ValaisCD Dataset

High-Resolution Aerial Change Detection (Switzerland, 2017–2023)

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

---

🗺️ Overview

ValaisCD is a high-resolution change detection dataset built from SwissTopo SWISSIMAGE 10 cm aerial imagery, covering several urban and peri-urban regions of the canton of Valais, Switzerland.
It provides pairs of aerial images captured in 2017 and 2023, along with automatically generated building-change labels derived from the SwissTLM3D topographic vector database.

---

📦 Data Summary

• Source imagery:
  • SWISSIMAGE 10 cm aerial photographs
  • Original resolution: 0.1 m/pixel
  • Downsampled to 0.5 m/pixel for dataset construction
• Geographic coverage:
  • Martigny
  • Sion
  • Sierre
  • Brig
    (all located in the canton of Valais, Switzerland)
• Temporal coverage:
  • 2017 and 2023
• Patch extraction:
  • Original tiles: 10,000 × 10,000 pixels
  • Converted into non-overlapping 256 × 256 pixel patches
• Label source:
  • Building-change annotations derived from the SwissTLM3D landscape model
  • Labels correspond to differences between 2017 and 2023 building footprints
• Task: Binary change detection (building changes)

---

🧹 Dataset Pruning & Quality Filtering

Despite the detailed SwissTLM3D vector labels, some discrepancies remain due to temporal misalignment between image capture and map updates.

To address this, ValaisCD introduces an automatic pruning pipeline that filters out mislabeled or noisy samples:

1. Train a supervised CD model (FC-Siam-Diff) on a development subset.
2. Run the model on the rest of the dataset.
3. Identify samples with significant false positives or false negatives.
4. Remove these samples, under the assumption that they contain annotation errors.
5. Preserve a fixed ratio of 5% changed samples to ensure distributional balance.
6. Apply the filtering independently to changed and unchanged samples.

📍 Development region

• Brig is used as the development set for training the filtering model.

📍 Final geographic split (after pruning)

• Train: Sion (70%)
• Validation: Sierre (20%)
• Test: Martigny (10%)

This split ensures geographical independence and avoids spatial leakage.

---

🔍 How Pruning Quality Was Evaluated

Dataset quality was evaluated indirectly:

• Train FC-Siam-Diff models on datasets with progressively more aggressive pruning.
• Evaluate all models on a fixed test set.
• Observe performance trends as noisy samples are removed.

Findings:

• Moderate pruning improves model performance, confirming removal of mislabeled or ambiguous samples.
• Excessive pruning hurts performance, due to loss of informative examples.
• Optimal size for ValaisCD is 5,000 samples, balancing label quality and dataset sufficiency.

This curated version (5000 samples) is the official version used in the associated 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 ValaisCD 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},
}