GEOID-Flood: A Large-Scale Multi-Modal Benchmark Dataset for Flood Segmentation

AI summary

Overview

• Research area: Computer vision / remote sensing — semantic segmentation of floods using satellite imagery, and benchmarking of geospatial foundation models.
• Technical level: Intermediate. The paper is readable by anyone familiar with image segmentation basics, but it assumes some familiarity with SAR imagery, optical remote sensing, and evaluation metrics such as IoU and F1.
• Scope in one sentence: The paper introduces GEOID-Flood, a global multi-modal flood segmentation dataset built from Copernicus Emergency Management Service activations, and uses it to benchmark geospatial foundation models against conventional encoders for flood and water-body mapping.

What This Paper Is About

Operational flood mapping needs to work under cloud cover, at night, and over wide areas, which makes radar (Sentinel-1 SAR) the natural data source — but a single SAR image cannot easily tell a permanent river apart from newly flooded terrain, because both return similarly dark backscatter. Resolving that ambiguity requires before-and-after imagery, co-registered optical context, and labels that explicitly separate permanent water from transient flooding.

The paper's goal is to supply a large benchmark that provides all of these at once, and then use it to test whether geospatial foundation models actually beat simpler, conventionally pretrained encoders on flood segmentation, which training designs work best, and whether training on this dataset transfers to unseen flood events.

Key Contributions

1. GEOID-Flood dataset: a global flood segmentation benchmark of 14,282 tiles at 1024×1024 pixels, covering 219 flood events across 65 countries from January 2016 to March 2026, with 1,141,749 km² of flood-affected terrain, and four co-registered modalities at 10 m ground sampling distance: pre- and post-event Sentinel-1 in both GRD and RTC format (VV/VH), a pre-event Sentinel-2 Level-2A composite (12 spectral bands resampled to 10 m), and the Copernicus GLO-30 DEM.
2. A dedicated permanent water layer derived by training a lightweight model on the Earth Surface Water dataset from annual AlphaEarth Foundations embeddings, so that labels distinguish background, permanent water, flooded water, and invalid pixels — a separation the authors state no prior public flood dataset provides.
3. A reproducible training and evaluation protocol spanning three scenarios of increasing complexity: single-image binary water-body segmentation, paired two-pass training with an explicit flood-change loss, and paired single-pass fusion with either early (channel stacking) or mid (separate branches merged by subtraction) fusion.
4. An extensive backbone benchmark of foundation models (TerraMind tiny/small/base/large, DOFA base/large, OlmoEarth-B, SSL4EO-ResNet50, Satlas Swin-B) against ImageNet-pretrained encoders (ResNet-50/101, ConvNeXt-Tiny/Base, Swin-Tiny/Base), plus a cross-dataset generalization study against Kuro Siwo, MMFlood, WorldFloods v2, and Sen1Floods11.

Main Findings

• Foundation models hold only a modest edge. Under a shared protocol, every model except the frozen Satlas Swin-B falls within a 0.04 range of binary IoU (0.844–0.884). Finetuned TerraMind-L reaches the best single-image result (IoU bin 0.884, F1 bin 0.936), but Swin-T reaches 0.873 IoU while being nearly an order of magnitude smaller (32 M vs. 323 M parameters).
• Finetuning helps the weak more than the strong. Satlas Swin-B improves from 0.751 to 0.861 IoU when finetuned, while already-robust backbones gain little. The authors conclude that with a strong shared decoder, the backbone is not the bottleneck for binary water segmentation.
• Single-image flood detection is weak. The best flooded-water IoU in the single-image benchmark is 0.484, showing the flood class is much harder than generic water segmentation.
• Temporal pairing alone does not help. In the paired two-pass setting, the flood-change loss matches the post-hoc baseline on binary water when frozen but trails on flooded water; finetuning lifts flooded-water IoU only marginally over the baseline (0.486 vs. 0.479).
• Explicit fusion helps only when finetuned. Both fusion variants sit near or below the baseline when frozen, with early fusion suffering most. Finetuned early fusion with pre-event optical and post-event SAR gives the best flooded-water IoU (0.521) and binary water F1 (0.942). Replacing pre-event SAR with pre-event Sentinel-2 matches stacking both modalities (S1+S2).
• Optical context is the most valuable input. In the modality ablation with a frozen TerraMind-B on pre-event tiles, pre-event Sentinel-2 adds the largest gain (+0.015 IoU bin, to 0.946), GRD consistently edges out RTC (0.931 vs. 0.922 IoU bin), and the DEM stays within noise of the baseline (0.931 to 0.934, within ±0.003 standard deviation).
• GEOID-Flood transfers best to unseen events. On the temporally disjoint held-out set, models trained on GEOID-Flood lead in both regimes: frozen, F1 0.911 and IoU flood 0.590 versus Kuro Siwo at 0.887 / 0.568; finetuned, F1 0.917 and IoU flood 0.601. Among external sources, Kuro Siwo transfers best on binary water and WorldFloods v2 on the three-class task (F1 avg 0.816).
• Cross-dataset finetuning is dataset-dependent. Finetuning clearly helps Sen1Floods11 but destabilizes MMFlood, whose flood/non-flood imbalance makes full finetuning harder (IoU flood 0.512 to 0.380).
• Coverage is geographically skewed. Europe dominates with 140 of the 219 events.

Methodology in Plain English

The authors start from Copernicus Emergency Management Service Rapid Mapping activations, where each activation is a flood event containing one or more Areas of Interest, and each area has vector flood delineations produced over the days following the disaster. For each event–area pair they pick a single product, prioritizing those derived from Sentinel-1 and Sentinel-2 so that labels and training imagery share the same sensor and resolution, and manually inspecting and correcting every selected label. They subdivide each area into regular 10,240 m bounding boxes aligned to UTM grids and retrieve Sentinel-1 GRD and RTC imagery for the pre- and post-event dates. Sentinel-2 is retrieved only for the pre-event period, since flood-time cloud cover makes a clear post-event optical acquisition unlikely; they build a median composite over a window of three weeks from the event date using at most three S2-L2A acquisitions, and also download the GLO-30 DEM resampled to 10 m. A cloud mask is produced with OmniCloudMask, and a validity mask marks usable pixels as the intersection of the area boundary, image footprint, and tile bounding box.

Because CEMS products map flood extent but not permanent water, they train a lightweight model on the Earth Surface Water dataset using AlphaEarth Foundations embeddings to produce a dedicated 10 m permanent water layer, then merge everything into four classes: background, permanent water, flooded water, and invalid. Bounding boxes with missing modalities, excessive cloud, or imagery inconsistent with the label are discarded, and 1,333 candidate areas reduce to 1,055 valid event–area pairs.

For evaluation, the 14,282 tiles are split at event level with target proportions of 70/10/20 percent (8,938 train, 1,241 validation, 2,674 test), with adjacent or overlapping areas forced into the same split to prevent leakage; a further 1,429 tiles from events after January 2026 (EMSR857–EMSR871, 83 event–area pairs, February–March 2026) form a temporally disjoint held-out set for cross-dataset tests. Models are trained for 20 epochs with AdamW and a cosine-annealed learning rate in TerraTorch v1.1 with PyTorch Lightning, using a U-Net decoder in every configuration, and are scored on both binary water-body segmentation and three-class flood detection using F1 and IoU, with IoU flood as the primary reference metric.

Why This Matters

Impact on research. The paper argues that existing flood datasets rarely combine bi-temporal SAR with co-registered optical imagery and event-level splits, so the value of geospatial foundation models for flood mapping has been largely untested. GEOID-Flood fills that gap with the widest spatial extent and longest acquisition window reported to date, standardized SAR at two processing levels (GRD and RTC) to remove cross-dataset normalization problems, and an explicitly separated permanent water class. The headline result — that training design matters more than the encoder — is a useful corrective for a field that often assumes bigger pretrained models automatically win.

Real-world applications.

• Rapid flood extent mapping for emergency response, where mapping must stay reliable under cloud cover and at night.
• Separating genuine new inundation from permanent rivers and reservoirs, which matters for damage assessment and for avoiding false alarms.
• Cross-region and cross-event monitoring, since event-level splits and a temporally disjoint test set support claims about generalization rather than memorization.
• Climate adaptation and risk planning, given the projected rise in flood frequency and severity cited by the authors.

Industry relevance. The paper directly addresses operational deployment concerns: standardization of SAR processing levels, event-level rather than random splits to avoid spatial leakage, a held-out test set for realistic transfer measurement, and a practical model choice. The authors adopt TerraMind-B for the remaining experiments because it loses negligible performance relative to the top model at roughly a third of the parameters (101 M vs. 323 M) — a trade-off relevant to anyone running inference at scale.

Future Directions

• Better use of the released modalities. The authors note that the DEM and RTC layers added no measurable gain with the encoders tested, but retain them in the released dataset on the assumption that purpose-built or more recent architectures may exploit them.
• Improving the flooded-water class. Flooded water is the rarest class and remains the hardest throughout, with single-image flooded-water IoU peaking at 0.484; closing this gap is the central open problem.
• Reducing label noise and improving permanent water layers. The paper acknowledges that labels inherit residual noise from CEMS delineations and from the deep-learning-derived permanent water layer.
• Broadening geographic balance and evaluation scope. Coverage skews toward Europe (140 of 219 events), and the cross-dataset comparison is scored against GEOID-Flood's own labels — the authors anchor that claim on binary water delineation, where source conventions converge, and leave broader evaluation open.

Target Audience

Researchers and practitioners in remote sensing and computer vision who work on flood or water segmentation; developers evaluating geospatial foundation models for downstream Earth-observation tasks; and teams building operational or humanitarian flood-monitoring systems who need a large, multi-modal, leakage-controlled benchmark with standardized SAR preprocessing. Readers interested in dataset construction methodology — particularly around permanent water separation, cloud handling, and event-level splitting — will also find the construction pipeline useful.

Authors’ abstract

Geospatial foundation models aim to learn representations that transfer across regions and sensors, yet evaluating them on specific tasks requires large, high-quality, multi-modal benchmarks that measure how well such models extract value from data. Concerning flood mapping, existing datasets rarely combine bi-temporal SAR and co-registered optical imagery at scale, leaving the value of foundation models for this downstream task largely untested. We introduce GEOID-Flood, a large-scale multi-modal flood segmentation benchmark, derived from Copernicus Emergency Management Service activations, spanning 219 events across 65 countries over ten years. The dataset provides more than 14,000 tiles with co-registered pre- and post-event Sentinel-1, in GRD and RTC format, pre-event Sentinel-2 composite, and DEM, including manually validated labels that separate background from permanent water and flooded water. Using this benchmark, we evaluate foundation models against conventional encoders across single-image, multi-temporal, and multi-modal protocols. We report three main findings: foundation models offer a consistent but modest advantage; optical-SAR fusion with finetuning best resolves transient flooding; and models trained on GEOID-Flood transfer to unseen events better than those trained on existing datasets. Dataset and code available at https://github.com/links-ads/geoid-flood.