SPHERE: Benchmarking YOLO vs. CNN on a Novel Dataset for High-Accuracy Solar Panel Defect Detection in Renewable Energy Systems

Abstract

Solar panels are critical for renewable electricity generation, yet defects significantly reduce power output and risk grid instability, necessitating reliable AI-driven defect detection. We propose the SPHERE (Solar Panel Hidden-Defect Evaluation for Renewable Energy) method for such cases. This study compares deep learning models for classifying solar panel images (broken, clean, and dirty) using a novel, proprietary dataset of 6079 images augmented to enhance performance. The following three models were evaluated: YOLOv8-m, YOLOv9-e, and a custom CNN with 9-fold cross-validation. Pre-trained models (e.g., VGG16 and ResNet) were assessed but outperformed by YOLO variants. Metrics included accuracy, precision–recall, F1-score, sensitivity, and specificity. YOLOv8-m achieved the highest accuracy (97.26%) and specificity (95.94%) with 100% sensitivity, excelling in defect identification. YOLOv9-e showed slightly lower accuracy (95.18%) but maintained high sensitivity. The CNN model demonstrated robust generalization (92.86% accuracy) via cross-validation, though it underperformed relative to YOLO architectures. Results highlight YOLO-based models’ superiority, particularly YOLOv8-m, in balancing precision and robustness for this classification task. This study underscores the potential of YOLO frameworks in automated solar panel inspection systems, offering enhanced maintenance and grid stability reliability. This contributes to advancing AI applications in renewable energy infrastructure, ensuring efficient defect detection and sustained power output. The dataset’s novelty and the models’ comparative analysis provide a foundation for future research in autonomous maintenance solutions.

Keywords

• solar panel classification
• deep learning
• YOLOv8
• YOLOv9
• convolutional neural networks
• k-fold cross-validation