Mitigating dataset imbalance using image-based stable diffusion and feature-level SMOTE for solar panel classification with CNNs

Imbalanced datasets pose a significant challenge in developing predictive models for solar panel fault detection, particularly for identifying dust accumulation on photovoltaic (PV) modules. This study addresses class imbalance in 842 polycrystalline silicon (poly-Si) PV module images (502 clean, 340 dusty) from the publicly available SolNET dataset collected across dusty regions of Bangladesh. All images were processed as 3-channel RGB inputs and resized to 224 × 224 pixels for CNN training (camera metadata and capture parameters such as irradiance, temperature, and distance are not provided in the dataset release). Two complementary augmentation strategies were employed: feature-level SMOTE (Synthetic Minority Over-Sampling Technique) and image-level Stable Diffusion augmentation. Stable Diffusion generated 162 photorealistic dusty-panel images, retained through a two-stage validation protocol combining visual screening and feature-space similarity testing (FID = 63.86, KID = 0.0220, pHash duplicates = 0 %), ensuring realism and diversity. Three convolutional neural networks (VGG-16, ResNet50, and MobileNetV3) were evaluated using stratified 10-fold cross-validation, with performance assessed by accuracy, F1-score, Cohen’s κ, and Matthews correlation coefficient (MCC). Results showed consistent improvements; for example, VGG-16 accuracy increased from 84.6 % (imbalanced) to 93.5 % (SMOTE) and 97.6 % (Stable Diffusion), while ResNet50 achieved 98.9 % with Stable Diffusion. Cross-validation confirmed robust performance (VGG-16: 95.6 % ± 1.7 % accuracy, κ = 0.91 ± 0.02, MCC = 0.93 ± 0.02), and Wilcoxon signed-rank testing verified statistical significance (p < 0.05). Constraints include reliance on a single public dataset, binary dust/clean labeling, uniform resizing, and restriction to poly-Si modules, which may limit generalization to other PV technologies and field conditions. Overall, the findings indicate that combining feature-level oversampling with diffusion-based image augmentation reduces false negatives and improves sensitivity to dusty panels, motivating further validation on larger, multi-technology PV datasets.