Recognition of Forest Damage from Sentinel-2 Satellite Images Using U-Net, RandomForest and XGBoost

Automatic analysis of satellite images is an actual problem today. This paper compares methods for semantic segmentation of forest damage in different regions of the European territory of Russia and the Urals based on multi-temporal Sentinel-2 satellite imagery. In this paper, it is selected the most informative features among the different spectral bands, their differences and the vegetation indices based on them. Three segmentation methods are compared: random forest, gradient boosting and U-net architecture convolutional neural network. The Dice coefficient is used for evaluation. As a result, the U-Net CNN model was the highest with a coefficient of 78.69% compared to 55.85% and 54.06% respectively for the best RF and XGBoost models.