High-Resolution Remote Sensing Image Segmentation Based on Spatial-Frequency Feature Enhancement and Dual Attention Fusion

Abstract

High-resolution remote sensing image segmentation is a core task in remote sensing interpretation, which faces challenges such as complex distribution of ground objects, significant scale differences and blurred edges. Existing methods are relatively single, mostly focusing only on spatial feature extraction, and there are problems of low efficiency in multi-scale and cross-domain feature fusion. To address these issues, this study proposes a multi-module fusion network (WSDA-Net), which integrates the wavelet feature enhancement module (WFEM), the spatial detail extraction module (SDEM), and the dual attention fusion module (DAFM). Firstly, WFEM utilizes the multi-directional features of Haar wavelet decomposition to enhance the frequency domain features respectively, improving the segmentation accuracy of small objects and edge regions. Secondly, SDEM enhances the perception of spatial structures such as the shapes and contours of ground objects by mining spatial information and local details in images. Furthermore, the designed DAFM efficiently integrates self-attention (SAB) and cross-attention (CAB) through a dual-branch structure, providing more discriminative feature representations for the final high-resolution remote sensing image segmentation. Finally, experimental results on the public remote sensing image datasets of Vaihingen, Potsdam and WHDLD show that the proposed WSDA-Net on the three datasets achieves mean Intersection over Union (mIoU) values of 78.87%, 80.32%, and 63.92% respectively. Compared with the mainstream remote sensing model RS3Mamba based on the Mamba architecture, this method achieves improvements of 1.22%, 1.14%, and 1.51% respectively. Experimental results verify the network effectively improves high-resolution remote sensing image segmentation performance, with stronger robustness against object occlusion and background interference.