RGB-thermal drone imaging for semantic segmentation and anomaly detection in photovoltaic panels

Photovoltaic (PV) systems need regular checks to ensure they work well and safely. Manual inspections are often slow, expensive, and not always reliable. In this work, we present a new dataset made of RGB and thermal images taken by drones flying over real solar panel installations. The images include detailed annotations for two main tasks: identifying the position of the panels (semantic segmentation) and detecting problems such as hotspots (anomaly detection). We tested several deep learning models on this dataset and compared their performance. Our results show that using both RGB and thermal images helps the models detect issues more accurately, even in difficult lighting or weather conditions. This approach can support faster and more efficient inspections of solar plants using AI. The dataset, along with the code and pre-training models, are available at https://github.com/ictlab-dev/PV30.