Beauchamp et al. (2025) Multiscale neural assimilation scheme for high-resolution sea surface temperature reconstruction from satellite observations

Identification

• Journal: Scientific Reports
• Year: 2025
• Date: 2025-11-14
• Authors: Maxime Beauchamp, Ioanna Karagali, Guisella Gacitúa, Jacob L. Høyer, Maxime Ballarotta, Ronan Fablet
• DOI: 10.1038/s41598-025-23682-9

Research Groups

• Danish Meteorological Institute, Copenhagen, Denmark
• IMT Atlantique, Plouzané, France
• CLS (Collecte Localisation Satellites), Ramonville-Saint-Agne, France

Short Summary

This study develops an enhanced multiscale neural assimilation scheme, 4DVarNet, incorporating a variational autoencoder (VAE) for probabilistic sea surface temperature (SST) reconstruction in the North and Baltic Seas. The method significantly improves accuracy and resolves finer spatial scales (33–45 km) compared to traditional optimal interpolation (59–69 km), while also providing robust uncertainty quantification.

Objective

• To address the lack of spatial and temporal granularity in existing SST reanalysis products for dynamic coastal areas like the North and Baltic Seas.
• To enhance the 4DVarNet deep learning framework by incorporating a self-attention embedded variational autoencoder (VAE) as a prior model, enabling probabilistic reconstruction and efficient uncertainty quantification for high-resolution SST analysis.

Study Configuration

• Spatial Scale: North and Baltic Seas, covering a domain for which the DMI-OI product has a 0.02° × 0.02° (approximately 2.2 km × 2.2 km at 50°N) resolution. The proposed method resolves spatial scales down to 33–45 km. Training uses random patches of 240 × 240 pixels, and inference uses overlapping patches cropped to 200 × 200 pixels.
• Temporal Scale: Daily outputs for the entire year 2021 (test dataset). Data assimilation windows of 7 days. Training dataset spans less than 2 years (April 2019 to December 2020).

Methodology and Data

• Models used:
  • 4DVarNet: An end-to-end deep learning framework integrating variational data assimilation with machine learning.
  • Enhanced 4DVarNet: Incorporates a self-attention embedded variational autoencoder (VAE) as a prior model for probabilistic reconstruction.
  • Multiscale neural variational approach: Combines a coarse approximation (0.2° × 0.2°) with high-resolution anomaly reconstruction (0.02° × 0.02°).
  • UNet-like parameterization for the prior operator and Long Short-Term Memory (LSTM) network for the solver.
  • DMI Optimal Interpolation (DMI-OI) as a baseline for comparison.
• Data sources:
  • Satellite observations:
    • Copernicus Marine Service (CMEMS) Near-Real-Time (NRT) Level 4 (L4) SST product (SSTBALSSTL3SNRTOBSERVATIONS010007b12) at 0.02° × 0.02° resolution.
    • Level-2 Pre-processed (L2P) and Level-3 Uncollated (L3U) satellite data from AVHRR, SEVIRI, VIIRS, and SLSTR (used as input for DMI-OI).
    • Sentinel 3 SLSTR SST (independent satellite dataset for validation).
  • In-situ observations: Met Office Hadley Centre Integrated Ocean Database (HadIOD) version 1.2.0.05, including drifting buoys, ship data, mooring buoys, and Argo data, collected at night time for the year 2021.
  • Covariates: Standardized longitude and latitude, binary land mask, error variance of the DMI-OI first-guess field, and high-resolution bathymetry from EMODnet Digital Bathymetry (1/16 arc-minute resolution).

Main Results

• Improved Accuracy: The 4DVarNet scheme reduced the Root Mean Square Error (RMSE) from 0.43 K (DMI-OI) to 0.41 K for the full North and Baltic Seas domain over 2021.
• Enhanced Spatial Resolution: The proposed method resolved smaller spatial scales down to 33–45 km, significantly outperforming DMI-OI which resolved scales of 59–69 km.
• Uncertainty Quantification: The VAE-based extension enabled efficient posterior sampling and uncertainty modeling, yielding Continuous Ranked Probability Score (CRPS) values typically between 0.1–0.2, compared to approximately 0.4 for DMI-OI.
• Performance Consistency: Demonstrated consistent improvements across seasonal and regional analyses, particularly strong in spring and summer.
• Computational Efficiency: Generating the 2021 test year of daily data took approximately 15 minutes on a single NVIDIA A100 GPU. Training the model for 500 epochs took approximately 8 hours on the same GPU.

Contributions

• Development of a neural variational scheme that leverages multiscale inputs for high-resolution SST reconstruction in the North and Baltic Seas.
• Implementation of a self-supervised training procedure involving random removal of observational patches to effectively simulate realistic data gaps.
• Integration of a pretrained generative prior model (Variational Autoencoder) within the 4DVarNet framework, enabling efficient posterior sampling and robust uncertainty modeling.
• Demonstrates a scalable and computationally efficient pathway toward operational, uncertainty-aware SST products tailored for dynamic coastal and high-variability regions.

Funding

• French ANR OceaniX (ANR-19-CHIA-0016)
• CPER AIDA GPU cluster (supported by the Regional Council of Brittany, Brest Métropole, and Fonds Européen de DEveloppement Régional (FEDER))
• EU Horizon Europe projects EDITO Model Lab (Grant 101093293)
• EU Horizon Europe projects AI4PEX (Grant 101137682)
• GENCI-IDRIS (Grant 2021-101030)

Citation

@article{Beauchamp2025Multiscale,
  author = {Beauchamp, Maxime and Karagali, Ioanna and Gacitúa, Guisella and Høyer, Jacob L. and Ballarotta, Maxime and Fablet, Ronan},
  title = {Multiscale neural assimilation scheme for high-resolution sea surface temperature reconstruction from satellite observations},
  journal = {Scientific Reports},
  year = {2025},
  doi = {10.1038/s41598-025-23682-9},
  url = {https://doi.org/10.1038/s41598-025-23682-9}
}