Extracting Wildfire Scars from Satellite Imagery

Updated: Apr 9, 2021

ENVI, ArcMap 10.7, USGS EarthExplorer

In a climate ever inching toward the hottest on record, conditions are increasingly becoming more favorable for wildfire propagation. California and Australia are no strangers to wildfire season and each year seems more intense than the last. The story is regrettably familiar northward in the provinces of Alberta and British Columbia.

In September, 2017, Waterton Lakes and the western surrounding Rocky Mountain forests suffered from one of the worst wildfires ever experienced in the area, named the Kenow Fire. High peaks and low valleys created formidable fire corridors, scorching dense forests into ash.

By representing the burned extents of the Kenow Fire, many observations can be made including changes in wildlife habitats and migration pathways, or urban development and mitigation. Multi-spectral imagery analysis provides an accurate representation of the burned extents by leveraging the spectral profiles of vegetation in the infrared spectrum. Here, I detail the workflow in ENVI including image processing and analysis.

Obtaining Satellite Imagery

The USGS Earth Explorer website is an incredible resource for accessing various platform imagery. For this project, Sentinel or Landsat imagery is appropriate as they both have the four bands required for burn ratio analysis. Fortunately, Landsat 8 data is free of any cloud cover and other distracting features. I'll be using LS8 data from August 12, 2017 as my pre-fire image and LS8 data from September 29, 2017 as my post-fire image.

The data is downloaded and unzipped to reveal a collection of TIF files for each band.

Imagery Preparation

Layer Stacking: Load all the required bands into the data manager and open the Layer Stacking tool. This tool is used to combine separate band files into one multispectral .dat file.

For burn ratio purposes, the bands of interest are band 5 (NIR) and band 6 (SWIR), but I'll keep the MSS bands as well.

Reorder the raster bands in the input list to ascending order or the bands will be improperly identified in the stacking process.

Resize: Landsat 8 images are extremely large data files. The resize tool allows us to extract a subset of samples and lines to focus the analysis. Use the ROI tool and draw an area around the Kenow fire and use it as the defining boundaries in the Resize tool.

Projection: In many cases, a projection will be applied to conform both images. In this case, the LS8 data has the same projection.

Below is a swipe-over video of the Color-Infrared band combination for both pre- and post-fire imagery highlighting the presence of vegetation. Near infrared is heavily reflected by healthy vegetation and appears red in this combination imagery.

The fire extents are made very clear using this band combination. Next, I will extract spectral values in cells within the burned area to calculate total area burned. The resolution of LS8 data is 30m, therefore the accuracy of the analysis will be limited to 30m cell size.

Calculating Burn Ratios

In order to extract the burned areas from the post-fire image, a normalized burn ratio will be calculated. Healthy vegetation reflects wavelengths in the near infrared (NIR) spectrum, as shown by red hues above. Notice the burned areas lack the red hue, indicating the absence of reflective vegetation. Conversely, healthy vegetation absorbs energy in the shortwave infrared (SWIR) while burned areas reflect these wavelengths.

To visualize these diverging spectral profiles, I'll create two regions of interest in ENVI and use the Endmember Collector tool.

With the ROIs saved within the post-fire image, search for the tool "Endmember Collector." Select the post-fire image and import the ROIs attached to the image. Click Plot.

Endmember Collection Spectra plot shows the range in spectral values within each ROI for all bands in the image data. Bands 1, 2 and 3 are visible light spectra, Band 4 is near infrared, and Bands 5 and 6 are shortwave infrared wavelengths.

Notice the significant difference in each of the profiles inside the NIR and SWIR wavelength ranges. Pre-fire healthy vegetation strongly reflects NIR, resulting in a high value at band 4. Burned areas strongly reflect SWIR, resulting in a high value in band 5. We'll leverage this difference in spectral profiles in the normalized burn ratio (NBR) calculation.

The NBR equation is as follows. I'll use the Band Math tool to perform the calculation on both the pre-fire and post-fire imagery.

In the Band Math tool, create the NBR equation using variables b1 and b2, and be sure to include the conversion of values to a float data type. The next window will request the b1 and b2 definitions. Select b2 = Band 5 (SWIR). Designate a save name and location and click OK. ENVI will now process the imagery and create a new single-band raster that houses the NBR value for each cell in the extent.

Display the NBR imagery and explore the data values. Healthy vegetation is represented by high NBR values or brighter cells.

Burn Severity

Each NBR raster calculated for both pre- and post-fire imagery can be used to analyze the change in values over time, thus resulting in a burn severity value. How heavily burned are the areas? Is there evidence of regrowth in the area?

By calculating the NBR on the same area one year later, vegetative regrowth can be assessed and impacts to wildlife habitats and migration patterns can be observed.

Burn severity is then derived from the difference between the pre-fire NBR values and the post-fire NBR values. Returning to the Band Math tool, our expression will simply be b1-b2, where b1 will be paired with the pre-fire NBR band 1, and b2 will be paired with the post-fire NBR band 1.

dNBR values greater than 0.1 indicate increasing burn severity, as shown in red in the image below. dNBR values less than -0.1 indicate potential regrowth. All values within -0.1 to 0.1 are unburned. The histogram tool displays the range in dNBR as shown below.

Extracting Burned Areas

The dNBR raster contains values representing regrowth and burns, which can be extracted using a mask and used to calculate total burned area. Search for the tool "Build Raster Mask" and select the dNBR raster. The most appropriate method for limiting the data to the burned areas is to input a minimum value of 0.1 based on the dNBR raster.

Apply the new mask raster to the dNBR to effectively remove all cell data outside the burned areas. Save the dNBR file as a .ENVI and select the mask to be applied.

NBR analysis does not extract just burned areas. Clouds, snow, and other distracting features in the imagery can cause noise that must be removed prior to calculating area statistics. To remove the noise, I brought the saved TIFF file into ArcMap for some post-processing.

ArcMap Post-Processing

Once loaded into ArcMap, I convert the TIFF mask to a set of vector polygons, enter an edit session and delete all of the polygons with a 0 value. By loading the original post-fire satelite image, I can delete all the false polygons that may have been associated with shadows, clouds, or other non-burned features. It is important to note that the actual burn sites were very well represented in the mask.

Once cleaned up, save edits and exit edit mode. Add a new field to the attribute table and calculate the area.

The total area affected by the Kenow fire is 362 sqkm

Next, I'll explore how NBR can be used to identify areas of regrowth years after the Kenow fire.

Stay tuned!

References and Data Sources:

Landsat Data: https://earthexplorer.usgs.gov/


Emily Gillis

GIS Specialist in the making

© 2021 Emily Gillis

  • LinkedIn