Open registry of 1.14M+ rooftop-solar detections across France, queryable by natural language.
MCP Server: io.github.gabrielkasmi/deeppvmapper
This MCP server exposes an open registry of 1.14M+ rooftop-solar detections across France, allowing those records to be queried using natural language. The server is identified by the slug io-github-gabrielkasmi-deeppvmapper and is described as a natural-language-access interface to that dataset.
🛠️ Key Features
Open registry of 1.14M+ rooftop-solar detections
Coverage includes France
Natural-language querying
🚀 Use Cases
Searching or filtering rooftop-solar detections within France
Retrieving detection records via natural-language questions
⚡ Developer Benefits
Query access designed around natural language
Uses a clear dataset scope (rooftop-solar detections in France) for targeted retrieval
⚠️ Limitations
Available description does not specify tooling methods, parameters, or output schema beyond natural-language querying
Return the DeepPVMapper/OpenPVMapper registry's documented data-quality characteristics: estimated detection recall, which fields are model-derived estimates vs. structural/observed fields, the source-encoding table, the recommended confidence threshold, confidence signals, and licensing/liability terms. Call this before advising how much to trust a result for a specific use case (e.g. exploratory research vs. a commercial or regulatory decision) — pair it with the quality_summary attached to search_detections / aggregate_detection_capacity results, which reflects the specific query rather than the registry as a whole.
Get installed rooftop-PV capacity (total kWp) and system counts for one or all French départements, from the DeepPVMapper/OpenPVMapper registry. This is a fixed, pre-computed department-wide aggregate with no other filters — use aggregate_detection_capacity instead if you need a filtered subset (e.g. only cross-validated detections). Data quality note: this is a detection dataset, not an exhaustive inventory (estimated recall ~0.6). A missing detection does not mean no PV installation exists there. kwp, surface, tilt and azimuth are model estimates, not surveyed values. Call get_data_quality_reference for the full picture before assessing fitness for a specific use case.
Parameters2
dpt
string
optional
French département code, e.g. "33" for Gironde. Omit for all départements.
top_n
integer
optional
If set and dpt is omitted, return only the top N départements by installed capacity.
Raw schema
{
"type": "object",
"properties": {
"dpt": {
"description": "French département code, e.g. \"33\" for Gironde. Omit for all départements.",
"type": "string"
},
"top_n": {
"description": "If set and dpt is omitted, return only the top N départements by installed capacity.",
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 96
}
},
"$schema": "https://json-schema.org/draft/2020-12/schema",
"additionalProperties": false
}
get_department_yearly_stats
Get yearly system counts and capacity by département, based on first-seen imagery year. Useful for tracking apparent PV deployment growth over time. Data quality note: this is a detection dataset, not an exhaustive inventory (estimated recall ~0.6). A missing detection does not mean no PV installation exists there. kwp, surface, tilt and azimuth are model estimates, not surveyed values. Call get_data_quality_reference for the full picture before assessing fitness for a specific use case.
Parameters1
dpt
string
optional
French département code, e.g. "33". Omit for all départements.
Raw schema
{
"type": "object",
"properties": {
"dpt": {
"description": "French département code, e.g. \"33\". Omit for all départements.",
"type": "string"
}
},
"$schema": "https://json-schema.org/draft/2020-12/schema",
"additionalProperties": false
}
search_detections
Search individual rooftop-PV detections by département, commune (INSEE code), estimated capacity range, and/or cross-validation confidence (min_vintages, cross_validated). Returns a bounded list of detection records plus a quality_summary for the returned sample (use get_detections_in_bbox instead for map/spatial queries with geometry). Data quality note: this is a detection dataset, not an exhaustive inventory (estimated recall ~0.6). A missing detection does not mean no PV installation exists there. kwp, surface, tilt and azimuth are model estimates, not surveyed values. Call get_data_quality_reference for the full picture before assessing fitness for a specific use case.
Parameters8
dpt
string
optional
French département code, e.g. "33".
insee
string
optional
INSEE commune code.
min_kwp
number
optional
Minimum estimated installed capacity, in kWp.
max_kwp
number
optional
Maximum estimated installed capacity, in kWp.
min_vintages
integer
optional
Minimum number of distinct imagery vintages (years) the installation was independently detected in. Use 2+ as a persistence/confidence signal, since a one-off detection in a single vintage is more likely to be a transient artifact.
cross_validated
boolean
optional
If true, only include detections confirmed by at least two independent sources (e.g. the automated DeepPVMapper pipeline plus OpenStreetMap or the FRPV reference dataset), not just a single pipeline. This is a stronger confidence signal than min_vintages.
quality_filter
boolean
required
If true (default), only include detections with frpv_proba >= 0.1, the threshold recommended in the data contract for a good precision/recall trade-off.
limit
integer
required
Maximum number of records to return (max 200).
Raw schema
{
"type": "object",
"properties": {
"dpt": {
"description": "French département code, e.g. \"33\".",
"type": "string"
},
"insee": {
"description": "INSEE commune code.",
"type": "string"
},
"min_kwp": {
"description": "Minimum estimated installed capacity, in kWp.",
"type": "number"
},
"max_kwp": {
"description": "Maximum estimated installed capacity, in kWp.",
"type": "number"
},
"min_vintages": {
"description": "Minimum number of distinct imagery vintages (years) the installation was independently detected in. Use 2+ as a persistence/confidence signal, since a one-off detection in a single vintage is more likely to be a transient artifact.",
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 9007199254740991
},
"cross_validated": {
"description": "If true, only include detections confirmed by at least two independent sources (e.g. the automated DeepPVMapper pipeline plus OpenStreetMap or the FRPV reference dataset), not just a single pipeline. This is a stronger confidence signal than min_vintages.",
"type": "boolean"
},
"quality_filter": {
"default": true,
"description": "If true (default), only include detections with frpv_proba >= 0.1, the threshold recommended in the data contract for a good precision/recall trade-off.",
"type": "boolean"
},
"limit": {
"default": 20,
"description": "Maximum number of records to return (max 200).",
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 200
}
},
"required": [
"quality_filter",
"limit"
],
"$schema": "https://json-schema.org/draft/2020-12/schema",
"additionalProperties": false
}
aggregate_detection_capacity
Compute the total estimated installed capacity (kWp) and count of detections matching a set of filters — département, commune, capacity range, and cross-validation across sources (cross_validated) or imagery vintages (min_vintages) — plus a quality_summary for the summed sample. Unlike get_department_capacity_stats, which is a fixed pre-computed département-wide aggregate with no other filters, this tool sums a live filtered subset, up to max_rows detections. Example: "installed capacity in Gironde confirmed by at least two sources" -> dpt="33", cross_validated=true. Data quality note: this is a detection dataset, not an exhaustive inventory (estimated recall ~0.6). A missing detection does not mean no PV installation exists there. kwp, surface, tilt and azimuth are model estimates, not surveyed values. Call get_data_quality_reference for the full picture before assessing fitness for a specific use case.
Parameters8
dpt
string
optional
French département code, e.g. "33".
insee
string
optional
INSEE commune code.
min_kwp
number
optional
Minimum estimated installed capacity, in kWp.
max_kwp
number
optional
Maximum estimated installed capacity, in kWp.
min_vintages
integer
optional
Minimum number of distinct imagery vintages (years) the installation was independently detected in. Use 2+ as a persistence/confidence signal, since a one-off detection in a single vintage is more likely to be a transient artifact.
cross_validated
boolean
optional
If true, only include detections confirmed by at least two independent sources (e.g. the automated DeepPVMapper pipeline plus OpenStreetMap or the FRPV reference dataset), not just a single pipeline. This is a stronger confidence signal than min_vintages.
quality_filter
boolean
required
If true (default), only include detections with frpv_proba >= 0.1, the threshold recommended in the data contract for a good precision/recall trade-off.
max_rows
integer
required
Cap on the number of matching detection rows fetched to compute the capacity sum. If the true match count exceeds this, total_kwp is a partial lower bound and `truncated` is true — increase max_rows or narrow the filters (e.g. add dpt or insee) for an exact total.
Raw schema
{
"type": "object",
"properties": {
"dpt": {
"description": "French département code, e.g. \"33\".",
"type": "string"
},
"insee": {
"description": "INSEE commune code.",
"type": "string"
},
"min_kwp": {
"description": "Minimum estimated installed capacity, in kWp.",
"type": "number"
},
"max_kwp": {
"description": "Maximum estimated installed capacity, in kWp.",
"type": "number"
},
"min_vintages": {
"description": "Minimum number of distinct imagery vintages (years) the installation was independently detected in. Use 2+ as a persistence/confidence signal, since a one-off detection in a single vintage is more likely to be a transient artifact.",
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 9007199254740991
},
"cross_validated": {
"description": "If true, only include detections confirmed by at least two independent sources (e.g. the automated DeepPVMapper pipeline plus OpenStreetMap or the FRPV reference dataset), not just a single pipeline. This is a stronger confidence signal than min_vintages.",
"type": "boolean"
},
"quality_filter": {
"default": true,
"description": "If true (default), only include detections with frpv_proba >= 0.1, the threshold recommended in the data contract for a good precision/recall trade-off.",
"type": "boolean"
},
"max_rows": {
"default": 5000,
"description": "Cap on the number of matching detection rows fetched to compute the capacity sum. If the true match count exceeds this, total_kwp is a partial lower bound and `truncated` is true — increase max_rows or narrow the filters (e.g. add dpt or insee) for an exact total.",
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 20000
}
},
"required": [
"quality_filter",
"max_rows"
],
"$schema": "https://json-schema.org/draft/2020-12/schema",
"additionalProperties": false
}
get_detections_in_bbox
Get rooftop-PV detections within a geographic bounding box (WGS84 lon/lat), including footprint geometry. Intended for map-style spatial queries over a small area. Data quality note: this is a detection dataset, not an exhaustive inventory (estimated recall ~0.6). A missing detection does not mean no PV installation exists there. kwp, surface, tilt and azimuth are model estimates, not surveyed values. Call get_data_quality_reference for the full picture before assessing fitness for a specific use case.
Parameters5
min_lon
number
required
min_lat
number
required
max_lon
number
required
max_lat
number
required
max_count
integer
required
Maximum number of detections to return (max 500 here; the underlying API defaults to 2000, capped lower to keep responses manageable for an LLM).
Raw schema
{
"type": "object",
"properties": {
"min_lon": {
"type": "number"
},
"min_lat": {
"type": "number"
},
"max_lon": {
"type": "number"
},
"max_lat": {
"type": "number"
},
"max_count": {
"default": 100,
"description": "Maximum number of detections to return (max 500 here; the underlying API defaults to 2000, capped lower to keep responses manageable for an LLM).",
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 500
}
},
"required": [
"min_lon",
"min_lat",
"max_lon",
"max_lat",
"max_count"
],
"$schema": "https://json-schema.org/draft/2020-12/schema",
"additionalProperties": false
}
get_community_activity
Get a snapshot of ongoing community contribution activity on the map: how many corrections are currently pending moderation, their breakdown by action type (add / modify / delete), and the most recent submissions (timestamp, action, target). Submitted free-text comments are intentionally excluded from this tool. This reflects unmoderated, unverified user activity, not the registry itself — do not present it as confirmed detection data.
Parameters1
recent_limit
integer
required
How many of the most recent pending contributions to list.
Raw schema
{
"type": "object",
"properties": {
"recent_limit": {
"default": 15,
"description": "How many of the most recent pending contributions to list.",
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 50
}
},
"required": [
"recent_limit"
],
"$schema": "https://json-schema.org/draft/2020-12/schema",
"additionalProperties": false
}
get_community_activity_in_area
Find pending, unmoderated community contributions relevant to a specific area, to answer "has anyone flagged anything here that is not in the registry yet?" Two independent filters: dpt finds pending edits/deletions (action=modify/delete) targeting existing detections in that département; a full bounding box (min_lon/min_lat/max_lon/max_lat) finds pending new additions (action=add) whose proposed footprint centroid falls inside it. Combine with get_department_capacity_stats / search_detections / get_detections_in_bbox for the confirmed registry picture, and present this separately and clearly as unverified, pending community input — not confirmed detection data. Free-text comments are never included.
Parameters7
dpt
string
optional
Filter pending edits/deletions of EXISTING detections (action=modify or delete) to this département. Does not apply to proposed new additions (action=add), which have no département recorded on the pending item itself — use the bbox parameters for those.
min_lon
number
optional
min_lat
number
optional
max_lon
number
optional
max_lat
number
optional
scan_limit
integer
required
How many recent proposed additions (action=add) to scan for a bbox match. Only relevant when all four bbox parameters are given — there is no server-side spatial index on pending items, so matching is done by fetching this many of the most recent ones and checking their centroid against the box.
recent_limit
integer
required
Maximum number of matching items to return per category.
Raw schema
{
"type": "object",
"properties": {
"dpt": {
"description": "Filter pending edits/deletions of EXISTING detections (action=modify or delete) to this département. Does not apply to proposed new additions (action=add), which have no département recorded on the pending item itself — use the bbox parameters for those.",
"type": "string"
},
"min_lon": {
"type": "number"
},
"min_lat": {
"type": "number"
},
"max_lon": {
"type": "number"
},
"max_lat": {
"type": "number"
},
"scan_limit": {
"default": 500,
"description": "How many recent proposed additions (action=add) to scan for a bbox match. Only relevant when all four bbox parameters are given — there is no server-side spatial index on pending items, so matching is done by fetching this many of the most recent ones and checking their centroid against the box.",
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 2000
},
"recent_limit": {
"default": 30,
"description": "Maximum number of matching items to return per category.",
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 100
}
},
"required": [
"scan_limit",
"recent_limit"
],
"$schema": "https://json-schema.org/draft/2020-12/schema",
"additionalProperties": false
}
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.
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
Builds per-department auxiliary files (buildings, plants, communes) into temp/. Skipped if already present — safe to rerun after a crash.
Classification
Tiles loaded fully in memory. InceptionV3 classifies 299×299 patches; positives saved as GeoTIFFs to temp/segmentation/.
Segmentation
FCN/DeepLab segments each positive patch. LAMB93 polygons extracted, sorted by tile, merged into pseudo-arrays.
Aggregation
pypvroof extracts tilt/azimuth/kWp per polygon. Building filter applied. Results written to outputs_dir.
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/):
file
description
arrays_{dpt}.geojson
Detected PV polygons in WGS84
characteristics_{dpt}.csv
Per-installation registry: surface (m²), tilt (°), azimuth (°), kWp, city code, lat, lon
aggregated_characteristics_{dpt}.csv
City-level aggregation: count, total kWp, avg surface, avg kWp
arrays_characteristics_{dpt}.geojson
Polygons enriched with all characteristics
Only residential-scale installations (1.7–36.1 kWp) located on buildings are retained.
Configuration reference
parameter
default
description
temp_dir
temp
Working directory. Deleted on success, kept on crash.
outputs_dir
data
Final outputs directory.
cls_threshold
0.4
Classification confidence threshold
cls_batch_size
512
Patches per GPU batch (classification)
decode_workers
3
Concurrent JP2 decode processes feeding the GPU — tune to your real CPU quota, not host core count
decode_stagger_s
35
Gap between initial decode submissions, to avoid lockstep bursty waits — rule of thumb: decode_time / decode_workers
seg_threshold
0.46
Segmentation binarization threshold
seg_batch_size
64
Images per GPU batch (segmentation)
filter_building
True
Discard detections not on a building
tilt_method
lut
pypvroof tilt method (lut or constant)
azimuth_method
bounding-box
pypvroof azimuth method
ic_method
clustered
pypvroof installed-capacity regression type
tiles_list
(empty)
Optional tile subset for partial runs
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.
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.
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:
bash
#!/usr/bin/env bashset -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 || truerm -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
bibtex
@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}
}