Synthesizing images with aligned masks using text-to-image based generative AI for robust PV segmentation

The robust detection of photovoltaic (PV) panels across multiple regions and backgrounds has traditionally relied on aligned remote-sensing imagery with manually labelled annotations, particularly in supervised learning. However, it faces challenges in collecting training datasets, including large volumes, accessibility concerns, and inconsistent quality. To efficiently produce paired PV images and masks, this study introduces SynthPV, a Text-to-Image based Generative AI (GenAI) approach that incorporates diverse backgrounds. By training on less than 1 % of real-world data, it can simultaneously generate highly aligned PV masks. The experimental results from three distinct scenarios using the Heilbronn, Jiaxing, and BDAPPV datasets demonstrate that synthetic image-mask pairs could significantly enhance PV segmentation performance. Notably, Intersection over Union (IoU) values increase by an average of 13.09 % and by 20.55 % for the Jiaxing dataset in particular. Cross-validation using synthetic data from Heilbronn with real datasets from Jiaxing and BDAPPV shows a further IoU improvement compared to solely real data. These findings underscore the effectiveness and robustness of the proposed GenAI method, offering a feature-adaptive and integrated approach to visual data augmentation that significantly enhances PV segmentation accuracy, thereby enabling more efficient and scalable applications in solar energy system analysis and deployment.