2Player: A General Framework for Self-Supervised Change Detection via Cooperative Learning

Manon Béchaz1, Emanuele Dalsasso1, Ciprian Tomoiagǎ2, Marcin Detyniecki2, Devis Tuia1,
ISPRS Journal of Photogrammetry and Remote Sensing 1ECEO Laboratory, EPFL, Switzerland 2AXA GDAI Research, Switzerland
Paper Code Hugging Face Clean HRSCD Hugging Face ValaisCD

What if we could turn any change detection architecture into an unsupervised model ?

Nowadays, Change Detection delivers strong performance, but still relies on huge labeled datasets. And without labels, unsupervised methods fall behind, often because they do not take advantage of the expressive architectures used in supervised models.

But what if it did not have to be this way? 2Player bridges the gap by transforming any CD model into an unsupervised one.

By letting a change detector and a reconstruction network guide each other in a self-supervised loop, with an additional Geographical Correspondence Module providing structural cues, 2Player achieves high-quality change detection without labels, combining the best of both worlds: supervised capacity and unsupervised scalability. Tested on four high-resolution datasets, including a new cleaned-up version of HRSCD and our new ValaisCD, 2Player achieves state-of-the-art results for label-free change detection.

Methodology at a glance

Method step

Overview:
2Player is a cooperative self-supervised framework built around two interacting models. A change detector proposes potential change maps, while a recon- struction network predicts how the scene should look if no change happened. Through an iterative feedback loop, the two models refine each other, enabling reliable change detection without labeled data.

Datasets and Filtering Method

Many popular change-detection datasets, like HRSCD and our newly introduced ValaisCD, suffer from coarse or inconsistent labels, often caused by automatic annotation errors from outdated vector maps. To ensure reliable evaluation, we introduce a simple but effective filtering strategy: we train a standard CD model on a small subset, use it to spot mislabeled or noisy samples, and discard those that consistently produce false positives or false negatives. This pruning significantly boosts data quality without sacrificing diversity, yielding refined versions of HRSCD and ValaisCD. We also test our method on two additional benchmarks, LEVIR-CD and WHU-CD. These datasets provide a solid foundation for evaluating 2Player under fair, label-reliable conditions.

Modalities overview
Pruning strategy significantly improves dataset quality across HRSCD and ValaisCD: Removing training samples improves performance, confirming that we are removing low-quality, i.e. mislabeled or noisy examples which hinder the learning process.

Results

1. Change Detection Results

We benchmark 2Player against a wide range of classical and state-of-the-art unsupervised methods, showing it produces cleaner, more reliable change maps with fewer false positives and better handling of irrelevant variations like seasonal or illumination changes.

2. Reconstruction Results

Player 2 produces visually convincing reconstructions of unchanged regions, and the reconstruction error maps highlight actual changes, demonstrating the effectiveness of the cooperative framework.

Modalities overview
Modalities overview

3. Two signals, one goal: learning complementary guidance

In 2Player, the change detection model, Player 1, is supervised by two complementary signals: the reconstruction error from Player 2 and the structural dissimilarity map from the Geographical Correspondence Module. The carousel below illustrates their complementarity. The reconstruction error highlights all deviations from expected appearance, including both true changes and irrelevant variations like seasonal shifts. The dissimilarity map, in contrast, focuses on structural transformations, suppressing false positives caused by appearance-only changes. In some cases, only one of the two signals correctly identifies a change, showing their complementarity.

4. Can it really be adapted to any change detection architecture?

A key objective of this work is to demonstrate that the 2Player framework is not tied to a specific change detection backbone. To verify this, we instantiated Player 1 using three widely adopted architectures: FC-Siam-Diff, SNUNet, and BIT. As shown in our results, 2Player integrates smoothly with all architectures and consistently delivers competitive performance. These findings confirm that 2Player is a flexible and architecture-agnostic framework, capable of turning a broad range of change detection models into unsupervised methods.

Modalities overview

Takeaways

  • Cooperative learning between two models: 2Player couples a change detection model with a reconstruction-based model, enabling each to guide the other through a self-supervised training loop.
  • Architecture-agnostic design: The framework can transform any existing change detection architecture into an unsupervised model, preserving its expressive capacity without requiring labels.
  • Dual complementary supervision: Reconstruction errors and dissimilarity maps provide two distinct but complementary supervision signals that jointly drive self-supervised learning.
  • Focus on structural change: By leveraging structural dissimilarity maps, 2Player becomes robust to irrelevant variations such as seasonal or illumination changes and concentrates on true structural transformations.
  • State-of-the-art performance: Across four high-resolution datasets, 2Player achieves state-of-the-art results among unsupervised methods.

BibTeX


@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},
}

}
@inproceedings{bechaz2025self,
  title={Self-supervised Change Detection via Cooperative Learning: A Two-Player Model},
  author={B{\'e}chaz, Manon and Dalsasso, Emanuele and Tomoiag{\u{a}}, Ciprian and Detyniecki, Marcin and Tuia, Devis},
  booktitle={2025 Joint Urban Remote Sensing Event (JURSE)},
  pages={1--4},
  year={2025},
  organization={IEEE},
  doi={10.1109/JURSE60372.2025.11076078}
}