Abstract
Sea Surface Temperature (SST) retrieved from satellites is a crucial variable for weather prediction and climate monitoring, yet satellite-based SST observations are often hindered by orbital constraints and cloud cover-especially in regions like the North and Baltic Seas where high-resolution data is essential to support operational monitoring and forecasting. Existing SST reanalysis products frequently lack the spatial and temporal granularity needed for these dynamic coastal areas. To address this, we explore an advanced data-driven approach to SST reconstruction using 4DVarNet, an end-to-end deep learning framework that integrates variational data assimilation with machine learning. While 4DVarNet has shown promising results in reconstructing SST fields, it lacks a robust mechanism for uncertainty quantification. We enhance the framework by incorporating a self-attention embedded variational autoencoder as a prior model, enabling probabilistic reconstruction and efficient sampling. The overall scheme includes a multiscale neural variational approach, self-supervised training, and the integration of pre-trained generative models. The proposed method demonstrates improved accuracy, resolving smaller spatial scales down to 33-45 km versus 59-69 km with current operational capabilities, faster convergence, and better representation of small-scale ocean features, offering a realistic path toward operational deployment of neural variational data assimilation techniques for high-resolution SST analysis in challenging regions.