A Multi-Level Attention-Fused Large Kernel Network for Remote Sensing Image Segmentation

In this paper, we propose a Multi-Level Attention and Large Kernel Network (MLA-LKNet) for remote sensing image segmentation. The MLA-LKNet introduces a dilated large kernel re-parameterization strategy, leveraging multiple parallel convolution branches with varying dilation rates, which are merged into a single kernel during inference. This approach effectively captures multi-scale and sparse pattern relationships while maintaining high efficiency. In the shallow layers, a channel attention mechanism is embedded to highlight local details and key semantics, while in the deep layers, simplified non-local attention is applied to capture long-range dependencies and global context. Additionally, a multi-scale feature fusion module integrates information from different resolutions and semantic levels, significantly improving segmentation accuracy. Experiments on the public LoveDA dataset demonstrate that MLA-LK Net achieves outstanding performance across multiple segmentation tasks, surpassing several state-of-the-art methods.