LS-UNet for Semantic Segmentation of Fine-Resolution Remotely Sensed Images

Semantic segmentation of remotely sensed images is a fundamental task for earth observation. While studies have focused on remote sensing imagery from oblique perspectives for broader land cover, these images present significant object size variations, complicating segmentation accuracy. Existing CNN-based and Transformer-based methods either struggle with long-range dependencies or exhibit high computational complexity. Addressing these issues, we present LS-UNet, utilizing a Large Selective Kernel Network (LSKNet) as the encoder for dynamic spatial receptive field adjustments, and CNN - Mamba hybrid blocks (SS-Conv-SSM) as the decoder for a balance between performance and computational efficiency. Our Multi-Scale Feature Integration (MSFI) module enhances feature representation through cross-scale interaction. Experiments on the UAVid dataset demonstrate that LS-UNet outperforms current methods in segmentation metrics while maintaining low computational complexity.