Quantitative photovoltaic soiling estimation via multi-modal multi-dataset late fusion using modified EfficientNet-b0

Soiling significantly degrades photovoltaic energy efficiency, making accurate quantification imperative for optimal energy production. Therefore, non-contact PV soiling quantification methods are needed to replace expensive and non-scalable conventional approaches, but they often struggle to achieve robust quantification across diverse real-world conditions. Addressing this limitation, this work acquires three imaging datasets (indoor, outdoor, and thermal) and employs a deep learning framework based on EfficientNet-b0 with late-fusion strategies for PV soiling quantification. Indoor, outdoor, and thermal datasets are acquired using a Nikon P-1000 camera (focal lengths: 24–185 mm), a DJI Matrice drone (altitudes: 6–30 m), and Zenmuse H20T sensor, respectively, for sand levels ranging from 10 to 380 g. Modified EfficientNet-b0 is trained independently on each dataset before fusing outputs using five techniques: regularized neural network, attention-based fusion, ensemble fusion, cross-modal consistency, and principal component analysis-based fusion. All two-dataset combinations and three-dataset integration are investigated. Results show that the indoor dataset delivers the strongest individual performance, while indoor-outdoor fusion outperforms other two-dataset combinations and the best individual network. Moreover, three-dataset fusion achieves the best overall predictive performance, yielding a 49.5% improvement over the best individual model and 30.8% over the best two-dataset fusion. Lastly, 5-fold cross-validation with 3 repeats using stratified soiling-level grouped partitioning confirms ensemble fusion superiority with a mean RMSE of 16.58 g (95% confidence interval: [14.79, 18.37]), demonstrating robust generalization. Ablation studies show that a 64-dimensional feature space balances performance and complexity, dropout of 0.3 provides effective regularization, and single fully connected hidden layer outperforms deeper architectures.