A comprehensive building-wise rooftop photovoltaic system detection in heterogeneous urban and rural areas: application to French territories

Abstract

With the rapid expansion of Rooftop Photovoltaic (RPV) systems, accurately identifying the location of these installations has become essential for urban planning, grid management, and socio-economic analysis. However, existing European datasets of RPV systems are often limited in both spatial coverage and precision, especially in regions with diverse architectural styles. This study presents a novel methodology for identifying RPV systems by employing a convolutional neural network (CNN) trained on high-resolution aerial imagery and building registry data. Alternatively to traditional tile-based methods, we propose a building-by-building approach, ensuring that each building is individually assessed. The model was trained and validated on five French departments representing a variety of roofing materials and urban typologies. It demonstrates a high correlation between predicted and registered RPV systems, though detection performance varies with roofing materials—achieving better accuracy on tiled roofs than slate roofs. When applied to the entire metropolitan French territory, the model processed images of more than 40 million buildings, identifying approximately 600,000 RPV systems. The results’ accuracy is evaluated, taking into account factors such as data quality and local urban characteristics. All data and the model are publicly available for further research and applications. • Building-by-building approach to detect rooftop PV systems from aerial images. • Comprehensive analysis of RPV systems across metropolitan French territories. • High detection accuracy influenced by roofing material and local urban specificity. • Performance of neural network model improves with heterogeneous training data. • Open-access datasets and tools for reproducible PV identification research.

Topics & Concepts