HiFRD-SegFormer: An Industrial Building Roof Classification Method Based on an Improved SegFormer Model

Accurate identification of the photovoltaic development status, material, and form of industrial building roofs is essential for delineating roofs available for development and conducting differentiated photovoltaic potential assessments. This study proposes HiFRD-SegFormer, an industrial building roof classification method integrating hierarchical supervision and learnable full-resolution decoding to identify PV-installed roofs, metal pitched roofs, and non-metal flat roofs from high-resolution remote sensing imagery through a single forward pass. A coarse-to-fine architecture combined with coarse-only supervision for ambiguous roofs prevents them from interfering with the classification of the latter two categories, while retaining their reliable non-PV roof information. A learnable full-resolution decoder improves roof-region recovery, boundary delineation, and class discrimination. Experiments on a Shanghai dataset show that HiFRD-SegFormer achieves the best overall performance among several representative segmentation models, with IoUs of 87.53%, 89.73%, and 78.05% for the three classes. Coarse-only supervision balances classification accuracy and roof-area completeness, while the learnable decoder improves Overall IoU by 0.98 pp over SegFormer-B2. Applied to eight districts of Shanghai, the method identifies 79.02 km2 of industrial building roofs, comprising 20.31, 38.90, and 19.81 km2 of the three classes. These results provide a data basis for PV potential assessment and development planning of industrial building roofs.