Comparative Estimation of Different Architectures of Convolutional Neural Network for Semantic Segmentation of Forest Cover Disturbances from Multidate Satellite Images

Abstract—Algorithms based on using convolutional neural networks are the most efficient for semantic segmentation of images, including recognition of forest cover disturbances from satellite images. In this study, we consider applicability of different modifications of the U-net architecture of the convolutional neural network for recognizing logging, burned, and windthrow areas in forests from multitemporal and multiseasonal Sentinel-2 satellite images. The estimation is carried out on three test sites that differ significantly in the characteristics of forest stands and forest exploitation. The highest accuracy (average F-measure of 0.59) was obtained from the U-net basic model, while the models that showed the best results during training (Attention U-Net and MobilNetv2 U-Net) did not improve segmentation on independent data. The resulting accuracy estimates are close to those previously published for forests with a substantial proportion of selective logged areas. The main factors determining the segmentation accuracy are characteristics of disturbances themselves (the area of cut blocks and their type). Substantial differences are also revealed between images taken in different seasons of the year, with the maximum segmentation accuracy based on winter pairs of images. The disturbance area obtained by summertime pairs of images and different-season images is substantially underestimated by the models. The dominant species in the forest stand has a less significant effect. However, for two of the three test sites, the maximum accuracy was observed in dark coniferous forests; the minimum accuracy, in deciduous forests. No statistically significant effect of slope illumination on segmentation accuracy based on winter pairs of images has been revealed. The accuracy of segmentation of burnt areas which was estimated using an example of 14 large forest fires in 2021–2022 turns out to be is unsatisfactory, which is probably due to the varying degrees of damage to the forest cover in the burnt areas.