README.md

metadata

title: DeepPVMapper
emoji: 🗺️
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 6.17.3
app_file: app.py
pinned: false
license: cc-by-4.0

DeepPVMapper — interactive demo

Detect and characterize rooftop solar (PV) installations anywhere in France, directly from IGN aerial imagery (20 cm resolution).

What it does

Search a city or click the map to select a zone (100 m to 1 km wide), then press Detect. The app fetches the orthophotos for the zone, runs a classification model to find patches containing PV panels, segments the panels in those patches, and merges the resulting polygons into individual installations. For each installation it estimates the surface (m²), the installed capacity (kWp) and the azimuth. Detected polygons are displayed on the map (click them for details) and the full result can be downloaded as GeoJSON.

Detection runs on CPU — expect roughly a minute for a full 1 km² zone.

Models & method

The classification and segmentation models are trained on the BDAPPV dataset and available at gabrielkasmi/bdappv-models. Installation characteristics (surface, capacity, azimuth) are estimated with the methods of pypvroof, using constant tilt (20°), bounding-box azimuth and a constant capacity coefficient (no DEM). Estimates on individual installations are therefore approximate; they are most meaningful aggregated over a zone.

Mapping larger areas

This Space is a demonstrator, capped at ~1 km² per run. To map a city, a department or more, use DeepPVMapper directly — it is the full pipeline this demo showcases, designed for large-scale runs (GPU support, tiled processing, registry construction).

Troubleshooting

• "Detection failed: 502 / 400 ... data.geopf.fr" — the public IGN Géoplateforme imagery service occasionally fails under load. The app retries automatically; if a run still fails, just press Detect again, it is transient and nothing is wrong on your side.
• No installation detected — zones with no rooftop PV are common; try a commercial area, a recent housing development, or a solar farm.
• Layout looks broken — best viewed on a desktop browser; the floating panels are not designed for mobile screens.

Links

• Large-scale pipeline: DeepPVMapper
• Models: gabrielkasmi/bdappv-models
• Dataset: gabrielkasmi/bdappv (paper)
• PhD thesis: pastel.hal.science/tel-04909303