Semantic segmentation of remote sensing images is widely applied in such fields as environmental change monitoring and automotive driving assistance.Remote sensing images exhibit great intra-class variability and small inter-class differences at the semantic object level，which limits the accuracy of segmentation models and consumes computational resources.To address the problems，this paper proposes a lightweight semantic segmentation method for remote sensing images based on Transformer and depth-wise separable convolution.Firstly，a weight-adaptive multi-head self-attention mechanism is introduced to model the long-range pixel associations in a global scale，capturing rich contextual information.Secondly，stacked layers of Depth-wise Separable Convolutions (DSCs) are constructed to reduce the loss of spatial detail information with low computational complexity.Additionally，a Feature Aggregation Module (FAM) is designed by using a Linear Attention (LA) mechanism to merge global scene information and spatial detail information.The tests on datasets of Vaihingen and Potsdam show that the proposed method achieves an overall segmentation accuracy of 92.6% and 92.1% respectively with GFLOPs of only 11.5.It not only effectively enhances segmentation precision，but also significantly reduces computational complexity.