DARWIN: a Geospatial Foundation Model for Marine Intelligence
Organizations: Ocean OS
Abstract
We present DARWIN, a geospatial foundation model for marine intelligence. Recent Earth-observation foundation models have concentrated on terrestrial domains. The marine realm poses distinct challenges: missingness, multi-modal physical and biogeochemical variables, sparse labels, and dynamics that are transient, multi-scale, and non stationary over time. DARWIN adapts a masked-autoencoder backbone to these properties through missingness-aware encoding, expanded spatial and temporal masking, and Fourier-style positional embeddings, and is pretrained on 26 years of global Copernicus reanalysis together with GEBCO bathymetry. On predicting chlorophyll-a, the pretrained backbone attains the lowest monthly relative error in every month of 2022, outperforming both no-encoder baselines and an identically architected random-init encoder. At 10% of the supervised training data, DARWIN already surpasses the random-init model trained on 100% of the data. Embedding analysis and qualitative reconstructions show that DARWIN captures coherent oceanographic structure rather than memorising input statistics, providing a scalable backbone for downstream marine applications such as anomaly detection, species distribution modelling, and environmental monitoring.
Figures & tables
| Category | Product Name | Processing Level |
|---|---|---|
| Physics | Physics Reanalysis | L4 (Reanalysis) |
| Biogeochemistry | Biogeochemistry Hindcast | L4 (Hindcast) |
| Biomass | Low and Mid Trophic Levels Biomass | L4 (Hindcast) |
| Ocean Colour | Colour (Copernicus-GlobColour) | L4 (Interpolated) |
| Wind | Monthly Mean Sea Surface Wind and Stress | L4 (Satellite-Model) |
| Train fraction | Train | Val | Test |
|---|---|---|---|
| (100 %) | |||
| (10 %) | |||
| (1 %) | |||
| (0.1 %) |
Appendix figures & tables24 assets
Supplementary material from the paper’s appendix.
Appendix
| Variable | Description | Unit |
| Physics | ||
| Temperature | Sea Water Potential Temperature | [°C] |
| Salinity | Sea Water Salinity | [1e-3] |
| Sea surface height | Sea Surface Height Above Geoid | [m] |
| Velocity (zonal) | Eastward Sea Water Velocity | [m s -1 ] |
| Velocity (meridional) | Northward Sea Water Velocity | [m s -1 ] |
| Category | Variables |
|---|---|
| Included Static (5) | ocean_mask , min_bathy , max_bathy , mean_bath , std_bath |
| Category | Variables |
|---|---|
| Included Dynamic (26) | bottomT , thetao , so , zos , uo , vo , usi , vsi , eastward_wind , northward_wind , eastward_stress , northward_stress , wind_stress_magnitude , wind_speed , o2 , ph , sithick , siconc , mlotst , fe , no3 , po4 , si , spco2 , KD490 , SPM |
| Category | Variables |
|---|---|
| Omitted Dynamic (21) | DIATO , DINO , GREEN , HAPTO , MICRO , NANO , PICO , PROCHLO , PROKAR , chl , phyc , npp , nppv , zeu , zooc , mnkc_epi , mnkc_hmlmeso , mnkc_lmeso , mnkc_mlmeso , mnkc_mumeso , mnkc_umeso |
| Hyperparameter | Value |
|---|---|
| Optimiser | AdamW |
| Peak learning rate | |
| Minimum learning rate | |
| Adam | |
| Weight decay | |
| Learning rate schedule | Linear warmup, cosine decay |
| Hyperparameter | DARWIN | DARWIN-NoColour |
|---|---|---|
| Dynamic input features | ||
| Static input features | ||
| Chip spatial size | ||
| Chip timesteps | ||
| Patch size (space space time) | ||
| Values per patch embedding |
| Setting | Value |
|---|---|
| Random masking ratio | |
| Temporal-slice masking ratio | |
| Spatial-tube masking ratio | |
| Dropout, absolute spatial positional enc. | |
| Dropout, scale positional enc. | |
| Dropout, absolute temporal positional enc. |
| Hyperparameter | Value |
|---|---|
| Optimiser | AdamW |
| Peak learning rate | |
| Minimum learning rate | |
| Adam | |
| Weight decay | |
| Learning rate schedule | Linear warmup, cosine decay |
| Hyperparameter | Linear | MLP | Attention head | DARWIN decoder |
|---|---|---|---|---|
| Hidden dimension | — | |||
| Depth (transformer blocks) | — | — | ||
| Attention heads | — | — | ||
| MLP expansion ratio | — |
| Model | Full MAE | Encoder only |
|---|---|---|
| DARWIN FM (with colour) | 34.8 M | 26.7 M |
| DARWIN FM (No-Colour) | 7.6 M | 6.7 M |
| Head | Backbone | Total | Trainable |
|---|---|---|---|
| Encoder-based heads (config_random_init / config_pretrained_init ) | |||
| Linear probe ( mode: linear ) | frozen | 6.72 M | 0.00231 M |
| Linear probe ( mode: linear ) | unfrozen | 6.72 M | 6.72 M |
| 2-layer MLP ( mode: mlp , hidden 512) | frozen | 6.85 M | 0.136 M |
| 2-layer MLP ( mode: mlp , hidden 512) | unfrozen | 6.85 M | 6.85 M |
| Vanilla attention head ( mode: attention_head , depth 1) | frozen | 7.57 M | 0.858 M |
| Hyperparameter | XGBoost | LightGBM |
|---|---|---|
| n_estimators | 1875 | 1701 |
| learning_rate | 0.0899 | 0.2011 |
| max_depth | 8 | 6 |
| min_child_weight / samples | 29 | 31 |
| subsample | 0.898 | 0.647 |
| colsample_bytree | 0.850 | 0.821 |
Geospatial Foundation Models
Geospatial foundation models turn satellite, map, and Earth-system data into reusable representations for spatial analysis and prediction.
See all 135 papers