SolarTherm-HG: A High-Resolution Thermal Dataset for Advancing Deep Learning Models in Photovoltaic Homography

SolarTherm-HG is the first large-scale open-source thermal imaging dataset specifically designed to advance deep learning models for thermal homography in photovoltaic (PV) systems. The dataset comprises high-resolution thermal images of PV panels captured under diverse and controlled conditions, including varying levels of cleanliness (from clean panels to four stages of dirt and mud buildup), four distinct heights and angles, and four different times of the day to reflect natural temperature variations. Each image underwent preprocessing to address common artifacts such as shadows and sun glare. A dedicated dataset generator supplements SolarTherm-HG, creating homography pairs with varying degrees of skewness and distortion to support deep learning applications. The dataset captures real-world challenges in thermal imaging, such as geometrical diversity, environmental variability, and spatial alignment, making it well-suited for advancing homography techniques. SolarTherm-HG includes 12,460 raw thermal images captured at 640×512 resolution. For deep learning applications, images were patched to 320×256, expanding the dataset to 49,840 thermal images for enhanced usability. To validate the utility of SolarTherm-HG, preliminary testing was conducted using classical feature-based methods, including ORB and SIFT, as well as a deep learning-based model, HomographyNet. These experiments demonstrated the dataset’s viability for robust benchmarking and training of homography models, establishing its potential as a benchmark for advancing traditional and deep learning techniques. Consequently, the dataset enables advancements in thermal homography for PV panel analysis, improving fault detection and enhancing predictive maintenance for sustainable energy systems.