Research on Land Cover Classification in the Yinbei Irrigation District of Ningxia by Integrating CNN and CSwin Transformer Models

To address the challenge of existing deep learning-based remote sensing intelligent interpretation methods in simultaneously capturing both local and global information when processing complex surface features in the ecologically fragile areas of Northwest China, this study selects the Yinbei Irrigation District in Ningxia as the research area and constructs a dataset based on 2020 Landsat-8 imagery. A novel remote sensing image semantic segmentation model, CCSTformer, is proposed. The model introduces innovative improvements in two main aspects: First, a dual-branch encoder architecture based on convolutional neural networks and CSwin Transformer is constructed. By introducing a coupled attention fusion module as the bridge between the dual branches, effective integration of global and local features is achieved from the perspective of attention interaction, significantly improving the segmentation accuracy of small target features. Second, a multi-scale feature fusion decoder is designed to further integrate spatial detail features and global contextual information at different scales, enhancing the classification accuracy of surface features. Experimental results demonstrate that CCSTformer can better distinguish edge information of surface features under complex conditions, achieving an overall accuracy and MIOU of 91.41% and 0.8428, respectively, outperforming other comparative models.