hadrilec/satellite-pictures-classification-ign-france

Dataset Overview and Labels

---

Methodology

The dataset was constructed from satellite imagery using geospatial references obtained from IGN geoservices and OpenStreetMap (OSM). The methodology consisted of two main stages:

1. Geodata preparation
   To ensure spatial consistency, a three-step procedure was applied:

   • Definition of spatial boundaries, such as departments, communes, or islands.
   • Intersection with thematic datasets, including agricultural parcels, building footprints, or forest polygons.
   • Filtering of results to retain only the geometries relevant to each land-cover class.

   This process was implemented for the four target categories:

   • Forest: Public forest areas were extracted from IGN BDTOPO_V3:foret_publique, the Fontainebleau Forest is the selected area.
   • Field: Agricultural areas were derived by intersecting the Eure-et-Loir department boundary (ADMINEXPRESS-COG-CARTO.LATEST:departement) with agricultural parcels from RPG.LATEST:parcelles_graphiques.
   • Urban: Built-up areas were delineated by intersecting commune boundaries (ADMINEXPRESS-COG-CARTO.LATEST:commune and ADMINEXPRESS-COG-CARTO.LATEST:arrondissement_municipal; Annemasse, Toulouse, Lyon, Vichy, Antony; Paris 2nd arrondissement) with building footprints from CADASTRALPARCELS.PARCELLAIRE_EXPRESS:batiment.
   • Sea: Maritime regions were obtained by first delineating island and commune boundaries from OSM (Belle-Île-en-Mer, Quiberon, Île de Porquerolles). These polygons were then filtered using IGN BDTOPO_V3:limite_terre_mer to retain only offshore areas.

2. Satellite image extraction
   Based on the geodata polygons, a list of tiles was generated, each corresponding to one satellite image. For each category, 10,000 tiles were randomly selected. The associated images were then downloaded at the highest zoom level available from the IGN ORTHOIMAGERY.ORTHOPHOTOS database.

---

Data Availability

All datasets used in this study are publicly available. Geospatial data, including administrative boundaries, cadastral parcels, forest areas, and land–sea limits, were obtained from IGN geoservices (https://geoservices.ign.fr/). Complementary geographic information on islands and communes was sourced from OpenStreetMap (https://www.openstreetmap.org/).

Model

a CNN model using the dataset is available here: https://huggingface.co/hadrilec/satellite-pictures-classification-ign-france-model

Stats

• train: 32000 images (sea: 8000, forest: 8000, urban: 8000, field: 8000)
• val: 4000 images (sea: 1000, forest: 1000, urban: 1000, field: 1000)
• test: 4000 images (sea: 1000, forest: 1000, urban: 1000, field: 1000)

Usage

from datasets import load_dataset

ds = load_dataset("hadrilec/satellite-pictures-classification-ign-france")