Flood Delineation

Introduction

Floods are the most frequent and widespread weather-related disaster, affecting more people worldwide than any other natural hazard, and both their frequency and severity are rising with a warming climate. Responding to them depends on one thing above all: knowing, quickly and over wide areas, where the water actually is.

Satellite radar is the natural instrument for this. Synthetic Aperture Radar (SAR) sees through clouds and works at night, precisely the conditions that accompany a flood and that blind optical sensors. But radar alone is ambiguous: open water is dark in Sentinel-1 backscatter, and a single acquisition cannot tell a river that has always been there from a field that was dry last week. Resolving that ambiguity is what separates a water map from a flood map, and it is what emergency responders actually need.

This solution addresses the problem by combining pre-event and post-event Sentinel-1 acquisitions with a pre-event Sentinel-2 optical reference and terrain data, and by predicting three distinct classes: background, permanent water, and flooded water. Rather than inferring flooding by subtracting one water mask from another after the fact, the models learn the distinction directly.

Underpinning it is GEOID-Flood, the benchmark dataset we built and released for this task: 219 Copernicus Emergency Management Service (CEMS) Rapid Mapping activations across 65 countries and a decade of acquisitions (2016–2026), decomposed into more than 14,000 co-registered tiles covering 1.14 million km² of flood-affected terrain, with manually validated labels. It is the widest spatial extent and the longest acquisition window reported for a flood segmentation dataset to date.

In figure: all layers of a GEOID-Flood tile, side by side, for three events. Flooded water is shown in cyan, permanent water in blue, invalid pixels in gray.

This work builds on our earlier MMFlood dataset, which paired Sentinel-1 imagery with elevation data and OpenStreetMap hydrography over 95 Copernicus EMS flood activations. GEOID-Flood extends that line of research to far wider coverage, bi-temporal SAR in two processing levels, co-registered optical imagery, and a permanent water layer derived at the native 10 m resolution rather than inherited from a coarser global product.

Key Features

• Flood vs. permanent water: Predicts three classes instead of a single water mask, so transient inundation is separated from rivers, lakes and reservoirs that were already there, the distinction that matters for damage assessment.
• Bi-temporal SAR: Uses co-registered pre-event and post-event Sentinel-1 acquisitions, in both GRD and terrain-corrected RTC form, to reason about change rather than about a single snapshot.
• Optical–SAR fusion: Early- and mid-fusion architectures combine radar with a cloudless pre-event Sentinel-2 composite; in our experiments, fusion with finetuning best resolves the transient flood class.
• Foundation models, rigorously benchmarked: Geospatial foundation models (TerraMind, OlmoEarth, DOFA) and other remote-sensing pretrained encoders (Satlas Swin-B, SSL4EO-S12 ResNet-50) are compared head-to-head with conventional ImageNet-pretrained encoders under one shared protocol, frozen and finetuned.
• Demonstrated generalization: Models trained on GEOID-Flood transfer to unseen, out-of-period flood events better than those trained on Sen1Floods11, WorldFloods, Kuro Siwo or MMFlood.
• Leakage-free evaluation: Splits are assigned per area of interest, with touching AoIs kept together and a temporally disjoint held-out test set, so no flood event straddles training and test.

In figure: the 219 activations span 65 countries across ten years.

Technologies Used

• Remote Sensing Imagery: Pre- and post-event Sentinel-1 GRD and RTC (VV/VH), a pre-event Sentinel-2 L2A composite, and the Copernicus GLO-30 DEM, all co-registered at 10 m.
• Reference labels: Copernicus EMS Rapid Mapping delineation products, manually inspected and corrected.
• Pretrained encoders: geospatial foundation models TerraMind, OlmoEarth and DOFA, plus SatlasPretrain and SSL4EO-S12 backbones, evaluated both frozen and finetuned against ImageNet-pretrained ResNet, ConvNeXt and Swin.
• Computer Vision: U-Net decoders on every backbone, plus paired two-pass models with a flood-change loss and early/mid fusion architectures with a native three-class head.
• Training frameworks: TerraTorch and PyTorch Lightning, driven entirely through declarative YAML configurations.
• Cloud Masking: Per-pixel clear/thin/thick cloud labels from OmniCloudMask.
• Permanent water layer: Derived at 10 m from AlphaEarth Foundations embeddings, avoiding the systematic gaps that 30 m global products leave in narrow rivers and small water bodies.

Use Cases

• Rapid Emergency Response: Delineate flood extent as soon as a post-event Sentinel-1 acquisition becomes available, regardless of cloud cover or time of day, to support civil protection during an ongoing event.
• Damage and Exposure Assessment: Because permanent water is labelled separately, the flooded class maps directly onto the terrain that was actually inundated, and can be intersected with settlement, road and cropland layers to quantify exposure.
• Insurance and Risk Modelling: Build consistent, event-level inundation footprints across a decade of historical activations for loss estimation and portfolio exposure analysis.
• Water Body Monitoring: Run the binary water head on its own to track surface water extent over time, independent of any flood event.
• Model Benchmarking: Use GEOID-Flood as an evaluation suite to test how well a new geospatial foundation model handles a hard, imbalanced, multi-modal downstream task.

Benefits

• Separates transient flooding from permanent water, rather than conflating the two
• Operates under cloud cover and at night through SAR
• Built on open Copernicus data, end to end
• Global in coverage and validated on events held out in both space and time

Live Demo

Explore the practical results of our AI solution through this interactive demo. The application lets you browse flood events and inspect the delineated water extent against the underlying Sentinel-1 and Sentinel-2 imagery.

Try it out

Try out the solution in real time on Hugging Face Spaces:

👉 Launch Demo

Integration

The full training and evaluation code, together with the experiment configurations used in the paper, is openly available at links-ads/geoid-flood. The pipeline is driven entirely through TerraTorch YAML configs, so adapting it to a different backbone, modality stack or area of interest is a configuration change rather than a code change.

The dataset is published on the Hugging Face Hub at links-ads/geoid-flood, packaged as Cloud-Optimized GeoTIFFs with full documentation of every layer and label value. A small sample subset is available for inspecting the data without a large download.

Publications

The dataset, the benchmark protocol and the full set of results are described in “GEOID-Flood: A Large-Scale Multi-Modal Benchmark Dataset for Flood Segmentation”, accepted at the TerraBytes workshop of ECCV 2026. The earlier MMFlood dataset, on which this line of work builds, was published in IEEE Access.

(Chiriaco et al., 2026) (Arnaudo et al., 2022)

Acknowledgements

This study was carried out in the context of the SIU (CUP I53D24000060005) and REHUBS (grant number 101214051) projects.