Fast and automatic solar module geo-labeling for optimized large-scale photovoltaic systems inspection from UAV thermal imagery using deep learning segmentation

Abstract

Photovoltaics (PV) are among the primary renewable and clean energy sources deployed globally, requiring frequent and efficient inspections to maintain optimal performance. Inspections conducted with unmanned aerial vehicles (UAVs) are optimized for large-scale installations, but they generate large datasets that are time-consuming to label manually. This study introduces a novel automatic geo-labeling approach of raw Infrared (IR) UAV imagery. It includes adaptive thresholding and edge refinement with photogrammetric data to automatically detect and geo-localize individual solar modules, without the need for manual annotation. The resulting auto-labeled images are then used to train and compare Deep Learning detectors for module-level segmentation. Among the tested models, Yolov7 demonstrated the best performance, achieving a mean Average Precision (mAP) of 98.33 %, with an inference time of only 15 ms, proving its suitability for real-time applications and integration within the inspection process. The dataset used for training includes UAV-based IR imagery from both ground-mounted and roof-mounted photovoltaic systems. The proposed approach accelerates the data preparation process and enables accurate on-site module localization, making it highly practical for automated PV inspection pipelines.