Intelligent monitoring of photovoltaic panel cleaning status: Fine-Scale dust accumulation estimation using hyperspectral data and mixed-pixel model

Dust accumulation on the surface of photovoltaic panels is one of the key factors affecting the operational performance of PV systems. Hyperspectral remote sensing, with its high-dimensional and fine spectral resolution, shows great potential in PV pollution monitoring. However, our observations indicate that, for dust-covered PV panels, each spectral pixel often contains a composite signal generated by both the PV surface and the deposited dust. This microscopic mixed-pixel phenomenon is often overlooked, limiting existing monitoring approaches—based on single-spectrum analysis or traditional image processing—to macroscopic identification, and making it difficult to reveal the intrinsic relationship between dust density and spectral response. To address this issue, this study proposes a dust monitoring method for PV panels that integrates mixed-pixel spectral unmixing, enabling high-precision monitoring of different dust accumulation states. A spectral-index-optimized non-negative least squares method was first applied to estimate the abundances of the PV panel and dust components. The resulting abundance information was then incorporated into a dust density inversion model, and multiple machine learning algorithms were used to systematically compare the effects of different preprocessing strategies and sensitive-band extraction methods on model performance. The results showed that incorporating mixed-pixel abundance features significantly improved model accuracy and stability (R2 = 0.9615, RMSE = 10.0092 g/m2). When the dust density reached 46.21 g/m2, the reflectance in the 440–675 nm range exhibited an inflection point, changing from a decreasing to an increasing trend. The corresponding dust proportion (64 %) slightly exceeded 50 %, making dust the dominant contributor to the spectral signal and indicating a transition in the dust layer from single to multiple scattering. This study provides theoretical support and technical foundations for intelligent PV soiling monitoring and optimized cleaning scheduling.