Efficient Vision Mamba with Dynamic Feature Fusion for Remote Sensing Image Semantic Segmentation

The improvement in imaging resolution and the increase in imaging swath of remote sensing satellites have enabled the acquisition of vast and complex remote sensing image data. Effectively and comprehensively extracting target information to enhance segmentation accuracy and interpretation efficiency has become a research hotspot in the field of remote sensing semantic segmentation. Although current segmentation algorithms based on CNN and Transformer frameworks have made significant progress in performance, achieving a good balance between segmentation accuracy and computational complexity remains challenging, limiting their widespread application in practical tasks to a certain extent. To address this issue, this paper introduces a state-space model and proposes an efficient vision Mamba for remote sensing image semantic segmentation, abbreviated as EVMDNet. By designing an efficient vision state space block (EVSSBlock), the method effectively integrates local and non-local features while improving computational efficiency through efficient 2D scanning (ES2D) strategy. Furthermore, to avoid the insufficient utilization of features and weak discriminative ability caused by the direct addition of skip connections, a dynamic feature fusion block (DFFBlock) is designed to adaptively adjust the weights of different features for efficient feature fusion. Experimental results on classical remote sensing image datasets demonstrate that the proposed method not only improves segmentation accuracy but also reduces computational complexity to a certain extent, outperforming existing state-of-the-art segmentation algorithms.