PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Gaetano Chiriaco, Luca Barco, Andrea Bragagnolo, Claudio Rossi, Edoardo Arnaudo

arXiv:2608.02315v1cs.CV

TL;DR

Flood mapping needs benchmarks that combine temporal context, diverse sensors, and labels distinguishing permanent from transient water. GEOID-Flood provides such a benchmark and evaluates foundation models across segmentation protocols, finding modest encoder gains, strongest results from optical-SAR fusion, and better transfer to unseen events.

  • Problem

    Single Sentinel-1 images ambiguously represent permanent and flooded water, motivating temporally informed, well-annotated benchmarks for evaluating generalizable representations.

  • Method

    GEOID-Flood combines co-registered bi-temporal Sentinel-1, pre-event Sentinel-2, DEM, and validated water labels, then benchmarks foundation and conventional encoders across segmentation protocols.

  • Results

    Foundation models hold a modest edge, optical-SAR fusion performs best, and GEOID-Flood training transfers better to unseen events than evaluated existing datasets.

  • Takeaways & Limitations

    GEOID-Flood offers a multi-modal benchmark for comparing flood-mapping models across sensors, temporal settings, and geographic transfer.

  • Takeaways & Limitations

    Cross-dataset transfer is scored against GEOID-Flood labels, so its claim is anchored to binary water delineation where source conventions converge.

Abstract

from arXiv · show

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.

1 Introduction

GEOID-Flood addresses limitations in flood-mapping benchmarks by combining bi-temporal SAR, co-registered optical imagery, DEM data, and labels distinguishing transient flooding from permanent water. The benchmark supports controlled evaluation of foundation models, temporal modeling, modality choices, and transfer to unseen events.

  • Motivation and problem: Flood mapping requires rapid, wide-area monitoring that remains reliable under cloud cover and at night, while SAR alone struggles to distinguish permanent water from new inundation.Pre/post-event context and labels separating flooding, background soil, and permanent water are therefore necessary.
  • Motivation and problem: Existing datasets rarely combine bi-temporal SAR, co-registered optical imagery, and event-level splits at continental scale for foundation-model evaluation.This leaves the value of geospatial foundation models for flood segmentation under-explored.
  • Dataset contribution: GEOID-Flood covers 1 141 749 km2 of flood-affected terrain across the widest reported spatial extent and longest acquisition window so far (2016– 2026).It is derived from Copernicus Emergency Management Service activations and is nearly double the largest prior dataset.
  • Dataset contribution: GEOID-Flood jointly provides bi-temporal Sentinel-1 in GRD and RTC variants, Sentinel-2, DEM, and a dedicated permanent-water layer.The labels separate transient flooding from background and permanent water, enabling multi-sensor flood segmentation analysis.
  • Evaluation framework: The study provides a reproducible protocol spanning single-image binary segmentation through multitemporal multi-class segmentation, plus backbone comparisons between foundation models and conventional encoders.The analysis examines pretraining, temporal modeling, modality selection, and generalization to unseen out-of-period events.

2 Related Work

Existing flood datasets lack consistent multi-modal, bi-temporal coverage, standardized SAR processing, and reliable water-class distinctions. GEOID-Flood addresses these limitations through co-registered Sentinel-1/Sentinel-2 data, standardized GRD and RTC products, and manually curated annotations.

  • Multi-modal and temporal coverage: Existing datasets vary in modality and temporal coverage: optical datasets lack SAR, SAR-centric datasets lack optical imagery, and Sen1Floods11 has only single-time paired acquisitions.Datasets combining SAR and optical imagery either lack pre-event SAR or remain largely unpaired.
  • Multi-modal and temporal coverage: 2,675 corresponding S2 images accompany 21,602 S1 tiles in STURM-Flood, illustrating incomplete cross-modal pairing.CAU-Flood instead pairs pre-event S2 with post-event S1 only.
  • Annotations: Existing labels range from automatically derived or weakly supervised masks to manual delineations, while none of the reviewed multi-modal datasets distinguishes flooding from permanent water.Sen1Floods11 manually labels only 446 of 4,831 tiles, under 10%.
  • Processing and evaluation: SAR datasets use heterogeneous processing levels and value ranges without common normalization, and random tile-level splits can risk spatial leakage between training and test sets.Examples include raw GRD and terrain-corrected RTC products.
  • GEOID-Flood design: GEOID-Flood jointly provides co-registered pre- and post-event Sentinel-1, a cloudless pre-event Sentinel-2 reference, standardized GRD and RTC products, and manually curated CEMS ground truth.Its SAR-prioritized annotations are designed to sidestep cloud-induced optical noise.

3 The GEOID-Flood Dataset

GEOID-Flood is a large-scale, multi-modal flood-segmentation benchmark spanning 219 events across 65 countries, built from co-registered Sentinel-1, Sentinel-2, and elevation data. Its reproducible pipeline combines event-based acquisition, validity and quality filtering, and labels that distinguish permanent from flooded water.

  • Dataset scope: GEOID-Flood spans 219 flood events across 65 countries and a decade of acquisitions, capturing diverse climates, land cover, and sensor conditions.The dataset is derived from CEMS Rapid Mapping activations and includes manually filtered flood masks.
  • Data modalities: The benchmark co-registers Sentinel-1 GRD and RTC imagery with VV and VH polarizations, a pre-event Sentinel-2 composite, and the Copernicus GLO-30 DEM at 10 m resolution.Sentinel-1 is available before and after events, whereas Sentinel-2 is retrieved for the pre-event period only because post-event optical acquisitions are often cloud-obscured.
  • Construction pipeline: The construction pipeline uses five automated, reproducible stages: spatial partitioning, data retrieval, validity-mask generation, label composition, and quality filtering.Tiles are formed as regular 10 240 m square bounding boxes aligned to UTM grids, while invalid or incomplete boxes are discarded.
  • Label composition: Permanent-water separation is a core design decision because CEMS products map flood extent but not permanent water, so GEOID-Flood provides a dedicated 10 m permanent-water layer.The layer is trained using the Earth Surface Water dataset and annual AlphaEarth Foundations embeddings.
  • Dataset statistics: The dataset contains 1 055 valid event-AoI pairs from 1 333 candidate areas and covers 219 events from January 2016 to March 2026 across six continents.The most recent events form a temporally disjoint held-out set for cross-dataset experiments.
  • Dataset splits: The event-level partition yields 8 938 training tiles, 1 241 validation tiles, and 2 674 testing tiles, with adjacent or overlapping AoIs constrained to the same split.AoIs are stratified by continent with target proportions of 70/10/20 % for training, validation, and test.

4 Methodology

The methodology frames flood mapping as pixel-wise semantic segmentation and evaluates models through three increasingly complex scenarios: single-image, paired two-pass, and paired single-pass fusion. Targets and losses are adapted to whether imagery is processed independently or jointly, while invalid pixels are excluded from training.

  • Segmentation task: Flood mapping uses three classes—background, permanent water, and flooded water—with target remapping differing between single-image and multi-image settings.Single-image targets merge water categories or relabel flooded pixels depending on acquisition time, whereas jointly processed pre/post imagery retains all three classes.
  • Segmentation task: Pixels outside the CEMS analysis area and other invalid pixels are excluded from the loss.
  • Experimental scenarios: The experiments progress from single-image benchmarking across the full encoder zoo to paired temporal and multi-modal studies using one selected backbone.The first scenario ranks foundation models against conventional ImageNet-pretrained encoders; later scenarios focus on temporal and fusion choices.
  • Single-image: Single-image training treats pre- and post-event crops independently, uses one forward pass per tile, and applies cross-entropy to binary water-body labels.Its binary predictions also provide the post-hoc three-class baseline for explicit flood modelling.
  • Paired, two-pass: Paired two-pass training processes co-registered pre/post images separately, sums their binary cross-entropy terms, and adds a flood-change loss for post-event water absent pre-event.The added binary cross-entropy term targets pixels that are water post-event but not pre-event.
  • Paired, single-pass (fusion): Paired single-pass fusion jointly processes both images with a full three-class target, comparing early channel stacking against mid-fusion feature subtraction before a shared decoder.Mid fusion subtracts pre-event features from post-event features element-wise.

5 Experiments

Experiments evaluate binary water-body segmentation and multiclass flood detection across single-image, paired, fusion, modality-ablation, and cross-dataset protocols. Results show modest foundation-model gains, stronger flood detection from finetuned optical-SAR fusion, and superior transfer from GEOID-Flood training.

  • Experimental setup: All configurations use a standard U-Net decoder and evaluate binary water segmentation alongside three-class flood detection on the test split.For single-image and paired models, flooded water is derived from pre-/post-event binary-mask differences; scenario (iii) predicts three classes directly.
  • RQ1: Encoder comparison: 0.844–0.884 binary IoU spans nearly all scenario-(i) encoders, while finetuned TerraMind-L leads at IoUbin 0.884 and F1bin 0.936.Swin-T reaches IoU 0.873 with 32 M parameters versus TerraMind-L’s 323 M, matching or outperforming every finetuned foundation model.
  • RQ2: Temporal modeling: Finetuning paired and fusion models improves explicit flood-change modeling, but scenario-(ii) finetuning raises IoUflood only marginally from 0.479 to 0.486.Frozen fusion variants remain near or below baseline, whereas finetuning lifts both fusion approaches above it, indicating gains from end-to-end change learning.
  • RQ3: Multi-modal fusion: Finetuned early fusion achieves the best flooded-water IoU of 0.521 and F1bin 0.942, while pre-event optical inclusion performs similarly whether replacing pre-event SAR or supplementing bi-temporal SAR.The comparison is reported for S2 → S1 and S1 + S2 input arrangements.
  • RQ3: Modality ablation: +0.015, to 0.946, is the gain from adding pre-event Sentinel-2 in the frozen TerraMind-B modality ablation, whereas GRD edges RTC at IoUbin 0.931 versus 0.922.DEM remains within noise of the baseline, changing performance from 0.931 to 0.934 within ±0.003 standard deviation.
  • RQ4: Cross-dataset generalization: GEOID-Flood is the strongest training source for transfer to unseen events in both frozen and finetuned regimes, with external datasets retaining the same overall ordering across tasks.Among external sources, Kuro Siwo transfers best on binary water, while WorldFloods v2 leads the three-class task at F1avg 0.816; independently derived flood metrics support genuine transfer gains.

6 Conclusion · A Models and Optimization · A.1 Encoder zoo

GEOID-Flood provides a large-scale multimodal flood benchmark, while its evaluation finds that training design matters more than encoder choice and that tested RTC and DEM layers add no measurable gain. The appendix specifies a controlled TerraTorch/PyTorch Lightning setup and compares geospatial foundation models with ImageNet-pretrained encoders under frozen and finetuned settings.

  • 6 Conclusion: GEOID-Flood is introduced as a large-scale multimodal flood benchmark derived from Copernicus EMS Rapid Mapping activations.
  • 6 Conclusion: Training design matters more than encoder choice: foundation models hold only a modest edge, while temporal pairing alone does not help.
  • 6 Conclusion: Optical context helps only before the event, and tested RTC and DEM layers add no measurable gain.The conclusion notes that purpose-built architectures may yet exploit RTC and DEM layers.
  • 6 Conclusion: Releasing all modalities with a dedicated permanent-water layer provides a benchmark for addressing these modeling gaps.
  • A Models and Optimization: All appendix runs use TerraTorch v1.1 with PyTorch Lightning, bf16-mixed precision, a fixed global seed, and shared optimization, augmentation, and tiled-inference protocols.Models differ only in encoder, decoder family, learning-rate regime, and scenario-specific task head.
  • A.1 Encoder zoo: The encoder zoo includes TerraMind v1, DOFA, OlmoEarth, SSL4EO-ResNet50, Satlas Swin-B, and ImageNet-pretrained ResNet, ConvNeXt, and Swin variants.
  • A.1 Encoder zoo: Geospatial foundation models are evaluated with frozen features or joint encoder-decoder finetuning, whereas ImageNet-pretrained encoder variants are included as comparators.Frozen settings update only the decoder and head; finetuned settings update encoder and decoder jointly.

A.2 Decoder and segmentation head … B Permanent Water Layer Generation

The segmentation system pairs foundation-model encoders with a U-Net decoder and dropout head, trains under mode-specific AdamW schedules, and evaluates tiled predictions. Permanent-water masks are generated from annual geospatial embeddings to distinguish stable water from transient flooding without requiring cloud-free Sentinel-2 imagery at inference.

  • A.2 Decoder and segmentation head: Foundation-model encoders feed a U-Net decoder with channel widths [512, 256, 128, 64] and a segmentation head using dropout 0.3.
  • A.2 Decoder and segmentation head: Transformer token outputs are reshaped into spatial feature maps and converted through learned interpolation into the decoder’s four-level feature pyramid.
  • A.3 Optimization: All models train for 20 epochs with AdamW and cosine-annealed learning rates reaching ηmin = 10−6.
  • A.3 Optimization: Finetuned foundation models use a decoder learning rate of 5 × 10−5 and a 10× lower encoder rate of 5 × 10−6, with weight decay 0.1.
  • A.3 Optimization: Training uses pixel-wise cross-entropy, ignores index 255 pixels, and applies 256×256 crops at stride 128 with D4 augmentation.
  • A.4 Inference and tiling: At test time, each 1024×1024 tile is partitioned into a 4×4 grid of non-overlapping 256×256 windows scored in single forward passes.
  • B Permanent Water Layer Generation: Permanent-water masks are generated per scene from annual geospatial embeddings rather than an external product, avoiding the need for cloud-free Sentinel-2 imagery at inference.
  • B Permanent Water Layer Generation: Annual multitemporal composites provide a temporally robust permanent-water imprint that differs qualitatively from transient flood signals in the same year.

B.1 Training data

Training and evaluation use the globally distributed ESW Sentinel-2 dataset, tiled into non-overlapping patches, alongside annual AEF and TESSERA embeddings over matching areas and acquisition years. These embeddings aggregate yearly observations into compact representations that distinguish stable permanent water from transient signals and artifacts.

  • Training data: ESW provides binary water/non-water labels for 95 globally distributed Sentinel-2 Level-2A scenes acquired in 2019.The original train/test split is retained.
  • Training data: Matching areas and acquisition years are retrieved from AEF 64-dimensional and TESSERA 128-dimensional annual embeddings at 10 m resolution.Both embedding sources are globally available over approximately 2017–2025, with TESSERA covering a variable range around that period.
  • Training data: Each annual embedding aggregates multi-temporal observations into a compact representation in which permanent water differs from ephemeral floods, seasonal moisture variation, and cloud-shadow artifacts.Stable patterns are associated with rivers, lakes, reservoirs, and coastal lagoons.

B.2 Methodology

The methodology uses lightweight decoders operating directly on embedding tensors and compares them with reference baselines that do not require embedding features.

  • Embedding decoders: Two lightweight decoders operate directly on embedding tensors: a linear probe and a shallow convolutional decoder.The linear probe is a 1×1 convolution mapping embeddings directly to logits, testing linearly accessible information without spatial aggregation.
  • Embedding decoders: The linear probe contains 65 parameters for AEF and 129 parameters for TESSERA.These parameter counts quantify the minimal decoder capacity used for each embedding.
  • Reference baselines: Reference baselines include training-free JRC-GSW and a trained DeepLabV3+ model with a ResNet-50 backbone.JRC-GSW uses Landsat-derived per-pixel water-occurrence statistics, binarised at ≥75% for permanent water.

B.3 Results

On the Earth Surface Water test set, embedding-based models strongly delineate permanent water, with linear probes already performing well and spatial decoding benefiting AEF more than TESSERA. AEF-MLP outperforms conventional spectral modeling and training-free JRC-GSW, while qualitative results show transfer to detailed water features without spectral inputs.

  • Embedding-based results: Linear probes achieve 0.883 F1 for AEF and 0.902 F1 for TESSERA using fewer than 130 parameters, indicating nearly linearly separable permanent-water signals.These probes use only a single 1×1 convolution without spatial aggregation.
  • Embedding-based results: Adding a spatial decoder increases AEF by +8.0 F1 points, from 0.883 to 0.963, but TESSERA by only +0.9 points, from 0.902 to 0.911.The passage attributes AEF’s larger gain to its compact 64-dimensional space and TESSERA’s smaller gain to its 128-dimensional contextual representation.
  • Baselines: JRC-GSW achieves F1 = 0.874 without ESW training, surpassing DeepLabV3+ and trailing the best model by 8.9 points.JRC-GSW operates at 30 m, whereas the embeddings produce predictions at 10 m.
  • Baselines: DeepLabV3+ reaches only 0.773 F1, reflecting limited transferability from training on relatively few scenes to globally distributed areas of interest.The comparison is presented as a realistic, resource-comparable regime rather than a tight spectral upper bound.
  • Qualitative results: AEF-MLP correctly delineates rivers, narrow channels, and coastal features despite never observing spectral bands from the input scene.Qualitative examples are shown in Fig. 6.

B.4 Inference Pipeline

The inference pipeline retrieves year-matched AEF embeddings for each flood-event area, decodes permanent water with AEF-MLP, and applies hysteresis thresholding to produce 10 m binary masks. Applied across 219 events, it provides finer-resolution permanent-water layers than JRC Global Surface Water without requiring a multi-decadal Landsat archive.

  • Inference Pipeline: AEF-MLP produces binary permanent-water masks at 10 m resolution from embeddings matched to each flood event’s year.Events before 2017 use the earliest available AEF year as a proxy because permanent water is stable over multi-year periods.
  • Inference Pipeline: Hysteresis thresholding uses p ≥0.5 seeds and p ≥ 0.3 for spatial extension during mask binarisation.
  • Inference Pipeline: 219 flood events were processed to produce permanent-water layers at 10 m resolution.The pipeline covered all valid AoIs from Copernicus EMS activations EMSR151–EMSR871.
  • Inference Pipeline: 10 m layers are three times finer than the JRC Global Surface Water product’s 30 m resolution and avoid requiring a multi-decadal Landsat archive.
Loading 2608.02315v1…