SD4EO: AI-based synthetic solar panel dataset on urban areas

This dataset has been created as part of the deliverables for ESA’s SD4EO project. It consists of aerial images of urban areas that mimic certain regions inside Poitiers, Bordeaux and Toulouse.The images have been synthetized with a generative diffusion model, conditioned with schematic maps, and later augmented to include solar panels in the most sunlighted portions of the building roofs.

Each entry has 4 types of files:

A PNG image containing the binary mask with pixel-by-pixel segmentation of areas where solar panels have been installed.

A PNG image which shows a variant of the synthetic image, each featuring differently placed panels with white support structures.

A TXT file containing the locations of axis-aligned bounding boxes in YOLOv8 format (also compatible with YOLOv5). These are included only if panels have been added to the image, with one file per panel variant, to facilitate YOLO training without needing to modify its Dataset class for file reading.

An NPZ file storing the coordinates of the bounding boxes aligned with the solar panels (not with the axes), intended for use in more advanced object detection models like YOLOv8.1 or Mask R-CNN.

The dataset includes 46,872 synthetic images, which have been further enhanced by adding solar panels in strategic locations. The complete dataset with metadata has been subdivided into 18 ZIP files (each containing a different subfolder), totalling 6.5GB. This division was made to avoid overloading the storage system. By distributing the 156,127 files across multiple folders, we prevent potential issues on users' computers related to exceeding the maximum number of inodes in the file system and/or the operating system.

The SD4EO Project is funded by the ESA’s FutureEO programme under contract no. 4000142334/23/I-DT and supervised by ESA Φ-lab.