Building Types Classification Using MHGNN from High-resolution Satellite Images

Traditional methods for detecting urban building functional types using high-resolution remote sensing images are mainly based on hand-designed features, ignoring spatial and structural features, resulting in difficulties in recognizing buildings with similar shapes. To solve this problem, this study constructs a multi-hop graph neural network (MHGNN) that integrates high-level structural and semantic information to recognize urban building functional types from high-resolution images. Experiments show that the proposed method outperforms traditional graph neural building classification methods. Our proposed method, MHGNN, demonstrates superior performance compared to standard GNNs such as MLP, GIN and GAT in multiregion building type classification tasks. The approach achieves state-of-the-art metrics, including IoU of 74.98%, F1-Score of $\mathbf{8 5. 3 9 \%}$, and OA of $\mathbf{8 9. 5 6 \%}$ - outperforming baseline methods by significant margins.