Diagnosing Shortcut Learning in CNN-Based Photovoltaic Fault Recognition from RGB Images: A Multi Method Explainability Audit

Convolutional neural networks are increasingly used for photovoltaic fault recognition from RGB imagery, yet high benchmark accuracy can mask shortcut learning induced by heterogeneous backgrounds, viewpoints and class imbalance. Using the Kaggle "PV Panel Defect Dataset" dataset, we compare five architectures (Baseline CNN, VGG16, ResNet50, InceptionV3 and EfficientNetB0) through a complementary explainability pipeline: LIME superpixel surrogates (with kernel-weighted R2 fidelity), occlusion sensitivity (functional relevance under localized masking) and Integrated Gradients (IG) validated by deletion-insertion curves. To reduce reliance on subjective saliency inspection, we quantify localization and concentration using IoU@Top10% against consistent proxy defect masks, Shannon entropy and Hoyer sparsity, and we summarize IG faithfulness with a Faithfulness Gap (AUC_insertion - AUC_deletion) and an accuracy-faithfulness consistency score at class level. ResNet50 attains the best predictive performance (82.3% accuracy), while EfficientNetB0 provides the strongest overall evidence faithfulness (mean Faithfulness Gap ~ 0.019) and stable, panel-centered attributions. InceptionV3 frequently yields diffuse relevance, and VGG16 produces highly concentrated but occasionally brittle hotspots. Bird-drop and Snow-covered show the most consistent alignment between accuracy and faithful evidence, whereas Clean and the two damage classes remain vulnerable to context cues (e.g., borders and background textures). The results support integrating quantitative explainability diagnostics into PV model selection and dataset curation to mitigate shortcuts and improve trustworthiness in vision-based PV monitoring.