融合局部感知与自适应门控的HGA-Mamba遥感图像语义分割模型OA

To address the challenges in semantic segmentation of high-resolution remote sensing images,such as blurred edge details,missed detection of small objects,and difficulties in multi-scale feature fusion,and to overcome the limitations of the original Mamba architecture in processing complex urban-rural scenes,a hybrid gated attention Mamba(HGA-Mamba)remote sensing image semantic segmentation model fusing local perception and adaptive gating was proposed.First,a hybrid feature encoding(Hybrid)module was constructed by introducing a parallel convolution branch to enhance the model's ability to capture textures and edges,thereby improving boundary segmentation effects.Second,an adaptive gated fusion(AGF)module was designed to dynamically adjust the weights of global and local features based on spatial content,mitigating feature conflicts between large-scale and small-scale objects.Finally,a squeeze-and-excitation(SE)channel attention mechanism was embedded to filter key features and suppress background noise.The results showed that the mean intersection over union(mIoU)of HGA-Mamba model reached 42.10%,increasing by 5.16 percentage points compared to the original visual Mamba(VMamba)model.The pixel accuracy improved by 3.48 percentage points,and the mean pixel accuracy improved by 4.67 percentage points.HGA-Mamba model effectively enhanced convergence speed and training stability,solved the problem of detail loss in complex scenes,and provided technical support for refined segmentation of high-resolution remote sensing images.