Citation
Colorado wildfire : using landsat imagery to assess burn severity of the West Fork Complex fire and hydrologic impact to the Rio Grande Basin

Material Information

Title:
Colorado wildfire : using landsat imagery to assess burn severity of the West Fork Complex fire and hydrologic impact to the Rio Grande Basin
Creator:
McCarley, Ryan
Place of Publication:
Denver, Colo.
Publisher:
Metropolitan State University of Denver
Publication Date:
Language:
English

Record Information

Source Institution:
Metropolitan State University of Denver
Holding Location:
Auraria Library
Rights Management:
Copyright [name of copyright holder or Creator or Publisher as appropriate]. Permission granted to University of Colorado Denver to digitize and display this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.

Downloads

This item has the following downloads:


Full Text
Colorado Wildfire: Using Landsat Imagery to Assess Burn Severity of the West Fork Complex Fire and Hydrologic Impact to the Rio Grande Basin by Ryan (Travis) McCarley
An undergraduate thesis submitted in partial completion of the Metropolitan State University of Denver Honors Program
May 2014
Thomas Davinroy Dr. Stella Todd Dr. Megan Hughes-Zarzo
Honors Program Director
Primary Advisor
Second Reader




Colorado Wildfire:
Using Landsat Imagery to Assess Burn Severity of the West Fork Complex Fire and Hydrologic Impact to the
Rio Grande Basin
T. Ryan McCarley Senior Honors Thesis Defended: 9 May 2014


Table of Contents
ABSTRACT_______________________________________________________________1
INTRODUCTION___________________________________________________________1
BACKGROUND_____________________________________________________________3
Linking Burn Severity to Runoff______________________________________4
Classifying Burn Severity____________________________________________5
OBJECTIVE_____________________________________________________________10
METHODS_______________________________________________________________10
Landsat Burn Severity Analysis______________________________________10
SLC Correction______________________________________________________13
Hydrologic Analysis_________________________________________________17
RESULTS_______________________________________________________________20
DISCUSSION OF RESULTS_________________________________________________22
Burn Severity Error and Discussion__________________________________22
Hydrologic Analysis Errors and Discussion___________________________25
CONCLUSION____________________________________________________________25
REFERENCES____________________________________________________________26
List of Figures
FIGURE 1 LOCATION OF THE WEST FORK FIRES WITHIN THE RIO GRANDE WATERSHED_5
FIGURE 2 REFLECTANCE CURVES FOR BURNED GROUND AND HEALTHY VEGETATION_____5
FIGURE 3 LANDSAT 7 BAND DESIGNATIONS_____________________________________5
FIGURE 4 LANDSAT 7 DIFFERENCE NORAMLIZED BURN RATIO _____________________11
FIGURE 5 LANDSAT 8 NORMALIZED BURN RATIO ________________________________11
FIGURE 6 US FOREST SERVICE THRESHOLDS FOR DIFFERENCED NORMALIZED BURN RATIO 12
FIGURE 7 LANDSAT 8 NORMALIZED BURN RATIO SPLIT INTO 50 CLASSES___________13
FIGURE 8 HISTOGRAMS OF ISO CLASSES CORRESPONDING TO BURN SEVERITY________14
FIGURE 9-ACCURACY MAP OF CLASSIFIED LANDSAT 8 IMAGE________________________16
FIGURE 10 DIGITAL ELEVATION MODEL OF WATERSHED________________________17
FIGURE 11 HYDRAULIC SOIL CODES FOR WATERSHED__________________________18
FIGURE 12 LANDUSE CODES FOR WATERSHED ________________________________18
FIGURE 13 CURVE NUMBER GRID FOR WATERSHED_____________________________19
FIGURE 14 FINAL LANDSAT 7 CLASSIFICATION MAP__________________________20
FIGURE 15 FINAL LANDSAT 8 CLASSIFICATION MAP__________________________20
FIGURE 16 COMPARISON OF FINAL MAP WITH US FOREST SERVICE MAP__________21
List of Tables
TABLE 1 BURN CLASSIFICATION THRESHOLDS DEVELOPED FOR THE LANDSAT IMAGE_14
TABLE 2 POSSIBLE VALIDATION VALUES_____________________________________15
TABLE 3-ACCURACY OF VALIDATION BY PERCENT IMAGE__________________________16
TABLE 4 CURVE NUMBER LOOKUP TABLE______________________________________19
List of Equations
EQUATION 1 NORMALIZED BURN RATIO______________________________________6
EQUATION 2 DIFFERENCED NORMALIZED BURN RATIO__________________________7
EQUATION 3 DIFFERENCED NORMALIZED BURN RATIO USING OFFSET_____________7
EQUATION 4-RELATIVIZED DIFFERENCED NORMALIZED BURN RATIO________________7
EQUATION 5 RELATIVIZED BURN RATIO_____________________________________8


Abstract
In the summer of 2013 the West Fork Complex Fire raged in southern Colorado. Burning over 109,000 acres the fire has had a significant impact on the hydrology of the area, especially the Rio Grande watershed (RWEACT 2013; Yochum and Norman 2014). Throughout the literature on hydrologic responses to fire, fire severity has been found to be the most significant variable (Moody et al. 2008). The aim of this study was to use remote sensing to estimate additional runoff that could be generated from the burned area. Because Landsat data is widely available at no cost, it is a useful tool for this analysis. There are several methods of determining burn severity, which exploit the difference in spectral reflectance between healthy vegetation and bare ground. A review of the literature in this report attempts to determine the differences between techniques. Using Landsat 7 imagery from September 2012 and 2013, the Differenced Normalized Burn Ratio (dNBR) was calculated. Adjustments were made for the Scan Line Corrector (SLC) error that occurs in Landsat images since 2003. Gaps in data caused by this error were classified by using a Landsat 8 image trained to the classes that were determined by the Landsat 7 dNBR. Compared to a similar burn severity maps by Yochum and Norman (2014), Verdin et al. (2013) and RWEACT (2013), which used ground-truthed U.S. Forest Service data, there appeared to be some difference. This study showed a higher burn severity, which would lead to greater runoff response.
Keywords: Remote Sensing, Burn Severity, Normalized Burn Ratio, West Fork Complex Fire, Post-Fire Runoff
Introduction
On June 5th 2013 a lightning strike ignited the West Fork Fire northeast of Pagosa Springs, Colorado. Another strike to the northwest ignited the Papoose Fire. Together these fires and two others in the area became known as the West Fork Complex Fire (RWEACT 2013). By July 12th, the fires had consumed 109,000 acres total (RWEACT 2013; Yochum and Norman 2014) and 90,000 in the Rio Grande watershed alone (RWEACT 2013). A location map of the fires within the watershed is shown in figure 1.
While this study focuses on the science of post-fire runoff and the processes of
satellite imagery assessment, the social significance of the project is the potential for better remediation. Each year millions of dollars are spent on rehabilitation of burned areas in order to protect human settlements or return the environment to a desired
1


ecological state. Often the most destructive result of fire is flooding and debris flows.
By assessing burn severity using Landsat data, remediation teams can be quickly informed what areas need the most attention. The use of Landsat imagery is substantially cheaper than areal or ground assessments and can be performed within weeks after a fire. As the frequency and intensity of fire increases in the Western U.S. the use of this type of assessment is likely to become increasingly relied upon because of its cost efficiency and relative effectiveness.
..Sources: Esri, DeLorme, HERE, 'TomTom, Intermap, Increment P Corp., GEBCO, USQ^RAO,
^Saguache
Antelope
Park
G e j ve
Gox.-
'ountati

Q*f£n fDrgT t-*?8§ ir&tedfTtta/a 'Atxuntafpr ' -s
f&ne&p f/fountain
A G A R T A M O U N T At M S
3591m
: !
-; 7
I
Legend
] Watershed Outline I Fire Perimeter
I Kilometers
Sources: Esri, DeLorme, USGS, NF^^ources: Esri, deLonne, HERE, TomTonjiJntermap,. iicrement P Corp., GEBCO, USGS, FAO, NPS, NRCAN. GeoBase, IGN, Kadaster NL, Ordnance Survey, Esri la pan, METI, Esri China (Hong Kong), swisstopg, and the GIS User Community -.J ^
Figure 1: Location of the West Fork Fires within the Rio Grande Watershed
2


Background
Most of the area of the fire is characterized by Engelmann spruce (Picea engelmannii), White fir {Abies concolof) and Subalpine fir {Abies lasiocarpa). Lower elevations at the edge of the fire contain Ponderosa pine {Pinus ponderosa) and Douglas fir {Pseduotsuga menziesii). The highest elevations of the fire are at timberline. This area is home to various forbs, shrubs and willows (USFS 2013). According to data available from the Rio Grande National Forest, the watershed above South Fork,
Colorado is approximately 1,330 square miles and has an average slope of 25.9%. Additionally, this area receives an average of 30.5 inches of precipitation per year (NCAR 2014).
In the literature burn severity and fire severity are often distinguished from one another. Soverel et al. (2010) describes fire severity as the direct and immediate effect of combustion on vegetation and soil, while they describe bum severity as the ecological changes of an area after a fire. Gitas et al. (2009) provides similar definitions, although adding that an aspect of bum severity is ecologic recovery. Miller and Thode (2006) also term bum severity as the effect of a fire on an ecosystem. Parks et al. (2014) suggests that burn severity is the degree of change to vegetation and soil caused by fire. Throughout this paper burn severity will be the preferred term because it encompasses the full range of ecological effects. From a holistic approach it is easy to relate the term burn severity to the range of impacts extending beyond simply the immediate effect of combustion.
The next few sections describe some of the current research and knowledge on post-fire runoff and classifying bum severity.
3


Unking Bum Severity to Runoff
One of the most prolific soil problems that result from wildfire is water repellency. This issue leads to increased runoff and flooding, which can endanger buildings and watersheds. That is why one of the main goals of the Burned Area Emergency Response (BAER) teams is to stabilize soil and protect downstream risks (Bobbe et al. 2002).
The cause of water repellency is hydrophobic, long-chained, organic compounds in the earth that prevent the soil under the ash layer from absorbing water (Moody and Ebel 2012). These organic compounds are generated as a direct result of the heating that occurs during a fire. Generally the conditions required for this process are a residence time of at least 5-10 minutes at 175-280 C or greater (Lewis et al. 2008). The longer and the hotter the fire, the more water will be evaporated off the soil. A wildfire can drive water out of macro pores, micro pores, and possibly even hydrated clay minerals (Moody and Ebel 2012). The hotter the fire, the deeper the soil repellency may exist (Lewis et al. 2008). Therefore a higher burn severity is likely to indicate greater hydrophobicity. In addition, the loss of vegetation correlating with high burn severity means that plants will retain less water and sediment. This is why BAER teams often assess burn severity using either soil hydrophobicity or the amount of remaining needles (for coniferous forests) (Bobbe et al. 2002).
In order to rehabilitate hyper-dry soil there must be rewetting. Studies of the Hayman fire in Colorado showed high water repellency for first two years and quickly decreasing levels of repellency thereafter (Robichaud et al. 2013). This would indicate
4


that the need for remediation is most pressing for the first few years after a fire, but following that period natural recovery can be expected to take place. Artificial rewetting requires either high water pressure or extended contact time. One common approach to extending contact time is to create earth depressions or shelves on hill slopes where water will settle and eventually infiltrate. It has been shown that even short, intense rainfall is unlikely to infiltrate because of insufficient contact time, so artificial rewetting is highly beneficial immediately following a fire (Moody and Ebel 2012).
Classifying Bum Severity
Throughout the literature there are numerous ways to characterize bum severity. Of these, Landsat based methods are often favored because of their large geographic coverage, extensive archive and cost effectiveness (Bobbe et al. 2002). Landsat has also been found to perform favorably to other satellite platforms when validated with ground measurements (Parks et al. 2014).
Landsat burn assessment seeks to exploit the spectral difference between Near Inferred (NIR) and Shortwave Inferred (SWIR) wavelengths. Figure 2 shows the reflectance curves of both burned ground and healthy vegetation displayed along the electromagnetic spectrum. The arrows correspond to the greatest differences between these two curves, which are NIR and SWIR reflectance. NIR is reflected well by healthy vegetation because of the plants chlorophyll, while it is not reflected much by barren ground or burned areas. Bare soil, dead wood and burned areas are highly reflective to SWIR, but not to NIR (Miller and Thode 2006; USFS 2014).
5


The Landsat 7 satellite detects light wavelengths reflected in the NIR between 0.77 and 0.90 micrometers and designates them as band 4. It detects light wavelengths reflected in the SWIR between 2.09 and 2.35 micrometers and designates them as band 7. All of Landsat 7s band designations can be seen in figure 3. For comparison the NIR band for Landsat 8 detects wavelengths 0.85-0.88 micrometers, while its SWIR detects
wavelengths 2.11-2.29 micrometers (USGS, 2014).
Figure 2: Reflectance Curves for Burned Ground and Healthy Vegetation
Using the difference between NIR and SWIR a metric was developed to detect spatial difference between healthy vegetation and burned ground. This method is known as the Normalized Burn Ratio (NBR). NBR is calculated by taking NIR values minus SWIR values all divided by NIR plus SWIR values (Bobbe et al. 2002; FIREMON 2004;
McKinley 2004; Miller and Thode 2006; Gitas et al. 2009;
Equation 1: Normalized Burn Ratio
Soverel et al. 2010; Parks et al. 2014). Equation 1 shows
_ CNIR SWIR) the calculation for NBR using Landsat 7.
NBR ~ (NIR + SWIR)
This measurement can be used alone as single scene NBR (SS NBR). In this case the NBR image is clustered into 50 classes and then grouped into burn severities based on available ground information. The National Land Cover Dataset can also be used to
6


adjust for previously barren areas that did not change (Bobbe et al. 2002). At its highest points the West Fork Complex Fire burned up to timberline. In these areas bare rock would reflect SWIR light similarly to burn areas so this type of adjustment would be necessary (Miller and Thode 2006; USFS 2014). An advantage to this method is the simplicity of using one image. A disadvantage is that there must be some ground reference to establish thresholds for the burn severity classes.
There are three other primary metrics that have been developed in addition to single scene NBR (SS NBR) that take into account pre-fire and post-fire images. The advantage to all of these methods is that objects that did not change, such as bare rock, can be accounted for directly in the calculation. The disadvantage to all of these is that factors such as phenology, sun angle, cloud cover and precipitation cannot be exactly matched between images (Miller and Thode 2006; Parks et al. 2014).
The three duel-image methods are Differenced Normalized Burn Ratio (dNBR), Relativized Differenced Normalized Bum Ratio (RdNBR) and Relativized Burn Ratio (RBR). In each case the NBR images are converted from their raw brightness values into at-satellite reflectance to try to account for some of the natural image differences. Equation 2 shows dNBR, which is calculated by taking the pre-fire NBR image and
subtracting it by the post-fire NBR image.
Equation 2: Differenced Normalized Burn Ratio
(Bobbe et al. 2002; FIREMON 2004; McKinley
dNBR=dNBR
prefire
- dNBR,
postfire
2004; Miller and Thode 2006; Soverel et al.
2010). In some cases the
Equation 3: Differenced Normalized Burn Ratio using
Offset
dNBRoffset equation, shown
dNBR'offset method = (dNBR 1000) dNBR,
offset
equation 3, can be used to better
7


account for image differences. This is the standard dNBR multiplied by 1000 all minus the dNBRoffset, where the dNBRoffset value is the average dNBR value of homogenous areas outside the fire perimeter. (Parks et al. 2014). Equation 4 shows RdNBR, which is calculated by taking the dNBR divided by the square root of the absolute value of the prefire NBR. The pre-fire
Equation 4: Relativized Differenced Normalized Burn Ratio
Equation 5: Relativized Burn Ratio
NBR in the denominator is
RdNBR =
dNBR
yj\NBRvrefire\
RBR =
dNBR
(NBRprefire + 1.001)
divided by 1000 if NBR has been converted into an
integer format (Miller and Thode 2006; Soverel et al. 2010; Parks et al. 2014). Equation
5 shows RBR, which is determined by the dNBR divided by result of the pre-fire NBR
plus 1.001. The 1.001 ensures that the denominator will never be zero (Parks et al.
2014).
In order to validate the effectiveness of these methods they are compared to ground measurements, most frequently the Composite Bum Index (CBI). CBI was developed within the framework of FIREMON (fire effects monitoring and inventory protocol) project. It is a qualitative measure whose values range from 0 (not burned) to 3 (completely burned). Although there is fluctuation between fires and because of various factors, a majority of R2 correlation between NBR/dNBR and CBI is greater than 0.55 (Gitas et al. 2009).
There is an ongoing debate regarding the most effective NBR related metric. In a
study of fires throughout the Sierra Nevada Mountains, Miller and Thode (2006) determined that RdNBR performed better than dNBR, particularly for heterogeneous landscapes and those with sparse vegetation. For instance using dNBR severely burned
8


grassland will look similar to its pre-bumed conditions, but severely burned forest will show a large amount of change. This is because RdNBR measures relative change, while dNBR measures absolute change (Miller and Thode 2006; Soverel et al. 2010; Parks et al. 2014). Because of this RdNBR is thought to be more accurate at classifying high burn severity. However, in homogeneous forests the two methods are likely equivalent in their accuracy (Miller and Thode 2006).
In a study performed in the Canadian Rockies, Soverel et al. (2010) found dNBR to be more accurate than RdNBR. Using the same CBI thresholds that Miller and Thode (2006) used RdNBR was determined to be 65.2% accurate compared to 70.2% for dNBR. Contrary to the previous findings RdNBR did not improve even in areas with heterogeneous or sparse vegetation.
Because there has been no consensus whether dNBR or RdNBR is a better metric Parks et al. (2014) developed a third measurement known as RBR. In this case fires from across the Western U.S. were sampled and severity thresholds were developed separately for each fire. The results of the study showed that on average RBR was most accurate compared to CBI, followed by RdNBR and then NBR.
However, performing an ANOVA analysis of the accuracies suggested that there was no statistically significant evidence at the 95% confidence level to support that there is a difference in accuracy between any of the three methods used in this study. Overall it appears that the three methods achieve similar results, particularly for homogeneous forests such as exist in the West Fork Complex.
In summary there is generally only a small difference in accuracy between the three methods. While improvement in all methods should be made, remote sensing of
9


burn severity is still preferable to the previous method of assessment, which was to draw the different classes on paper while flying over the burned area (McKinley 2004).
Objective
The objective of this study was to use Landsat 7 imagery in order to assess the burn severity of the West Fork Complex Fire. The second objective was to use the burn severity map to model runoff in the Rio Grande Watershed. During the development process solving issues with Landsat 7s Scan Line Corrector (SLC) error became a primary goal as well.
Methods
Landsat Bum Severity Analysis
. dNBR was selected as the method of calculation because it accounts for the burn-like reflectance of alpine areas. In addition, total change versus relative change was not seen as an issue for the type of forest being studied, so a relativized method was unnecessary. The Landsat images were collected from the USGS EarthExplorer website (http://earthexplorer.usgs.gov/). Images that contained the extent of the fire were identified as row 34 path 34, a unique identifier for the spatial location of Landsat imagery. Images at this location were selected that contained the least cloud cover obscuring the fire area. Since Landsat 8 has only been operation since 2013 all photos were considered. Landsat 7 photos were considered from 2010 until the time of collection. These were sorted until the best images were determined. From Landsat 7 the September 23rd, 2012 and the September 25th, 2013 images were used. From Landsat 8
10


the October 3rd, 2013 image was used. It was determined there were no useable pre-fire photos for the Landsat 8 platform.
For both sets of data the ENVI image analysis software was used to calculate NBR. The band math tool was used to implement the NBR calculation from equation 1. For the Landsat 7 images the NBR brightness values were converted into reflectance then run through the dNBR calculation from equation 2. These NBR and dNBR images were converted into TIFF files so that further analysis could be performed in ArcGIS. Once there a polygon was digitized representing the fire perimeter. Everything in the images outside the perimeter was removed using the ArcGIS mask function. The Landsat 7 dNBR and the Landsat 8 NBR are shown in figure 4 and figure 5 respectively.
Figure 4: Landsat 8 Normalized Burn Ratio
11


Figure 5: Landsat 7 Difference Normalized Burn Ratio
For the Landsat 7 dNBR the ArcGIS reclassify function was used to symbolize the raster as 10 classes. Seven classes represented distinct thresholds used by the U.S. Forest Service. These values are shown in figure 6. Two classes were values falling
above or below the Forest Service thresholds, and the last class represented the no data values caused by the SLC error. From this analysis a map showing low, low-moderate, high-moderate and high severity classes was generated. This output is shown in the results section under figure 14.
SEVERITY LEVEL ANBR RANGE
Enhanced Regrowth. High -500 to-251
Enhanced Regrowth. Low -250 to -101
Unbutned -100 to +99
Low Severity +100 to +269
Moderate-low Severity +270 to +439
Moderate-high Severity +440 to +659
High Severity +660 to +1300
Figure 6: U.S. Forest Service Thresholds for Differenced Normalized Burn Ratio
12


The Landsat 8 NBR was not immediately used to generate a single scene NBR because the value of the thresholds could not be determined. However, this image later became useful for correcting the SLC error.
SLC Correction
Based on the number of no data pixels within the fire perimeter there was just under 30% of the image missing due to the SLC error. It seemed reasonable that in order to successfully accomplish the secondary objective of creating a hydrologic model there would have to be a continuous bum severity map.
A number of methods were employed to try to fill the gaps in data. Among the trials and errors one attempt involved converting the raster to polygons, then using those polygons as training regions (ROIs). Next the no data values from the raster were classified using the Maximum Likelihood Classifier. This method uses the ROIs and distance to automatically determine the best class for a pixel. Unfortunately the gaps tended to classify all as one value.
Finally it was determined that the Landsat 7 dNBR could be used to generate thresholds in order to create a SS NBR from the Landsat 8 image. This was accomplished by first classifying the Landsat 8 NBR image into 50 classes using the Iso Cluster Unsupervised Classification tool in ArcGIS. The result is shown in figure 7.
Next, the polygons representing each of the seven Forest Service defined Landsat 7 classes were used to extract the Iso Cluster pixels falling into the same spatial extent with the mask tool. The attribute table from each resulting raster was exported into R, where the data could be evaluated. Using R, histograms were developed for each Landsat
13


Figure 7: Landsat 8 Normalized Burn Ratio Split into 50 Classes
7 severity class with the 50 classes on the x-axis and frequency of occurrence on the y-axis. Histograms are not shown for classes less than unbumed because it was irrelevant for determining burn thresholds. The resulting R histograms are shown in figure 8. After evaluating the mean class initially, it was later determined that the median class was a better measure due to the extreme skew of the data. Using the medians of each class thresholds were calculated as the middle value between each median. Table 1 shows the thresholds used.
Lastly, the ArcGIS reclassify function was used to sort the image into high severity, high-moderate severity, low-moderate severity, low severity and unbumed. The final output is shown in the results section under figure 15.
14


High Severity
L
20 30
Iso Classes
Low Severity
nr>n-m-i-n4T-rTT-H-rri^^
10 20 30
Iso Classes
Moderate High Severity
20 30
Iso Classes
Unburned

J
10 20 30 40
Iso Classes
Moderate Low Severity
j^rrTlTlTTlTTTlTTTinTinTTHTTT-r-rvn-T-r-r-rT-T-rrrri n I
10 20 30 40 50
Iso Classes
Figure 8: Histograms of Iso Classes Corresponding to Burn Severity
Table 1: Burn Classification Thresholds Developed for the Landsat Image
Classification Median Threshold
High Severity 2 O 2.5
Moderate High Severity 3 2.5 S
Moderate Low Severity 13 S 22.5
Low Severity 32 22.5 39
Unburned 46 39 50
Table 2: Possible Validation Values
VALUE NEW CLASS NEW CLASS TRUE CLASS
o High High 1 Med High High
-1 High Med High O Med High Med High
2 High Med Low -1 Med High Med Low
-3 High Low -2 Med High Low
A High Unbumed -3 Med High Unbumed
2 Med Low High 3 Low High
1 Med Low Med High 2 Low Med High
O Med Low Med Low 1 Low Med Low
-1 Med Low Low O Low Low
-2 Med Low Unbumed -1 Low Unbumed
Unbumed
Unbumed
Unbumed
Unbumed
Unbumed
High
Med High Med Low Low
Unbumed
15


Later on validation was performed to assess the accuracy of this technique. This was done by using reclassify to match the values of the Landsat7 raster and the Landsat 8 raster so that: l=high, 2=med_high, 3=med_low, 4=low and 5=unburned. Next the ArcGIS raster calculator was used to subtract the Landsat 7 image from the Landsat 8 image. The resulting values ranged from 4 to -4, with 0 indicating a perfect match and values of |4| indicating a poor match. Each integer unit is representative of being off by one class. Lastly, positive values corresponded to a tendency to underestimate severity, while negative values related to the overestimation of severity. A break down of possible values is shown in table 2.
Figure 9 shows the validation map, with areas of blue corresponding to perfectly accurate classification between the Landsat 7 and Landsat 8 images. In the map green indicates areas that were misclassified by only one class, so for instance a high burn classed as high-moderate, or unburned classed as low burn. The percentages for each validation value are shown in table 3. In summary 40.7% of the pixels matched perfectly, while 81.1% of the pixels were within one class. Negative values were indicative of overestimating severity, which judging by table 3 occurred slightly more frequently than underestimation. Understanding that perfect matching is difficult, 81.1% within one class is acceptable for concluding that the Landsat 8 image was successfully classified using the Landsat 7 thresholds. While the no-data gaps could not be included in the validation, it is assumed that the accuracy can be interpolated across the image.
Table 3: Accuracy of Validation by Percent of Image
-4 4038 1.4
-3 5881 2.0
-2 20103 6.9
-i 68827 23.6
0 118917 40.7
i 48980 16.8
2 16669 5.7
3 6050 2.1
4 2687 0.9
292152
16


Misclassified by 4 classes
Figure 9: Accuracy Map of Classified Landsat 8 Image
Hydrologic Analysis
The method for determining runoff began with creating a curve number (CN) grid. The CN is a value between 0 and 100, with 0 being the most permeable and 100 being the least. Using these values an input can be added and the amount runoff calculated (50% for CN 50, 100% for CN 100, etc.) (Yochum and Norman 2014).
The U.S. Corp of Engineers developed an ArcGIS extension known as HEC-GeoHMS. Using this model a CN grid can be calculated for any area using a digital elevation model (DEM), a soil layer showing hydraulic soil classes A through D and a
17


land cover layer. The DEM was taken from the National Elevation Dataset (Data Credit:
Figure 10: Digital Elevation Model of Watershed
Figure 21: Hydraulic Soil Codes for Watershed
Hydrologic SoilCode
A
B
I c
D
_I Kilometers 60

National Elevation Dataset http://ned.usg s.gov/), the soil layer was obtained from the U.S. Department of
Agricultures web soil survey (Soil Survey Staff 2014)and the land use data set was acquired from the National Land Cover Database (Fry etal. 2011).
18


The DEM, hydraulic soil types and land cover are shown in figures 10, 11 and 12 respectively.
Figure 12: Land Use Codes for Watershed
Once all of these files are complied and projected correctly the HMS-GeoHEC model synthesizes them into a CN grid using a CN lookup table, which is provided by the user. The lookup table is shown in table 4. The values in the table were estimated from
various sources. The output is shown in the results section under figure 13. Additional analysis relating the CN grid and the bum severity map were not performed due to time constraints and lack of defined methodology about the process.
Table 4: Curve Number Lookup Table
OBJECTID* LUValue Desription A B C D
1 1 Water 10 10 10 10
2 2 Developed 57 72 81 86
3 3 Forest 30 58 71 78
4 4 Open Fields 67 77 83 87
19


Results
Using the methods described above three outputs were completed to meet the objectives of this study. The first is the Landsat 7 dNBR with severity classes shown in figure 14. Second is the continuous Landsat 8 NBR that was classified from the Landsat 7 image, shown in figure 15. Lastly, shown in figure 13 is the CN grid generated from the HMS-GeoHEC extension.
jjS'SiK'erton
Fire Perimeter
CN Grid
Value
- High : 100
2191 njJ

I I______________I I_____________I " I Kilometers
0 1b 30 60
Pagos^ptfrces: Esri, DeLorme, HERE, TeTnTdtriltffiterpttap' inbreme'nt P Corp., GEBCO, 13SGS, FAO, NPS, NRCAN, l- Sptin§f4Base, JGN, Kadaster NL, OrdrijfTOeflsfrr^eff Esri Japan, METI. Esri China (Hong Kong), swisstopo, and the Gis\)ser Community
Figure 13: Curve Number Grid for Watershed
20


High Burn Severity f
High Moderate Burn Severity
Low Moderate Burn Severity ' b
Low Burn Severity ' "VJ % L.
A * Jr\ j-jSstt?* "i,
Figure 15: Final Land sat 8 Classification Map
21


There could many reasons for the differences between the two maps. The U.S. Forest Service used Landsat imagery for their analysis, but the date(s) of the image is not known. In addition their image was ground-truthed after the image processing (Belz 2014). Differing image dates could have led to differences in phenology. This study used images from late September, which in these higher altitude regions can tend to be well into fall. An image from late July or early August is likely to look different because vegetation is in a different growth stage. Ground-truthing the image was another source of discrepancy. In this process the Forest Service visually verified the areas and changed severity classes where appropriate. However, often this process involves a visual estimation of the area rather than quantitative measurement of burn severity. So while ground-truthing may account for much of the discrepancy between the two images it is difficult to say that one image is more correct than the other.
Another issue is the sensor differences between Landsat 7 and Landsat 8. There is a slight difference in wavelength range detected by each satellite in its NIR and SWIR bands. While this is likely a very small discrepancy there may be some impact on classification differences between Landsat 7 and Landsat 8. This study used Landsat 8 NBR with thresholds derived from Landsat 7 dNBR, but it is unknown which platform the Forest Service used. Another unknown is the thresholds and methods used for classification. Various sources indicate that these Forest Service maps likely use dNBR, but this hasnt been confirmed.
There are a few other issues within an unknown impact to the final result. Because dNBR is a measure of total change it seems reasonable that it might overestimate burn severity in a heavily forested area like the West Fork Complex. This
23


would support the ground-truth changes made by the Forest Service. However, the spruce beetle also affected much of the area. One study suggested that 88% of the burned area was continuous areas of beetle kill (USFS 2013). These dying or dead trees would have not reflected NIR as well as healthy trees. Therefore total change expected by dNBR might have been smaller than previously suggested. While all of these factors may have played a role, this study is unable to offer a quantifiable error associated with each of them.
The errors that were observed in the SLC correction process may be due to a few factors. Some unknown issues are similar to those described above. These include differing image dates and difference in spectral detection between satellites. Another issue is with comparing NBR to dNBR. Because dNBR accounts for the pre-fire image there are going to be some parts of the image classified very differently. While these errors were not quantified individually, collectively they result in the variation that can be see in the histogram of Iso Cluster classes by Landsat 7 severity class, figure 8. Lastly, there was error in the final classification of the Iso Cluster classes. From the histograms median was chosen to create the thresholds using the middle value between medians. There may be a better method since the true thresholds probably arent exactly in the middle of each median.
In order to resolve the aforementioned errors ground based validation should be performed. This could be incorporated into a future phase of this project. The study would likely involve using CBI plots in relatively homogenous classification areas (at least 3x3 pixels). Once this is complete the values could be compared using an accuracy
24


assessment matrix. If required for future hydrologic study the severity map could be updated to reflect ground measurements.
Hydrologic Analysis Errors and Discussion
The error in the CN grid is primarily due to the values assigned in the CN lookup table. If this project were to be continued then more robust values should be generated based on known attributes of the specific area.
Much more analysis remains to be done regarding the hydrologic analysis of the West Fork Complex Fire on the Rio Grande basin. For this study some attempt was made at using the CN grid as well as incorporating the bum severity grid, but these attempts failed. Eventually time constraints made it unreasonable to finish this objective. If the project was completed values could be validated with a USGS gauge at South Fork, Colorado, where the watershed empties. Other studies have looked at potential runoff and debris flows for this fire, and these could be used for a rough comparison as well. Yochum and Norman (2014) suggest that in many areas a 10-year precipitation event could lead to a 50 or 100-year flood. The results of this study are likely to vary somewhat since the bum severity map used by Yochum and Norman (2014) was from the Forest Service and would have therefore underestimated burn severity in comparison.
Conclusion
The West Fork Complex Fire in 2013 burned 90,000 acres of forest in the Rio Grande River basin (Yochum and Norman 2014). This was just one of many fires across
25


the Western U.S. As fire size and frequency increase the ability to successfully assess burn severity and respond the threat it creates will become more important.
Landsat imagery has been shown to be successful in this venture. While there are a number of methods for calculating burn severity, it appears that there is no significant difference in accuracy between methods. This study used dNBR because it accounts for before-after images and because a measure of total change was appropriate for a homogenous coniferous forest. The method of correcting the SLC error was developed specifically for this study and has worked well for this site. The differences between this study and burn severity maps produced by the U.S. Forest Service suggests greater burn severity, and thus more runoff than had been predicted by groups using that data. Lastly, the successful implementation of these procedures for the West Fork Complex Fire suggests that this study could be adapted to assess burn severity for other fires.
References
Belz, Steve. 2014. [Email communication], March 19.
Bobbe, Thomas; Finco, Mark V.; Quayle, Brad; Lannom, Keith. 2002. Field measurements for the training and validation of bum severity maps from spaceborne, remotely sensed imagery. Joint Fire Science Program-2001-2. Salt Lake City, UT: U.S. Department of Agriculture, Forest Service, Remote Sensing Applications Center. 14 p.
Fire Research and Management Exchange System (FIREMON). 2004. FIREMON BR cheat sheet V4 [Online], Available: http://burnseverity.cr.usgs.gov/pdfs/LAv4_BR_ CheatSheet.pdf [2014, February 3],
Fry J.; Xian G.; Jin S.; Dewitz J.; Homer C.; Yang L.; Bames C.; Herold N.; Wickham J. 2011. Completion of the 2006 National Land Cover Database for the Conterminous United States. PE&RS. 77(9): 858-864.
26


Gitas, Ioannis Z.; de Santis, Angela; Mitri, George H. 2009. Remote sensing of burn severity. In: Chuvieco, E., ed. Earth observations of wildland fires in Mediterranean ecosystems. Berlin: Springer: 129-148.
Lewis, S.A.; Robichaud, P.R.; Frazier, B.E.; Wu, J.Q.; Laes, Denise Y.M. 2008. Using hyperspectral imagery to predict post-wildfire soil water repellency. Geomorphology. 95: 192-205.
McKinley, Randy. 2004. Evaluation of gap-filled SLC-off imagery for burn severity mapping [Online], Sioux Falls, SD: U.S. Geological Survey, EROS Data Center. Available: http://landsat.usgs.gov/documents/bum_severity_mapping_slcoff_ mckinley.pdf [2014, March 22],
Miller, Jay D.; Thode, Andrea E. 2006. Quantifying bum severity in a heterogeneous landscape with a relative version of the delta Normalized Bum Ratio (dNBR). Remote Sensing of Environment. 109: 66-80.
Monitoring Trends in Burn Severity (MTBS). 2013. Methods [Online], Available: http://www.mtbs.gov/methods.html [2014, March 17],
Moody, J.A.; Ebel, B.A. 2012. Hyper-dry conditions provide new insight into the cause of extreme floods after wildfire. Catena. 93: 58-63.
National Center for Atmospheric Research (NCAR) Staff (Eds). 2014. The Climate Data Guide: PRISM High-Resolution Spatial Climate Data for the United States: Max/min temp, dewpoint, precipitation [Online], Available:
https://climatedataguide.ucar.edu/climate-data/prism-high-resolution-spatial-climate-data-united-states-maxmin-temp-dewpoint [2014, March 17],
Rio Grande Watershed Emergency Action Coordination Team (RWEACT) [Online], 2013. Available: http://www.rweact.org/ [2013, November 25],
Soil Survey Staff. 2014. Web Soil Survey. U.S. Department of Agriculture, Natural Resources Conservation Service. Available: http://websoilsurvey.nrcs.usda.gov/. Accessed [2014, February 13],
Soverel, Nicholas O.; Perrakis, Daniel D.B.; Coops, Nicholas C. 2010. Estimating burn severity from Landsat dNBR and RdNBR indices across western Canada. Remote Sensing of Environment. 114: 1896-1909.
Parks, Sean A.; Dillon, Gregory K.; Miller, Carol. 2014. A new metric for quantifying burn severity: The Relativized Burn Ratio. Remote Sensing. 6: 1827-1844.
Robichaud, P.R.; Lewis, S.A.; Wagenbrenner, J.W.; Ashmun, L.E.; Brown, R.E. 2013. Post-fire mulching for runoff and erosion mitigation Part I: Effectiveness at reducing hillslope erosion rates. Catena. 105: 75-92.
27


U.S. Forest Service (USFS). 2013. Burned-area report [West Fork Complex Fire], FS-2500-8 [Online], U.S. Department of Agriculture, Forest Service. Available: http://www.fs.usda. gov/Intemet/FSE_DOCUMENTS/stelprdb5431267.pdf [2013, November 25],
U.S. Forest Service (USFS). 2014. BARC Frequently asked questions [Online], U.S. Department of Agriculure, Forest Service, Remote Sensing Applications Center. Available: http://www.fs.fed.us/eng/rsac/baer/barc.html [2014, February 3],
U.S. Geological Survey (USGS). 2014. Landsat Missions [Online], Available: http://landsat.usgs.gov/ [2014, February 3],
Verdin, Kristine L.; Dupree, Jean A.; Stevens, Michael R. 2013. Postwildfire debris-flow hazard assessment of the area burned by the 2013 West Fork Fire Complex, southwestern Colorado. Open-File Report 2013-1259. U.S. Geological Survey. 30 p.
Yochum, Steven E.; Norman, John. 2014. West Fork Complex Fire: Potential increase in flooding and erosion. Colorado State Office: U.S. Department of Agriculture, Natural Resource Conservation Service. 26 p.
28


Discussion of Results
Burn Severity Error and Discussion
Aside from the gaps in data caused by the SLC issue, the bum severity analysis went very well. Although in comparison there are some discrepancies between this map and maps used by the U.S. Forest Service. A Forest Service map is used by many sources including Yochum and Norman (2014), Verdin et al. (2013) and RWEACT (2013). While a quantitative validation could not be performed, the two were compared visually. The maps side by side with are shown in figure 16. The high-moderate and
Figure 16: Comparison of Final Map with U.S. Forest Service Map
low-moderate classes were lumped for better comparison. Overall there seem to be some correlations with the high severity areas, although this study tended to classify more area as high burn severity. There seems to be some general agreement with moderate and low severity classes. The unbumed class seems to match very well between images.
22


Full Text

PAGE 1

Colorado Wildfire: Using Landsat Imagery to Assess Burn Severity of the West Fork Complex Fire and Hydrologic Impact to the Rio Grande Basin by Ryan (Travis) McCarley An undergraduate thesis submitted in partial completion of the M etropolitan State University of D enver Honors Program May 2014 Thomas Davinroy Dr. Stella Todd Dr. Megan Hughes Zarzo Primary Advisor Second Reader Honors Program Director

PAGE 3

Colorado Wildfire: Using Landsat Imagery to Assess Burn Severity of the West Fork Complex Fire and Hydrologic Impact to the Rio Grande Basin T. Ryan McCarley Senior Honors Thesis Defended : 9 May 2014

PAGE 4

Table of Contents ABSTRACT ________________________________ ________________________________ __ 1 INTRODUCTION ________________________________ ______________________________ 1 BACKGROUND ________________________________ _______________________________ 3 L INKING B URN S EVERITY TO R UNOFF ________________________________ _______________ 4 C LASSIFYING B URN S EVERITY ________________________________ ____________________ 5 OBJECTIVE ________________________________ ________________________________ 10 METHODS ________________________________ ________________________________ __ 10 L ANDSAT B URN S EVERITY A NALYSIS ________________________________ ______________ 10 SLC C ORRECTION ________________________________ ___________________________ 13 H YDROLOGIC A NALYSIS ________________________________ _______________________ 17 RESULTS ________________________________ ________________________________ ___ 20 DISCUSSION OF RESULT S ________________________________ ____________________ 22 B URN S EVERITY E RROR AND D ISCUSSION ________________________________ __________ 22 H YDROLOGIC A NALYSIS E RRORS AND D ISCUSSION ________________________________ ____ 25 CONCLUSION ________________________________ _______________________________ 25 REFERENCES ________________________________ _______________________________ 26 List of Figures FIGURE 1 LOCATION OF THE WEST FORK FIRES WITHIN TH E RIO GRANDE WATERSH ED ______ 5 FIGURE 2 REFLECTANCE CURVES F OR BURNED GROUND AND HEALTHY VEGETATION _______ 5 FIGURE 3 LANDSAT 7 BAND DESIG NATIONS ________________________________ _____________ 5 FIGURE 4 LANDSAT 7 DIFFERENCE NORAMLIZE D BURN RATIO ___________________________ 11 FIGURE 5 LANDSAT 8 NORMALIZED BURN RA TIO ________________________________ _______ 11 FIGURE 6 US FOREST SERVICE TH RESHOLDS FOR DIFFERE NCED NORMALIZED BURN RATIO 12 FIGURE 7 LANDSAT 8 NORMALIZED BURN RA TIO SPLIT INTO 50 CL ASSES __________________ 13 FIGURE 8 HISTOGRAMS OF ISO CLASSES CORRESPO NDING TO BURN SEVERI TY ___________ 14 FIGURE 9 ACCURACY MAP OF CLAS SIFIED LANDSAT 8 IMA GE ____________________________ 16 FIGURE 10 DIGITAL ELEVATION MO DEL OF WATERSHED ________________________________ 17 FIGURE 1 1 HYDRAULIC SOIL CODES FOR WATERSHED ________________________________ __ 18 FIGURE 12 LANDUSE CODES FOR WA TERSHED ________________________________ ________ 18 FIGURE 1 3 CURVE NUMBER GRID FO R WATERSHED ________________________________ ____ 19 FIGURE 1 4 FINAL LANDSAT 7 CLASSIFICATION MAP ________________________________ _____ 20 FIGURE 1 5 FINAL LANDSAT 8 CLAS SIFICATION MAP ________________________________ _____ 20 FIGURE 1 6 COMPARISON OF FINAL MAP WITH US FOREST S ERVICE MAP __________________ 21 List of Tables TABLE 1 BURN CLASSIFICATION THRESHOLDS DEVELOPED FOR THE LANDSAT IMAG E ______ 14 TABLE 2 POSSIBLE VALIDATION VALUES ________________________________ ______________ 1 5 TABLE 3 ACCURACY OF VALIDATI ON BY PERCENT IMAGE _______________________________ 16 TABLE 4 CURVE NUMBER LOOKUP TABLE ________________________________ _____________ 19 List of Equations EQUATION 1 NORMALIZED BURN RATI O ________________________________ ________________ 6 EQUATION 2 DIFFERENCED NORMALIZ ED BURN RATIO ________________________________ ___ 7 EQUATION 3 DIFFERENCED NORMALIZ ED BURN RATIO USING OFFSET _____________________ 7 EQUATION 4 RELATIVIZED DIFFERENCED NORMALIZ ED BURN RATIO _______________________ 7 EQUATION 5 RELATIVIZED BURN RAT IO ________________________________ ________________ 8

PAGE 5

1 Abstract In the summer of 2013 the West Fork Complex Fire raged in southern Colorado. Burning over 109,000 acres the fire has had a significant impact on the hydrology of the area, especially the Rio Grande watershed (RWEACT 2013 ; Yochum and Norman 2014 ) Throughout the literature on hydr ologic responses to fire, fire severity has been found to be the most significant variable (Moody et al. 2008). The aim of this study was to use remote sensing to estimate additional runoff that could be generated from the burned area. Because Landsat da ta is widely available at no cost, it is a useful tool for this analysis. There are several metho ds of determining burn severity, which exploit the difference in spectral reflectance between healthy vegetation and bare ground. A review of the literature in this report attempts to determine the d ifferences between techniques. Using Landsat 7 imagery from September 2012 and 2013, the Differenced Normal ized Burn Ratio (dNBR) was calculated. Adjustments were made for the Scan Line Corrector (SLC) error that occurs in Landsat images since 2003. Gaps in data caused by this error were classified by using a Landsat 8 image trained to the classes that were determined by the Landsat 7 dNBR. Compared to a similar burn severity maps by Yochum and Norman (2014) Verdin et al. (2013) and RWEACT (2013), which used ground truthed U.S. Forest Service data there appeared to be some difference. This study showed a higher burn severity which would lead to greater runoff response. Keywords: Remote Sensing, Burn Sever ity, Normalized Burn Ratio, West Fork Complex Fire, Post Fire Runoff Introduction On June 5 th 2013 a lightning strike ignited the West Fork Fire northeast of Pagosa Springs, Colorado. Another strike to the northwest ignited the Papoose Fire. Together these fires and two others in the area became known as the West Fork Complex Fire (RWEACT 2013). By July 12 th the fire s had consumed 109,000 acres total (RWEACT 2013 ; Yochum and Norman 2014) and 90,000 in the Rio Grande watershed alone (RWEACT 2013). A location map of the fires within the watershed is shown in figure 1. While this study focuses on the science of post fire runoff and the processes of satellite imager y assessment, the social significance of the project is the potential for better remediation. Each year millions of dollars are spent on rehabilitation of burned areas in order to protect human settlements or return the environment to a desired

PAGE 6

2 ecological state. Often the most destructive result of fire is flooding and debris flows. By assessing burn severity using Landsat data, remediation teams can be quickly informed what areas need the most attention. The use of Landsat imagery is substantially chea per than areal or ground assessments and can be performed within weeks after a fire. As the frequency and intensity of fire increases in the Western U.S. the use of this type of assessment is likely to become increasingly relied upon because of its cost e fficiency and relative effectiveness. Figure 1 : Location of the West Fork Fires within the Rio Grande Watershed

PAGE 7

3 Background Most of the area of the fire is characterized by Engelmann spruce ( Picea engelmannii ), White fir ( Abies concolor ) and Subalpine fir ( Abies lasiocarpa ). Lower elevations at the edge of the fire contain Ponderosa pine ( Pinus ponderosa ) and Douglas fir ( Pseduotsuga menziesii ). The highest elevations of the fire are at timberline. This area is home to v arious forbs, shrubs and willows (USFS 2013). According to data available from the Rio Grande National Forest, the watershed above South Fork, Colorado is approximately 1,330 square miles and has an average slope of 25.9%. Additionally, this area receive s an average of 30.5 inches o f precipitation per year (NCAR 2014). In the literature burn severity and fire severity are often distinguished from one another. Soverel et al. (2010) describes fire severity as the direct and immediate effect of combustion o n vegetation and soil, while they describe burn severity as the ecological changes of an area after a fire. Gitas et al. (2009) provides similar definitions, although adding that an aspect of burn severity is ecologic recovery. Miller and Thode (2006) al so term burn severity as the effect of a fire on an ecosystem. Parks et al. (2014) suggests that burn severity is the degree of change to vegetation and soil caused by fire. Throughout this paper burn severity will be the preferred term because it encomp asses the full range of ecological effects. From a holistic approach it is easy to relate the term burn severity to the range of impacts extending beyond simply the immediate effect of combustion. The next few sections describe some of the current resea rch and kno wledge on post fire runoff and classifying burn severity.

PAGE 8

4 Linking Burn Severity to Runoff One of the most prolific soil problems that result from wildfire is water repellency. This issue lead s t o increased runoff and flooding, which can endanger buildings and watersheds. That is why one of the main goals of the Burned Area Emergency Response (BAER) teams is to stabilize soil and protect downstream risks (Bobb e et al. 2002). The cause of water repellency is hydrophobic, long chained, organic compounds in the earth that prevent the soil under the ash layer fro m absorbing water (Moody and Ebel 2012). These organic compounds are generated as a direct result of the heating that occurs during a fire. Generally the conditions required for this process are a residence time of at least 5 10 minutes at 175 280 ¡C or greater (Lewis et al. 2008). The l onger and the hotter the fire, the more water will be evaporated off the soil. A wildfire can drive water out of macro pores, micro pores, and possibly even hydra ted clay minerals (Moody and Ebel 2012). The hotter the fire, the deeper the soil repellenc y may exist (Lewis et al. 2008) Therefore a higher burn severity is likel y to indicate greater hydrophobicity. In addition, the loss of vegetation correlating with high burn severity means that plants will retain less water and sediment This is why BAER teams often assess burn severity using either soil hydrophobicity or the amount of remaining needles (for coniferous forests) (Bobbe et al. 2002) In order to rehabilitate hyper dry soil there must be rewetting. Studies of the Hayman fire in Colorado showed high water repellency for first two years and quickly decreasing level s of repellency thereafter (Robichaud et al. 2013). This would indicate

PAGE 9

5 that the need for remediation is most pressing for the first few years after a fire, but following that period natural recovery can be expected to take place. Artificial rewetting req uires either high water pre ssure or extended contact time. One common approach to extending contact time is to create earth depressions or shelves on hill slopes where water will settle and eventually infiltrate. It has been shown that even short, intense rainfall is unlikely to infiltrate because of insufficient contact time, so artificial rewetting is highly benefici al immediately following a fire (Moody and Ebel 2012) Classifying Burn Severity Throughout the literature there are numerous ways to characterize burn severity. Of these, L andsat based methods are often favored because of their large geographic coverage, extensive archive and cost effective ness (Bobbe et al. 2002). Landsat has also been found to perform favorably to other satellite platforms when validated with ground measurements (Parks et al. 2014) Landsat burn assessment seek s to exploit the spectral difference between Near Inferred (NIR) and Short wave Inferred (SWIR) wavelengths Figure 2 shows the reflectance curves of both burned ground and healthy vegetation displayed along the electromagnetic spectrum. The arrows correspond to the greatest differences between these two curves, which are NIR and SWIR reflectance. NIR is reflected well by healthy vegetation because of the plant's chlorophyll, while it is not reflected much by barren ground or burned areas. Bare soil dead wood and burned areas are highly reflective to SWIR, but not to NI R ( Miller and Thode 2006; USFS 2014).

PAGE 10

6 The Landsat 7 satellite detects light wavelengths reflected in the NIR between 0.77 and 0.90 micrometers and designates them as band 4. It detects light wavelengths reflected in the SWIR between 2.09 and 2.35 m icrometers and designates them as band 7 All of Landsat 7's band designations can be seen in figure 3 For comparison the NIR band for Landsat 8 detects wavelengths 0.85 0.88 micrometers while its SWIR detects wavelengths 2.11 2.29 micrometers (USGS, 2 014). Using the difference between NIR and SWIR a metric was developed to detect spatial difference between healthy vegetation and burned ground. This method is known as the Normalized Burn Ratio (NBR). NBR is calculated by taking NIR values minus SWIR values all divided by NIR plus SWIR values ( Bobbe et al. 2002; FIREMON 2004; McKinley 2004; Miller an d Thode 2006; Gitas et al. 2009; Soverel et al. 2010; Parks et al. 2014). Equation 1 shows the calculation for NBR using Landsat 7. This me asurement can be used alone as single scene NBR (SS NBR). In this case the NBR image is clustered into 50 classes and then grouped into burn severities based on available ground information. The National Land Cover Dataset can also be used to F igure 2 : Reflectance Curves for Burned Ground and Healthy Vegetation F igure 3 : Landsat 7 Band Designations Equation 1: Normalized Burn Ratio

PAGE 11

7 adjust for previously barren areas that did not change (Bobbe et al. 2002). At its highest points the West Fork Complex Fire burned up to timberline. In these areas bare rock would reflect SWI R light similarly to burn areas so this type of adjustment would be necessary ( Miller and Thode 2006; USFS 2014). An advantage to this method is the simplicity of using one image. A disadvantage is that there must be some ground reference to establish thresholds for the burn severity classes. There are three other primary metrics that have been developed in addition to single scene NBR (SS NBR) that take into account pre fire and post fire image s The advantage to all of these methods is that objects that did not change, such as bare rock, can be accounted for directly in the calculation The disadvantage to all of these is that factors such as phenology, sun angle, cloud cover and p recipitation cannot be exactly matched between images (Miller and Thode 2006; Parks et al. 2014). The three duel image methods are Differenced Normalized Burn Ratio (dNBR), Relativized Differenced Normalized Burn Ratio (RdNBR) and Relativized Burn Ratio (RBR). In each case the NBR images are converted from their raw brightness values into at satellite reflectance to try to account for some of t he natural image differences. Equation 2 shows dNBR which is calculated by taking the pre fire NBR image and subtracting it by the post fire NBR image. ( Bobbe et al. 2002; FIREMON 2004; McKinle y 2004; Miller and Thode 2006 ; Soverel et al. 2010 ). In some cases the dNBR offset equation shown equation 3 can be used to better Equation 3: Differenced Normalized Burn Ratio using Offset Equation 2: Differenced Normalized Burn Ratio

PAGE 12

8 account for image differences. This is the standard dNBR multiplied by 1000 all minus the dNBR offset where the dNBR offset value is the average dNBR value of homogenous areas outside the fire perimeter. (Parks et al. 2014). Equation 4 shows RdNBR which is calculated by taking the dNBR divided by the square roo t of the absolute value of t he pre fire NBR. The pre fire NBR in the denominator is divided by 1000 if NBR has been converted into an integer format ( Miller and Thode 2006; Soverel et al. 2010; Parks et al. 2014 ). Equation 5 shows RBR, which is determined by the dNBR d ivided by result of the pre fire NBR plus 1.001. The 1.001 ensures that the denominator will never be z ero (Parks et al. 2014). In order to validate the effectiveness of these methods they are compared to ground measurements, most frequently the Composite Burn Index (CBI). CBI was developed within the framework of FIREMON (fire effects monitoring and inventory prot ocol) project. It is a qualitative measure whose values range from 0 (not burned) to 3 (completely burned). Although there is fluct u ation between fires and because of various factors, a majority of R 2 correlation between NBR/dNBR and CBI is greater than 0.55 (Gitas et al. 2009). There is an ongoing debate regarding the most effective NBR related metric. In a study of fires throughou t the Sierra Nevada Mountains Miller and Thode (2006) determined that R dNBR performed better than dNBR, particularly for heterogeneous landscapes and those with sparse vegetation. For instance using dNBR severely burned Equation 4: Relativized Differenced Nor malized Burn Ratio Equation 5: Relativized Burn Ratio

PAGE 13

9 grassland will look similar to its pre burned conditions, but severely burned forest will show a large amount of change. This is because RdNBR measures relative change, while dNBR measures ab solute change (Miller and Thode 2006; Soverel et al. 2010; Parks et al. 2014). Because of this Rd NBR is thought to be more accurate at classifying high burn severity. However, in homogeneous forests the two methods are likely equivalent in t heir accuracy (Miller and Thode 2006). In a study performed in the Canadian Rockies Soverel et al. (2010) found dNBR to be more accurate than RdNBR. Using the same CBI thresholds that Miller and Thode (2006) used RdNBR was determined to be 65.2% accurate compared to 70.2% for dNBR. Contrary to the previous findings RdNBR did not improve even i n areas with heterogeneous or sparse vegetation. Because there has been no consensus whether d NBR or RdNBR is a better metric Parks et al. (2014) developed a third measurement known as RBR. In this case fires from across the Western U.S. were sampled and severity thresholds were developed separately for each fire. The results of the study showed that on average RBR was most accurate compared to CBI, followed by RdNBR and then NBR. However, performing an ANOVA ana lysis of the accuracies suggested that there wa s no statistically significant evidence at the 95% confidence level to support that there is a difference in accuracy between any of the three methods used in this study. Overall it appears that the three methods achieve similar results, particula rly for homogeneous forests such as exist in the West Fork Complex. In summary there is generally only a small difference in accuracy between the three methods. While improvement in all methods should be made, remote sensing of

PAGE 14

10 burn severity is still pre ferable to the previous method of assessment, which was to draw the different classes on paper while flying over the burned area (McKinley 2004). Objective The obj ective of this study was to use Landsat 7 imagery in order to assess the burn severity of the West Fork Complex Fire The second objective was to use the burn severity map to model runoff in the Rio Grande Watershed. During the development process solving issues with Landsat 7's Scan Line Corrector (SLC) error became a primary goal as well. Methods Landsat Burn Severity Analysis dNBR was selected as the method of calculation because it accounts for the burn like reflectance of alpine areas. In addition, total change versus relative change was not seen as an issue for the type of forest being studied, so a relativized method was unnec essary. The Landsat images were collected from the USGS EarthExplorer website (http://earthexplorer.usgs.gov/). Images that contained the extent of the fire were identified as row 34 path 34, a unique identifier for the spatial location of Landsat imag ery. I mages at this location were selected that contained the least cloud cover obscuring the fire area Since Landsat 8 has only been operation since 2013 all pho tos were considered Landsat 7 photos were considered from 2010 until the time of collecti on. These were sorted until the best images were determined. From Landsat 7 the September 23 rd 2012 and the September 25 th 2013 images were used. From Landsat 8

PAGE 15

11 the October 3 rd 2013 image was used. It was determined there were no useable pre fire ph otos for the Landsat 8 platform. For both sets of data the ENVI image analysis software was used to calculate NBR. The band math tool was used to implement the NBR calculation from equation 1 For the Landsat 7 images the NBR brightness values were c onverted into reflectance then run through the dNBR calculation from equation 2 These NBR and dNBR images were converted into TIFF files so that further analysis could be performed in ArcGIS. Once there a polygon was digitized representing the fire peri meter. Everything in the images outside the perimeter was removed using the ArcGIS mask function. The Landsat 7 dNBR and the Landsat 8 NBR are shown in figure 4 and figure 5 respectively. Figure 4: Landsat 8 Normalized Burn Ratio

PAGE 16

12 For the Landsat 7 dNBR the ArcGIS reclassify function was used to symbolize the raster as 10 classes. Seven classes represented distinct thresholds used by the U.S. Forest Service These values are shown in figure 6 Two classes were values falling above or below the Forest Service thresholds, and the last class represented the no data values caused by the SLC error. From this analysis a map showing low, low moderate, high moderate and high severity classes was generated. This output is shown i n the results section under figure 14 F igure 6 : U.S. Forest Service Thresholds for Differenced Normalized Burn Ratio Figure 5 : Landsat 7 Difference Normalized Burn Ratio

PAGE 17

13 The Landsat 8 NBR was not immediately used to generate a single scene NBR because the value of the thresholds could not be determined. However, this image later became useful for correcting the SLC error. SLC Correction Based on the number of no data pixels within the fire perimeter there was just under 30% of the image missing due to the SLC error. It seemed reasonable that in order to successfully accomplish the secondary objective of creating a hydrologic model there would have to be a continuous burn severity map. A number of methods were employed to try to fill t he gaps in data. Among the tria ls and errors one attempt involved converting the raster to polygons, then using those polygons as training regions (ROIs). Next the no data values from the raster were classified using the Maximum Likelihood Classifier. This method uses the ROIs and distance to automatically determine the best class for a pixel. Unfortunately the gaps tended to classify all as one value. F inally it was determined that the Landsat 7 dNBR could be used to generate thresholds in order to create a SS NBR from the Landsat 8 image. This was accomplished by first classifying the Landsat 8 NBR image into 50 classes using the Iso Cluster Unsupervis ed Classification tool in ArcGIS. The result is shown in figure 7 Next, the polygons representing each of the seven Forest Service defined Landsat 7 classes were used to extract the Iso Cluster pixels falling into the same spa tial extent with the mask tool. The attribute table from each resulting raster was exported into R, where the data could be evaluated. Using R, histograms were developed for each Landsat

PAGE 18

14 7 severity class with the 50 classes on the x axis and frequency of occurrence on the y axis. Histograms are not shown for classes less than unburned because it was irrelevant for determining burn thresholds. The resulting R histograms are shown in figure 8 Afte r evaluating the mean class initially, it was later determined that the median class was a better measure due to the extreme skew of the data. Using the medians of each class thresholds were calculated as the middle value between each median. Table 1 sho ws the thresholds used. Lastly, the ArcGIS reclassify function was used to sort the image into high severity, high moderate severity, low moderate severity, low s everity and unburned. The final output is shown in the results section under figure 15 F igure 7 : Landsat 8 Normalized Burn Ratio Split into 50 Classes

PAGE 19

15 Table 1 : Burn Classification Thresholds Developed for the Landsat Image Table 2 : Possible Validation Values F igure 8 : Histograms of Iso Classes Corresponding to Burn Severity

PAGE 20

16 Later on validation was performed to assess the accuracy of this technique. This was done by using reclassify to match the values of the Landsat7 raster and the Landsat 8 raster so that: 1=high, 2=med_high, 3=med_low, 4=low and 5=unburned. Next the ArcGIS raster calculator was used to subtract the Landsat 7 image from the Landsat 8 image. The resulting values ranged from 4 to 4, with 0 indicating a perfect match and values of |4| indicating a poor match. Each integer unit is representative of being off by one class. Lastly, positive values corresponded to a tendency to underestimate severity, while negative values related to the overestimation of severity. A break down of possible values is shown in table 2 Figure 9 shows the validation map, with areas of blue corresponding to perfectly accurate classification between the Landsat 7 and Landsat 8 images. In the map green indicates areas that were misclassified by only one class, so for instanc e a high burn classed as high moderate, or unburned classed as low burn. The percentages for each validation value are shown in table 3. In summary 40.7% of the pixels matched perfectly, while 81.1% of the pixels were within one class. Negative values w ere indicative of overestimating severity, which judging by table 3 occurred slightly more frequently than underestimation. Understanding that perfect matching is difficult, 81.1% within one class is acceptable for concluding that the Landsat 8 image was successfully classified using the Landsat 7 thresholds. While the no data gaps could not be included in the validation, it is assumed that the accuracy can be interpolated across the image. Table 3 : Accuracy of Validation by P ercent of Image

PAGE 21

17 Hydrolo gic Analysis The method for determining runoff began with creating a curve number (CN) grid. The CN is a value between 0 and 100, with 0 being the most permeable and 100 being the least. Using these values an input can be added and the amount runoff calculated (50% for C N 50, 100% for CN 100, etc.) (Yochum and Norman 2014). The U.S. Corp of Engineers developed an ArcGIS extension known as HEC GeoHMS Using this model a CN grid can be calculated for any area using a digital elevation model (DEM), a soil layer showing h ydraulic soil classes A through D and a F igure 9 : Accuracy Map of Classified Landsat 8 Image

PAGE 22

18 land cover layer. The DEM was taken from the National Elevation Dataset (Data Credit: National Elevation Dataset http://ned.usg s.gov/), the soil layer was obtained from the U.S. Department of Agricultures web soil survey (Soil Survey Staff 2014) and the land use data set was acquired from the National Land Cover Database (Fry et al. 2011). F igure 10 : Digital Elevation Model of Watershed Figure 2 1 : Hydraulic Soil Codes for Watershed

PAGE 23

19 The DEM, hydraulic soil types and land cover are shown in figures 10 11 and 12 respectively. Once all of these files are complied and projected correctly the HMS GeoHEC model synthesizes them into a CN grid using a CN lookup table, which is provided by the user The lookup table is shown in table 4 The values in the table were estimated from various sources. The output is shown in the results section under figure 13 Additional analysis relating the CN grid and the burn severity map were not performed due to time constraints and lack of defined methodology about the process. F igure 12 : Land Use Codes for Watershed Table 4 : Curve Number Lookup Table

PAGE 24

20 Results Using the methods described above three outputs were completed to meet the objectives of this study. The first is the Landsat 7 dNBR with severity classes shown in figure 14 Second is the continuous Landsat 8 NBR that was classified from the Landsat 7 image, shown in figure 15 Lastly, shown in figure 1 3 is the CN grid generated from the HMS GeoHEC extension. Figure 1 3 : Curve Number Grid for Watershed

PAGE 25

21 F igure 14 : Final Landsat 7 Classification Map F igure 15 : Final Landsat 8 Classification Map

PAGE 26

22 Discussion of Results Burn Severity Error and Discussion Aside from the gaps in data caused by the SLC issue, the burn severity analysis went very well. Although in comparison there are some discrepancies between this map and maps used by the U.S. Forest Service A Forest Service map is used by many sources including Yochum and Norman (2014), Verdin et al. (2013) and RWEACT (2013). While a quantitative validation could not be performed, the two were compared visually. The maps si de by side with are shown in figure 16 The high moderate and low moderate classes were lumped for better comparison. Overall there seem to be some correlations with the high severity areas, although this study tended to classify more area as high burn s everity T here seems to be some general agreement with moderate and low severity classes. The unburned class seems to match very well between images. F igure 16 : Comparison of Final Map with U.S. Forest Service Map

PAGE 27

23 There could many reasons for the differences between the two maps. The U.S. Forest Service used Lands at imagery for their analysis, but the date(s) of the image is not known. In addition their image was "ground truthed" a fter the image processing (Belz 2014). Differing image dates could have led to differences in phenology. This study used images from late September, which in these higher altitude regions can tend to be well into fall. An image from late July or early August is likely to look different because vegetation is in a different growth stage. Ground truthing the image was another source of discrepancy. In this process the Forest Service visually verified the areas and changed severity classes where appropriate. However, often this process involves a visual estimation of the area rather than quantitative measu rement of burn severi ty. So while ground truthing may account for much of the discrepancy between the two images it is difficult to say that one image is more correct than the other. Another issue is the sensor differences between Landsat 7 and Landsat 8. There is a slight difference in wavelength range detected by each satellite in its NIR and SWIR bands. While this is likely a very small discrepancy there may be some impact on classification differences between Landsat 7 and Landsat 8. This study used Landsat 8 NBR with thresholds derived from Landsat 7 dNBR, but it is unknown which platform the Forest Service used. Another unknown is the thresholds and methods used for classification. Various sources indicate that these Forest Service maps likely use dNBR, but this ha sn't been confirmed. There are a few other issues within an unknown impact to the final result. Because dNBR is a measure of total change it seems reasonable that it might overestimate burn severity in a heavily forested area like the West Fork Complex. This

PAGE 28

24 would support the ground truth changes made by the Forest Service. However, the spruce beetle also affected much of the area. One study suggested that 88% of the burned area was contin uous areas of beetle kill (USFS 2013). These dying or dead trees would have not reflected NIR as well as healthy trees. Therefore total change expected by dNBR might have been smaller than previously suggested. While all of these factors may have played a role, this study is unable to offer a quantifiable error assoc iated with each of them. The errors that were observed in the SLC correction process may be due to a few factors. Some unknown issues are similar to those described above. These include differing image dates and difference in spectral detection between s atellites. Another issue is with comparing NBR to dNBR. Because dNBR accounts for the pre fire image there are going to be some parts of the image classified very differently. While these errors were not quantified individually, collectively they result in the variation that can be see in the histogram of Iso Cluster classes by Land sat 7 severity class, figure 8 Lastly, there was error in the final classification of the Iso Cluster classes. From the histograms median was chosen to create the threshol ds using the middle value between medians. There may be a better method since the true thresholds probably aren't exactly in the middle of each median. In order to resolve the aforementioned errors ground based validation should be performed. This coul d be incorporated into a future phase of this project. The study would likely involve using CBI plots in relatively homogenous classification areas (at least 3x3 pixels). Once this is complete the values could be compared using an accuracy

PAGE 29

25 assessment mat rix. If required for future hydrologic study the severity map could be updated to reflect ground measurements. Hydrologic Analysis Errors and Discussion The error in the CN grid is primarily due to the values assigned in the CN lookup table. If this project were to be continued then more robust values should be generated based on known attributes of the specific area. Much more analysis remains to be d one regarding the hydrologic analysis of the West Fork Complex Fire on the Rio Grande basin. For this study some attempt was made at using the CN grid as well as incorporating the burn severity grid, but these attempts failed. Eventually time constraints made it unreasonable to finish this objective. If the project was completed values could be validated with a USGS gauge at South Fork, Colorado, where the watershed empties. Other studies have looked at potential runoff and debris flows for this fire, and these could be used for a rough comparison as well. Yochum and Norman (2014) suggest that in many areas a 10 year precipitation event could lead to a 50 or 100 year flood. T he results of this study are likely to vary somewhat since the burn severity map used by Yochum and Norman (2014) was from the Forest Service and would have therefore underestimated burn severity in comparison. Conclusion The West Fork Complex Fire in 2013 burned 90,000 acres of forest in the Rio Grande River basin (Yochum and Norman 2014 ). This was just one of many fires across

PAGE 30

26 the Western U.S. As fire size and frequency increase the ability to successfully assess burn severity and r espond the threat it creates will become more important. Landsat imagery has been shown to be successful in this venture. While there are a number of methods for calculating burn severity it appears that there is no significant difference in accuracy b etween methods This study used dNBR because it accounts for before after images and because a measure of total change was appropriate for a homogenous coniferous forest. The method of correcting the SLC error was developed specifically for this study an d has worked well for this site. The differences between this study and burn severity maps produced by the U.S. Forest Service suggests greater burn severity, and thus more runoff than had been predicted by groups using that data. Lastly t he successful implementation of these procedures for the West Fork Complex Fire suggests that this study could be adapted to assess burn severity for other fires. References Belz, Steve. 2014. [Email communication]. March 19. Bobbe, Thomas; Finco Mark V.; Quayle, Brad; Lannom, Keith. 2002. Field measurements for the training and validation of burn severity maps from spaceborne, remotely sensed imagery. Joint Fire Science Program 2001 2. Salt Lake City, UT: U.S. Department of Agriculture, Forest S ervice, Remote Sensing Applications Center. 14 p. Fire Research and Management Exchange System (FIREMON). 2004. FIREMON BR cheat sheet V4 [Online]. Available: http://burnseverity.cr.usgs.gov/pdfs/LAv4_BR_ CheatSheet.pdf [2014, February 3]. Fry J.; Xian G.; Jin S.; Dewitz J.; Homer C.; Yang L.; Barnes C.; Herold N.; Wickham J. 2011. Completion of the 2006 National Land Cover Database for the Conterminous United States. PE&RS. 77(9): 858 864.

PAGE 31

27 Gitas, Ioannis Z.; de Santis, Angela; Mitri, Geor ge H. 2009. Remote sensing of burn severity. In: Chuvieco, E., ed. Earth observations of wildland fires in Mediterranean ecosystems. Berlin: Springer: 129 148. Lewis, S.A.; Robichaud, P.R.; Frazier, B.E.; Wu, J.Q.; Laes, Denise Y.M. 2008. Using hyperspect ral imagery to predict post wildfire soil water repellency. Geomorphology 95: 192 205. McKinley, Randy. 2004. Evaluation of gap filled SLC off imagery for burn severity mapping [Online]. Sioux Falls, SD: U.S. Geological Survey, EROS Data Center. Availabl e: http://landsat.usgs.gov/documents/burn_severity_mapping_slcoff_ mckinley.pdf [2014, March 22]. Miller, Jay D.; Thode, Andrea E. 2006. Quantifying burn severity in a heterogeneous landscape with a relative version of the delta Normalized Burn Ratio (dNB R). Remote Sensing of Environment. 109: 66 80. Monitoring Trends in Burn Severity (MTBS) 2013. Methods [Online] Available: http://www.mtbs.gov/methods.html [2014, March 17]. Moody, J.A.; Ebel, B.A. 2012. Hyper dry conditions provide new insight into the cause of extreme floods after wildfire. Catena 93: 58 63. National Center for Atmospheric Research (NCAR) Staff (Eds). 2014. The Climate Data Guide : PRISM High Resolution Spatial Climate Data for the Uni ted States: Max/min temp, dewpoint, precipitation [Online]. Available: https://climatedataguide.ucar.edu/climate data/prism high resolution spatial climate data united states maxmin temp dewpoint [2014 March 17]. Rio Grande Watershed Emergency Action Coo rdination Team (RWEACT) [Online] 2013. Available: http://www.rweact.org/ [2013, November 25]. Soil Survey Staff. 2014. Web Soil Survey. U.S. Department of Agriculture, Natural Resources Conservation Service. Available: http://websoilsurvey.nrcs.usda.gov/ Accessed [2014, February 13]. Soverel, Nicholas O.; Perrakis, Daniel D.B.; Coops, Nicholas C. 2010. Estimating burn severity from Landsat dNBR and RdNBR indices across western Canada. Remote Sensing of Environment. 114: 1896 1909. Parks, Sean A.; Dillon, Gregory K.; Miller, Carol. 2014. A new metric for quantifying burn severity: The Relativized Burn Ratio. Remote Sensing. 6: 1827 1844. Robichaud, P.R.; Lewis, S.A.; Wagenbrenner, J.W.; Ashmun, L.E.; Brown, R.E. 2013. Post fire mulching for runoff and erosion mitigation Part I: Effectiveness at reducing hillslope erosion rates. Catena 105: 75 92.

PAGE 32

28 U.S. Forest Service (USFS) 2013. Burned area report [West Fork Complex Fire]. FS 2500 8 [Online] U.S. Department of Agri culture, Forest Service Availa ble: http://www.fs.usda. gov/Internet/FSE_DOCUMENTS/stelprdb5431267.pdf [2013, November 25]. U.S. Forest Service (USFS). 2014 BARC Frequently asked questions [Online] U.S. Department of Agriculure, Forest Service, Remote Sensing Applications Center. A vailable: http://www.fs.fed.us/eng/rsac/baer/barc.html [2014, February 3]. U.S. Geological Survey (USGS) 2014. Landsat Missions [Online]. Available: http://landsat.usgs.gov/ [2014, February 3]. Verdin, Kristine L.; Dupree, Jean A.; Stevens, Michael R. 2 013. Postwildfire debris flow hazard assessment of the area burned by the 2013 West Fork Fire Complex, southwestern Colorado. Open File Report 2013 1259. U.S. Geological Survey. 30 p. Yochum, Steven E.; Norman, John. 2014. West Fork Complex Fire : Potential increase in flooding and erosion. Colorado State Office: U.S. Department of Agriculture, Natural Resource Conservation Service. 26 p.