Remote Sensing

Abstract

In addition to the immediate loss of biodiversity, wildfires can affect the landscape in which they occur for years after being contained and extinguished. Detecting areas that have been burned has become a prevalent application of remote sensing techniques, often by calculating the Normalized Burn Ratio (NBR) before and after a fire. Another index has more recently been developed called the Normalized Burn Ratio-SWIR (NBR-SWIR) that aims to avoid errors from water bodies being present and can be calculated without ground truth data. This project applies the NBR-SWIR index to the Eagle Creek Fire in Oregon in 2017. Landsat-8 images were acquired one year before and one year after the fire and then topographically corrected. The differences between both the NBR and the NBR-SWIR before and after the fire were calculated. The NBR results were used to create a burn severity map following established ranges. As the NBR-SWIR does not have established ranges, the Iterative Self-Organizing Data Analysis unsupervised classification was applied to classify the images into burned and unburned areas. Finally, a difference map was generated to visually assess how the two indices differ. The NBR-SWIR index overall classified the fire boundary as containing 4.2% more burned area than the NBR index. As a typical range of values for the NBR-SWIR is not yet established, the NBR-SWIR as applied in this study was not a suitable approach on its own in the absence of ground truth data. 

Methods

Preprocessing (Performed with R in RStudio)

Landsat-8/9 Collection 2 Level-2 images were downloaded for the area (Path: 046, Row: 028). Capture dates for the images were from the same month approximately one year before and one year after the fire, so as not to introduce seasonal differences (08/12/2016 and 08/18/2018). Two adjacent 1 arc-second digital elevation models (DEM) from the Shuttle Radar Topography Mission were also acquired and merged before the analysis (USGS). The DEM was projected to the same coordinate system as the Landsat images, and then all three rasters were cropped to an area of interest around the fire. As the area of interest contains considerable topographic variation from Mt Hood in the southeast down to the Columbia River Gorge, topographic corrections were applied to the Landsat images. The slope and aspect, calculated from the DEM, along with the sun azimuth and elevation angles at the time of Landsat image acquisition were used to apply a topographic correction to lessen the effects of the illumination differences (Figures 1 and 2). As this resulted in each band of the images containing negative values, the absolute value of the minimum value of each band was added to each cell in order to shift each band to positive values and to be able to calculate the burn indices.

Methods

Processing (Performed with R in RStudio)

The Normalized Burn Ratio (NBR) uses the near-infrared (NIR) and shortwave-infrared (SWIR2) bands to distinguish between burned and unburned areas. From pre-fire to post-fire, in typical applications the NIR reflectance decreases while SWIR2 increases. The bands are subtracted from each other and then normalized by their overall brightness, as given in the following formula.

NBR = (NIR - SWIR2) / (NIR + SWIR2)

Once the NBR is calculated for the pre-fire and post-fire images (Figure 3), the difference is taken to be able to quantify the change between unburned and burned areas. This difference is referred to as the Delta Normalized Burn Ratio (ΔNBR) as calculated below.

ΔNBR = NBRpre-fire - NBRpost-fire

Low positive to negative values indicate unburned areas or areas of regrowth, while positive values indicate burned areas. These values can be classified into burn severity levels developed by the USGS (Key and Benson, 2006). Using the area of one cell, the area of each burn severity level was calculated (Table 1). From these classifications, a burn severity map can be created in order to estimate the ecological change caused by the fire (Figure 6). A relatively new index proposed for Landsat-8 images by Liu et al. (2020) utilized both SWIR bands to increase the spectral separation between burned and unburned areas. This fire index is named the Normalized Burn Ratio SWIR (NBR-SWIR) and calculated as follows.

NBR-SWIR = (SWIR2 - SWIR1 - 0.02) / (SWIR2 + SWIR1 + 0.1)

Subtracting 0.02 from the numerator is intended to reduce the changes in water to close to zero, while adding 0.1 in the denominator attempts to avoid amplification of atypical water features. The pre-fire and post-fire indices (Figure 4) are similarly subtracted as below to highlight changes.

ΔNBR-SWIR = NBR-SWIRpre-fire - NBR-SWIRpost-fire

The ΔNBR-SWIR does not have the same range of values as the ΔNBR. It thus cannot be readily classified into the same burn severity levels. Alternate comparison methods between the two indices were instead explored below.

Methods

Index Comparison (Performed with ArcGIS Pro)

As there was not enough information known to be able to classify the ΔNBR-SWIR into burn severity levels, instead both ΔNBR-SWIR and ΔNBR were classified into burned and unburned areas. The Iterative Self-Organizing Data Analysis (ISODATA) classification was applied to each difference image. This unsupervised classification does not rely on training data and iteratively merged clusters into the two categories of burned and unburned (Figure 5). From these classified images, the boundary of the visual burned area was delineated in order to isolate the area of the fire from errors in the classification. The areas of burned and unburned classifications were then calculated for each index (Table 2).

In order to visually assess the differences between the two indices, a difference map was generated (Figure 8). Cells that only the NBR index classified as burned are shown, as well as those cells that only the NBR-SWIR index classified as burned. It also contains the remaining cells where both indices were in agreement. Finally, areas of each category were also calculated (Table 3).

Discussion

The range of calculated ΔNBR values aligned with the typical ranges given by Key & Benson (2006) in Table 1. The values were readily able to be classified into interpretable severity levels that could be displayed as in Figure 6.

From Table 2, the NBR-SWIR index overall classified the fire boundary as containing 4.2% more burned area. This is supported by Table 3, in which 4.2% of the fire boundary area was only classified as burned in the NBR-SWIR index. This is also visualized in Figure 7, in which the NBR-SWIR index classified more burned area along the southern extent of the fire boundary (where the fire originated) as well as on the steeper slopes to the northeast. Less than 1% of the fire boundary area was classified as burned in only the NBR index, and these areas appear to be more along the western edges of the fire extent where the channels that slope down into the Gorge begin to form.

From this analysis, the NBR-SWIR was not a suitable index in the absence of ground truth data. The range of calculated NBR-SWIR values does not align with the range of NBR values, and there are not enough studies as of yet to determine what the expected NBR-SWIR range should be. It would be worth comparing the results of the ISODATA classification for the areas outside of the fire boundary (Figure 5). There are numerous errors in both classification, some of which may be differences in moisture in agricultural fields to the west and urban areas to the east. The NBR-SWIR classification has visually fewer errors on the slopes of Mt Hood, but it actually has more errors along the Columbia River. Part of the rationale behind its development was that it should produce fewer errors from water bodies (Liu et al. 2020). Utilizing an ISODATA unsupervised classification to produce just two classifications was likely not the ideal method on its own. In another study to automate burn area mapping, Pulvirenti et al. (2020) also utilized an ISODATA classification. However, this was as an exploratory method to determine inputs for automatic thresholding. When first describing the NBR-SWIR index, Liu et al. (2020) instead used the maximum between-class variance (Otsu) algorithm to divide images into burned and unburned areas. Finally, instead of trying to directly compare two different indices, further analysis could explore merging multiple indices to better highlight burned areas as in Vhengani et al. (2015).

Conclusion

The Columbia River Gorge has continued to change since the Eagle Creek Fire, and further study can reveal the extent of renewal and recovery. As the field of remote sensing continues to progress, so too can our ability to monitor and analyze environmental phenomena. Field investigations and ground truthing methods remain a critical component of burn area mapping. The NBR-SWIR index was not a suitable approach on its own without a more robust classification and thresholding methods. As both indices can be derived from the same data source, it seems sensible to at least begin with calculating the NBR before proceeding with any additional analysis.

References

Alcaras, E., Costantino, D., Guastaferro, F., Parente, C., & Pepe, M. (2022). Normalized Burn Ratio Plus (NBR+): A New Index for Sentinel-2 Imagery. Remote Sensing, 14(7), Article 7. https://doi.org/10.3390/rs14071727

Calhoun, N. C., Burns, W. J., Hayduk, S. H., Staley, D. M., & Kean, J. W. (2019). Post-fire rockfall and debris-flow hazard zonation in the Eagle Creek Fire burn area, Columbia River Gorge, Oregon: A tool for emergency managers and first responders. 581–588. https://pubs.usgs.gov/publication/70203873

Key, C. H., & Benson, N. C. (2006). Landscape Assessment (LA). In: Lutes, Duncan C.; Keane, Robert E.; Caratti, John F.; Key, Carl H.; Benson, Nathan C.; Sutherland, Steve; Gangi, Larry J. 2006. FIREMON: Fire Effects Monitoring and Inventory System. Gen. Tech. Rep. RMRS-GTR-164-CD. Fort Collins, CO: U.S. Department of Agriculture, Forest Service, Rocky Mountain Research Station. p. LA-1-55, 164. https://www.fs.usda.gov/research/treesearch/24066

Liu, S., Zheng, Y., Dalponte, M., & Tong, X. (2020). A novel fire index-based burned area change detection approach using Landsat-8 OLI data. European Journal of Remote Sensing, 53(1), 104–112. https://doi.org/10.1080/22797254.2020.1738900

Pulvirenti, L., Squicciarino, G., Fiori, E., Fiorucci, P., Ferraris, L., Negro, D., Gollini, A., Severino, M., & Puca, S. (2020). An Automatic Processing Chain for Near Real-Time Mapping of Burned Forest Areas Using Sentinel-2 Data. Remote Sensing, 12(4), Article 4. https://doi.org/10.3390/rs12040674

USGS. (n.d.). EarthExplorer. https://earthexplorer.usgs.gov/ 

Vhengani, L., Frost, P., Lai, C., Booi, N., van den Dool, R., & Raath, W. (2015). Multitemporal burnt area mapping using Landsat 8: Merging multiple burnt area indices to highlight burnt areas. https://doi.org/10.1109/IGARSS.2015.7326740