A foundation model built for the ocean: DARWIN beats a random-init encoder trained on all labels using just 10% of them
Synopsis
Earth-observation foundation models have mostly been trained on land, while ocean data are sparse, multi-modal and non-stationary, with no standard benchmarks. DARWIN adapts a masked autoencoder to marine conditions and is pretrained on 26 years of Copernicus reanalysis plus GEBCO bathymetry; on 2022 chlorophyll-a prediction it attains the lowest monthly relative error in every month, with a yearly mean of 13.4% against 14.5% for the best no-encoder baseline and 16.5% for an identically architected random-init encoder. With only 10% of supervised training data, pretrained DARWIN's test MAE is already below that of the random-init model trained on 100%.
Interpretation
The marine domain now has a purpose-pretrained geospatial foundation model, DARWIN, which brings yearly mean relative error on chlorophyll-a prediction to 13.4%, below the 14.5% of the best no-encoder baseline and the 16.5% of an identically architected random-init encoder. Earlier Earth-observation foundation models were developed and benchmarked mainly on land, the ocean lacks standardised benchmarks, and existing ocean-colour models treat satellite imagery as static snapshots, ignoring temporal dynamics and omitting physical and biogeochemical variables. On the independent 2022 test year, pretrained DARWIN has the lowest monthly relative error in all twelve months, the random-init encoder the highest, and the no-encoder baseline in between.
The model extends the masked autoencoder to ocean data by feeding each pixel a missingness indicator and a land indicator, and by adding whole-month masking and spatial column masking so the model must infer hidden parts from seasonal continuity and surrounding water. A standard masked autoencoder uses only random masking, and reconstruction can be dominated by interpolation from neighbouring pixels, which does not teach variable interdependencies or temporal dynamics. Pretraining uses 47 variables from 312 monthly steps of Copernicus L4 reanalysis for 1998 to 2023 plus GEBCO bathymetry, with samples as 360 by 360 km spatiotemporal chips spanning 12 months.
The pretrained weights pay off most under low labels: with 10% of training data DARWIN's test MAE is below that of the random-init model trained on 100%, and at 0.1% (871 training samples) it still beats every compared model. The random-init encoder needs the full label set to reach its own best level, and the no-encoder deep-learning baselines trail pretrained DARWIN at every data fraction. The data-efficiency experiment retrains the pretrained encoder, the random-init encoder and the best no-encoder baselines on nested training subsets while validation and test sets stay fixed.
Perspective
The demonstrated scope is dense chlorophyll-a prediction over the global ocean on an 8 km grid at monthly means, with 1998 to 2021 used for training and validation and the whole of 2022 held out as an independent test set. Pretraining inputs are surface variables from Copernicus L4 reanalysis and GEBCO bathymetry, with no subsurface observations. The authors propose next steps of incorporating direct satellite observations and in-situ measurements into pretraining and systematically building downstream task benchmarks.
The authors list three qualifications: the 8 km grid and monthly cadence constrain the resolvable phenomena and likely underlie the comparatively weaker temporal reconstructions during pretraining; the MAE objective spends capacity on pixel-level reconstruction that may not match the abstraction useful downstream; and downstream validation covers only a single surface variable. They also note that pretraining relies on L4 reanalysis products, so the model learns what operational systems estimate the ocean to be rather than what sensors directly observe.
