Improved DeepLabV3+ Remote Sensing Image Semantic Segmentation Method Based on Large Selective Kernel Attention Module

Aiming to address the high computational complexity, content complexity, scale differences, and uneven object distribution in remote sensing images, which lead to incomplete segmentation and low accuracy, an improved DeepLabv3+ model using a large-scale nucleus selection attention mechanism is proposed. The model uses an optimized MobileNetV2 backbone to reduce parameters, incorporates a large-scale nucleus selection attention module in the encoder to capture long-range information, and optimizes the atrous spatial pyramid activation function. Verification on the LoveDA dataset shows significant improvements in mIoU and accuracy compared to the basic DeepLabv3+ model, demonstrating the efficacy of the proposed network.