SSAtt-SolNet: An efficient model for dusty solar panel classification with Sparse Shuffle and Attention mechanisms

This study introduces SSAtt-SolNet, a novel deep learning approach designed to detect dusty solar panels, thereby improving the efficiency and reliability of solar photovoltaic systems. The proposed model uses MobileNetV3 as its backbone to balance accuracy and computational efficiency. We introduced a novel sparse shuffle block that combines depth-separable convolution with a shuffle layer to improve model performance. We also incorporated an attention mechanism in the classification layer to selectively focus on relevant features while minimizing noise interference. This lightweight approach was evaluated on two public and one self-collected dataset containing a total of 10,118 images. The model was benchmarked against eight SOTA models in image classification and dusty solar panel detection using four metrics: accuracy, model parameters, model size, and floating point operations (FLOPs). The experimental results showed that our approach outperformed all baseline models, achieving the smallest standard deviation over five folds (99.68 ± 0.3%). Furthermore, the proposed model had the smallest size, the fewest parameters, and the minimum GFLOPs (0.1005). The paired t-test confirmed that the accuracy of our model is statistically significantly higher than all baseline models at the 95% confidence level. These results suggest that our proposed model is feasible for use in environments with limited computing resources.