Comparative analysis of neural network models for felling mapping in summer satellite imagery

The study aimed to improve the efficiency of detecting and mapping felling using satellite imagery, in order to identify violations of environmental regulations. Traditional remote sensing data interpretation methods are labor-intensive and require high operator expertise. To automate the satellite image interpretation process, numerous approaches have been developed, including those leveraging advanced deep machine learning technologies. The presented work conducted a comparative analysis of convolutional and transformer neural network models for the segmentation of felling in summer Sentinel-2 satellite imagery. The convolutional models evaluated included U-Net++, MA-Net, 3D U-Net, and FPN-ConvLSTM, while the transformer models were SegFormer and Swin-UperNet. A key aspect was the adaptation of these models to analyze pairs of multi-temporal, multi-channel satellite images. The data preprocessing, training sample generation, and model training and evaluation procedures using the F1 metric are described. The modeling results were compared to traditional visual interpretation methods using GIS tools. Experiments on the territory of the Khanty-Mansiysk Autonomous Okrug showed that the F1 accuracy of the different models ranged from 0.409 to 0.767, with the SegFormer transformer model achieving the highest performance and detecting felling missed by human interpretation. The processing time for a 100 × 100 km2 image pair was 15 minutes, 16 times faster than manual methods — an important factor for large-scale forest monitoring. The proposed SegFormer-based felling segmentation approach can be used for rapid detection and mapping of illegal logging. Further improvements could involve balancing the training dataset to include more diverse clearing shapes and sizes as well as incorporating partially cloudy images.