A machine learning burned area extracting method based on dual-polarization Sentinel-1 images

Aiming at the problem that it is difficult to perform accurate burned areas mapping with a single polarization feature, we propose a machine learning burned area extraction method based on dual-polarization sentinel-1 images, which aims to make full use of the parameter changes in backscattering intensity, polarization decomposition, and texture features caused by forest fires. In this paper, we first construct a feature parameter dataset for the polarization feature and texture feature changes caused by forest fires. On this basis, the optimal combination of features that best reflects the changes in the burned area is determined by combining feature importance and correlation analysis. We used a random forest (RF) model for burned area mapping. For verifying the feasibility of the model and algorithm, Sentinel-1 images before and after the forest fire were selected for the Xichang fire in 2020 to carry out experiments. The experimental results show that the accuracy of combustion area extraction of the method in this paper is 93.68%, and the Kappa coefficient is 0.7955.