THED-PV: A High-Resolution Multi-Perspective Thermal Imaging Dataset for Photovoltaic Homography Estimation

This dataset, THED-PV, is specifically designed to advance deep learning models for thermal homography in photovoltaic (PV) systems. It includes high-resolution thermal images of PV panels captured under diverse and controlled conditions. The images were taken from various heights, angles, and times of day to reflect the natural variations in temperature and environmental conditions, including panel cleanliness from clean to heavily soiled states. These controlled conditions make the dataset ideal for training and benchmarking homography estimation models in PV inspection tasks. The dataset consists of 12,460 raw thermal images at a resolution of 640×512, which were preprocessed to mitigate common artifacts like shadows and sun glare. It is further augmented with synthetic homography pairs generated through controlled geometric perturbations, expanding the dataset to 99,680 homography pairs. This rich set of data, including both raw and preprocessed images, is accompanied by environmental metadata such as irradiance, temperature, humidity, and soiling levels. THED-PV was designed to facilitate advancements in the field of thermal homography estimation, enabling precise fault detection and predictive maintenance in photovoltaic systems. The dataset can be used to train deep learning models or validate classical homography estimation methods, and it includes all necessary tools and documentation for reproducibility and further research.