Hotspot Detection in Solar Panels Based on UAV Thermal Imaging via Deep Learning

This research explores the assessment of hotspot detection in solar panels, utilizing UAV thermal images, via a deep learning approach. The primary objective of the study is to replace the traditional way of monitoring and detecting solar panel hotspots via manual inspection with the UAV thermal imaging and deep learning approach. This work presents an automated setup using the YOLO algorithm to scan detailed thermal images captured over solar farms, incorporating GIS tools to detect and classify hotspot patterns and determine their exact location in the field. In this research, a thermal drone was used for the first time in the largest operating solar farm in Nepal in order to capture images of the solar field and investigate solar panels for defective cells and create an orthomosaic image of the entire area. Training samples were obtained and model development was performed using orthophotos, and the generated orthomosaic/individual images/videos were subjected to a YOLO model to detect the hotspots and faults. The model performs exceptionally well in detecting various solar panel defects—Bypass Diode Error, Multiple Hotspots, Panel Failure, and Single Hotspot—achieving high F1, precision, and recall values. The model exhibits excellent overall performance, with precision near 0.9 indicating minimal false positives, recall at 0.85 capturing most true hotspots, and mAP@0.5 reaching 0.95 for superior detection accuracy at moderate overlap thresholds. Strong mAP@0.5–0.95 (0.8) and stable generalization underscore its robustness across varied conditions, supporting reliable deployment in real-world PV monitoring. These findings establish a replicable model for PV monitoring, advancing sustainable solar infrastructure worldwide.