Automatic on Field Detection and Localization of Defective Solar Photovoltaic Modules from Orthorectified RGB UAV Imagery

In the maintenance framework of solar photovoltaic (PV) installations, modules’ defect detection, identification and on field localization play a key role in preserving the reliability and efficiency of the electrical power generation. Remotely sensed imagery by means of Unmanned Aerial Vehicles (UAVs) is actively used in this context as it allows faster, cost-effective and contactless characterization of modules’ surface together with large-scale deployment. We develop herein an end-to-end approach to detect, identify and locate on field defects on PV installations based on RGB imagery acquired by UAVs. The approach is fundamentally designed for large-scale applications and comprises: (1) A photogrammetric image acquisition and post-processing phase that produces one orthorectified and georeferenced support covering the entire inspected site; (2) A module extraction phase that yields the individual images of modules; and (3) A deep learning-based defect detection stage using a fine-tuned instance of the YOLOv4 architecture. The approach was developed, validated and tested using a dataset collected from two large-scale PV sites comprising 35 305 modules. The developed defect detector scored a mean Average Precision (mAP) of 83% and 73% respectively on the validation and test sets.