Improving photovoltaics hotspot and dirt accumulation detection with computer vision: A systematic YOLO model comparison based on field aerial images

Photovoltaics represent a significant renewable energy source, and substantial investments exist in large-scale PV plants. The accumulation of dust and other deposits on the PV panels can result in the formation of hotspots, which drastically impair the performance of the systems and cause damage to the PV cells. The use of aerial imagery provides a valuable tool for monitoring photovoltaic plants to identify hotspots and dirt accumulation. However, the employment of effective machine learning algorithms is essential for the detection of such abnormalities. Τhe implementation of a monitoring system is proposed, based on the You Only Look Once (YOLO) computer vision algorithm to address this need. The performance of various versions of YOLO models (YOLOv5, YOLOv7, YOLOv8, YOLO-NAS, and YOLOv9) was evaluated using two custom datasets generated for the detection of hotspots and bird droppings, respectively. The datasets comprise aerial thermal and optical images collected via a drone. Additionally, optimal sets of parameters were identified for each individual YOLO model. YOLOv5 demonstrated exceptional performance, achieving a mean Average Precision of 99.4% on the ’BIRD DROPPING’ dataset and 99.3% on the ’HOT SPOT’ dataset. Notably, for the ’HOT SPOT’ dataset, YOLOv8 achieved the highest recall at 99.4%, showcasing its capability in accurately identifying hotspots.