A Semantic Segmentation Method for Remote Sensing Images Based on an Improved TransDeepLab Model

Due to the various types of land cover and large spectral differences in remote sensing images, high-quality semantic segmentation of these images still faces challenges such as fuzzy object boundary extraction and difficulty in identifying small targets. To address these challenges, this study proposes a new improved model based on the TransDeepLab segmentation method. The model introduces a GAM attention mechanism in the coding stage, and incorporates a multi-level linear up-sampling strategy in the decoding stage. These enhancements allow the model to fully utilize multi-level semantic information and small target details in high-resolution remote sensing images, thereby effectively improving the segmentation accuracy of target objects. Using the open-source LoveDA large remote sensing image datasets for the validation experiment, the results show that compared to the original model, the improved model’s MIOU increased by 2.68%, aACC by 3.41%, and mACC by 4.65%. Compared to other mainstream models, the model also achieved superior segmentation performance.