Identification of Logged and Windthrow Areas from Sentinel-2 Satellite Images Using the U-Net Convolutional Neural Network and Factors Affecting Its Accuracy

The results of detection (segmentation) of forest disturbances (logged and windthrow areas) based on Sentinel-2 satellite images with convolutional neural networks of U-net architecture in different regions of the European territory of Russia and the Urals are presented. The volume of the training sample was over 17 thousand objects. Overall, both logged and windthrow areas are detected with satisfactory accuracy (the average F-measure is over 0.5). At the same time, substantial differences in detection accuracy were found depending on the characteristics of both disturbances themselves and the affected forest cover. Thus, the maximum accuracy was achieved for tornado-induced windthrow areas, due to their geometric features. The dependence of windthrow detection accuracy on the species composition of damaged forests is not obvious and requires clarification; at the same time, the average area of damaged forest sites has a substantial effect on it. The maximum F-measure calculated for logged areas detected on test pairs of Sentinel-2 images reaches 0.80, which is substantially higher than in previously published studies with the U-net model. The maximum accuracy is typical for large clear-cuts in mixed and dark coniferous forests, while selective logged areas in deciduous forests are characterized by lowest one. The accuracy for wintertime and summertime pairs of images is substantially higher than for multiseasonal pairs. Also, the accuracy strongly varies for different types of logged areas. Thus, forest roads on summertime images are detected with the lowest producer’s accuracy, while logged areas on wintertime images are detected with highest one.