Calculating the Effects of the Tubbs Fire in Napa, Sonoma, and Lake counties, California using Binary Change, Normalized Difference Vegetation Index, and Normalized Burn Ratio

Introduction

The purpose of this remote sensing analysis is to determine the effects of the Tubbs Fire that occurred in California in 2017 using binary change, normalized difference vegetation index (NDVI), and normalized burn ratio (NBR). The Tubbs Fire, named the most destructive California wildfire, started on October 8th around 9:43 PM near Calistoga, California (Griggs et al., 2017). Due to high winds from the northeast, the fire quickly spread. The damage from the fire was felt strongest in Santa Rosa, a city southwest from the fire’s origin. During the time this fire was occurring, two other major fires, the Atlas and Nuns Fires, were burning areas to the southeast of Santa Rosa. By the time the fire was contained 32 people were killed and about 5,500 structures were destroyed (Jansen, 2017, Krishnakumar et al., 2017). As of now the cause of the fire is suspected to be a downed power line, but investigations are still continuing. 

The Tubbs fire occurred in the central coast bioregion of California. The ecosystems in the area consist of grassy savannas, oak woodlands, and chaparral (Forests Forever, 2018). Due to previous fire suppression in much of the western United States, the majority of these ecosystems have a build-up of dead plant material that would normally be burned away each year in low-intensity fires. This was also the case when the Tubbs Fire occurred. In addition, the fire picked up in speed and intensity once it met the town of Santa Rosa and was able to burn many of the homes. 

Two other major factors played a role in the Tubbs Fire: humidity and winds. If the winds are high, as was the case during the Tubbs Fire, then more fuel (oxygen) is brought to the fire which helps it continue to burn. High winds can also carry embers over a mile causing the fire to spread faster and out-pace fire barriers. When the humidity is low, fuel sources such as dead vegetation are drier and thus easier to ignite. The night of October 8th the humidity was 15% (Funke, 2018). The Tubbs Fire was a combination of low humidity, high winds, dry air, and plenty of fuel sources that caused it to be the most destructive fire in California wildfire history. 

In this analysis a subset of the Tubbs Fire perimeter from a Landsat 8 satellite image was created to focus the following three functions to determine areas affected by the fire. Binary change determines which areas in an image experienced a change in reflectance and then creates an output image of those areas. In addition to the function in Erdas, two models of binary change were completed as well to create outputs that could be used in ArcMap. NDVI is a function that uses the near infrared (NIR) and red wavelengths of the electromagnetic spectrum to determine which areas have healthy vegetation. NBR uses NIR and short-wave infrared (SWIR) wavelengths to determine which areas had the highest burn intensity. The combination of all these functions highlighted which areas were most heavily affected by the Tubbs Fire (figure 1).

Figure 1. Area burned by the Tubbs Fire in 2017 located in the coastal mountain region of California.

Methods

Data

The two primary images for this analysis were obtained from the United States Geological Survey (USGS) GloVis site. Both images were from the Landsat 8 satellite with one image taken October 1st, 2017 and the other December 7th, 2017 correlating to before the fire started and after it was contained (United States Geological Survey, 2018). 

Remote Sensing Analysis

The first step in the analysis was to obtain the data from the USGS. Two images were imported: an image of Sonoma, Napa, and Lake Counties before the Tubbs Fire (October 4, 2017) and an image of the same area after the Tubbs Fire was contained (December 7, 2017). 

Figure 2. GloVis site with the Landsat 8 images to be downloaded for analysis.

Next, bands 2 through 7, or blue through short-wave infrared (SWIR) were imported from TIFF files to IMG files in Erdas Imagine (figure 3).

Figure 3. Import function in Erdas Imagine used to bring in TIFF files and convert them to IMG files.

Once the images were imported, the layer stack function was run which took bands 2 through 7 and stacked them to make one composite image (figure 4).

Figure 4. Layer stack function in Erdas Imagine.

A subset of the Tubbs Fire from the images were obtained using a shapefile of a polygon digitized from a Tubbs Fire feature class (figure 5).

Figure 5. Subset of the October and December Landsat 8 images using the subset tool in Erdas Imagine.

The subset of the images was analyzed further using binary change detection, NDVI, and Normalized Burn Ratio Area. 

Binary Change Detection

In the binary change detection function, the difference between the two subsetted images was computed using subtraction and specifically with band 5 which correlates to the near infrared (NIR) (figure 6). 

Figure 6. Binary change output showing areas that had a significant change after the Tubbs Fire. The darker areas are those with more change.

This image is the corrected version of the original output. When the process was first run, a pansharpened image was used. This in itself is not a bad process to run. However, when the pansharpen function was used, the multiplicative method was employed. This changes the radiometric properties of the image from the original. When the subsequent models were run, they resulted in almost no areas with significant changes. Once this error was recognized, a non-pansharpened image was used and the process ran correctly.

To look more closely at the areas that changed, two models were run on the NIR bands of each image. In model one, the pixel values were subtracted from each other and a constant of 127 was added to remove any negative values (figure 7).

Figure 7. Model one in model builder. The model is the same as binary change but it adds a constant to remove any negative values.

Another model was created to show the pixels that changed between October 4th and December 7th, 2017 using the mean pixel value and the standard deviation from the first model’s results (figure 9). In this case, the mean was 42.856 and the standard deviation was 64.651. The threshold for which pixels changed between the two time periods and which did not was defined as the mean pixel value plus three times the standard deviation. This was the threshold because in model one a constant of 127 was added to remove any negative values in the histogram. The mean plus three times the standard deviation was used to determine the threshold because data beyond three standard deviations is significantly different from the mean, approximately 1% of the data (figure 8). 

Figure 8. Model two in model builder. This model creates two values: a '1' if the pixel changed based on the threshold calculated in model one or '0' if the pixel did not change.

Then the results from model two were used to create a map in ArcMap showing which areas had a significant change based on the NIR wavelength. 

Normalized Difference Vegetation Index

The NDVI function was relatively simple. The function was run two times, once on the October 4th subset image and once on the December 7th subset image. The parameters were set to Landsat Multispectral for each run (figure 9).

Figure 9. NDVI function in Erdas Imagine.

Normalized Burn Ratio

The NBR ran the same as the NDVI with the same parameters but only the post-fire image was run through the function (figure 10).

Figure 10. NBR function in Erdas Imagine.

Results

Figure 11. Results of the binary change detection.

Figure 11 shows the results of the binary change detection function. The Tubbs Fire caused the most significant changes on the slopes of the mountainous terrain areas. The concentration of changes in these areas could be due to the winds which pushed the fire southwest. The changed areas also follow the main path of the fire and are less prevalent on the perimeter of the fire where conditions were probably less severe than in the center of the fire’s path.

Figure 12. Results of the NDVI function before and after the Tubbs Fire. The left map is before the fire and the right map is after the fire.

Figure 12 shows that there was a significant decrease in vegetation health after the Tubbs fire. The decrease is visible as an increase in red or yellow areas (unhealthy or very unhealthy vegetation) and a decrease in green areas (healthy vegetation). The main area of decrease was the center of the area of interest in the mountainous region. This function was performed on band 5 (NIR) because healthy vegetation reflects NIR radiation and unhealthy vegetation does not. Another smaller area of change is the very top of the perimeter. In the left map, the area is mostly healthy vegetation. In the right map, the area is more red and yellow indicating unhealthy or very unhealthy vegetation. 

Figure 13. Burn Severity map of the area affected by the Tubbs Fire.

Figure 13 shows which areas had the highest burn severity in the area affected by the Tubbs Fire denoted as red. The areas with the highest burn severity are in the mountainous terrain of the area of interest along the main path of the fire. These results mirror the findings in the NDVI and the binary change outputs. 

Conclusions

Overall the binary change, NDVI, and NBR functions proved useful in identifying areas that were affected by the Tubbs Fire. The binary change detection showed changes in the landscape within the Tubbs Fire perimeter which isn’t surprising considering it has been labeled the most destructive wildfire in California history. 

The NDVI function indicated the highest losses in healthy vegetation cover were in the center of the fire perimeter starting in the north and moving south. The NBR function mirrored the spatial patterns found in the binary change and NDVI image indicating fire intensity was highest in the center of the mountainous region where the main path of the fire was. 

These methods do have some limitations. Both NDVI and NBR used the NIR band which is most useful in detecting changes in vegetation. The most destructive aspect of the Tubbs Fire was its effect on Santa Rosa, a city in the southwest corner of the Tubbs Fire perimeter. Because cities tend to have less vegetation, these processes did not pick up on the changes that occurred in the area. Different bands would have to be employed to assess changes the fire caused in Santa Rosa. 

This analysis shows that basic remote sensing processes such as binary change, NDVI, and NBR are helpful in determining areas affected by wildfires. However, these processes used the NIR band to assess vegetation changes. This meant that urban areas were not assessed nearly as well as vegetated areas. In future studies other bands should be studied using different processes to assess changes in urban areas caused by wildfires. Despite this shortcoming this analysis has proved remote sensing to be a useful tool in hazard monitoring, especially regarding wildfires. 

References

California Department of Forestry and Fire Protection. (2018, January 12). Top 20 Most Destructive California Wildfires. Retrieved from http://www.fire.ca.gov/communications/downloads/fact_sheets/Top20_Destruction.pdf

Forests Forever. (2018). Central Coast Bioregion. Retrieved from http://www.forestsforever.org/archives_resources/cabioregions/centralcoastbioregion.html

Griggs, T., Lai, R.K.K, Park, H., Patel, J.K., & White, J. (2017, October 12). Minutes to Escape: How One California Wildfire Damaged So Much So Quickly. New York Times. Retrieved from https://www.nytimes.com/interactive/2017/10/12/us/california-wildfire-conditions-speed.html

Humboldt State University. (2017). Normalized Burn Ratio. [Website page]. Retrieved from http://gsp.humboldt.edu/OLM/Courses/GSP_216_Online/lesson5-1/NBR.html

Jansen, B. (2017, October 13). California Fires Have Killed 32 People; here’s a Look at Some of the Victims. USA Today. Retrieved from https://www.usatoday.com/story/news/2017/10/13/california-fires-victims- killed/762312001/

Krishnakumar, P., Fox, J., & Keller, C. (2017, October 23). Here’s Where More than 7,500 Buildings were Destroyed and Damaged in California’s Wine Country Fires. LA Times. Retrieved from http://www.latimes.com/projects/la-me-northern-california-fires-structures/

S Funke. (2018, January 9). Fire Ecology of Non-Scientists: The Fire Triangle & Fire Behavior. [Blog Post]. Retrieved from https://www.pepperwoodpreserve.org/2018/01/09/fire-ecology-for-non-scientists-the-fire-triangle-fire-behavior/

United States Geological Survey. (2018). Landsat 8 Satellite Imagery [satellite image]. Retrieved from https://glovis.usgs.gov/

Remote Sensing Lab 8

Introduction

Determining land use for parcels of land is important in a variety of areas such as real estate and construction. Often the cost to send out a team to survey the area and determine the land use is too time consuming or the cost is too high. Remote sensing allows analysts to determine the land use from satellite data and provide that information to clients. In this lab, the goal is to gain experience in measuring and interpreting spectral reflectance signatures of various surfaces using satellite imagery. After collecting the signatures they are analyzed to determine whether their spectral signatures verify the land use type identified on a reference image. The second portion of the lab involves monitoring the health of vegetation and soils using band ratio techniques. Vegetation monitoring is useful for a wide variety of professions such as farmers and conservation scientists.

Methods

Part 1: Spectral Signature Analysis

The first part of this lab involved collecting spectral signatures of several different surfaces from a satellite image. Google Earth was used as a reference to identify the specific land cover type of areas on the satellite image (figure 1).

Figure 1. Google Earth and satellite image used to collect spectral signatures.

 Once the area of interest was determined, a polygon was drawn within the area. Then the mean pixel values within the polygon were inputted into the signature editor where plots of each surface’s spectral signature could be obtained (figure 2).

Figure 2. Spectral signatures obtained using polygons on the satellite image.

Once all 12 signatures were obtained, the highest and lowest reflectance bands were recorded for each surface shown in figure 3.

Figure 3. Spectral signatures with highest and lowest reflectance.

Part 2: Vegetation Health Monitoring

In this part of the lab the health of vegetation was determined using the normalized difference vegetation index (NDVI).

NDVI = NIR – Red/NIR + Red

The NDVI tool was opened in Erdas Imagine with the specified input image, the sensor set as Landsat 7 Multispectral and NDVI set as the function (figure 4).

Figure 4. NDVI function.

Finally, a band ratio using the ferrous mineral ratio was implemented to monitor the spatial distribution of iron contents in soils within Eau Claire and Chippewa Counties.

Results

From the signature tool spectral signature curves for each of the twelve features were obtained.

Figure 5. Standing water spectral reflectance curve.

Figure 6. Moving water spectral reflectance curve.

Figure 7. Deciduous forest spectral reflectance curve.

Figure 8. Evergreen forest spectral reflectance curve.

Figure 9. Riparian vegetation spectral reflectance curve.

Figure 10. Crops spectral reflectance curve.

Figure 11. Dry soil spectral reflectance curve.

Figure 12. Wet soil spectral reflectance curve.

Figure 13. Rock spectral reflectance curve.

Figure 14. Asphalt highway spectral reflectance curve.

Figure 15. Airport runway spectral reflectance curve.

Figure 16. Concrete surface spectral reflectance curve.

Figure 17. Spectral reflectance curves for all 12 surfaces.

1.       While all the curves are slightly different in figure 17, there are a couple of trends that are apparent in the curves. There are six curves that follow the same general trend but at different respective reflectance levels: airport runway, dry soil, wet soil, rock, asphalt, and concrete surface. Rock had the highest reflectance out of the five and the asphalt highway had the lowest overall reflectance. All these surface have the same general trend because they are all similar surfaces (flat and made from minerals). The rock had the highest reflectance because the minerals were light colored which increased the reflectance. The asphalt had the lowest reflectance because of the dark coloring of the surface which absorbs more wavelengths. The second trend was the two water curves which were similar to each other because they both depicted water, but the standing water had a higher reflectance because it had a flatter surface that reflected energy better. Finally, the Crops, deciduous forest, riparian vegetation, and evergreen forest all had similar curves. These are all similar in overall shape because they all have some form of vegetation in them and therefore they will exhibit the same overall curve trend even though the level of reflectance may vary. The two water signatures varied the most in trend from the other surface because its properties are so different. In addition, the vegetation surfaces and the ground surfaces are both located on land and thus share commonalities that the water surfaces do not. The vegetation spectral curves differed in the NIR or band 4 from the 6 hard surfaces because plants reflect more energy to avoid protein deformation while the hard surfaces absorbed this wavelength.

Figure 18. Vegetation cover of Eau Claire and Chippewa Counties.

 Figure 18 is a map of the vegetation cover of Eau Claire and Chippewa Counties. This map was based on the NDVI function which measures the health of vegetation using the red band and NIR band. The northeastern portion of the map is dark green which indicates high amounts of vegetation. The eastern half of the map is lighter in color which indicates lesser amounts of vegetation. The lack of vegetation in these areas is related to cities such as Eau Claire, roads, and crop fields that may not have crops currently growing.

Figure 19. Ferrous mineral distribution in Eau Claire and Chippewa Counties.

Figure 19 shows the distribution of ferrous minerals in Eau Claire and Chippewa Counties.  In the map overall, there is a higher concentration of ferrous minerals in the eastern half compared to the western and northwestern portion. Dark orange sections of the image denote high concentrations of ferrous minerals, which would appear in the image where there is exposed soil. Upon closer inspection, it is apparent that higher ferrous mineral concentrations coincide with exposed crop fields and exposed soil surfaces. This is more common in urban areas such as Eau Claire and Chippewa Falls. The light orange areas correspond to heavily vegetated areas that block the ground surface which covers the ferrous minerals. This is more common in the eastern portion of the map where there is more vegetation cover. 

Both portions of this lab utilized spectral signature curves to determine land cover characteristics such as land cover type, vegetation cover, and ferrous mineral concentrations. The ability to use remote sensing to determine land characteristics is a useful tool that can save time, money, and resources. Remote sensing can also analyze areas that are inaccessible to ground studies. Upon completion of this lab basic skills in spectral signature analysis were obtained. 

Sources

Satellite image is from Earth Resources Observation and Science Center, United States Geological Survey