CLM-UNet: A Remote Sensing Image Semantic Segmentation Method Incorporating Attention Mechanism

This study proposes an enhanced U-Net model incorporating a CLM attention module, designed to improve semantic segmentation performance in remote sensing images of the Circum-Tarim region. Traditional U-Net structures face challenges in balancing the extraction of global and local features, especially in complex, multi-scale remote sensing data. The CLM module addresses this by dynamically adjusting the focus on key channel features, thereby suppressing irrelevant information and improving segmentation accuracy. Ablation experiments demonstrate the effectiveness of the CLM module, showing that CLM-UNet outperforms the standard U-Net in key metrics, including an increase of 6.07% in mean Intersection over Union (MIOU) and a 0.19% improvement in overall accuracy (OA). Additionally, the CLMU-Net shows superior performance in distinguishing between desert and urban bare land, with MIOU values 4.75% and 3.19% higher than U-Net for buildings and vegetation, respectively. Comparative analysis with mainstream models in other semantic segmentation fields shows that CLM-UNet achieves better generalization, with relatively higher evaluation metrics for MIOU, OA, and G-means, confirming its robustness in complex environments.