A Shape and Size Free-CNN for Urban Functional Zone Mapping With High-Resolution Satellite Images and POI Data

Urban functional zone (UFZ) refers to the spatial aggregation of similar human activities in urban areas, and its category information has significant implications for city planning and layout. Existing studies have incorporated high spatial resolution (HSR) remote sensing images with social sensing data to obtain UFZ patches for classification and identification purposes. While deep learning techniques have proven effective in remote sensing image classification, two challenges arise when applying them to UFZ classification: irregular shapes and inconsistent sizes, making it difficult to input UFZ patches into deep learning models directly. To address these challenges, this study proposes an end-to-end model, known as the shape and size free convolutional neural network (SSF-CNN), to automatically classify UFZ patches of varying sizes and irregular shapes. First, the SSF-CNN adopted a novel network, named hierarchical attentional residual network (Res-HANet), which embeds a hierarchical group convolution (HGC) module and attention mechanisms to learn multiscale features from fused image blocks of four different sizes. Then, a mask layer is followed to filter the deep features and preserve the original information of irregular UFZs. The proposed method was applied to classifying UFZs in Zhuhai and Guangzhou, Guangdong, China. The evaluation results showed that SSF-CNN achieved an overall accuracy (OA) of 87.85% for the Zhuhai dataset and 90.49% for the Guangzhou dataset, significantly better than existing methods. In addition, ablation experiments confirm the effectiveness of components in the SSF-CNN. Overall, the results suggest that the proposed method has great potential for large-scale UFZ mapping.