WaterBench: Evaluating Geospatial Foundation Models for Coastal and Marine Tasks

Overview

WaterBench is a benchmark and evaluation protocol for assessing GFMs on high-resolution (10m) radar and optical satellite imagery (Sentinel-1, Sentinel-2) across two downstream task families: image-level regression and classification (e.g., water quality, bathymetry, oil-slick detection) and pixel-level segmentation (e.g., mangroves, seagrass). We also include 300m Sentinel-3 ocean-color imagery as contextual information. WaterBench spans multiple coastal regions and seasons, defining fixed splits for in-distribution (ID) evaluation and held-out tests that are spatial (new geographies) and, where applicable, temporal (new years).

Total size: ~175 GB (44 tar archives across 6 tasks)

Tasks

Task	Resolution	Modalities	Target / Classes	Metrics	Train	Val	Test	Spatial	Temporal
OceanState	10 km	S1, S2, S3	Continuous (SST, SSS, SSH, CHL, O2)	RMSE, R²	1,400	200	400	200	200
GEBCO	500 m	S1, S2	Continuous (depth)	RMSE, R²	5,000	800	1,600	800	800
Oil Slick	10 m	S1, S2, S3	Binary (oil/non-oil)	Acc, F1	900	150	300	150	---
ACA	10 m	S1, S2	Multi-class (benthic: 7, geomorphic: 12)	Acc, mIoU	8,211	2,719	2,658	879	---
Seagrass	10 m	S1, S2	Binary (presence/absence)	Acc, mIoU	832	114	222	124	---
GMW	10 m	S1, S2	Binary (mangrove extent)	Acc, mIoU	11,012	1,835	3,670	1,820	---

Modalities

Task Descriptions

Image-Level Tasks

• OceanState (global, 2017-2023): Multi-target regression predicting five ocean variables (SST, SSS, SSH, CHL, O2) from Sentinel-1/2/3 imagery. Labels from Copernicus Marine Service (CMEMS). Spatial OOD: Polar biome. Temporal OOD: 2022-2023.

• GEBCO (global, 2025 grid): Depth regression from satellite imagery in shallow coastal waters. Labels from GEBCO 2025 global bathymetry grid. Spatial OOD: Mediterranean Sea. Temporal OOD: 2022-2023.

• Oil Slick (global, 2017-2023): Binary classification of oil slicks vs. lookalikes. Labels from GlobalOSD-SAR dataset. Spatial OOD: Mediterranean Sea.

Pixel-Level Tasks

• ACA (global, 2018-2020): Coral reef segmentation into geomorphic zones and benthic classes following Allen Coral Atlas taxonomy. Spatial OOD: Great Barrier Reef + Bermuda.

• Seagrass (Maldives, 2021): Binary segmentation of seagrass meadows in shallow lagoons. Spatial OOD: Northern/Southern atolls.

• GMW (global, 1996-2020): Binary segmentation of mangrove extent along tropical/subtropical coasts from Global Mangrove Watch. Spatial OOD: Sundarbans + Everglades + Niger Delta.

Data Format

All images are stored as GeoTIFF files with LZW compression, packaged in uncompressed .tar archives.

Sentinel-2 (images_s2)

• 12 bands: B02, B03, B04, B05, B06, B07, B08, B8A, B11, B12, B01, B09
• 224x224 pixels at 10m resolution (resampled from native resolutions)
• Surface reflectance values scaled to [0, 1]

Sentinel-1 (images_s1)

• 2 bands: VV, VH (backscatter)
• 224x224 pixels at 10m resolution

Sentinel-3 (images_s3) — OceanState, Oil Slick only

• 16 OLCI bands (Oa01-Oa12, Oa16-Oa18, Oa21)
• 120x120 pixels at 300m resolution

Labels

• OceanState/GEBCO: Continuous values in metadata.csv columns
• Oil Slick: Binary label in metadata.csv
• ACA: Per-pixel GeoTIFF masks (benthic: 7 classes, geomorphic: 12 classes)
• Seagrass/GMW: Per-pixel binary GeoTIFF masks

Repository Structure

data/
├── OceanState/
│   ├── metadata.csv
│   ├── metadata.json
│   ├── splits/{random,geographic,temporal}/{train,val,test}.txt
│   ├── OceanState-images_s2.tar
│   ├── OceanState-images_s1.tar
│   └── OceanState-images_s3.tar
├── GEBCO/
│   ├── metadata.csv, metadata.json, splits/
│   ├── GEBCO-images_s2-{00,01}.tar
│   └── GEBCO-images_s1.tar
├── OilSlick/
│   ├── metadata.csv, metadata.json, splits/
│   ├── OilSlick-images_s2-{00,...}.tar
│   ├── OilSlick-images_s1-{00,01}.tar
│   └── OilSlick-images_s3.tar
├── ACA/
│   ├── metadata.csv, metadata.json, splits/
│   ├── ACA-images_s2-{00,...}.tar
│   ├── ACA-images_s1.tar
│   ├── ACA-labels_benthic.tar
│   └── ACA-labels_geomorphic.tar
├── Seagrass/
│   ├── metadata.csv, metadata.json, splits/
│   ├── Seagrass-images_s2.tar
│   ├── Seagrass-images_s1.tar
│   └── Seagrass-labels.tar
└── GMW/
    ├── metadata.csv, metadata.json, splits/
    ├── GMW-images_s2-{00,...}.tar
    ├── GMW-images_s1-{00,01}.tar
    └── GMW-labels.tar

Download

Full dataset

huggingface-cli download ayushprd/WaterBench --repo-type dataset --local-dir WaterBench

Single task (e.g., ACA)

huggingface-cli download ayushprd/WaterBench --repo-type dataset --include "data/ACA/*" --local-dir WaterBench

Extract archives

cd WaterBench/data/ACA
for f in *.tar; do tar xf "$f"; done

Splits

Each task provides pre-defined train/val/test splits as text files listing sample IDs (one per line).

• splits/random/ — Standard i.i.d. split
• splits/geographic/ — Spatial OOD: held-out geographic regions in test set
• splits/temporal/ — Temporal OOD: held-out years in test set (OceanState, GEBCO only)

Loading splits in Python

from pathlib import Path

task_dir = Path("WaterBench/data/OceanState")
train_ids = task_dir.joinpath("splits/random/train.txt").read_text().strip().split("\n")
val_ids = task_dir.joinpath("splits/random/val.txt").read_text().strip().split("\n")
test_ids = task_dir.joinpath("splits/random/test.txt").read_text().strip().split("\n")

License

This dataset is released under the CC-BY-4.0 license.