Enhanced burn severity estimation using fine resolution ET and MESMA fraction images with machine learning algorithm

Abstract

Successful post-fire management depends on accurate burn severity maps that are increasingly derived from satellite data, replacing field-based estimates. Post-fire vegetation and soil changes, besides modifying the reflected and emitted radiation recorded by sensors onboard satellites, strongly alters water balance in the fire affected area. While fire-induced spectral changes can be well represented by fraction images from Multiple Endmember Spectral Mixture Analysis (MESMA), changes in water balance are mainly registered by evapotranspiration (ET). As both types of variables have a clear physical meaning, they can be easily understood in terms of burn severity, providing a clear advantage compared to widely-used spectral indices. In this research work, we evaluate the potential of Landsat-derived ET to estimate burn severity, together with MESMA derived Sentinel-2 fraction images and important environment variables (pre-fire vegetation, climate, topography). In this study, we use the random forest (RF) classifier, which provides information on variable importance allowing us to identify the combination of input variables that provided the most accurate estimate. Our study area is located in Central Portugal, where a mega-fire burned >450 km2 from 17 to 24 June 2017. We used the official burn severity map as ground reference. The RF algorithm identified ET as the most important variable in the burn severity model, followed by MESMA char fractions. When both ET and MESMA char fraction image were used as RF inputs, burn severity estimates reached higher accuracy than if only one of them was used, which suggests their potential synergetic interaction. In particular, when environmental variables were used in addition to ET and char fraction, the highest accuracy for burn severity was reached (κ = 0.79). Our main conclusion is that post-fire fine resolution ET is a useful and easily understandable indicator of burn severity in Mediterranean ecosystems, in particular when used in combination with a MESMA char fraction image. This novel approach to estimate burn severity may help to develop successful post-fire management strategies not only in Mediterranean ecosystems but also in other ecosystems, due to ease of generalization.