Objectives As an essential component of urban renewable energy, the assessment of rooftop photovoltaic (PV) potential depends on the accurate identification of structural characteristics. In particular, determining rooftop area and orientation is crucial for reliable PV capacity evaluation. Current mainstream approaches employ deep learning models to extract rooftop area and orientation from aerial imagery for PV resource assessment. However, variations in solar azimuth angles during image acquisition lead to differences in rooftop shadow distribution, resulting in significant apparent discrepancies among rooftops with the same orientation. Consequently, deep learning models may misclassify rooftop structural features.

Methods To address this core limitation, a novel framework for robust solar energy potential estimation is proposed. First, the advanced semantic segmentation architecture DeepLabv3+ is employed to simultaneously delineate rooftop boundaries and classify each rooftop region into primary orientation categories from input aerial imagery. Second, recognizing the susceptibility of visual-based orientation classification to shadow-induced errors, the framework incorporates a refinement stage. This stage integrates photovoltaic geographical information system (PVGIS) data, a validated source for location-specific historical and modeled solar radiation data. A multi-orientation quantitative integration strategy is employed, which utilizes the PVGIS to calculate the theoretical annual solar irradiation. By weighting and synthesizing these PVGIS-calculated values, the strategy effectively mitigates errors caused by visual misclassification.

Results The results on the roof information dataset indicate that DeepLabv3+ outperforms other models and improves the performance of the proposed framework. The quantitative integration strategy narrows the relative error in solar potential estimation. For 4, 8, and 16 orientation categories, the proposed framework reduces estimation errors to 0.15%, 0.76%, and 1.07%, respectively. By excluding disturbances of flat roofs, relative errors are further reduced to 0.04%, 0.49%, and 0.68%, respectively, which confirming the effectiveness of our framework.

Conclusions By combining high-resolution rooftop segmentation and orientation with geospatial irradiation modeling and a quantitative mechanism to resolve orientation ambiguity, the proposed framework effectively mitigates a primary source of error in aerial image-based solar potential assessment. This significantly enhances estimation accuracy, providing a more reliable tool for urban-scale renewable energy planning and investment decisions.