The Classification of Building Functional Types Using A Multi-hop GNN Constrained with Street Boundaries From High-resolution Satellite Images

Existing methods for recognizing building functions from high-resolution satellite images insufficiently capture spatial structural relationships between buildings, particularly in densely populated urban areas. This issue reduces the distinguish ability of complex building types when relying solely on semantic image features. To address this issue, we propose a Multi-hop graph neural network (Multi-hop GNN) to classify building functional types from high-resolution remote sensing images constrained by open street boundaries. Our method models spatial topological relationships between buildings as graph-structured data, where buildings serve as graph nodes extracted through semantic segmentation of high-resolution remote sensing images using a CNN-Transformer. Unlike traditional node-wise GNNs, our method is motivated by the multi-hop neighborhood concept, which transforms the modeling of spatial interactions between nodes into interactions between different hop neighborhoods. Moreover, to effectively capture the spatial structures of building objects, we apply a boundary constraint using open street data, which naturally delineates land use areas and refines the spatial context for representing spatial structural information. In addition, an adaptive graph channel attention mechanism and a node feature regularity enhancement module are developed to further enhance the fusion of multi-scale structural features for building objects, leading to further improvement in classification accuracy. Experiments on two recently released building functional type classification datasets, i.e. UBT and TBT, demonstrate that our method outperforms existing CNN- and Transformer-based methods, achieving the highest OA of 96.28%, F1 of 94.33% and MIoU of 89.61% on the Fuzhou study area of the UBT dataset. We also evaluated the classification performance between different structural feature extraction backbones, including GCN, GIN, and GraphSAGE. Results showed that our Multi-hop GNN performed better than these methods. These findings justify that the proposed Multi-hop GNN is an effective method for deriving building functional type information from remote sensing images. Related codes are available at https://github.com/caixiaojiao/Multi-hop-GNN/tree/master .