Prototype AI drone detects solar-panel anomalies

At 18 m to 30 m above a photovoltaic (PV) installation, a custom hexacopter moved through autonomous flight paths while two cameras searched the panels below. The Polish-Spanish research team tested it over a small residential site and a large industrial-scale facility; the researchers classified modules as hotspots, dust, sand, bird droppings or clean. The aircraft and sensor payload used dedicated power supplies, enabling approximately 25 minutes of flight time.

The method does not treat the thermal camera as a simple thermometer. The researchers use thermal imagery to spot relative visual patterns, then check whether an anomaly remains consistent across several frames captured from different viewing angles. The payload combined a FLIR Boson 640 radiometric thermal camera with a MicaSense Altum, which simultaneously captures RGB, thermal and multispectral imagery.

The team built its dataset from 1,950 proprietary flight images, 190 images from the public Kaggle “Solar Panel Images” repository and 260 laboratory images showing controlled contamination. After filtering blurred, overexposed and noisy frames, the researchers segmented individual modules and compared four convolutional neural network (CNN) architectures. EfficientNet-B0 performed best, with precision of 0.903, recall of 0.907, a macro F1 score of 0.904 and a weighted F1 score of 0.927. F1 measures the balance between precision and recall.

The results were uneven by anomaly type. Hotspots and clean modules each received precision, recall and F1 scores of 1.000. Sand reached an F1 score of 0.907, bird droppings 0.889, while dust was the hardest to identify, with precision of 0.711, recall of 0.751 and an F1 score of 0.732. The campaigns took place during daylight, at ambient temperatures from 12 C to 22 C, mostly under clear or partly cloudy skies, with wind speeds below 12 m/s. The reported figures therefore describe this dataset and these test conditions, not every solar installation.

Concretely, the prototype could give residential and industrial PV operators an automated way to flag suspicious modules for closer inspection, if later testing confirms the performance outside the researchers' campaigns. Its modular design could also be extended to coordinated multi-UAV operation and additional sensor fusion, the team says. For now, the work is a research framework published in Measurement, developed by scientists from Rzeszow University of Technology and Valencia Polytechnic University, rather than a deployed inspection service.