Physics-guided unmixing of spaceborne hyperspectral data for effective photovoltaic area estimation

Monitoring large-scale photovoltaic (PV) deployment from space is critical for asset management and regulatory compliance, yet it requires robust estimation of effective light-collecting areas amidst complex backgrounds. To address spectral confusion and sub-pixel variability in diverse layouts (e.g., rooftop, water-based, and agrivoltaic PV), we propose a physics-guided unmixing framework utilizing spaceborne hyperspectral data. First, we generate EnMAP-aligned synthetic scenes to emulate realistic sub-pixel radiative transfer conditions. Then, we introduce PTV-Mamba, a deep unmixing model that instantiates an extended linear mixing model with a per-pixel illumination scaling term and a physics-tied decoder. By explicitly disentangling illumination-driven brightness changes from material variability, the model recovers precise PV abundance fields. In controlled simulations, PTV-Mamba consistently outperforms classical and deep baselines. On real EnMAP imagery in Taiwan, it achieves a low PV abundance RMSE of 0.093, with the estimated effective PV area (11.23 ha) closely matching high-resolution references (11.16 ha). Furthermore, cross-scene application to a heterogeneous agrivoltaic region in China yields physically plausible estimates, demonstrating robust generalizability. Overall, these results highlight the potential of physics-guided informatics for deriving interpretable, engineering-grade PV area estimates from satellite spectroscopy.