A new model for understanding the ocean
DARWIN uses 26 years of data to help monitor marine environments.
Researchers have created DARWIN, a geospatial foundation model for marine intelligence. This model is designed to work with data from the ocean.
The ocean covers most of the Earth. It helps regulate the global climate, store carbon, and support biodiversity. However, humans understand parts of the deep sea less than the surface of Mars. Traditional models for ocean analysis often struggle because labeled datasets are rare and expensive to collect. Most existing models also lack the ability to work across different marine areas.
To build DARWIN, researchers used 26 years of global Copernicus reanalysis data and GEBCO bathymetry. They used a masked-autoencoder backbone. This backbone was modified with missingness-aware encoding, expanded spatial and temporal masking, and Fourier-style positional embeddings.
In tests for predicting chlorophyll-a, DARWIN achieved the lowest monthly relative error for every month in 2022. It performed better than a random-init encoder with the same architecture. It also performed better than baselines with no encoder. When using only 10% of supervised training data, DARWIN surpassed a random-init model trained with 100% of the data.
A foundation model is a large model trained on massive datasets. These models can be adapted to different tasks through fine-tuning or few-shot/zero-shot learning. Self-supervised learning is a method where models learn from unlabelled data. This approach uses the internal structures of the data to learn useful representations.
Analysis shows that DARWIN captures coherent oceanographic structure instead of just memorizing input statistics. This provides a scalable backbone for tasks like species distribution modelling, environmental monitoring, and anomaly detection.
Key facts
- DARWIN was trained on 26 years of global Copernicus reanalysis and GEBCO bathymetry.
- The model achieved the lowest monthly relative error for chlorophyll-a in every month of 2022.
- DARWIN outperformed a random-init model using only 10% of supervised training data.