HSDFormer: an improved transformer for remote sensing image semantic segmentation

The extraction of building locations is crucial in the field of remote sensing, commonly applied in tasks such as emergency response, urban planning, and environmental monitoring. Existing methods often employ convolutional models such as ResNet and U-Net. however, these models struggle to capture long-range features of geographical objects, limiting their practical effectiveness. Meanwhile, Transformer based models face challenges due to their quadratic computational complexity. Additionally, the characteristics of remote sensing images, including dense arrangement, small scale, and class imbalance, pose difficulties for traditional patch-merging modules in downsampling, leading to the loss of significant building information. To address such limitations, we present the HSDFormer model in this paper. The model achieves a reduction in the computational complexity of self-attention calculation through a straightforward sequential reduction process. Additionally, our model has optimized the downsampling process specifically for remote sensing semantic segmentation tasks, effectively minimizing the loss of crucial information during layer-wise feature extraction. We evaluate the HSDFormer on the WHU Building Dataset and Satellite Dataset II (East Asia). The results demonstrate the model's effectiveness in accurately segmenting building locations, surpassing performance compared to other baseline models.