Attention-Audited Lightweight CNNS For Reliable Dust Detection on Solar Panels

for subsequent comparisons with attention mechanisms and data augmentation experiments. It also highlights clear directions for improvement, such as threshold calibration or cost-sensitive adjustments. Grad-CAM is employed to visualize the model's discriminative basis: it generates a heatmap by computing gradient weights of the target class with respect to the feature maps of the final convolutional layer via backpropagation, thereby intuitively highlighting the regions to which the model pays the most attention in an image. Aligning this interpretability outcome with quantitative metrics (training/validation curves, ROC/PR, confusion matrix) allows for a clearer explanation of the performance source and primary error patterns of the Baseline. Based on four example cases, for images that are 'Clean and correctly predicted (0.953)' and 'Dusty and correctly predicted (0.999/1.000)', the Baseline’s highlighted regions primarily fall on the panel surface grid and areas with noticeable local texture variations, indicating that the model indeed leverages cell/ribbon structures and surface texture differences for judgment. In the example of 'Clean but predicted as Dusty (0.593)', the highlighted regions spread more distinctly to areas outside the panel or edge-related zones (e.g., roof/sky highlights, locations with strong boundary contrast). This attention pattern appears to capture lighting reflections, background structures, or imaging artifacts rather than being solely driven by dust features on the panel surface, which easily triggers misclassification of clean samples as dusty. Figure 14. Baseline Grad-CAM IJEMT Figure 15. Baseline Grad-CAM Combined with the training process and overall discriminative ability of the Baseline, the training and validation curves show continuously rising accuracy and synchronously decreasing loss, indicating that the network can stably converge and learn transferable discriminative features. On the test set, the ROC-AUC reaches 0.9908 and the AP of the PR curve is 0.9886, demonstrating that under adjustable thresholds, the model possesses strong discriminative capability, achieving a good trade-off between high true positive rates and low false positive rates across different thresholds. Under the discrete decision at a fixed threshold (thr=0.5), the confusion matrix shows that the model rarely misses Dusty samples (Dusty?not Dusty = 1), but false positives for the clean class are relatively concentrated (not Dusty?Dusty = 104). This aligns with the phenomenon observed in Grad-CAM: when the model partially directs attention to backgrounds, strong reflections, or high-contrast edge areas, it is more likely to push 'clean but highly reflective/complex background' samples toward the Dusty category. Overall, the Baseline exhibits strong ranking discriminative ability and high recall for the Dusty class, but the primary error stems from false positives on the clean class. This error pattern is supported by corresponding evidence from the interpretability visualizations, providing clear guidance for subsequent efforts to suppress background/reflection interference through attention placement, data augmentation strategies, or threshold selection. 3.1.3 M1 performance on the dataset After introducing the custom attention mechanism and adjusting hyperparameters in the M1 model, the model demonstrates more stable and efficient convergence trends on both the training and validation sets. The training and validation accuracy curves show that the model's accuracy improves rapidly in the early stages, then steadily approaches a saturated state, with the final validation accuracy consistently maintained above 0.95, slightly higher than the training accuracy. This indicates that the model possesses good generalization capability without significant overfitting. Both the training and validation loss curves exhibit a consistent declining trend and IJEMT remain relatively close throughout the training process. Although the validation loss shows minor fluctuations, its overall level remains stable and continues to decrease, indicating that the model effectively reduces classification errors during iterations and maintains stable convergence. Overall, the M1 model not only achieves improved accuracy after the introduction of the attention mechanism but also demonstrates stronger stability and convergence efficiency in training dynamics, laying a solid foundation for subsequent model enhancement and comparative analysis. Figure 16. M1 Acc Figure 17. M1 loss On the test set, the ROC curve of M1 rises almost vertically along the left axis and extends horizontally near the top region, closely hugging the upper-left corner overall. This distribution indicates that the model achieves a high True Positive Rate while maintaining a low False Positive Rate, meaning it effectively detects 'Dusty' samples while minimizing the misclassification of 'not Dusty' samples as 'Dusty.' Compared to the diagonal baseline of a random classifier, the curve is significantly higher across the entire range, demonstrating that the model's discriminative ability does not rely on a specific threshold but remains consistently superior as the threshold varies. IJEMT The reported AUC of 0.9986 indicates that M1 possesses very strong overall discriminative capability in the binary classification task, approaching the performance of an ideal classifier, which reflects its reliable generalization and discriminative performance on the test set. Figure 18. M1 ROC The Precision–Recall curve of M1 remains consistently close to the upper boundary of the graph over a wide range of recall values, indicating that it maintains a precision level near 1.0 across most recall intervals. This suggests that while the model detects 'Dusty' samples, the predictions labeled as 'Dusty' are highly reliable, with a low proportion of false positives. In other words, when the model outputs a positive classification, its predictions demonstrate strong stability. As recall further approaches 1.0, the curve shows a noticeable decline at its tail, reflecting that when the threshold is further lowered in an attempt to 'capture as many positive samples as possible,' precision begins to be affected by an increasing number of misclassified samples. This is a common precision–recall trade-off phenomenon in binary classification tasks. The reported AP of 0.9980 indicates that the model achieves very high overall detection quality for the positive class across different threshold settings. It not only covers the majority of true positive samples but also maintains a low cost of false positives across most intervals, thereby demonstrating the comprehensive performance advantage of M1 in positive-class identification on the test set. IJEMT Figure 19. M1 Precision–Recall curve With the test set threshold set at 0.5, the confusion matrix of M1 shows that the model distinguishes between the two classes with relatively stable performance. For the not Dusty class, 325 samples are correctly classified as not Dusty, while 29 are misclassified as Dusty. For the Dusty class, all 258 samples are correctly classified as Dusty, with no instances of Dusty being misclassified as not Dusty (FN=0). Correspondingly, the overall accuracy is 95.26% (583/612). From a class- specific perspective, the model achieves a recall of 100% for the Dusty class, indicating effective suppression of missed detections for dusty samples. However, the precision for the Dusty class is approximately 89.90% (258/287), revealing that the primary source of error stems from conservatively misclassifying some clean panels as Dusty (FP=29). This error pattern suggests the model adopts a strategy of 'prioritizing over-reporting to minimize missed reports' regarding dust risk. In an operational context, this reduces potential losses from undetected contamination but may incur occasional unnecessary cleaning triggers or alerts. IJEMT Figure 20. M1 confusio