Abstract: Forest cover information extraction is one of the essential tasks in forest remote sensing applications, which is of great significance for forest resource management, ecological environment protection, and climate change research. Traditional convolutional neural network-based methods can effectively extract local features, but they struggle to capture long-range dependencies and global context information. To address this issue, we propose a method for forest cover information extraction that fuses Transformer and UNet, referred to as DiUNet. This approach embeds Transformer modules into the UNet network to enhance its perception of long-range dependencies and global context information. Meanwhile, considering the fragmentation, irregularity, and inconsistent scale of forest cover information, our method enhances the model's ability to capture spatial information by using relative position encoding to increase the positional information, enabling the model to capture features at different levels and scales. We constructed a forest cover information dataset based on Landsat 8 and CDL data layers and conducted in-depth experimental analyses on this dataset. In the comparative experiments, DiUNet achieved the best results in accuracy, recall, F₁ score, intersection-over-union, and frequency-weighted intersection-over-union indices, which were 91.22%, 92.66%, 91.94%, 85.08%, and 81.65%, respectively. The model also performed well in generalization experiments. The experimental results show that the DiUNet method outperforms existing methods in forest cover information extraction and has high robustness and generalization capabilities.

Key words: semantic segmentation, UNet, Transformer, forest cover information