Detecting tree mortality with Landsat-derived spectral indices: Improving ecological accuracy by examining uncertainty

Satellite-derived fire severity metrics are a foundational tool used to estimate fire effects at the landscape scale. Changes in surface characteristics permit reasonably accurate delineation between burned and unburned areas, but variability in severity within burned areas is much more challenging to detect. Previous studies have relied primarily on categorical data to calibrate severity indices in terms of classification accuracy, but this approach does not readily translate into an expected amount of error in terms of actual tree mortality. We addressed this issue by examining a dataset of 40,370 geolocated trees that burned in the 2013 California Rim Fire using 36 Landsat-derived burn severity indices. The differenced Normalized Burn Ratio (dNBR) performed reliably well, but the differenced SWIR:NIR ratio most accurately predicted percent basal area mortality and the differenced normalized vegetation index (dNDVI) most accurately predicted percent mortality of stems ≥10 cm diameter at breast height. Relativized versions of dNBR did not consistently improve accuracy; the relativized burn ratio (RBR) was generally equivalent to dNBR while RdNBR had consistently lower accuracy. There was a high degree of variability in observed tree mortality, especially at intermediate spectral index values. This translated into a considerable amount of uncertainty at the landscape scale, with an expected range in estimated percent basal area mortality greater than 37% for half of the area burned (>50,000 ha). In other words, a 37% range in predicted mortality rate was insufficient to capture the observed mortality rate for half of the area burned. Uncertainty was even greater for percent stem mortality, with half of the area burned exceeding a 46% range in predicted mortality rate. The high degree of uncertainty in tree mortality that we observed challenges the confidence with which Landsat-derived spectral indices have been used to measure fire effects, and this has broad implications for research and management related to post-fire landscape complexity, distribution of seed sources, or persistence of fire refugia. We suggest ways to account for uncertainty that will facilitate a more nuanced and ecologically-accurate interpretation of fire effects. This study makes three key contributions to the field of remote sensing of fire effects: 1) we conducted the most comprehensive comparison to date of all previously published severity indices using the largest contiguous se