SolarVisionNet (SVN): Fusing Physics-Inspired Mask Features with EfficientNetV2 for Interpretable Solar Panel Condition Classification

Assessment of photovoltaic panel surfaces must cope with illumination variation, dust, snow, glare, cracks and bird droppings. This study presents SolarVisionNet (SVN), which fuses a seven-channel input (RGB plus physics-inspired glare, snow, bird-dropping and crack masks) with a 44-dimensional handcrafted descriptor through feature-wise linear modulation and adds an auxiliary head that predicts the masks. SVN is evaluated on a public six-class dataset under a leakage-aware protocol: because a naive random split places 38 of 131 test images next to a training near-duplicate, images are grouped into pseudo-sites, and results cover repeated seeds and grouped 5-fold cross-validation. The six-class model reaches 86.5 ± 3.3% accuracy and 86.8 ± 3.3% macro-F1, and the clean-versus-defective detector reaches 94.7% hold-out accuracy (90.4 ± 2.4% under cross-validation). In a nine-variant ablation, the physics-inspired components change macro-F1 by at most 2.7 points, within seed variability, and three of four ImageNet backbones trained with the same optimization recipe are equally accurate. Contrast-limited adaptive histogram equalization (CLAHE) lowers mean macro-F1 by 1.2 points over 50 paired runs, fixed-threshold rule overrides reduce macro-F1, and class-specific synthetic augmentation induces shortcut learning; the recommended configuration omits all three. SVN is a reproducible, leakage-audited baseline with inspectable glare, snow and bird-dropping masks.