SEGMENTATION OF FOREST FELLING USING A PAIR OF SENTINEL-2 SPACE IMAGESIN WINTER PERIOD

The paper presents a comparison of segmentation models for solving the problem of detecting forest felling in winter using a pair of Sentinel-2 satellite images. The comparison includes models based on convolutional neural networks from the segmentation models library developed for the python programming language. As data for model training we used images from 2018 to 2022 from open sources of the European Space Agency, which were taken over the territory of Khanty-Mansiysk Autonomous Okrug Yugra. These images were preprocessed to solve the following tasks: atmospheric correction of images, bringing image pairs to a single projection, slicing images into frames. Forest felling masks were manually generated since 2015 in the space services center of the Ugra Research Institute of Information Technologies. F1-measure was used to assess the quality of models, as it is required to evaluate whether the model finds all felling areas, how accurately the model finds felling areas, as well as F1-measure allows taking into account false positives of the model. The UNet++ model showed the best result with a score of 0.847. The other models showed similar results, which indicates the similarity of these models for forest harvest segmentation tasks.