DeepPVMapper

Large-scale rooftop PV detection pipeline for France. Classifies IGN aerial tiles with InceptionV3, segments positive patches with FCN/DeepLab, and extracts panel characteristics (surface, tilt, azimuth, installed capacity) via pypvroof.

🗺️ Explore the map and results →
🎮 Try the interactive demo →

Work carried out by Gabriel Kasmi as part of his PhD at Mines Paris-PSL (2020–2024).

MCP Server

DeepPVMapper's detection registry is also exposed as a public MCP (Model Context Protocol) server over Streamable HTTP, so any MCP-compatible client (Claude, etc.) can query the 1.14M+ rooftop-solar detections directly via natural language: search detections, aggregate installed capacity by department, explore an area, track yearly deployment, and check data-quality signals.

• Live endpoint: https://zelhliylrlktnasircwp.supabase.co/functions/v1/mcp/mcp

• Server source: supabase/functions/mcp (gh-pages branch)

• Registry listing: registry.modelcontextprotocol.io — io.github.gabrielkasmi/deeppvmapper

No install required — add the endpoint as a custom connector in any MCP client.

Related MCP server: MeteoControl MCP Server

Mapping data

Pre-computed detection results for French departments are available on Zenodo:

Installation

GDAL must be installed system-wide first, then via pip to match the running Python version:

# Ubuntu/Debian
apt-get install -y gdal-bin libgdal-dev libopenjp2-7

pip install "GDAL==$(gdal-config --version)"
pip install -r requirements.txt

GPU required. CUDA 11.8+, 8 GB VRAM minimum.

Data and model weights

Download model weights and runtime data from Zenodo:

Model weights are also available on Hugging Face:

The training dataset (BDAPPV):

Fill in the source paths in config.yml before running:

Usage

python main.py --dpt 06

--count sets tiles per classification batch (default 16 — reduce if OOM):

python main.py --dpt 06 --count 8

--config points to an alternative config file (useful for RunPod deployments):

python main.py --dpt 06 --config /workspace/config_runpod.yml

To process a subset of tiles (local testing), set tiles_list in config.yml:

tiles_list:
  - 01-2024-0850-6565-LA93-0M20-E080
  - 01-2024-0850-6570-LA93-0M20-E080

Force a full rerun (wipe prior progress):

python main.py --dpt 06 --clean

Pipeline

Four steps run sequentially inside main.py:

On success: temp/ is deleted automatically.
On crash: temp/ is kept. Rerun the same command to resume from where it stopped.

Outputs

Written to outputs_dir (default: data/):

Only residential-scale installations (1.7–36.1 kWp) located on buildings are retained.

Configuration reference

Contributing

Contributions are welcome — both code (performance, new imagery sources, models, building filters) and registry corrections via the interactive map, no coding required.

See CONTRIBUTING.md for the contribution areas, setup instructions and workflow. Issues labelled good first issue are the best entry points.

Known issues

GDAL is fragile to install, for two distinct reasons — and the fix below handles both.

1. The pip GDAL binding must match the system libgdal version exactly. pip install GDAL fails to build, or segfaults at import, if its version differs from the system library.

2. On images that ship a pre-installed python3-gdal apt package (common on RunPod/cloud GPU images), that package bundles its own osgeo/, which takes priority over the pip-installed one in sys.path — and its .so is often broken, regardless of what pip installs.

Run this in place of a plain pip install, e.g. right when deploying the pipeline, around the pip install -r requirements.txt step:

#!/usr/bin/env bash
set -e

# --- native GDAL lib (gdal-config must exist before pip can build the python binding) ---
apt-get update
apt-get install -y --no-install-recommends gdal-bin libgdal-dev

# --- purge python3-gdal if present: this apt package ships its own osgeo/,
#     which takes priority over the pip-installed one in sys.path, and its
#     .so is often broken ---
dpkg -l | grep -q python3-gdal && apt-get remove --purge -y python3-gdal || true
rm -rf /usr/lib/python3/dist-packages/osgeo

# --- python deps ---
pip install -r requirements.txt

# --- repin the pip GDAL binding to exactly match the native lib version
#     (requirements.txt only pins GDAL>=3.0, so pip can grab a newer
#     version than the one apt just installed -> ABI mismatch at import) ---
GDAL_VERSION=$(gdal-config --version)
pip install --no-cache-dir --force-reinstall "GDAL==${GDAL_VERSION}"

# --- check ---
python -c "
import torch
from osgeo import gdal
print('torch', torch.__version__, '| cuda', torch.cuda.is_available())
print('gdal', gdal.__version__, '| GTiff driver:', gdal.GetDriverByName('GTiff') is not None)
"

Citation

@phdthesis{kasmi2024enhancing,
  title={Enhancing the Reliability of Deep Learning Models to Improve the Observability of French Rooftop Photovoltaic Installations},
  author={Kasmi, Gabriel},
  year={2024},
  school={Universit{\'e} Paris sciences et lettres}
}

License

MIT