Abstract
The widespread adoption of distributed photovoltaic (PV) systems highlights the need for sophisticated segmentation technologies that can accurately identify PV panels, essential for calculating potential capacity and informing development strategies. Although artificial intelligence has significantly advanced the accuracy and reliability of PV panel segmentation, real-world complexities such as diverse panel types, installation methods, and varied backgrounds pose challenges to model adaptability and generalization. This research introduces a method that enhances PV panel segmentation by employing the enhanced Segment Anything Model, which has been extensively pre-trained using a comprehensive real-world dataset to incorporate multimodal semantic information, thus improving generalization. Additionally, a fine-tuning process has been integrated to better absorb critical features from the training data, increasing the model’s sensitivity to the unique characteristics of specific PV installations. Field tests in Heilbronn, Germany, confirm the method’s superior performance and flexibility, underscoring its potential to support strategic planning for large-scale PV deployment.
Keywords renewable energy, photovoltaic panel, computer vision, remote sensing, semantic segmentation, deep learning
Copyright ©
Energy Proceedings