Multiscale neural assimilation scheme for high-resolution sea surface temperature reconstruction from satellite...

Area of Science:

• Oceanography
• Climate Science
• Data Science

Background:

• Satellite-derived Sea Surface Temperature (SST) is vital for weather and climate monitoring.
• Orbital constraints and cloud cover limit satellite SST data, especially in dynamic coastal regions like the North and Baltic Seas.
• Existing reanalysis products lack the necessary spatial and temporal detail for these areas.

Purpose of the Study:

• To develop an advanced data-driven approach for high-resolution SST reconstruction.
• To enhance the 4DVarNet deep learning framework for improved SST analysis.
• To incorporate uncertainty quantification into SST reconstruction.

Main Methods:

• Utilized 4DVarNet, an end-to-end deep learning framework integrating variational data assimilation and machine learning.
• Enhanced 4DVarNet with a self-attention embedded variational autoencoder for probabilistic reconstruction and uncertainty quantification.
• Employed a multiscale neural variational approach, self-supervised training, and pre-trained generative models.

Main Results:

• Achieved improved accuracy in SST reconstruction, resolving smaller spatial scales (33-45 km) compared to current operational methods (59-69 km).
• Demonstrated faster convergence and better representation of small-scale ocean features.
• Enabled probabilistic reconstruction and efficient sampling for uncertainty quantification.

Conclusions:

• The proposed method offers a realistic pathway for the operational deployment of neural variational data assimilation for high-resolution SST analysis.
• The enhanced framework provides more accurate and detailed SST data for challenging coastal regions.
• This advancement supports improved operational monitoring and forecasting in critical marine environments.