GeoFlood-275

GeoFlood-275 is a global event-level benchmark for rapid event-induced flood inundation mapping from heterogeneous optical-SAR observations with terrain and land-cover information. It contains 275 flood event-AOI pairs worldwide and 133,555 512 x 512 GeoTIFF patch samples.

The benchmark follows an operational setting: pre-event Sentinel-2 optical imagery provides antecedent land-surface context, post-event Sentinel-1 SAR imagery provides all-weather crisis observations, and slope plus land-cover layers provide environmental context. Reference inundation maps are derived from Copernicus Emergency Management Service Rapid Mapping (EMSR) delineation products and represent event-induced inundation within the official EMSR AOI. Permanent or pre-existing surface-water bodies are not treated as target flood pixels under the adopted EMSR delineation definition.

This repository packages the benchmark as WebDataset TAR shards to keep the public Hugging Face dataset efficient to browse, download, and stream.

Dataset Paper

Toward Rapid Flood Mapping Anywhere via Terrain- and Land-Cover-Conditioned Optical-SAR Fusion

Jiepan Li, He Huang, Wenke Li, Linxin Li, Anqi Xie, Ruoru Ye, Lei Hu, Ting Hu, Wei He, and Liangpei Zhang.

Manuscript under revision for Remote Sensing of Environment.

Splits

GeoFlood-275 uses event-level temporal separation. Events from 2015-2025 are used for training/model development; events from 2026 onward are reserved for temporally separated validation and testing.

split	event-AOI pairs	patch generation	samples	shards	payload
train	234	512 x 512 patches, stride 256	125,552	86	342.57 GiB
validation	20	512 x 512 patches, stride 512	4,476	3	11.92 GiB
test	21	512 x 512 patches, stride 512	3,527	3	9.46 GiB
total	275	-	133,555	92	363.96 GiB

Patches with more than 70% invalid pixels were discarded during construction. Boundary patches are padded when necessary.

Repository Layout

README.md
LICENSE
NOTICE.md
CITATION.bib
data/
  train/geoflood-train-000000.tar
  validation/geoflood-validation-000000.tar
  test/geoflood-test-000000.tar
metadata/
  shard_manifest.jsonl
  sample_index_train.csv
  sample_index_validation.csv
  sample_index_test.csv

Each WebDataset sample contains files sharing the same sample key:

<sample_key>.s2.tif
<sample_key>.s1.tif
<sample_key>.s1pre.tif
<sample_key>.flood.tif
<sample_key>.dem.tif
<sample_key>.dynamic_landcover.tif
<sample_key>.esav200.tif
<sample_key>.slope_norm.tif
<sample_key>.json

Example sample key:

EMSR251_02ARENDAL_DEL_MONIT01_v2_0_0

Modalities

suffix	source layer	bands	dtype	no-data	notes
s2.tif	pre-event Sentinel-2 Level-1C	4	Float32	-100	B4, B3, B2, B8; stored in clipped reflectance scale, typically 0-6000
s1.tif	post-event Sentinel-1 GRD	2	Float32	-50	VV and VH backscatter in dB
s1pre.tif	auxiliary pre-event Sentinel-1 GRD	2	Float32	-50	optional layer for SAR-SAR or optical+SAR ablation experiments
flood.tif	EMSR-derived inundation reference	1	Byte	none	binary event-induced flood mask, 1 = flooded, 0 = non-flood
dem.tif	Copernicus DEM GLO-30	1	Float32	-9999	elevation layer aligned to Sentinel-1 reference grid
dynamic_landcover.tif	Dynamic World NRT land-cover composite	1	Int16	0	event-adjacent 14-day pre-event majority-vote land-cover labels
esav200.tif	ESA WorldCover v200 2021	1	Int16	0	static 10 m land-cover reference
slope_norm.tif	normalized slope derived from Copernicus DEM	1	Float32	0	slope clipped to 0-45 degrees and normalized to 0-1
json	sample metadata	-	JSON	-	split, source event id, tile coordinates, and original relative paths

The released GeoTIFF files are not pre-normalized training tensors. They preserve the benchmark raster values; normalization is applied in the training dataloader. The reference implementation is utils/dataloader.py in the GeoFloodNet/RapidFloodMapping codebase: Sentinel-1 is clipped to [-45, 25] dB and mapped to [-1, 1], Sentinel-2 is clipped to [0, 6000] and mapped to [-1, 1], DEM is scaled by /5000 then standardized with mean 0.5 and std 0.5, slope is clipped to [0, 1], ESA WorldCover raw codes are remapped to contiguous IDs, Dynamic World labels are kept as integer labels, and flood masks are mapped to {0, 1} when needed.

Loading

Install the common tooling:

python -m pip install -U datasets huggingface_hub rasterio webdataset

Stream with Hugging Face Datasets:

from datasets import load_dataset

repo_id = "jiepanli/GeoFlood-275"
data_files = {
    "train": f"hf://datasets/{repo_id}/data/train/*.tar",
    "validation": f"hf://datasets/{repo_id}/data/validation/*.tar",
    "test": f"hf://datasets/{repo_id}/data/test/*.tar",
}

ds = load_dataset("webdataset", data_files=data_files, streaming=True)
sample = next(iter(ds["train"]))
print(sample.keys())

Download selected shards and read GeoTIFF bytes with Rasterio:

from glob import glob
from huggingface_hub import snapshot_download
from rasterio.io import MemoryFile
import webdataset as wds

repo_id = "jiepanli/GeoFlood-275"
root = snapshot_download(
    repo_id=repo_id,
    repo_type="dataset",
    allow_patterns=["data/test/*.tar", "metadata/*", "README.md", "NOTICE.md", "CITATION.bib"],
)

urls = sorted(glob(f"{root}/data/test/*.tar"))
dataset = wds.WebDataset(urls)

sample = next(iter(dataset))
with MemoryFile(sample["s2.tif"]) as memfile:
    with memfile.open() as src:
        s2 = src.read()  # shape: bands x height x width

print(sample["__key__"], s2.shape)

Source Data And Attribution

GeoFlood-275 is built from independently retrieved and processed public geospatial products anchored by EMSR event metadata:

• Copernicus Emergency Management Service Rapid Mapping EMSR delineation products: https://emergency.copernicus.eu/mapping/ems/rapid-mapping
• Sentinel-2 Level-1C imagery from COPERNICUS/S2: https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S2
• Sentinel-1 GRD imagery from COPERNICUS/S1_GRD: https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S1_GRD
• Copernicus DEM GLO-30: https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_DEM_GLO30
• ESA WorldCover v200 2021: https://developers.google.com/earth-engine/datasets/catalog/ESA_WorldCover_v200
• Dynamic World V1/NRT land-cover labels: https://developers.google.com/earth-engine/datasets/catalog/GOOGLE_DYNAMICWORLD_V1

Please cite the GeoFlood-275 paper and acknowledge the above source products when using this dataset. Users are responsible for complying with the original source-data terms in addition to the GeoFlood-275 license.

Limitations

The EMSR-derived references are expert-interpreted operational products, not absolute hydrodynamic ground truth. Residual clouds and cloud shadows may remain in some pre-event Sentinel-2 observations because the EMSR-designated pre-event acquisition is retained rather than replaced by an independently selected cloud-free image. ESA WorldCover and Dynamic World are useful but imperfect land-cover references and should not be interpreted as deterministic flood/non-flood constraints.

Citation

@misc{li2026geoflood275,
  title = {Toward Rapid Flood Mapping Anywhere via Terrain- and Land-Cover-Conditioned Optical-SAR Fusion},
  author = {Li, Jiepan and Huang, He and Li, Wenke and Li, Linxin and Xie, Anqi and Ye, Ruoru and Hu, Lei and Hu, Ting and He, Wei and Zhang, Liangpei},
  year = {2026},
  note = {Manuscript under revision for Remote Sensing of Environment}
}