A multimodal deep learning framework of large-scale urban functional zone mapping in the Guangdong-Hong Kong-Macao greater bay area

Abstract

Mapping urban functional zones (UFZs) is of paramount importance in urban planning and construction. However, existing studies have predominantly focused on individual cities, which restricts their applicability to large-scale UFZ mapping. Although points of interest (POI) are commonly used to classify UFZs from remote sensing imageries, their unordered and uneven distributions hinder robust POI feature extrations. Inspired by the fact that LiDAR point clouds share these two properties, we proposed a novel POI-based method of large-scale UFZ mapping where POIs are treated as point clouds. A multimodal deep learning-based framework was adopted to extract deep features from high-spatial-resolution remote sensing images and POIs, in which a classic point cloud neural network was employed for POI feature extraction. Simultaneously, to fuse the heterogeneous features and mitigate two issues associated with POIs, the overlook of POI spatial relationships and bias of POIs among UFZs, a multi-head self-attention (MSA) layer with prompts was introduced, followed by a classification module for UFZ prediction. The proposed method was validated in the Guangdong-Hong Kong-Macao Greater Bay Area (GBA), China, achieving an overall accuracy of 84.90% and a kappa coefficient of 82.36%. Furthermore, ablation experiments confirmed the effectiveness of the MSA layer and prompts.