Deep solar PV refiner: A detail-oriented deep learning network for refined segmentation of photovoltaic areas from satellite imagery

To estimate electricity generation and evaluate the socio-economic effects of solar photovoltaic (PV) systems, it is critical to calculate the installed PV areas and quantify the installed capacity over a large region. Although general deep learning networks can be used to extract PV areas from satellite imagery, the capability of segmenting small and distributed ones with accurate and refined boundaries is still lacking. This is because significantly small foreground objects (i.e., PV areas) severely impeded by large and highly diverse background contexts, background objects having similar characteristics to PV modules are easily misclassified, and PV modules under various daylighting conditions present varying textures and colours. To overcome the challenges, this study proposes Deep Solar PV Refiner, a detail-oriented deep learning network, to enhance PV segmentation from satellite imagery. The proposed network advances the backbone by incorporating Split-Attention Network, combines Dual-Attention Network with Atrous Spatial Pyramid Pooling using four different structures, and integrates PointRend Network that refines PV boundary prediction. With transfer learning, a synthetic strategy, hybrid loss functions, and ablation experiments, the optimal network is obtained that outperforms the benchmark by 5%, 2%, 3%, 3%, and 2% for IoU, Accuracy, F1-score, Precision, and Recall, respectively. The network is also competitive with the state-of-the-art semantic segmentation networks and has a favourable generalization capability, with the mean IoU increasing by 0.63–11.18%. The new network effectively improves the capability of segmenting hard and small PV samples, which is deliverable to different areas and is significant for estimating the installed capacity of PV systems.