SemanticSeg4EO Tutorial Dataset — Land Cover Classification from BD Ortho IRC

Description

This dataset accompanies the SemanticSeg4EO framework and its QGIS plugin. It provides all the data and outputs needed to reproduce, step by step, a complete land cover classification workflow using deep learning on very high resolution aerial imagery.

The study area is located in Brittany, France. The source imagery is BD Ortho® at 20 cm resolution in IRC (Infrared–Red–Green) composite. Land cover labels are derived from OCS GE (Occupation du Sol à Grande Échelle) and comprise 5 thematic classes (coded 0 to 4).

The dataset covers the full pipeline — from raw inputs to final predictions — so that users can reproduce the tutorial exactly or use the data as a benchmark for their own experiments.

Contents

Raw inputs

• Image_BDOrtho_IRC.tif — BD Ortho IRC image (20 cm, 3 bands: infrared, red, green) used for training
• Labels_OCSGE.tif — Land cover mask (5 classes, values 0–4) derived from OCS GE
• Grid.shp (+ .shx, .dbf, .prj) — Polygon grid defining patch locations over the study area

Extracted patches

• Patch/ — Directory containing 1 435 georeferenced GeoTIFF patches (224 × 224 px), split into:
  • train/images/ and train/labels/ (≈ 70 %)
  • validation/images/ and validation/labels/ (≈ 20 %)
  • test/images/ and test/labels/ (≈ 10 %)

Trained model

• trained_models/model_best_iou.pth — Best IoU checkpoint (SegFormer-B2, 5 classes)
• trained_models/model_final_model.pth — Final model weights
• trained_models/model_metrics.json — Full training metrics history
• trained_models/model_training_plot.png — Training curves (loss, IoU)

Independent test data

• Image_test_BDOrtho_IRC.tif — Independent BD Ortho IRC image (20 cm, same specifications), not used during training
• Prediction_test.tif — Model prediction on the independent test image

Classes

Spatial reference

All rasters and vectors share the same coordinate reference system: RGF93 / Lambert-93 (EPSG:2154).

How to use

1. Install the SemanticSeg4EO QGIS Plugin and set up the external Python environment (see documentation).
2. Follow the tutorial in the plugin documentation.
3. Use the raw inputs to reproduce patch extraction, or start directly from the Patch/ folder for training.
4. Use the trained .pth model to reproduce the prediction on the independent test image, or apply it to your own BD Ortho IRC scenes.

Source data credits

• BD Ortho® — IGN (Institut national de l'information géographique et forestière), distributed under Licence Ouverte 2.0.
• OCS GE — IGN, Occupation du Sol à Grande Échelle, distributed under Licence Ouverte 2.0.

Related software

• SemanticSeg4EO (standalone framework): https://github.com/aleguillou1/SemanticSeg4EO
• SemanticSeg4EO QGIS Plugin: https://github.com/aleguillou1/SemanticSeg4EO_Qgis (Plug-in folder)

Citation

If you use this dataset, please cite:

Le Guillou, A. (2025). SemanticSeg4EO Tutorial Dataset — Land Cover Classification from BD Ortho IRC [Data set]. Zenodo. 10.5281/zenodo.18784043

Author

Adrien Leguillou Research Engineer — LETG, Université de Bretagne Occidentale adrien.leguillou@univ-brest.fr

License

This dataset is distributed under Creative Commons Attribution 4.0 International (CC BY 4.0).