Geospatial Segmentation and Vision AI for Rapid Detection and Classification of Faults in Rooftop Solar Panels

This research introduces an innovative approach for solar panel segmentation and anomaly detection utilising Vision AI, specifically the YoloV11 model, combined with Semantic Anything Model (SAM2) semantic geospatial segmentation techniques. Our study focuses on rapidly segmenting rooftop PV systems across various buildings within a university campus, achieving high processing speed and accuracy. The Vision AI model has been trained to identify and classify common faults, such as microcracks and dirt, that impact solar panel efficiency and lifespan. Using a dataset split into a Training Set (70%, 1,470 images), Validation Set (10%, 210 images), and Test Set (20%, 420 images), the AI model achieves a 77.6% accuracy rate in detecting anomalies. Moreover, the geospatial segmentation of high-resolution satellite imagery enables precise mapping and identification of solar panel arrays, facilitating targeted inspection, maintenance, and performance monitoring. This approach significantly enhances the management of extensive areas, such as large-scale university campuses, by providing a rapid and accurate system to assess solar panel conditions. By accurately segmenting and assessing solar panels, the proposed method delivers detailed insights into damage detection, optimal panel placement, and streamlined lifecycle management processes for photovoltaic installations, ultimately contributing to improved energy production and operational efficiency.