Georeferencing of photovoltaic modules from aerial infrared videos using structure‐from‐motion

To identify abnormal photovoltaic (PV) modules in large-scale PV plants economically, drone-mounted infrared (IR) cameras and automated video processing algorithms are frequently used. While most related works focus on the detection of modules, little has been done to automatically localize those within plant. In this work, we use incremental structure-from-motion obtain geocoordinates all a plant based visual cues measured GPS trajectory drone. addition, extract multiple IR images each module. Using our method, successfully map 99.3% 35,084 four one rooftop over 2.2 million module images. As compared previous extraction misses 18 times less (one 140 as eight). Furthermore, two or three rows can be processed simultaneously, increasing throughput reducing flight duration by factor 2.1 3.7, respectively. Comparison with an accurate orthophoto yields root mean square error estimated 5.87 m relative row 0.22 0.82 m. Finally, extracted visualize distributions temperatures anomaly predictions deep learning classifier map. temperature distribution helps disconnected strings, also find that its accuracy for anomalies reaches, even exceeds, seven out ten common types. The software is published at https://github.com/LukasBommes/PV-Hawk.