README.md

---
language:
- en
license: other
task_categories:
- image-segmentation
- time-series-forecasting
tags:
- earth-observation
- remote-sensing
- change-detection
- disaster-response
- building-damage-assessment
- conflict-zones
- multimodal
- sar
pretty_name: M2Change
---

# M2Change: A Multimodal-Temporal Benchmark for Building Damage Assessment in Conflict Zones

**Overview**
* M2Change is the first large-scale, publicly available benchmark dataset specifically curated for multimodal, multi-temporal building change detection in conflict zones.
* It addresses the critical data scarcity in battlefield scenarios, offering pre-event high-resolution (1m) optical imagery and post-event multi-temporal Sentinel-1 SAR (10m) time series to capture structural dynamics.
* The dataset encapsulates diverse typologies of modern warfare's impact on urban landscapes across two distinct conflict scenarios.

<p align="center">
  <img src="./overview.png" alt="M2Change Dataset Overview" width="100%">
</p>

**Dataset Highlights**
* **M2Change-CZ1**: Covers 17 major cities representing spatially extensive and scattered damage patterns. Pre-event references are from ESRI Wayback Imagery (2021), paired with SAR time series from 2022 to 2023. Positive pixels account for 3.20%.
* **M2Change-CZ2**: Focuses on a single, densely populated area subjected to concentrated destruction. Pre-event optical data is from Google Earth (2022), with post-event SAR series spanning 12 months from 2024 to 2025. Positive pixels account for 8.60%.
* **High-Quality Annotations**: Ground-truth labels were generated through a rigorous human-in-the-loop protocol involving UNOSAT assessment points, OpenStreetMap geometries, and expert validation.

**Usage & Access**
* The dataset is partitioned into non-overlapping training and testing sets at an 8:2 ratio using a geographic splitting strategy.
* To ensure responsible use and strictly confine this work to area-level humanitarian purposes, the dataset is governed by a controlled vetted access model.
* Code and resources are available at: [https://github.com/Weikan0425/M2Change/](https://github.com/Weikan0425/M2Change/)

**License & Ethics**
* The M2Change dataset and the related models are governed by a Responsible AI License (RAIL)[cite: 920]. [cite_start]This license explicitly prohibits any military, intelligence, or surveillance applications.
* To strictly confine this work to area-level humanitarian purposes, the dataset implements strict legal licensing, vetted access, and technical anonymization in alignment with the highest Responsible AI standards.
* The proposed framework is intended strictly as a decision-support tool[cite: 924]. [cite_start]Its outputs must complement, not replace, the judgment of qualified human experts within a human-in-the-loop workflow.
* Users must comply with the respective Terms of Service of Google Earth, ESRI, and Copernicus Sentinel data when using this dataset.

**Paper & Citation**
Details of M2Change and MTCNet can be found in our paper. If this dataset is useful for your research, please consider citing our work:

```bibtex
@article{wei2026beyond,
  title={Beyond bi-temporal and unimodal: A multimodal-temporal coupling network for change detection in conflict zones},
  author={Wei, Kan and Yao, Jing and Cui, Jiahui and Zhao, Xinyu and Wang, Lei and Vivone, Gemine and Ghamisi, Pedram},
  journal={Information Fusion},
  volume={135},
  pages={104402},
  year={2026},
  publisher={Elsevier}
}