Predicting Postfire Forest Mortality Using Remote Sensing Data and Machine Learning

The large number of wildfires and burned area that occur annually in Siberia make it necessary to develop tools for assessing and predicting the effects of wildfires on forests. The study uses open access data sets and machine learning methods to predict areas of postfire forest mortality, as well as to assess the significance of features that determine the amount of tree mortality. Using a set of thematic satellite products and the Random Forest method for the territory of Krasnoyarsk krai and the Republics of Khakassia and Tyva, the significance of a set of features in predicting the proportion of forest stand loss after fire was assessed and a Random Forest model was developed. The work used features describing both the forest and topographic conditions of the territory and the moisture content of forest fuels, as well as characteristics of wildfires and magnitude of changes in spectral properties of the surface caused by wildfires. The overall accuracy of the model was 0.84, and the F1-score was 0.76. The dNBR index showed the greatest relative importance in predicting the proportion of postfire tree mortality. In general, the features characterizing changes in the surface reflective properties after fire (spectral indices dNBR and dBAI, as well as reflectivity in several spectral ranges) determined more than 50% of variability in the estimates of postfire tree mortality.