tuned based on scale parameters, color, shape, smoothness, and
compactness. The scale parameter in Definiens is an abstract term which determines the
maximum allowed heterogeneity for resulting image objects (Definiens Professional 5
reference book). The algorithm of calculating the homogeneity/heterogeneity index for the
scale parameter was described in the paper by Benz et al., (2004). In that paper, a
heterogeneity index f was defined as the weighted combination of spectral heterogeneity
(Ahcolor) and shape heterogeneity (Ahshape): f = WC0|0r AhC0|0r + Wshape Ahshape, where WC0|0r
e [0, 1], Wshape e [0,1], and Wshape + WC0|0r = 1. While color heterogeneity still represents the
primary information contained in image data, shape heterogeneity in the heterogeneity
index calculation defines the textural homogeneity of the resulting image objects by
combining magnitudes of smoothness and compactness. In this trail detection study, the
smoothness criterion was set at a relatively high level, because DOQQs data is very
heterogeneous and trail features always have smooth borders. Actually, there are no
57


concrete rules for setting these criteria in the image segmentation process. A systematic
trial-and-error approach is employed in determining segmentation parameters. By
evaluating consistency between trail features displayed on the image and the segmented
linear objects in each run, the optimal parameters will be eventually derived. A portion of
the final segmentation results of the Level 2 are shown in the Figure 3.5.
The image is segmented at five different resolution levels in this study. Each level of
segmentation has its specific purpose. However, they are all used in facilitating the
detection of ATV objects at the high-resolution levels (Level 1 and Level 2).
3.4.1.1 Level 5 Segmentation
Fifth level segmentation is terrain-level segmentation. The objects delineated at this level
are like the fragmented patches caused by the ATV roads, terrain geomorphologic
variances, or land cover changes over the terrain (Figure. 3.6). In addition to possible uses
58


in the fragmentation index, the goal of level 5 segmentation is to furnish coarse object
feature information to level 4 classification operations. Level 5 object information will
significantly enhance the classification of a certain kinds of land uses at level 4, so that
water towers, developed areas, landfills, and ATV zones can be clearly identified.
3.4.1.2 Level 4 Segmentation
Due to the fact that ATV routes would never occur in certain kinds of land uses, such as in
power plants or in residential areas, fourth level image segmentation aimed to create
objects that could be clearly and specifically assigned to a land use class. The settings for
Level 4 were a scale parameter of 204, color/shape weighting factors of 0.6/0.4, and
compactness/smoothness of 0.8/0.2. The configuration of the segmented objects is shown
in Figure 3.7. In this level, elongated trail objects also corresponded to the major roads in
the real world.
59


Figure 3.7 Image Segmentation Result in Level 4
3.4.1.3 Third to First Level Segmentation
The third level to first level image segmentations aimed to generate objects corresponding
to existing trail and ATV features of various sizes in the real world. In addition to
configuring objects suitable for trail detection, third level segmentation aimed to identify
objects that could enhance ATV zone (bulb type) extraction. In this sense, smoothness
would not be the major concern for the third level segmentation. However, in order to
configure objects which are more trail formation oriented, more weight was put on the
smoothness criteria in level 2 and in level 1 segmentation. Therefore, in second level
segmentation, a smaller scale, 27, was used to produce finer objects. The shape/color and
compactness/smoothness weights were set at 0.7/0.3 and 0.1/0.9, respectively. The first
level segmentation, on the other hand, aimed to generate objects suitable for retrieving
narrower trails. The settings for the first level were scale parameter of 12, color/shape
weighting factors of 0.8/0.2, and compactness/smoothness of 0.3/0.7. The segmentation
result of the first level is shown in the Figure 3.8.
60


Figure 3.8 Image Segmentation Result of Level 1
3.4.2 Multi-Level Hierarchical Classification
In this hierarchical classification process (Figure 3.9), segmented objects were categorized
using a supervised fuzzy classification technique. By exchanging the range of feature
values into values between 0 and 1 in fuzzy membership, the fuzzy supervised
classification algorithm lets the users set up a threshold of feature value, which would then
form one of the rule sets for object classification. The formation of ATV trails across terrain
is multifaceted. Considering the interfering effects of shadow and bare ground existing
between the bushes, the extraction of trails from aerial photos is a difficult job and cannot
simply rely on a few membership expressions. In object classification operations, these
membership expressions were organized using operators such as and, or, and not. To
overcome the challenges of detecting road features from images full of interference and
the impacts of the inconsistent image quality, each level of classification contains a variety
of ATV zone categories, which comprise the full range of trail/bulb formations under
different situations of geomorphology and surroundings. For example, there are nine
categories of ATV disturbance in Level 3:
61


Segmentation
i
Layers
Objects with
Different
Resolution
Hierarchical
Classification
(Rule Setup)
I
1 Jl 1
Level 5
Classification
Level 4
Classification
Level 3
Classification
I
Classes furnish
information that helps
to classify LULC in .
Level 4
Classes of
Land Use Land
Cover
Various ATV
Classes of
Bulbs, Roads,
and Trails
Level 2
Classification
Classes of Roads
& Trails i
Level 1
Classification
Trail Classes (<
- 3 m)

^-/^Synthotic Bulb/Trail
Evaluation
Ancillary Data for
Attribute Data
Filling
Database
Management
Roads
Classification
Figure 3.9 Diagram of the Hierarchical Classification
62


1) ATV Bulbs 1
The objects assigned to this class are those objects whose super-objects were classified
as ATV zones in Level 4 but had relatively high red-band values and lower standard
deviations in near-infrared in Level 3. This class was also assigned to the objects whose
ETM+ band 5 was relatively high and whose standard deviation of the CIR Near Infrared
band was relatively low. TM band 5 is involved in extracting bulbs because the Landsat
band 5 (1.55 1.75 pm) is sensitive to variations in water content, both in leafy vegetation
and soil moisture. (Short et al., NASA, 2010)
2) ATV Bulbs 2
The class of ATV Bulbs 2 was created to extract objects that did not possess high red band
brightness value but had either relatively high brightness values in ETM+ band 5 or lower
standard deviations in the near infrared band of the CIR. Obviously, these objects were
wider and possessed lower NDVI.
3) Brighter Roads
Objects that can be classified as Brighter Roads are those that have high brightness
values in the red band of CIR and elongated shapes. NDVI of CIR was also an important
ancillary factor used to extract Brighter Road objects. In the field, these Brighter Roads
were found to be routes having high intensity of ATV use.
4) Darker Roads
Not all ATV trails look bright on the imagery. Instead, due to the quality of CIR DOQQ and
the variance of surface tone of the trails, some of the trail objects cannot be detected by
simply depending on the brightness value and the standard deviation of the brightness
values. The purpose of the creation of the class Darker Roads was to discover hidden
trails which blended with surroundings and were difficult to identify as other classes. The
extraction of this type of object depended on the rule sets that were emphasized on
elongated shape, consistency of the area represented by segments, higher contrast with
neighbors, low NDVI, and narrow object width.
5) Roads from Level 4
The purpose of the class Roads from level 4 was to categorize those objects whose
super-objects had been classified as routes in level 4 and possessed relatively high
brightness values of the red band of the CIR in level 3.
6) Roads_01
Most the trail objects found in the CIR were not extremely bright. Instead, with the
intermediate brightness values of the red band, this type of object was distinguished by its
consistency in color, elongated shape, low NDVI, near proximity to the edge of super-
objects, narrowness in width, high contrast to neighbors, or higher brightness values in
band 5 of LANDSAT ETM+. Thus, the class of Roads_01 was created to classify trail
features whose reflectance in the CIR red band was between 146 and 177.
7) Roads_02
The class Roads_02 was targeted at objects which were stretched with relatively high
contrast to their neighbors.
63


8) Roads_03
The purpose of the class Road_03 was to identify those routes which were evaluated as
undisturbed by the above classes due to shadows or other reasons, but actually were
inventoried as roads in the existing trail dataset or have been identified as roads in the field
campaign. There are not many of these types of roads, but they contain very important
information in observing ATV riders behavior.
9) Trails
The class trails serves to identify those routes narrower than 3.2 meters.
3.4.3 Feature Extraction
In this study, many object features were interchangeably used in detecting ATV trails and
bulbs, including mean value, standard deviation, maximum pixel value, mean of the inner
border, mean difference to darker neighbors, mean difference to super-objects, density,
area, elliptic fit, border index, radius of largest enclosed ellipse, shape index, relative inner
border to super-object, degree of skeleton branching, length/width of main line, standard
deviation of area represented by segments, width of main line, shape texture based on
standard deviation of the direction of the sub-objects, GLCM standard deviation, GLCM
correlation, average mean difference to neighbors of sub-objects, distances, and so on.
Among applied features, some are easy to be comprehended, while others can only be
understood with the mathematical functions, defined using the following equations.
Table 3.2 Selected features in object-based classification
Feature Equation and Explanation
Density V'#Pv
1 + V(VarX + VarY) Where V#Pv is the diameter of a square object with #PV pixels included within the classified image object and V(VarX + VarY) is used for the approximate diameter of the ellipse. This formula makes density an effective index in evaluating the objects compactness. Therefore, the closer the outline of an image object to a linear shape, the lower its density and more likely it is to be classified as a trail.
Elliptic Fit #{(x, y) e Pv: Ev(x,y) < 1} q> = 2* -1 #PV Where ev(x, y) is the elliptic distance at a pixel (x, y) and #PV is the total number of pixels contained in an object v. cp in this equation is the probability that the considered object fits into the created ellipse. If the considered object completely fits, cp would be equal to 1.

64


Table 3.2 (Cont)
the created ellipse. The trail features, with their elongated and smooth formations, are hard to fit into created ellipses and thus possess relatively small cp value. In fact, many evident ATV trail objects in the study area have the (p values close zero.
Relative Inner Border to Super-Object I b (v, u) u e N.Yv) bv Where Nu(v) is the neighbors of the object v that exist within the super-object and bv is the image object v border length. The value of the index ranges from 0 to 1. If the index value is 0, it means that the image object v and its super-object are the same object. If the index value is 1, the object v is thoroughly enveloped by its super-object. Due to the fact that trail networks often cut through the terrain, fragmenting the landscape into patches, trail objects segmented in level 2, 3 and 4 were often found to be located on the edge of the super-objects. In other words, the Relative Inner Border to Super-object feature of the trail objects is prone to be lower than 0.65.
Standard Deviation of Area Represented by Segments This feature is generated by calculating the standard deviation of all triangles created by the Delaunay triangulation within the object. The trail objects always have a very consistent width and therefore show relatively low variance in segmented areas with the Delaunay triangulation. Bulb objects, however, have comparatively high values of Standard Deviation of Area Represented by Segments.
GLCM Correlation The purpose of this index is to measure the linear dependency of the gray levels of neighboring pixels. The calculation is based on the gray level co- occurrence matrix (GLCM). The higher the GLCM correlation, the more consistent the texture of the considered object and the more likely it is to be deemed as an ATV-used object. I Pb j L V((CTi )*(Oj )) J ij=0 Where i and j are the row and column numbers,
65


Table 3.2 (Cont)
respectively. P,7 is the normalized value in the cell i,j. N in this equation is the number of rows or columns, py and ay is the GLCM mean and GLCM standard deviation, respectively. By calculating how often different combinations of pixel gray levels occur in an image (object), GLCM is one of the very effective ways employed by DEFINIENS software to calculate texture.
GLCM Standard Deviation N-1 <*= J I V ij=0 Where i and j are row and column numbers, respectively. Py is the normalized value in the cell i,j and py is the GLCM mean. N in the equation represents the number of rows or columns. Since GLCM is used in this feature calculation, GLCM standard deviation deals exclusively with combinations of reference and neighbor pixels. Therefore, it is not the same as the simple standard deviation feature of the objects. Instead, it is analogous to contrast. The application of GLCM standard deviation can supplement the simple standard deviation feature in detecting the ATV trails, especially when the target object is not in the clearest of situations.
Standard Deviation of Area Represented by Segments Divided by Width (SD_A_Div_W) This customized shape variable is very effective in identifying linear features. The smaller the value, the more elongated the shape. If the object a value of SD_A_Div_W less than 4.6, the object has a high potential to be detected as a trail class. This customized index is composed of two parts of the shape indices. The first portion of the index, standard deviation of area represented by segments, is obtained by calculating the standard deviation of all triangles created by the Delaunay triangulation for the object. On the other hand, the second portion, width, is obtained by calculating the average height h of all triangles intersected by the main line. An exception is triangles in which the height h does not cross one of the sides of the corresponding triangle. In this case the nearest side s is used to define the height.
66


3.4.4 Classification Rules
The hierarchical classification algorithm employed in this study can evaluate the
membership value of an image object to a list of selected classes. The class description is
composed of membership rules. One of the rule sets for classifying Roads_01" in Level 3
is demonstrated in the Table 3.3, as an example. The multilevel rules of class description
are organized by operators to imitate the intricate configuration of ATV trails and bulbs in
the real world.
Table 3.3 Rules used for the Roads 01 class
Roads_01 in the Level 3 Classification
o c GLCM Standard Deviation of CIR05 red band in all directions must be less than 26.8.
The mean CIR05 red band of the object must be greater than 156.
Mean Difference to Super Object should be greater than 20.05.
Mean of Landsat 7 Middle Infrared (band 5) should be larger than 99.
Mean of NDVI should be less than -0.061.
The object should not have a classification value of ATV Bulbs 1, of level 3.
The object should not exist within super-objects which have been classified as Developed High Intensity class in the Level 4.
The object should not exist within super-objects which have been classified as Developed Low Medium Intensity" class in the level 4.
The object should not exist within the super objects which have been classified as Industrial Use class in the level 4.
o d c The feature value of Distant Intensity Relative Border Products should be greater than 12.5.
Relative Inner Border to Super-object (1 Level Distance) must be less than 0.54.
73 C The feature Degree of Skeleton Branching must be less than 3.
The feature Mean Difference to Darker Neighbors with CIR05 red band should be greater than 52.7.
The feature Compactness must be greater than 3.6.
The customized index SD of Area Represented by Segments Divided by Width must be less than 4.3.
The feature Radius of Largest Enclosed Ellipse" must be less than 0.492.
The feature Relative Border to the class ATV Bulbs 1 should be less than 0.25.
The standard deviation of the ETM+ green band must be between 1.59 and 3.38.
The feature Std. Dev. of Area Represented by Segments must be between 47.8 and 94.3.
The width of main line must be between 4.13 and 9.4.
67


3.5
Results
3.5.1 Validation
The accuracy and reliability of the route inventory products were examined based on a
composite validation method. The composite reference data of routes are collected through
GPS field operations, ATV route digitization, existing BLM designated routes, and the
qualitative visual analysis. However, collection of sample data through field operations is
always time-consuming and expensive. The approach was adapted because of the
difficulty of randomly collecting sufficient reference data through GPS field operations,
particularly for such fine spatial resolution products (Stow et al., 2004). Based on the
observed recreational travel pattern of ATVs and the geomorphology of the study area, the
route samples were roughly stratified into three categories:
1) Trails and Bulbs. This category includes the major routes designated by BLM, the
articulated (or highly intensified used) trails detected by the object-based
classification method, and larger size bulbs. Generally, these trails and bulbs form a
transportation network across the terrain of the study area.
2) Dispersed Use. Outside the networks of category 1, there are still many trails and
small-sized bulbs, although not frequently used, which can attract ATV riders and
have the potential to extend their traces through time. For the convenience of later
KAPPA analysis, these routes are defined as the low usage intensity trail features
that situate outside the buffer zones of the category 1.
3) No Disturbance. The undisturbed areas are those sections where no significant signs
of ATV activity can be found. However, except those absolutely undisturbed
features, such as wetlands, high density shrub areas, and bare ground without any
signs of ATV activities, there are still some suspicious trail features with low usage
intensity that possibly exist among the bushes. Examination of these suspicious
areas will greatly increase impartiality in the accuracy assessment.
Trails and bulbs are the major veins that convey the greatest portion of traffic. Along the
buffer zones of the trail-and-bulb network, the ATV activity area has increased evidently in
size over the years. Many of these trails havent been designated by BLM, and yet they
either connect to important ATV play areas or possess the specific features that attract
ATV riders. In addition to the trail-and-bulb network and its vicinity, the widespread areas
fragmented by buffering zones of the trail-and-bulb network are the grounds of small-sized
bulbs and infrequently used trails. To verify the effectiveness of the object-based
classification algorithm in detecting ambiguous ATV trails, these infrequently used trail
features located in the fragmented areas represent the concentration of field validation
work. The bright features located along desert washes need to be checked for the trail
signs. Not all of the desert washes can be classified as ATV routes.
Ten validation sites were selected for field work. The size of these validation sites range
from 96,000 square meters to 285,000 square meters. In each validation site, around 12
validation points were randomly selected from the classified ATV trail features inventory, as
well as around three to six validation points that were selected from the classified no-
disturbed objects for field validation work. Unlike the randomly selected ATV trail features,
the non-disturbance validation points are deliberately chosen from suspicious trail features.
They may have a very low brightness value, but they possess trail-like shapes and
68


textures. The validation of these suspicious points in the field helps uncover possible
deficiencies in the classification rules setup in object-based classification operations. If the
misclassification is proportionally large, the rule sets for the object-based classification may
need to be further adjusted. In addition to the randomly selected validation points in the ten
validation sites, the GPS validation points obtained through the operations within the 100-
m buffer zones along the major trails and desert washes in 2008 is also used for validation
(Figure 3.16).
Figure 3.10 Ten Strategically Selected Validation Sites
3.5.2 Sign of ATV Activity
In the field, there are various signs that can be utilized to determine whether a validating
target is an ATV disturbance area. Major signs include ruts (Figure 3.11), major routes
(Figure 3.12), uprooted vegetation (Figure 3.13), connectives (Figure 3.14), and straight
vegetation edges (Figure 3.15).
69


Figure 3.11 Ruts
Figure 3.12 Major Routes
70


Figure 3.13 Uprooted Vegetation
Figure 3.14 Connective
71


Figure 3.15 Straight Vegetation Edges
Figure 3.16 GPS Validation Points Collected in 2008 with CIR 2005 Photo as Background
72


3.5.3 Accuracy Assessment
Accuracy is calculated based on the comparison of randomly selected features with
composite trail reference data. The composite trail reference data was assembled by either
GPS field operations or qualitative visual interpretation. Many trails are so obviously
presented on the image that the objects located within these trail formations can be
definitely categorized as ATV trails/bulbs and thus, field validation is not necessary to
evaluate their accuracy. However, to more objectively evaluate trail detection results, more
than 90% of the composite reference data consisted of true GPS field points. A confusion
matrix (predicted classes versus observed classes) was constructed to assess the
accuracy of the final trail inventory. Users accuracy, producers accuracy, overall
accuracy, and the Kappa coefficient will be computed and analyzed (Congalton 1991).
Kappa values take into account agreements that can occur by change (expected
agreement). In general, Kappa values of 0.75 and higher are considered good
classification results. KAPPA analysis yields a Khat statistic that is a measure of agreement
or accuracy (Rosenfield and Fitzpatrick-Lins, 1986). The Kha, statistic is computed by
, _ NXXn X (Xl+ X+j)
hat N2 Z(Xi+ X+i)
In order to execute KAPPA analysis and fit it into distinct ATV styles found in the field, the
ATV reference data was reorganized into three classes. These three classes were bulbs,
trails, and dispersed use. The calculated Khat and overall accuracy are listed in Table 3.4.
Table 3.4 Error matrix
Bulb Trail Dispersed Use Row Total
Bulb 13 2 1 16
Trail 0 45 0 45
Dispersed Use 1 0 38 39
Column Total 14 47 39 100

Khat 0.9349
Overall 0.9600
3.6 Conclusions and Discussion
3.6.1 Conclusions
This study developed a multi-level object-based classification approach for detecting ATV
trails in a semi-arid area adjacent to Las Cruces, New Mexico with high resolution DOQQ
and medium resolution Landsat imagery. The result indicates that multi-level object-based
classification methods can generate trail inventory with very high accuracy. Kappa value is
around 0.9 in this study. The greatest advantage to this method is the good detection
performance in heavily textured and obstructed environments and the ability to detect
elongated structures. It can be considered a supporting approach for creating trail
inventory to resolve the limits of trail detection as mentioned by Witztum and Stow (2004)
in their research analyzing direct impacts of recreational activity.
73


In the Las Cruces site, for example, the rule sets used for detecting trails in the open
meadow area were not necessarily competent in identifying trail objects located in scrub
areas. When the imagery used contains only limited spectral information, such as CIR
used in this study, detection work is even more challenging. The rule sets need to be
organized systematically, so that all situations of trail episodes can be covered. The points
below are the findings of this research regarding the application of the object-based
classification algorithm in trail detection for a large area.
1) Some object features are especially influential in detecting trails. These features
include density, degree of skeleton branching, length/width, elliptic fit, standard
deviation of area represented by segments, and standard deviation of the directions
of the sub-objects. In cooperation with variables such as NDVI, intensity, mean
difference to darker neighbors, and GLCM correlation of NIR at a direction of 45,
these features can be used as independent variables, in regression analysis or in
logistics regression analysis, to figure out their relationship with new ATV use of
2005.
2) Various classes of trails and bulbs need to be created in each level that emphasize
the different reasoning of the trail and bulb formations, to thoroughly detect trails and
bulbs formed within diverse surrounding environments. The iterative and hierarchical
segmentation process has proved to be an effective means of retrieving ATV trails
and bulbs systematically in this study. However, one of the disadvantages of object-
based classification was also revealed in this study. The segmentation results were
greatly influenced by the target, data quality, the context, and the parameters (Xie et
al. 2008). These segmentation results then subsequently impact classification
results. The boundary of the trails and bulbs at high resolutions will more or less be
diverged.
3) There are a variety of methods that have been explored for linear or polygonal
feature extraction, but most of them are specific to paved roads or buildings. They
are not necessarily suitable for application to ATV trail extraction. Some algorithms
may have to be modified. Some of the areas identified by this object-based
classification method as disturbances are actually only one or two ruts that were
found in the validation process. These ruts may have been traced long ago and have
never been used since then. Should areas like this count as disturbances? Principal
Component Analysis in the image preprocessing module has proved useful for
reducing this kind of interference. However, a threshold still needs to be set up in
order to produce trail inventory that fits jurisdiction application needs. Inserting rules
containing this threshold and supplementing them with principal component
information in the classification fuzzy logic rule set should distinguish ATV trails from
uncertain features. These infrequently used areas are not considered in the
classification scheme in this study because the main purpose of this study is to
detect ATV routes and form a traveling network.
4) If including no-disturbance objects, the kappa statistic for trails and bulbs detection is
around 0.7, with the prominent error being related to overdetection of no-
disturbance objects as a certain class trails or bulbs. Most of the over-detection
happened in open meadows or the areas with a high percentage of bare ground.
74


5) The detection difficulty caused by shadows and narrowness of the trails has been
partially resolved. Some disturbance features, such as ATV trails (<3 m), were not
separable from the bare ground existing among bushes. This issue could be
addressed with the multi-sensor multi-scale approach based on Markov Random
Field theory. In higher resolutions (CIR 1 m aerial photo or QuickBird 0.6 m multi-
spectral imagery), ATV trails appear as elongated heterogeneous regions. Different
from ATV trails, bare ground would always appear as a squiggly shape having
relatively lower brightness values within its super-object. Eventually, it is possible to
clearly segregate ATV trails, bare ground, and trails for other purposes.
6) This study was executed with the 2005 aerial photo of CIR, while ground truth work
was performed in 2008 and 2009. In several spots, inconsistencies between the
imagery and ground truth were discovered. Although some of the changes were
caused by the dumping of construction wastes, a large portion of this inconsistency
actually underlined the increase of narrow trails and the expansion of original trails.
This implies that the ATV users continue to explore no-disturbance areas and that
creation of a high-accuracy trail inventory is mandatory for finding unauthorized trails
at an early stage to facilitate policy intervention.
7) The algorithm of the object-based trail classification work described in this paper was
implemented with Definiens software installed on a workstation equipped with
Dual-Core Intel Xeon processors. The computer has 3 GB memory, but still, the
software easily crashes due to the fact that the object-based classification quickly
consumes the calculation capability of the computer. For example, for the study area
of 42 square kilometers, which is composed of around 180,000 objects in level 3, the
classification task can take as much as 15 hours until hierarchical classification has
converged. To efficiently execute object-based classification in a large area, a
computer equipped with a large amount of memory is crucial.
8) The object-based classification method applied in this study is a hierarchical
classification algorithm and requires substantial waiting time while setting and testing
rule sets, although it can generate an effective trail inventory. A recently released
tool (FeatureObjeX, PCI Geomatics Inc.) can lessen this challenge by using
machine learning technology. According to He et al. (2009), this tool seamlessly
blends automated feature extraction algorithms with interactive intelligent editing
tools to overcome possible interference caused by various topographies and
vegetative configurations. Attention should also be paid to enhancing understanding
of the nature and magnitude of potential uncertainties induced by imagery quality
and by the terrain itself; factors such as shadows and misclassified bare ground. It is
vital to take these uncertainties into account in order to identify small-size ATV
disturbances.
3.6.2 Recommendations for Future
One of the greatest constraints of our method is that it is entirely supervised. The creation
of the classification rule sets relies on the iteration of observations on many features of the
trail objects across the study area. For a large study area, this means that many tedious
trial-and-error classification operations need to be executed in order to obtain the most
75


appropriate classification rule sets. From the perspectives of data inventory and data
application, this study demonstrates a comprehensive method capable of accurately
detecting trails from obstructed surroundings and identifying features appropriate for trail
detection. These distinguished features can also be used to support the discovery of future
ATV activity growth areas. However, from the standpoint of automated trail detection, the
algorithm still has not been fully developed and much effort is needed improve the
deficiencies. The shortages of the algorithm depicted in this paper fall into two categories:
First, the method cannot automatically fill the gaps existing between the individual trail
objects to form a smooth transportation network. Second, the image segments do not
always correspond to the actual trail objects. This constraint is actually an intrinsic problem
descending from DEFINIENS.
Katartzis et al. (2001) developed an automatic method for grouping lines based on a
Markov random field model and maximum a posteriori probability criterion. This line
grouping process was defined as a global technique in their study. By integrating it with a
local analysis scheme, such as the hierarchical object-based trail detection method
described in this paper, this model-based line grouping technique should be able to
develop an automatic high-resolution trail inventory method. The integration of object-
based classification and the model-based line grouping technique is an area that should be
studied further. As for the problems in the second category, Akcay et al. (2008) developed
a new segmentation algorithm using neighborhood, spectral, and morphological
information of the objects to enhance correspondence between segments and actual
objects. To accurately evaluate the recreational traveling impacts on the environment, this
new segmentation algorithm is also an area that deserves our future endeavors. Actually,
eCognition Developer provides an ideal platform for integrating custom features (such as
the methods developed by Katartzis et al. (2001) and Akcay et al. (2008)) and the inherent
shapes, textures, and class features together.
While most of the studies strove to develop a fully automated algorithm for feature
extraction and still produce deviated or incomplete linear features, semi-automated
methods, such as the method used in this study or the method of template matching,
(Vosselman and Knecht, 1995) provides effective ways to reduce errors. The semi-
automated user interactive method can integrate different information to enhance
extraction accuracy. This can include either information of shape, texture, and topology or
information derived from data mining or statistical data analysis techniques. Sometimes,
manually editing extraction results is most efficient (Kim et al., 2004). The non-linear
characteristics of natural phenomena and human behaviors, which strongly influence ATV
disturbance formation, explain why fully automated methods always have deficiencies in
capturing entire ATV activity zones.
76


4. Evaluating Route Formation and Disturbance of
Off-Highway Travel Using Logistic Regression Method
4.1 The Phenomenon of OHVs
The environmental and social impacts caused by OHVs (Off Highway Vehicles) are varied
and many. While an individuals well being, through connecting with nature, others, and
self, is significantly enhanced through OHV recreational activities (Davenport et al., 2002;
Jang et al., 2000; Mann & Leahy, 2009), the impacts of OHV travel on the environment are
devastating. Such impacts include soil erosion, trail deterioration, vegetation density
decrease, water and air quality stress, noise pollution, habitat degradation, and social
conflicts among different types of recreation user groups (Stokowski and LaPointe, 2000).
Lovich and Bainbridge (1999) discovered that disturbances or habitat alteration associated
with OHV use caused reduction in wildlife populations. In a water erosion simulation study,
Ayala et al. (2005) found that the average annual sediment load from the stream crossing
is much higher than what is allowed by the USDA Forest Service-National Forests for
temporary roads, and that Off Road Vehicle (ORV) expeditions on steep hill-slope trail
sections contribute sediment directly to streams. An assessment of soil degradation
caused by recreational activities by Misak et al. (2002) in the Kuwait desert concluded that
compaction of soil due to pressure exerted on the soil by vehicles led to a significant
reduction in its porosity, permeability, and infiltration capacity.
In facing increasing OHV activities and a degrading environment, public land managers
now are trying to resolve a contentious issue: designating trails that can fit OHV riders
preferences and interests while mitigating potential adverse effects (Snyder et al., 2008).
Trail designation is complex and needs to consider riders preferences and the bearing
capacity of the environment. Fail to meet riders preference and behavior patterns, an
optimum land management objective would not be reached. To evaluate environmental
and social impacts induced by the proposed OHV network, which should consist of an
existing set of trails and future planned expansions, a checklist is normally used by local
governments. Lewis County in New York State, for example, created the Generic
Environmental Impact Statement (GEIS) as a template for future decision making, in order
to bring consistency and predictability to the trail expansion process. GEIS is intended to
organize and economize the Countys decision process; establish criteria for simplifying
future impact assessment, pursuant to related environmental quality acts; enhance sound
environmental planning by allowing consideration of mitigation and alternatives at an
earlier juncture where there is greater flexibility; providing early guidance on significance
determinations; and providing public disclosure of agency considerations used in
environmental decision making (Lewis County, NY, 2007). However, predicting future OHV
activity formation in terms of location based on the checklist format to fit policy alternatives
requirement such as complete trail closure or complete trail openness is still difficult.
The field campaign performed in Las Cruces, New Mexico in 2008 and 2009 witnessed the
complexity of formation logic in creating new OHV zones. This complexity can be seen
from the varying, complicated shapes of OHV zones that have been engraved through the
landscape after years of activity. With the shapes as either stretched trails or as clustered
77


plazas, the forms of OHV zones are in fact the representation of the two major behavior
types of riders. These two major behavior types correspond to 1) following closely on
existing roads to reach destinations or 2) exploring any interesting terrain that features
spectacular geomorphologic changes, trail or no trail. Though the riders behavior is
actually regulated by preference, topography, social rules, casual choice, and skill level,
the shapes of OHV zones is an implication of riders behavior patterns and can be
quantified. For example, behavior that involves traveling straight, crossing streams,
following ridges or canyons, and creating shortcuts is shown in the terrain with the shape of
stretched trails, whereas the behavior indicating riders that are testing skills can easily be
found in the areas with those spectacular geomorphologic changes. Other preferences,
such as routes that have scenic views, loop trails, or connected paths that allow for long-
distance riding (Snyder et al., 2008), can also be conformed to corresponding
shape/texture indicators. To apply these shape/texture indicators for the assessment of
preferences of off-road travelers, object-based land use change data generated by
Definiens software was used for this research project. These data were separated into
radiation, geomorphology, and route configuration categories.
Many types of contextual data can be extracted to evaluate OHV impacts. Groom et al.
(2007) calculated the density and survival of Astragalus magdalenae var. peirsonii based
on the navigation transect system to quantify the impact of OHVs. Tull and Brussard (2007)
used a Fluctuating Asymmetry index to quantify the responses of the western fence lizard
to the stress originated from OHV activities. While these methods can be efficiently applied
for quantitative impact assessment, they cannot be used to uncover the logic of OHV zone
formation. To imitate the semantic perception style of riders and to account for context
information, such as location, fragmentation, appearance/disappearance,
expansion/contraction, deformation, and texture, object-oriented analysis has been
extensively utilized by the researchers. For example, Gamanya et al. (2009) used spectral
information, vegetation indices, shape/texture features, relations to super and sub objects,
and a priori land cover classification system to detect land use change for the city of
Harare, Zimbabwe. The applications of object-based classification are very broad.
Bontemps et al. (2008) segmented composite time series daily SPOT-VEGETATION
images into objects. By creating the signature of reflectance and temporal correlation for
each object, They were able to compute the Mahalanobis distance between the signatures
of each object and eventually generate the probability of changed to unchanged of each
object for the Brazilian Amazonian forest in the time period of 2001 to 2004. However,
despite the fact that many object-oriented studies have been executed for land use land
cover change studies, it is rare for these studies to be utilized in investigating the formation
of trails, especially in the realm of interpreting the semantic meaning of the OHV objects.
4.2 Background
There are a variety of ways in which preferences for travel patterns can be evaluated:
through surveys (e.g. Crimmins, Colorado State Parks OHV Program, 1999), observation,
or land use data analysis (e g. Snyder, et al., 2008). These methods all have benefits and
negatives in assessing travel behavior in an open landscape context. Survey methods offer
benefits like an opportunity to collect rich information about individual users. Jurisdictions
usually use these methods to cooperate with local governments, private land owners, local
ATV clubs, state OHV associations, or individuals affiliated with the ATV industry to predict
and designate ATV trail networks periodically. However, there is a significant likelihood of
78


bias or inaccuracy in self-reporting. Observational methods have the potential advantage of
accurately depicting behavior. Contextually, however, there are significant disadvantages
that relate to sample sizethat is, the number of individuals, range of activities, and area
of landscape that can be feasibly observed. Stokowski and LaPointe (2000) pointed out
that research based on experimental designs or observational techniques to study activity
types and impacts have generally failed. As a matter of fact, no matter what survey or
observation method is used, many behavior-related trail expansion patterns are difficult to
numerically model. The expansion of trails represents the riders interactions to the
environmental configuration. For instance, when trails become disturbed, wet, or rutted,
people look for easier places to travel. Therefore, trails become wider through years as
more vegetated areas are disturbed and more habitats are altered.
On the other hand, land use data analysis methods, such as the method proposed in this
paper, does avoid many of the problems related to reported bias and sampling issues. In
the data analysis area, the most recent development has been the application of
knowledge representation and statistical theories. These novel techniques possess the
capability to resolve the complexities of non-linear patterns found in natural phenomena.
Rule-based systems, fuzzy logic, Bayesian models, logistic regression model, support
vector machine, and the neural network system are the methods used in many related
studies. Venterink and Wassen (1997) developed several empirical models for the
prediction of vegetation responses in relation to hydrological and hydro-geochemical
habitat states. However, they pointed out that empirical statistical models were usually
applicable only for restricted and intensively studied areas. By integrating fuzzy theory and
mathematical morphology, Mohammadzadeh et al. (2006) developed a two-stage method
of detecting roads from high resolution images. Their method involves the application of
the spectrum information (grey level) of the pixels, size and shape patterns (granulometry),
and fuzzy if-then rules. Snyder et al. (2008) used a GIS-based least-cost method in
bringing together ecological criteria, riders' preferences, and geographical factors to
generate the optimal trail system. The result of the use of these methods is the
transformation of the non-linear behavior of riders into examinable processes by using
either knowledge-based systems or statistical methods.
Furthermore, the land use data analysis method has been applied by scientists for change
detection purposes since the time-series of remote sensing imagery became available.
Commonly used methods, such as the Tasseled Cap Transformation, Principal Component
Analysis (PCA), and Image Differencing methods, have proved to be able to automatically
differentiate between changed and unchanged locations for general land use classes.
However, the identification of the boundary of changed and unchanged areas can be very
challenging (Zhang & Zhang, 2007), especially for specific land use changes such as road
extensions and the conversion of wetland to irrigated agricultural land. Lu et al. (2004)
categorized the techniques that have been applied into seven groups and listed their
advantages and disadvantages in the table. Among the listed methods, some have been
advanced further in recent years, showing promise for locating the boundary of changed
and unchanged areas. For example, a filter integrating Markov Random Field (MRF)
segmentation, mixed Markov model, and probability theory, developed by Benedek and
Sziranyi (2009), has proved to be a robust algorithm for change detection application
suitable for high level interpretation activities. The method developed by Desclee et al.
(2006), which combined the object-based method and the multivariate iterative trimming
procedures for identifying forest land cover change in eastern Belgium, is not only able to
79


enhance detection accuracy but also to conquer the different reflection problem existing in
the channels of the multi-date images.
These novel methods underline the efficiency of involving contextual information in the
prediction process of land use and land cover changes (e.g. Gamanya et al., 2009) and
ecological analyses (e.g. Tague & Band, 2004), especially in a high resolution mode.
Radiation, shape, texture, and class-related features embedded in the objects when filtered
using fuzzy logic can become fitting variables in exploring the relationship with OHV riders
behavior. Shape/texture variables can be as simple as basic morphological perceptions,
such as area, length/width, density, or radius of largest enclosed ellipse, or as
sophisticated as the features of various gray level co-occurrence matrices (GLCM), elliptic
fit, standard deviation of area represented by segments, relations to neighbor objects, and
so on. Though information extracted from object-based classification abounds, the travel
theory behind landscape shape/texture analysis is poorly developed. The complexity of
pre-processing and interpretation required for analysis on the revolution of amalgamated
spatial forms (rather than pixel units) makes the shape/texture-based travel theory
especially difficult to develop. Object-based classification methods provide an appropriate
framework for shape analysis. These methods typically involve grouping similar pixels into
simple objects based on clustering and other algorithms, measuring the shape/texture/
hierarchy attributes of these objects, and setting up rules to depict the relationships among
these objects and with the environment as well. These rules, being composed of if-then",
and, or, and not operators, must capture behavioral or collective social processes to
be applied to social analysis.
By integrating statistical and object-oriented techniques, these novel methods have
dramatically promoted understanding of the logic of land use change in addition to
enhancing the accuracy of change detection. Likewise, the incorporated method should be
able to enhance the discovery of the logic of OHV disturbances and thus become the
means of support for policymaking regarding recreational travel management in open
lands. Logistic regression analysis is especially suitable to a situation where the response
variable is qualitative and recorded as present or absent (Franklin 1995, Collingham et al.
2000, and Kunter et al., 2004). Therefore, the purpose of this research is to explore the
applicability of the integrated method in finding OHV riders' preferences as well as
predicting OHV zone formation. The geographical revolution of OHV disturbances is
emphasized in this study because it relates to riders behavior patterns.
4.3 Approach
4.3.1 Generalized Liner Model
Utilizing the information from the objects of 1996 as predictor variables, logistic regression
analysis is executed to explain the dichotomous OHV disturbance presence/absence
values of 2005. The predictor variables of 1996 are grouped into radiation, geomorphology,
and route configuration categories. The purpose of this research is to discover the riders
preference and behavioral patterns through the creation of a predicting model. In this way,
the zones opened to accommodate future OHV activities would be practical and induced
environmental impacts can be evaluable. Data about usage intensity and trail extent is
obtained through object-based hierarchical classification from 1996 and 2005 imagery,
including CIR and LANDSAT.
80


The dependent variable, the binary presence or absence of OHV disturbance, is designed
to be the indicator of the location of OHV zone expansion in 2005. Presence/absence data
are most commonly modeled using generalized linear models (GLMs) such as the logistic
regression analysis utilized in this study with independent errors (Hastie and Tibshirani,
1990). Since the study area is located within a semi-arid region, where the existence and
density of vegetation are significantly influenced by OHV use, the bareness and dryness of
soil is presented as the respective standard deviation of Near Infrared band and brightness
value in the red band from remotely sensed data. Route expansion is measured through an
object-based inventory of route presence in two years (1996 and 2005) with a comparison
between the years to identify areas of expansion and retraction.
4.3.2 Study Site and Data
The study site is adjacent to the western edge of Las Cruces, New Mexico, forming the
edge of city development. The Las Cruces site illustrates key landscape and policy factors
influencing OHV travel debates in semi-arid and arid regions. Urban development is
occurring on the western side of the study area, and the area is used for a variety of
industrial and public purposes including recreation, landfill, construction waste dumping,
and water storage, and is the location of a power plant and power lines. There are two
major north-south OHV routes and one major east-west OHV route across this landscape.
Most of the land is comprised of bushes and bare ground although a wetland area was
also found in the southeastern corner of the study area along the desert wash. The study
site is comprised of 1,373,650 square meters or 137.4 hectares.
Figure 4.1 The Study Site in Las Cruces, New Mexico A wasteland situated in the
southeastern corner of the state between state highway 70 and I-25. It is also
the eastern edge of the Las Cruces city urban development.
81


Figure 4.2 The Study Site (Hatched Purple Rectangle Area) on CIR Imagery East-west
running desert washes form the natural course for OHV riding. This is the
area where wetland comes into existence in wet seasons. The configuration
of north-south OHV routes cutting through east-west running valleys is typical
of major OHV zones in this area.
The data sets used in this research study were derived from a variety of sources including
pre-existing trail vector data obtained from local jurisdictions, LANDSAT TM5 satellite
images, 10-m Digital Elevation Model (DEM) images, and trail inventory data generated
using rule-based object-oriented classification techniques. The preexisting trail vector data
is used for the road centerline product. The LANDSAT TM5s were used for deriving
Normalized Difference Vegetation Index (NDVI) for two years (1996 and 2005). The 10-m
DEM is static elevation data used to test the hypothesis that changes in OHV activity were
due to topographic characteristics, such as slope or variability in elevation.
The key dataset is the OHV use inventory, a polygon-based dataset delineating OHV
disturbance boundaries. The polygons were generated by performing multi-level object-
based classification on USGS DOQs. The USGS DOQs meet National Map Accuracy
Standards at the 1:12,000 scale for 3.75-minute quarter quadrangles and at the 1:24,000
scale for 7.5-minute quadrangles. The object-oriented classification system includes four
different modules: data preprocessing; image segmentation based on five levels of spatial
resolution; rule sets creation for hierarchical classification; and a module for system
82


evaluation. The data preprocessing and image segmentation modules were used to
generate suitable objects for subsequent trail classification at different resolution levels.
Organized by and, or, not operators, the rule sets in the hierarchical classification
module were designed to classify the OHV objects in various terrain situations. The
features used in these rules include layer value, shape, texture, hierarchy, and class-
related attributes (i.e. relations to super-objects). Finally, field validation was accomplished
by choosing the random samples from ten validation sites. The 2005 trail inventory was
validated in the field with close to 90% accuracy rate. A full description of the object-based
data development method and validation can be referred to the Chapter 3.
In order to execute the statistical analysis, the derived variables from the classified OHV
objects were assigned to 5m 5m cells covering the study area. Thus, a data set with a
total of 54,946 records was created. The assignment involved the application of GIS
functions such as Intersection and Spatial Join. With respect to radiation variables, if a cell
contained more than one polygon, the average was calculated to represent the radiation
attribute of that cell. With respect to shape/texture variables, if a cell contained more than
one polygon, the maximum or minimum was identified for the shape/texture variable. The
location of the study site of the 54,946 cells is depicted figure 4.2.
Figure 4.3 Landscape Overlaid with 5m Cells Aggregation of the Attributes of 1 -m
Cells into 5-m Grids
Identification of many of the OHV activity zones in semi-arid areas are easily interrupted by
distracting factors such as shadows, undisturbed bare ground that exists among
vegetation, and deteriorating picture quality. These factors make the identification of small
trails very difficult, especially when the employed dataset has only limited spectrum
information. This issue was addressed in part by collecting data at the 1-m resolution, and
assembling it in the 5-m 5-m grid framework (Figure 4.3). The spatial variability of the
variables embedded within the 5-m cells is reduced because the attributes of the contained
83


1-m 1-m cells are thoroughly conveyed to 5-m cells. This method of aggregation reduces
the impact of those distracting factors. The techniques that can be employed to reduce the
effect of these factors are numerous. Similar methods can be found in the hydrology, soil,
and remote sensing research (e.g. Merlin et al., 2008; Karnieli et al., 2008; and Goncalves
et al., 2009). The attributes embedded in the 5m cells are described in the Table 4.1.
Table 4.1 Variable description
Name Mean St. Dev. Description
Intsty96 0.033 0.686 The use intensity of 1996 was derived from the standardized Red Bands (30% were contributed by LANDSAT Red Band and 70% were contributed by CIR Red Band). BVred ~ M(BVred) Insty96 = SD(BVred)
Where BVred is the brightness value of the red band, p(BVred) is the mean of red band of the study area, and SD(BVred) is the standard deviation of the red band of the study area. The red bands of CIR and LANDSAT were entered into the formula to find the respective intensity. The 70% of the intensity from CIR was then added to the 30% of the intensity from LANDSAT to obtain the OHV intensity.
Intsty05 0.036 0.634 OHV use intensity of 2005 was created with the same method as that of 1996.
OHV96_Ag 2.684 6.200 The number of 1 m 1 m cells, which had been classified as OHV disturbances in the original object-based classification operation; in the 5 m 5 m pixel in this study. The maximum is 25 and the minimum 0.
OHV05_lnd binary binary Indicates 2005 OHV use on 5m cells. If no 2005 use occurs on the 25 underlying 1m cells, the value is 0. If one or more cells show use, the value is 1.
NDVI96 -0.063 0.030 Normalized Difference Vegetation Index of 1996.
SD_NIR96 13.745 3.438 The standard deviation of Near Infrared of the 1996 CIR imagery shows the degree of surface consistency of the cell. The lower the value, the higher the possibility of OHV use.
84


Table 4.1 (Cont)
Demcurv96_5 8.180 12.729 Indicates the slope of landscape. The higher the value, the steeper the slope. This variable was created using the Spatial Analyst curvature function on 10m DEM.
Highelev96_5 binary binary Based on the existing OHV use corridors, the OHV use of 1996 was analyzed to determine the side of corridor that OHV disturbance occurred. A value of 1 would be assigned if the OHV disturbance cell was found on the side with higher elevation, while a value of 0 would be assigned the side of corridor with the lower elevation.
Washcat96 categorical categorical Hydrological analysis on 10m DEM (Spatial Analyst functions flow direction and flow accumulation) to determine the location of washes (flow accumulation greater than 400). The landscape is then categorized by distance from each wash: 1 is 0-50m from wash; 2 is 50-100m; 3 is 100-150m; 4 is150- 200m; 5 is >200m.
Density96 1.470 0.300 This index is used to measure the expansiveness or compactness of 1996 OHV zones. Theoretically, the density can be expressed by the area covered by the image object divided by its radius. In practice, the density index is calculated by dividing the diameter of the equivalent square shape of the same area with the diameter of the approximate ellipse surrounding the border of the image object. Based on this ellipse approximation, the radius of the image objects can be calculated using the covariance matrix. Therefore, the more an image object is shaped like a square, the higher its density and possibility of bulb formations. The less this value is, the higher the possibility that the objects are the trails in the real world. In Definiens , the expression for the density calculation as below. \'#P 1 + \(VarX+VarY)
85


Table 4.1 (Con't)
Where #PV is the number of the pixels of the image object and VarX/Y is the variance of the X and Y coordinates, used to calculate the diameter of the ellipse approximation.
Ellip_Fit96 0.282 0.198 In Definiens , the value of Elliptic Fit is derived by multiplying two to the ratio of the objects area inside the approximate ellipse to the whole area of the image object and then subtracting one from the derived. In the calculation of the ellipse the proportion of the length to the width of the Object is regarded. The mathematic base to estimate the approximate ellipse is described in the Definiens Developer 7 Reference Book. The formula to compute the Elliptic Fit can be represented with _ #{(X,Y) Pv:ev(X,Y)<1} 0 = 2* vpt -1 where £v(x,y) is elliptic distance at a pixel (x,y), Pv is the set of pixels of an image object v, and #PV is total number of pixels contained in Pv. The use of this index is to measure the magnitude of elongation and expansiveness. The feature value range is between 0 and 1. 1 represents complete fitting, whereas 0 represents 50% or fewer pixels fitting inside the ellipse. Therefore, the lower the value, the more trail-like the object concerned is. Conversely, the higher the value, the higher possibility the regarded object is an OHV bulb shape.
SD_Dir_Sub96 30.412 10.679 The standard deviation of the directions of the sub-objects is a supplementary index useful for identifying trail features in a disturbed environment, especially when the regarded object is not as bright in contrast to the surroundings but has an elongated
86


Table 4.1 (Cont)
shape, without too many branches. The sub- objects main directions are weighted by the symmetries of the respective sub-objects. The set of sub-object main directions by which the standard deviation is calculated is determined by the method depicted in the Definiens Developer 7 Reference Book. The lower the value, the higher the possibility that the object is an OHV trail.
Dir_Homo96 0.621 0.521 The Dir_Homo96 index is created first by calculating the ratio of the main direction of the concerned object to the main direction of the trail objects. The concerned objects are those located within 46 pixels (meters) from the identified trail objects, whereas the trail objects are those which have been identified as OHV routes in the super level. If the ratio is close to 1, the regarded object has a direction that is almost the same as the direction of the nearby identified route object. Secondly, the index is linearly redistributed based on the other OHV trail identification variables, such as brightness, density, and texture of the sub-object direction, L_divd_W, vegetation index, or the standard deviation of the Near Infrared. As not all of the objects having indices close to 1 calculated in the first step can be classified as trail objects, several SQLs are executed in the second stage to filter the Dir_Homo generated in the first step, so that the objects which have a high potential to be classified as trails are assigned higher values, whereas the objects with lower a possibility of such are assigned a lower value. Conversely, the objects which are allocated lower values in the first step do not definitively denote non-disturbance objects. For example, some of the objects which are perpendicular to or farther than 46 pixels away from the existing OHV trails can be categorized as disturbed objects. Again,
87


Table 4.1 (Cont)
Dir_Homo value generated in the first step needs to be re-calculated to give the objects higher values.
SD_A_Div_W96 7.422 2.430 This is a customized shape index and is used to identify linear features of 1996. Because the wide roads normally have more variability in their width than that of narrow roads, this index aims to create a universal criterion that can be used to evaluate wide and narrow roads. The smaller the value, the more elongated the shape. This customized index is composed of two parts of shape indices. The first portion of the index, standard deviation of area represented by segments, is obtained by calculating the standard deviation of all triangles created by the Delaunay Triangulation for the object. The second portion, width, is obtained by calculating the average height, h, of all triangles crossed by the main line. The ratio of the first portion to the second portion forms the value of this index.
L_Divd_W96 21.091 12.279 This is simply obtained by dividing the length of the concerned object by the width of the object. This is an index to observe the degree of elongation of the objects. The higher the value, the more elongated the shape. If the Red band value of the concerned object is also high, this elongated object has a high possibility of being an OHV trail.
4.4 Hypothesis
Based on field observations and the information above, hypothesis of OHV zone formation
is presented as following:
4.4.1 Radiation
Intensity of vegetation. Vegetative intensity can be measured with the evaluation of NDVI,
which captures greenness or chlorophyll levels. It is anticipated that OHV disturbance
expansion is associated with areas of less vegetation. Due to the fact that pre-existing
88