Utility and optimization of LANDSAT-derived burned area maps for southern California

Maps of burned area derived from Land Remote-Sensing Satellite (LANDSAT) system imagery may be more reliable than fire perimeter records based on ground observation or manual cartographic delineation. This study evaluates accuracy of LANDSAT-derived burned area maps associated with 19 fires (65 to 86,776 ha) within shrublands of southern California in the period 1996–2018. High spatial resolution aerial images collected soon after these 19 fires were used to verify burned fractions within LANDSAT ground resolution elements. Validated burned fractions were used to optimize classification thresholds applied to burn severity metrics based on LANDSAT Normalized Burn Ratio (NBR), Normalized Burn Ratio 2 (NBR₂), and Normalized Difference Vegetation Index (NDVI). The burn severity metrics included post-fire spectral index data (e.g., NBR), pre-fire to post-fire differences (e.g., dNBR), and relative differences (e.g., RdNBR). Optimized classifications of dNBR had the greatest overall accuracy (94%) compared to dNBR₂ (93%) and dNDVI (92%). Classifications based on standard difference metrics yielded burn area maps that were 1% more accurate than those based on relative-difference metrics for each spectral index. Classification products based solely on post-fire imagery were 85% to 90% accurate, depending on the spectral index utilized. In comparison, perimeter data from the California Fire and Resource Assessment Programme (FRAP) and Monitoring Trends in Burn Severity (MTBS) substantially overestimated burned extent (by about 44%). The MTBS severity maps excluding the ‘unburned/low severity’ class slightly overestimated burned extent (11%). Commission error in the FRAP data set was attributed to low cartographic detail and inclusion of internal unburned patches. Site-specific differences in unburned soil and vegetation fractions (within partly-burned areas) correlated strongly with overestimation of burned area in the optimized LANDSAT-derived maps; areal overestimates of 10%–15% resulted from data grain size rather than map commission error. The main innovation presented in this study is an empirical method to predict dNBR thresholds based on local (pre-fire) NDVI mean and variance, which produced burned area maps of 90% median accuracy. This LANDSAT-based method could support an efficient reconstruction of fire history in recent decades, which would be more comprehensive and accurate than available fire records for southern California.