Machine learning approach to burned area mapping for Southern California

Accurate representation of the location and amount of burned areas is vital to the understanding of patterns and impacts of fires. Some extant burned area maps appear to have high commission errors, which lead to an overrepresentation of burned area. The primary research objective of this study was to assess whether region-specific training data used for machine learning routines improve accuracy of burned area products for western San Diego County. We used training data derived from fine-scale aerial orthoimagery to create and compare three training sets, each with a different Landsat scale sub-pixel burn threshold: 20%, 50%, or 80%. Meaning either 20%, 50%, or 80% of a 30 m × 30 m pixel had to burn for the entire pixel to be classified as burned. High-resolution orthoimagery was used for the creation of the training/testing data as well as determining which sub-pixel threshold leads to more accurate burned area representation. These training data were input into a gradient-boosted regression model. We compared the burned area product from the region-specific gradient boosted model (L-GBRM) to the three products: Monitoring Trends in Burn Severity, Fire and Resource Assessment Program, and the Landsat Burned Area product. We found >20% sub-pixel burn threshold of a Landsat pixel yielded the most accurate classification results. We used a 50% sub-pixel burn threshold for the reference data to compare the results to since the burned area associated with it is closely aligned with the high-resolution orthoimagery determined burn area. The L-GBRM was the most accurate product while also mapping the smallest area burned, suggesting that the extant products have relatively high commission errors. Using region-specific training data achieved a higher accuracy than nationwide training data. Looking at sub-pixel burn thresholds for creating a burned area map could prove to make a more accurate map in terms of area burned represented.