FRPV - Aerial Imagery of French Rooftops

IMPORTANT INFORMATION: This version (V3) of the dataset is based on aerial imagery from 2024 (month depends on department) and cadastral data from January 2025. It features: updates for 27 departments: 01, 02, 04, 11, 17, 23, 24, 29, 30, 33, 34, 38, 40, 47, 56, 60, 62, 64, 66, 67, 68, 73, 80, 84, 87, 2A, 2B 7 new departments: 77, 78, 91, 92, 93, 94, 95 Data for other departments is not reuploaded in this version of the dataset. Please use the version selector and goes to V2 to access older data for these departments. This dataset contains images of the rooftops of French buildings, with a large portion of the images from metropolitan France available. Ultimately, it will include around 40,000,000 images, organized by department. This dataset is related to the scientific publication "Thebault, Nerot, Govehovitch, Ménézo - A comprehensive building-wise Residential Photovoltaic system detection in heterogeneous urban and rural areas: application to French territories" Applied Energy, 2025, doi.org/10.1016/j.apenergy.2025.125630 Aerial Land Imagery The aerial imagery used in this study comes from the Institut National de l'Information Géographique et Forestière (IGN), the French national geographic institute. These images are provided in 25 km² RGB tiles with a resolution of 20 cm. The tiles are organized by French department and are freely accessible as JP2 raster files BD Ortho - Institut National Géographique. All the imagery utilized in this project is less than four years old. The availability of department-level imagery is fundamental to our methodology, as both cadastral data processing and the analysis of CNN model predictions are performed at this geographic scale. Building Registry The location and geometry of each building were extracted from a national building registry distributed by the French Etalab project. For each French department, a single SHP file is provided, containing building geometries stored as polygon features. Data Post-Processing Both raster (aerial imagery) and vector (building registry) data were processed using PyQGIS via QGIS. The preprocessing of vector data follows several steps. First, polygons with an area smaller than 10 m² were discarded, as they typically represent small, likely non-residential buildings, which are unlikely to host PV panels. Next, a 4-meter buffer was applied to each polygon to account for the frequent spatial discrepancies between the building registry and the actual building locations. To include additional contextual information in each final image and accommodate these shifts, each polygon was replaced with its oriented rectangular bounding box, minimizing the area of the box. Finally, the X and Y coordinates and a department-based unique identifier were added to each polygon feature. Creation of Building Images Each building polygon was intersected with the corresponding aerial imagery raster to generate a cropped image. These images were saved to individual files. For example, the Herault department (34), one of the more populated regions of France, contains approximately 700,000 images, with an average image size of 120x120 pixels. Notably, 97.9% of these images are smaller than 250x250 pixels. Approximately 1.5% of buildings span multiple raster tiles, resulting in final images that do not fully capture the entire rooftop.