Abstract: To balance accuracy and efficiency in high-resolution remote sensing image semantic segmentation, this paper proposes a lightweight network named EMWMamba (Efficient Multiscale Wavelet-Fused Mamba). The network adopts a U-shaped architecture, combining the advantages of lightweight Transformer and Mamba frameworks. The encoder uses a ResT backbone to efficiently extract multi-scale features using low-parameter multi-head self-attention. The decoder introduces a feature-reconstruction wavelet Mamba module, which supplements frequency-domain weighting while acquiring spatial information, achieving frequency-spatial complementarity. Additionally, a wavelet feature fusion upsampling module is designed to improve the accuracy of information reconstruction. To further enhance the discriminative ability of learned features, the network incorporates a bidirectional cross-level feature aggregation auxiliary branch, which provides supplementary supervision during training to optimize feature representation. Experimental verification was carried out on two public remote sensing datasets, Vaihingen and Potsdam. Compared with the optimal comparison model, EMWMamba, with only 13.2 million parameters, achieves a 2.80 percentage point improvement in mean Intersection over Union (mIoU) and a 1.61 percentage point improvement in mean F1 score (mF1) on the Vaihingen dataset; On the Potsdam dataset, the mIoU and mF1 are improved by 1.68 and 1.09 percentage points respectively, realizing the optimal balance between accuracy and computational complexity.