A Mixture of Experts Model for Image Classification Based on High-Resolution Remote Sensing Image

To address the limitations in accuracy and reliability of remote sensing image classification models, this study innovatively integrates the Mixture of Experts (MoE)widely adopted in large-scale models - with the U-Net segmentation framework, proposing a novel UNetMoE model for remote sensing image segmentation. The UNetMoE employs a sparse architecture comprising two core components: 1) a gating system that extracts global image features and dynamically allocates weights to expert submodels, and 2) multiple parallel expert submodels specialized in mining localized detail features to execute category-specific classification. By training dedicated expert submodels for individual land-cover categories, the model achieves targeted feature extraction to enhance per-class recognition precision, while the gating system synthesizes outputs from all experts to optimize overall classification accuracy. Experimental results demonstrate that UNetMoE outperforms comparable models in classification accuracy. Our method achieves an accuracy of 87.66 %, surpassing the HRNet baseline by 1.75 %. Additionally, UNetMoE exhibits superior performance compared to parameter-matched dense models, with only a 1.18 % discrepancy in farmland classification, while exceeding dense models by 8.23 % in overall accuracy. UNetMoE delivers significantly higher accuracy across five major categories compared to single-expert implementations, albeit with a 9.21 % decline in precision for background class identification, suggesting strategic trade-offs in multi-expert consensus integration to maximize holistic performance.