A statistical approach to site-specific thresholding for burn severity maps using bi-temporal Landsat-8 images

Abstract

Burn severity mapping facilitates post-fire management and restoration and predicts surface erosion and landslide risk. Different severity levels are usually distinguished by fixed threshold values with remote sensing techniques. Since the climate, ecosystem, geology, and morphology control the destruction level of forest fires, site-specific class thresholds should be considered to discriminate severity classes precisely. Therefore, the purpose of this study is to produce an accurate burn severity map using spectral indices with site-specific thresholds for unburned, low, moderate and high severity classes. In this context, pre- and post-fire Landsat 8 images were used to produce bi-temporal burn severity indices such as normalized burn ratio (NBR), normalized burned index (NBI), normalized difference vegetation index (NDVI), and green optimized soil adjusted vegetation index (GOSAVI). An alternative classification method based on a statistical distribution-based clustering approach was employed on the differential indices to determine severity class thresholds. The proposed thresholds were validated by the composite burn severity index (CBI) ratings of the field sampling points. The overall classification accuracy was found to be between 50% and 92.5%. In addition, the results were compared with the thresholds published in the literature. Consequently, this methodology can be used as adaptive thresholding in similar ecological and morphological zone to determine the burn severity classes.