<![CDATA[This study presents the implementation of the YOLOv8s-cls model for automatic classification of solar panel surface conditions into six categories: Clean, Dusty, Bird-drop, Snow-Covered, Electrical-Damage, and Physical-Damage. A dataset comprising 619 images was used to train the modified YOLOv8s-cls architecture, spanning 50 epochs with a batch size of 16, input dimensions set to 128×128, and the AdamW optimizer applied throughout. The training was conducted on a CPU-only system, yet the inference benchmark was performed in a separate testing phase, yielding an average inference time of 0.032 seconds per image, indicating strong feasibility for real-time deployment. The achieved accuracies were 85.88% for Top-1 and 99.44% for Top-5 predictions, demonstrating robust performance in multi-class classification tasks. Nonetheless, some visual ambiguities remained between similar classes such as Dusty vs. Snow-Covered and Electrical-Damage vs. Physical-Damage. These results affirm the effectiveness of YOLOv8s-cls as a lightweight and adaptable deep learning solution for solar panel condition monitoring. Future enhancements are proposed, including targeted data augmentation, texture-based preprocessing, and deployment on GPU-accelerated or edge-optimized platforms to improve generalization and deployment flexibility in real-world settings.]]>
Copyrights © 2025
