Due to the various types of land cover and large spectral differences in remote sensing images, their high-quality semantic segmentation still faces severe challenges. This study proposes a new improved model based on the TransDeepLab segmentation method. The model introduces the GAM attention mechanism in the encoding stage, which can effectively reduce the information dispersion and enlarge the global interaction features; in the decoding stage, a multi-level linear upsampling strategy is designed to gradually amplify the multi-scale features extracted from the encoder and integrate them with the low-scale features to improve the segmentation effect of different shapes and sizes. The model makes full use of the multi-level semantic information and small target detail information in high-resolution remote sensing images, which can effectively improve the segmentation accuracy of target objects. Using open-source LoveDA large remote sensing image data sets for the validation experiment, the results show that compared to the original model, its Mean Intersection Over Union (MIOU) increased by 2.68 %, Average Pixel Accuracy (aACC), and Mean Pixel Accuracy (mACC) by 3.41% and 4.65%. Compared to other mainstream models, the model also achieved a better segmentation effect.