A Semantic Segmentation of Remote Sensing Images via Spatial–Frequency Encoding and Multiscale Feature Fusion

Semantic segmentation of remote sensing images plays a vital role in downstream applications such as land cover mapping and urban planning. While current mainstream methods leveraging Convolutional Neural Networks (CNNs) and Transformers are widely adopted, they often face a trade-off between effectively capturing long-range dependencies and maintaining computational efficiency. Recently, the Visual State Space (VSS) model, exemplified by Mamba, has gained attention for its ability to model long-range relationships with linear computational complexity, offering an effective solution to these limitations. In this work, we introduce an innovative semantic segmentation model for remote sensing images, termed WE-MambaNet. This framework enhances segmentation performance through a synergistic design that combines spatial-frequency collaborative encoding with multi-scale feature fusion. Specifically, we construct a dual-encoder architecture that complements spatial semantic extraction with a wavelet-guided frequency analysis, dynamically strengthening inter-channel feature dependencies. To further integrate cross-scale features and global semantic context, we design an efficient decoding pathway incorporating attention-constrained pyramid fusion and Mamba-based decoding blocks. Comprehensive testing on three open remote sensing datasets—ISPRS Vaihingen, ISPRS Potsdam, and LoveDA—demonstrates the superior performance of WE-MambaNet, achieving mIoU scores of 84.92%, 86.87%, and 53.46%, respectively. Both quantitative results and qualitative visual analyses confirm that our method surpasses current state-of-the-art methods in accuracy, robustness, and generalization.