EPFL-ECEO/HRSCD_clean

---
license: mit
task_categories:
- image-segmentation
language:
- en
pretty_name: HRSCD-Clean
---

# 📚 HRSCD-Clean Dataset

**Project page:** https://manonbechaz.github.io/2Player/

## 📝 Description

**HRSCD-Clean** is a refined and higher-quality version of the [original **HRSCD**](https://rcdaudt.github.io/hrscd/) remote-sensing
change detection dataset (Daudt *et al.*, 2019). The dataset contains **291 bi-temporal aerial
image pairs**, each at **10,000 × 10,000 px** and **0.5 m spatial resolution**, covering the
regions of **Rennes** and **Caen**, France. Each pair is accompanied by a binary change mask and segmentation masks for both images.

While the imagery is very high resolution, the original annotations suffer from **coarse and
sometimes imprecise labels for change**. To address this, HRSCD-Clean
introduces improved annotations based on the **BD TOPO®** vector database from the French
National Institute of Geographic and Forest Information (IGN).

## 🔧 Label Refinement

To obtain sharper and more reliable change maps:

- For pixels originally marked as *changed*, the coarse mask is replaced by precise
differences extracted from the BD TOPO® vector maps.
- Regions marked as *unchanged* remain unchanged.
- This improves spatial accuracy and significantly reduces false positives.
- Note that this refinement does **not** recover missing changes.
- The method relies on access to high-resolution national vector maps (e.g., BD TOPO®).

This results in a substantially cleaner dataset than the original HRSCD release.

## 🧹 Dataset Pruning

Despite refinement, some inconsistencies remain due to temporal mismatches between the imagery
and vector data production. To filter out noisy samples, we apply a supervised pruning strategy:

1. Train a **FC-Siam-Diff** model on the **Caen** region (development split).
2. Use the trained model to infer changes on the **Rennes** region.
3. Remove samples with high false-positive or false-negative predictions.
4. Maintain at least **5% changed samples**.
5. Apply pruning separately to training, validation, and test regions.
6. Split Rennes geographically into
- **70% train**,
- **10% validation**,
- **20% test**.

Through controlled experiments, the optimal training set size is determined to be:

**10,000 samples (this is the official pruned subset used in the paper.)**

Note that all the images, i.e. the non pruned dataset is provided here. The list of images of each kept pruned version of the dataset are provided for each split in respective `.pkl` files.

---
## 📚 Citation

If you use HRSCD-Clean or if you find this work helpful, please cite:

```
@article{bechaz_2player_2026,
title = {{2Player}: {A} general framework for self-supervised change detection via cooperative learning},
volume = {232},
issn = {0924-2716},
url = {https://www.sciencedirect.com/science/article/pii/S0924271625004630},
doi = {https://doi.org/10.1016/j.isprsjprs.2025.11.024},
journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
author = {Béchaz, Manon and Dalsasso, Emanuele and Tomoiagă, Ciprian and Detyniecki, Marcin and Tuia, Devis},
year = {2026},
keywords = {Change detection, Cooperative learning, Self-supervised learning, Very high-resolution imagery},
pages = {34--47},
}

```