Wang et al. (2026) Few-Shot change detection in optical and SAR remote sensing images for disaster response

Identification

• Journal: International Journal of Applied Earth Observation and Geoinformation
• Year: 2026
• Date: 2026-01-12
• Authors: Di Wang, Guorui Ma, Xiao Wang, Ronghao Yang, Yongxian Zhang
• DOI: 10.1016/j.jag.2026.105100

Research Groups

• State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan 430079, China
• State Key Laboratory of Target Vulnerability Assessment, Defense Engineering Institute, China
• School of Architecture and Civil Engineering, Chengdu University, Chengdu 610106, China
• Joint Innovation Center of Intelligent Unmanned Perception System, College of Earth and Planetary Sciences, Chengdu University of Technology, Chengdu 610099, China
• State Key Laboratory of Hydroscience and Engineering, Tsinghua University, Beijing 100084, China

Short Summary

This paper addresses few-shot change detection in optical and Synthetic Aperture Radar (SAR) images for disaster response, a critical task challenged by limited labeled samples and data distribution shifts. The authors propose a Dual-Stage Training framework for Change Detection (DSTCD) that leverages structural and semantic features transferred from an image registration source task, achieving superior performance and robustness in various disaster scenarios.

Objective

• To develop a robust framework for few-shot change detection in heterogeneous optical and SAR remote sensing images, specifically for disaster monitoring with extremely limited labeled samples (fewer than 20 image pairs).
• To mitigate the data distribution shift between public datasets and real disaster scenarios by enhancing feature representations through task-guided transfer of structural and semantic knowledge from an image registration task.
• To enable accurate identification of affected areas for efficient disaster response despite data scarcity and noise.

Study Configuration

• Spatial Scale: Images up-sampled by a factor of 10, then partitioned into non-overlapping 256 × 256 pixel patches for training. Real-world disaster images had resolutions of 15 meters (Xinxiang flood) and 1 meter (Noto earthquake).
• Temporal Scale: Utilizes multi-temporal images (pre-disaster optical and post-disaster SAR) for change detection. Few-shot conditions involved training with as little as 2% of total patches, typically fewer than 20 labeled image pairs.

Methodology and Data

• Models used:
  • Proposed: Dual-Stage Training framework for Change Detection (DSTCD), comprising a Change Detection Feature Encoder (CDFE) and a Task-Guided Feature Transfer Module (TGFTM), which includes a Structure Feature Transfer Block (STFTB) and a Semantic Feature Transfer Block (SEFTB).
  • Source Task Pre-training: ASpanFormer, a detector-free image registration network.
  • Comparison Models: M−UNet, HAFF, ISTA, HeteCD, GLCD-DA (heterogeneous change detection models); FC-Siam-diff, FC-Siam-conc, ICIFNet, DMINet, GASNet, HGINet (homologous change detection models).
• Data sources:
  • Optical and Synthetic Aperture Radar (SAR) remote sensing images.
  • Public Datasets: Four public optical and SAR image change detection datasets (Lv et al., 2022b) covering urban expansion and water expansion scenarios.
  • Source Task Dataset: MegaDepth dataset for heterogeneous image registration pre-training.
  • Real-world Disaster Datasets: Remote sensing imagery from the Xinxiang flood of 2021, China (8239 × 5537 pixels), and building damage from the Noto earthquake of 2024, Japan (6664 × 5451 pixels).

Main Results

• DSTCD significantly outperformed state-of-the-art methods in few-shot scenarios. Its average F1-score surpassed the second-best method by 6.98% in urban expansion scenarios and by 13.09% in water expansion scenarios.
• On the Shuguang urban expansion dataset, DSTCD achieved an F1-score of 76.15% and an IoU of 61.50%, improving over the second-best (HeTeCD) by 8.74% (F1) and 10.74% (IoU).
• On the Wuhan urban expansion dataset, DSTCD improved F1 by 5.21% and IoU by 6.12% over the second-best (ISTA).
• For water expansion datasets, DSTCD exceeded the second-best model on Yellowriver by 10.55% (F1) and 13.29% (IoU), and on Yellowriver2 by 15.62% (F1) and 18.5% (IoU).
• Robustness analysis showed DSTCD maintained exceptional performance under extreme data scarcity (e.g., 2% training samples) and consistently outperformed other models across varying sample sizes.
• In real-world disaster applications, DSTCD achieved an F1-score exceeding 80% on the Xinxiang flood dataset (0.95% improvement over second-best) and 51.01% on the challenging Noto earthquake dataset (5.59% improvement over second-best).
• Ablation studies confirmed the effectiveness of the Structure Feature Transfer Block (STFTB) for urban expansion and the Semantic Feature Transfer Block (SEFTB) for water expansion, as well as the multi-loss supervision strategy for balanced precision and recall.
• Paired t-tests demonstrated the extremely high statistical significance of DSTCD's performance improvements across all datasets.

Contributions

• Introduction of a dual-stage framework that leverages task-specific structural knowledge from image registration to enhance few-shot change detection, reducing sample requirements to 2% of typical dataset sizes.
• Proposal of a task-guided feature transfer module (TGFTM) within the DSTCD framework, which strategically leverages structural knowledge from the registration task to enable effective cross-task feature transfer, significantly enriching feature representations for the few-shot target change detection task.
• Comprehensive experimental validation on four public optical and SAR image change detection datasets across two distinct scenarios, including robustness analysis under varying training sample sizes and generalization ability evaluation through real-world disaster applications.

Funding

• Guangxi Science and Technology Major Project (grant No. AA22068072)
• Open Research Fund of State Key Laboratory of Target Vulnerability Assessment, Defense Engineering Institute, AMS (YSX2024KFPG005)

Citation

@article{Wang2026FewShot,
  author = {Wang, Di and Ma, Guorui and Wang, Xiao and Yang, Ronghao and Zhang, Yongxian},
  title = {Few-Shot change detection in optical and SAR remote sensing images for disaster response},
  journal = {International Journal of Applied Earth Observation and Geoinformation},
  year = {2026},
  doi = {10.1016/j.jag.2026.105100},
  url = {https://doi.org/10.1016/j.jag.2026.105100}
}