Sea surface temperature reconstruction has traditionally relied on deterministic methods that yield a single field without uncertainty estimates. Satellites cover nearly the entire ocean but have systematic errors in regions of strong currents; buoys are accurate but sparse. Existing machine learning approaches address gap-filling in satellite data and weather forecasting, but the problem of probabilistic global SST reconstruction with explicit handling of heterogeneous sources remained open. SAGE fits within the generative methods trend but focuses specifically on uncertainty and the fusion of satellite and in situ measurements.
SAGE uses a diffusion model trained on 39 years of OISST data (1982–2020) at 0.75°×0.9375° resolution. Training ran on four NVIDIA RTX 4090 GPUs for 400,000 steps. During reverse diffusion, each step computes a gradient of consistency with observations that guides the sampling. The progressive strategy uses FY-3D MWRI satellite data for the first 500 steps to set large-scale structure, and HadIOD in situ measurements for the last 500 steps to refine local details. The guidance strength parameter $ζ_t$ was tuned separately for each configuration. For forecasts, a Vision Transformer was used for 10-day SST prediction, and the coupled climate model ICM with SST nudging for seasonal forecasts.
Against independent HadIOD observations, the global root mean square error (RMSE) was 0.17°C for SAGE(Sat+Hadi), versus 0.59°C for OISST and 0.72°C for satellite-only SAGE. Ensemble uncertainty is higher in dynamically active regions (western boundary currents) and in areas with sparse observations. In 10-day SST forecasting, SAGE initialization reduced RMSE by 4.90% relative to OISST and outperformed the operational GIOPS system. In the seasonal forecast of the 2023 El Niño, the Niño 3.4 index correlation rose from 0.24 to 0.81, and RMSE dropped from 0.74°C to 0.30°C. Reconstruction of the 2023 marine heatwaves agrees well with OISST, with typical intensity differences within 0.1°C.
SAGE differs from related work by framing SST reconstruction as a Bayesian inference problem with explicit uncertainty modeling, rather than deterministic gap-filling. Unlike methods that simply interpolate satellite data, SAGE merges heterogeneous sources in a unified probabilistic framework and demonstrates measurable gains in forecasts. This is a step toward a new generation of ocean state estimation systems where initial condition ensembles become the norm, not the exception.
Extending SAGE to multivariate fields (salinity, currents) and higher resolution will require larger models and more computational resources. Latent diffusion models could reduce costs. Further research should test how improved reconstruction of SST extremes affects heat fluxes, atmospheric response, and coupled predictability. Evaluation over longer retrospective periods and across different climate regimes will strengthen conclusions.
This work will influence operational ocean monitoring, especially for tracking marine heatwaves, and weather and climate forecasting systems that use ensemble initial conditions.
Conduct multi-year retrospective forecasts to assess the robustness of gains; extend the model to multivariate ocean fields and higher resolution; integrate additional observation types (Argo, satellite salinity).
A key unresolved problem in oceanography is estimating the uncertainty of reconstructions from heterogeneous observations and its impact on predictability. SAGE directly addresses this by offering a probabilistic alternative to deterministic products and showing that accounting for uncertainty improves forecasts.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
This paper introduces SAGE, a diffusion-based generative model for probabilistic reconstruction of sea surface temperature. It is trained on historical OISST data and then conditioned on satellite and in-situ measurements using a progressive fusion strategy: first a coarse scale from satellite data, then local details from buoys. Validation on independent data shows RMSE decreasing from 0.59°C to 0.17°C, and initializing forecasts improves 10-day SST forecasts and seasonal El Niño predictions.
Where its strength is
SAGE combines heterogeneous observational sources within a unified probabilistic framework, explicitly modeling uncertainty. This yields not just a reconstruction but an ensemble of fields that can be used to improve forecasts.
Where it could go next
Worth trying
Why this may matter
Probabilistic SST reconstruction with explicit uncertainty handling could serve as a foundation for more reliable climate forecasts and early warning systems for marine heatwaves.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.53, against a usual 0.64. It looks like few have drilled nearby.
In 2023, global marine heatwave activity exceeded the 1982 average by more than three standard deviations.