To address the issues of significant scale differences in targets, sparse distribution of objects, and high similarity between certain classes in remote sensing images, an improved semantic segmentation method DKTU-Net for remote sensing image based on U-Net structure is proposed. Based on the U-Net framework, this method makes targeted improvements to the encoder and feature fusion structure. Firstly, the original convolutional encoder is replaced by a hybrid encoding structure that fuses convolution and Shunted Transformer, which can effectively capture cross-scale and cross-region global semantic dependencies while maintaining local detail sensitivity. Then, the detail-preserving contextual fusion module (DPCF) is adopted to fuse skip connections and upsampled features, and through a learnable fusion strategy, multi-scale features are adaptively combined to improve the problems of significant scale differences and sparse distribution of objects. Finally, the kernel selective fusion attention module (KSFA) is introduced to enhance the features after context fusion. This module can adaptively adjust the receptive field according to the needs of different classes, alleviating the problem of high similarity between easily confused classes. Experiments were conducted on the two high-resolution remote sensing datasets, ISPRS Vaihingen and ISPRS Potsdam. The results show that the proposed DKTU-Net, compared with the original U-Net network model, The MIoU and M-F1 on the ISPRS Vaihingen dataset increased by 6.52% and 4.87% respectively, and the MIoU and M-F1 on the ISPRS Potsdam dataset increased by 3.54% and 2.4% respectively It has effectively improved the problems existing in remote sensing image segmentation, such as large differences in target scale, sparse distribution of objects, and high similarity between easily confused categories.