Research map / Label-efficient and robust learning / Self-supervised and contrastive
Kuro Siwo: 33 billion $m^2$ under the water. A global multi-temporal satellite dataset for rapid flood mapping
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Kuro Siwo, a manually annotated global multi-temporal Sentinel-1 SAR flood mapping dataset (GRD and SLC products, 43 events) plus an unlabeled SSL pool and the BlackBench baseline suite.
Problem
Deep learning for rapid flash-flood mapping from SAR is limited by the lack of large, well-annotated, globally diverse datasets, and existing labels (e.g., CEMS) are often noisy and Europe-centric.
Approach
A manually photointerpreted, multi-temporal Sentinel-1 dataset with coregistered GRD and SLC products, evaluated on unseen test events across six continents, plus a large unlabeled SAR pool for self-supervised pretraining and a benchmark (BlackBench) of CNN, transformer, change-detection and recurrent baselines, including a MAE-pretrained ViT (FloodViT).
Key results
- UNet-ResNet50 is the best BlackBench model with 80.12% F1 for flood, 78.24% for permanent water (best single PW value 79.33% for UNet-ResNet101), and 83.85% F1 for water.
- FloodViT (frozen MAE encoder with light decoder) performs comparably to most segmentation models and outperforms the change detection baselines on GRD data (78.70% F1 flood).
- SLC-based UNet-ResNet18 reaches about 79.94% F1 on binary water detection, showing that unrefined complex SAR data can be classified.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use all available pre-event images rather than a shorter time series (ablation) | BlackBench semantic segmentation models on GRD · Kuro Siwo GRD test set | F1-score | Helped |
| Add DEM elevation information (ablation) | BlackBench models on GRD · Kuro Siwo GRD test set | F1-score | No clear effect |
| Add slope information derived from DEM (ablation) | BlackBench models on GRD · Kuro Siwo GRD test set | F1-score | No clear effect |
| Fine-tune the last 12 transformer layers in addition to the decoder (ablation) | FloodViT · Kuro Siwo GRD test set | F1-F (%): 78.70 -> 78.35 | Hurt |
| Fine-tune the last 12 transformer layers in addition to the decoder (ablation) | FloodViT · Kuro Siwo SLC test set | F1-F (%): 56.87 -> 57.62 | Helped |
| Use SLC data rather than GRD data | UNet-ResNet18 · Kuro Siwo test set | F1-W (%): 83.31 -> 79.94 | Hurt |
Surprising findings
- Original CEMS annotations gave IoU of only 51% and 48% against photointerpreted labels for permanent water and flood on one French event, and the authors' models reached under 70% on all metrics with them.
- Adding DEM information yields negligible or marginal gains across segmentation models.
Limitations
- Approximately half of the flood events are from Europe, giving imbalanced geographic coverage.
- Observed flood extents depend on Sentinel-1 revisit time and may not capture maximum flood extent.
- C-band SAR struggles under dense vegetation and in urban areas (double-bounce), and speckle complicates delineation.
Experimental setup
Datasets: Kuro Siwo labelled (Sentinel-1 GRD) Kuro Siwo SLC Kuro Siwo unlabelled SAR ImageNet SRTM 1 Sec DEM
Benchmarks: Kuro Siwo test set (10 unseen flood events) BlackBench
Metrics: F1-score per class mIoU F1 water
Baselines: U-Net DeepLabv3 UPerNet FC-EF-Diff FC-EF-Conc SNUNet-CD Changeformer ConvLSTM
Builds on: MAE UPerNet SNUNet-CD Changeformer FC-EF
Closest papers by meaning
- M3LEO: A Multi-Modal, Multi-Label Earth Observation Dataset Integrating Interferometric SAR and Multispectral Data (NeurIPS 2024 · similarity 0.85)
- GLH-Water: A Large-Scale Dataset for Global Surface Water Detection in Large-Size Very-High-Resolution Satellite Imagery (AAAI 2024 · similarity 0.84)
- DisasterM3: A Remote Sensing Vision-Language Dataset for Disaster Damage Assessment and Response (NeurIPS 2025 · similarity 0.83)
- FUSAR-GPT: A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery (CVPR 2026 · similarity 0.83)
- Flood Insights: Integrating Remote and Social Sensing Data for Flood Exposure, Damage, and Urgent Needs Mapping (AAAI 2024 · similarity 0.82)
- Off to new Shores: A Dataset & Benchmark for (near-)coastal Flood Inundation Forecasting (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.