Based on Trans-Unet Remote Sensing Image Segmentation Network Model

Remote sensing image segmentation technology, as a core task in computer vision, has significant application value in disaster monitoring and urban planning. However, traditional convolutional neural networks (such as U-Net) struggle to effectively model long-range dependencies due to their limited receptive fields; while Transformer, despite its ability to capture global context, falls short in handling low-level detail information. To address this, this paper proposes a TransUNet model that integrates Transformer and U-Net, using ResNet50 as the backbone network. By serializing feature maps, introducing position encoding, and incorporating self-attention mechanisms, the model achieves efficient fusion of global and local features. Experiments were conducted on the LandCover.ai dataset to compare the performance of U-Net, ResNet50, and TransUNet. The results show that TransUNet achieves an average intersection-over-union (MIoU) of 81.56%, an average pixel accuracy (MPA) of 88.37%, and an accuracy rate of 94.21%, significantly outperforming other models. The model demonstrates stronger robustness in complex scenarios, particularly showing significant improvements in segmenting boundary blurred areas and similar category regions, providing new optimization directions for remote sensing image segmentation tasks.