Integration of Ground Truth Data via Cloud Computing for Enhanced Burn Severity Mapping – An Example from Honduras

Abstract

Until now, most severity products are generated from a reclassification of dNBR index ranges. In this study, we focused on an automated global burn severity mapping approach. Using the catalogue of satellite imagery and the high-performance computing power of GoogleEarthEngine we propose an automated pipeline to generate severity maps of burned areas at a medium scale of 30 and 10m from the time series of Landsat and Sentinel2 images. Landsat-8 images available during 2020 and the dNBR spectral index were used to calculate the severity level of each pixel using a calibration model and linear regression adjustments, which were taken in the field from the CBI index in an app developed for field capture. A calibration approach was carried out to give the severity level of the final burned areas after several carefully designed logic filters on the normalized burn rate (NBR). This script focuses on the fires that occurred in Honduras in 2020. The regression model found a similar spatial distribution and strong correlation between the areas analyzed in the field and those generated from the dNBR. The preliminary global validation showed that the overall accuracy reached 53.85%. However, the adjustments through the correlation models im-proved the results, yielding an R2 of 0.93 for the quadratic model, 0.79 for the Exponential model and 0.72 for the linear model.