OceanVerse Dataset

OceanVerse is a comprehensive dataset designed to address the challenge of reconstructing sparse ocean observation data. It integrates nearly 2 million real-world profile data points since 1900 and three sets of Earth system numerical simulation data. OceanVerse provides a novel large-scale (∼100× nodes vs. existing datasets) dataset that meets the MNAR (Missing Not at Random) condition, supporting more effective model comparison, generalization evaluation and potential advancement of scientific reconstruction architectures. The dataset and codebase are publicly available to promote further research and development in the field of AI4Ocean.

Quick Start

One-Step Download

# Step 1: Install uv (skip if already installed)
pip install uv

# Step 2: Download the dataset
uvx hf download jingwei-sjtu/OceanVerse --repo-type=dataset --local-dir 'YOUR_LOCAL_DIR'

# If a proxy is required:
HTTP_PROXY='YOUR_HTTP_PROXY' HTTPS_PROXY='YOUR_HTTPS_PROXY' \
  uvx hf download jingwei-sjtu/OceanVerse --repo-type=dataset --local-dir 'YOUR_LOCAL_DIR'

Dataset Structure

After downloading, the local directory structure will be as follows:

YOUR_LOCAL_DIR/
├── CESM2_omip1/
│   ├── omip1_indices/          # split indices of CESM2-OMIP1 (random split, temporal split and spatial split)
│   ├── omip1_graph/            # Graph-structured data (model input)
│   └── omip1_ground_truth/     # Ground truth of CESM2-OMIP1
├── CESM2_omip2/
│   ├── omip2_indices/          # split indices of CESM2-OMIP2 (random split, temporal split and spatial split)
│   ├── omip2_graph/            # Graph-structured data (model input)
│   └── omip2_ground_truth/     # Ground truth of CESM2-OMIP2
├── GFDL_ESM4/
│   ├── GFDL_indices/           # split indices of GFDL-ESM4 (random split, temporal split and spatial split)
│   ├── GFDL_graph/             # Graph-structured data (model input)
│   └── GFDL_ground_truth/      # Ground truth of GFDL-ESM4
├── .gitattributes
└── README.md

Subfolder Description

Indices (*_indices/)

Each indices folder contains 7 split index files (all in .pt format) used to partition the data:

File	Description
random_split_seed1.pt	Random splits of observation data into training/validation sets (7:3 ratio) with seed 1
random_split_seed2.pt	Random splits of observation data into training/validation sets (7:3 ratio) with seed 2
random_split_seed3.pt	Random splits of observation data into training/validation sets (7:3 ratio) with seed 3
random_split_seed4.pt	Random splits of observation data into training/validation sets (7:3 ratio) with seed 4
random_split_seed5.pt	Random splits of observation data into training/validation sets (7:3 ratio) with seed 5
temporal_split.pt	Time-based split: first 70% of years for training, remaining 30% for validation
spatial_split.pt	Geography-based split across 5 ocean regions (Atlantic, Pacific, Indian, Polar, Enclosed Seas) in a 7:3 ratio

All index files are stored as PyTorch tensors (.pt). They identify which spatiotemporal locations have observations and are used to construct MNAR (Missing Not at Random) masks.

Graph (*_graph/)

Content	File Format	Purpose
Graph-structured data (one file per year)	.pt (PyTorch tensor)	Graph input for the model, containing spatiotemporal adjacency information

Ground Truth (*_ground_truth/)

Model	Filename	File Format	Description
CESM2-OMIP1	OMIP1_do.npy	NumPy array (.npy)	Dissolved oxygen ground truth
CESM2-OMIP2	OMIP2_do.npy	NumPy array (.npy)	Dissolved oxygen ground truth
GFDL-ESM4	GFDL_do.npy	NumPy array (.npy)	Dissolved oxygen ground truth

Dataset Description

The datasets are used to construct a virtual Earth with known dynamical behavior from the widely recognized numerical simulation ensemble CMIP6 (Coupled Model Inter-comparison Project Phase 6). In this study, the time span for CESM2-omip1 is from 1948 to 2009, for CESM2-omip2 it is from 1958 to 2018, and for GFDL-ESM4 it is from 1920 to 2014. The depth level is divided into33 layers ranging from 0 to 5500 meters.

This is the detailed information:

Model Name	CESM2-OMIP1	CESM2-OMIP2	GFDL-ESM4
Developing Institution	National Center for Atmospheric Research (NCAR)	National Center for Atmospheric Research (NCAR)	Geophysical Fluid Dynamics Laboratory (GFDL)
Spatial Range	Global scale	Global scale	Global scale
Spatial Resolution	1° × 1°	1° × 1°	1° × 1°
Temporal Range	1948-2009	1958-2018	1920-2014
Temporal Resolution	Yearly output	Yearly output	Yearly output
Simulation Assumptions	Fixed greenhouse gas concentrations or specific scenarios (e.g., RCP8.5)	Fixed greenhouse gas concentrations or specific scenarios (e.g., RCP8.5)	Fixed greenhouse gas concentrations or specific scenarios (e.g., SSP5-8.5)
Simulation Conditions	Global climate change scenarios, focusing on ocean carbon cycle and ecosystems	Global climate change scenarios, focusing on ocean carbon cycle and ecosystems	Global climate change scenarios, focusing on carbon cycle, ocean acidification, and ecosystem processes

Dataset Split

We partitioned the observational data into training and validation sets based on three distinct methodologies: random, time, and space.

1.Random Split

The training and validation sets are randomly divided in a 7:3 ratio, while the test data spans all years. We employed fixed random seeds to ensure the reproducibility of the dataset partitioning process:

def set_random_seed(seed):
    torch.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    np.random.seed(seed)
    random.seed(seed)
    torch.backends.cudnn.deterministic = True

2.Temporal Split

The observation data is divided into training and validation sets based on time. The first 70% of the years are used for training, while the remaining 30% are used for validation. Specifically, for the CESM2-omip1 dataset, the training period is from 1948 to 1991 (the first 44 years), and the validation period is from 1992 to 2009. For CESM2-omip2, the training period is from 1958 to 2000 (the first 43 years), and the validation period is from 2001 to 2018. For GFDL-ESM4, the training period is from 1920 to 1986 (the first 67 years), and the validation period is from 1987 to 2014. During the model testing phase, the evaluation focuses on the reconstruction results for all unobserved regions, reflecting the model's spatio-temporal global reconstruction performance.

3.Spatial Split

Based on the range mask from WOD (World Ocean Database), the global ocean is divided into five regions: the Atlantic Ocean, the Pacific Ocean, the Indian Ocean, the Polar Oceans, and the enclosed seas. The training and validation sets are then allocated in a 7:3 ratio according to these spatial regions.

Ocean Regions	Train	Validation
Atlantic Ocean	1. North Atlantic	1. Equatorial Atlant
	2. Coastal N Atlantic
	3. South Atlantic
	4. Coastal S Atlantic
Pacific Ocean	5. North Pacific	3. Equatorial Pac
	6. Coastal N Pac	4. Coastal Eq Pac
	7. South Pacific
	8. Coastal S Pac
Indian Ocean	9. North Indian	5. Equatorial Indian
	10. Coastal N Indian	6. Coastal Eq Indian
	11. South Indian
	12. Coastal S Indian
Polar Oceans	13. Arctic	7. Antarctic
Enclosed Seas	14. Baltic Sea	8. Mediterranean
	15. Red Sea	9. Black Sea
		10. Persian Gulf
		11. Sulu Sea

Evaluation

Our evaluation metrics are computed separately for the target variable (i.e.,dissolved oxygen) in the output vector. The ground truth is utilized to evaluate the output vector. Specifically, the Root Mean Square Error (RMSE) and the coefficient of determination (R²) are computed independently for each horizontal and vertical location.