Multimodal 4DVarNets for the reconstruction of sea surface dynamics from SST-SSH synergies

Abstract

The space-time reconstruction of ocean dynamics from satellite observations is a challenging inverse problem due to the associated irregular sampling of the sea surface. Satellite altimetry provides a direct observation of the sea surface height (SSH), which relates to the divergence-free component of sea surface currents. The associated sampling pattern prevents from retrieving fine-scale dynamics, typically below 10 days. By contrast, other satellite sensors provide higher-resolution observations of sea surface tracers such as sea surface temperature (SST). Multimodal inversion schemes then arise as appealing approaches. Though theoretical evidence supports the existence of an explicit relationship between sea surface temperature and sea surface dynamics under specific dynamical regimes, the generalization to the variety of upper ocean dynamical regimes is complex. Here, we investigate this issue from a physics-informed learning perspective. We introduce a trainable multimodal inversion scheme for the reconstruction of sea surface dynamics from multi-source satellite-derived observations, namely satellite-derived SSH and SST data. The proposed 4DVarNet schemes combine a variational formulation involving trainable observation and a priori terms with a trainable gradient-based solver. An observing system simulation experiment for a Gulf Stream region supports the relevance of our approach compared with state-of-the-art schemes. We report a relative improvement greater than 60% compared with the operational altimetry product in terms of root mean square error and resolved space-time scales. We discuss further the application and extension of the proposed approach for the reconstruction and forecasting of geophysical dynamics from irregularly-sampled satellite observations.

End-to-End Neural Interpolation of Satellite-Derived Sea Surface Suspended Sediment Concentrations

“…The evaluation of the interpolation metrics for the gradient of the SSSC fields in Table 5 supports the hypothesis that the improvement obtained with 4DVarNet relates to a better reconstruction of fine-scale patterns. Previous work with similar 4DVar based architecture shows a significant improvement of a high resolution spatial pattern [44]. Surprisingly, the metrics are much better for the OSE.…”

Machine learning with data assimilation and uncertainty quantification for dynamical systems: a review

“…This also applies to the shift from general-purpose DA pipelines to application-centric ones optimized for specific observing systems, states and/or diagnosis variables. Beyond applications on toy examples, recent demonstrations for the reconstruction of sea surface dynamics from satellite-derived observations [207] support the relevance of these schemes to advance the state-of-the-art for real DA problems. A key challenge is their application to complex spatial-temporal DA problems currently solved by operational DA systems in climate simulation, operational oceanography and weather forecast.…”

“…A key challenge is their application to complex spatial-temporal DA problems currently solved by operational DA systems in climate simulation, operational oceanography and weather forecast. In such contexts, we may emphasise the great flexibility in terms of state definition and model parameterisation opened by the end-to-end learning framework, including for instance augmented state [208], multimodal formulation [207] and uncertainty representation [209].…”

Surrogate modeling for the climate sciences dynamics with machine learning and data assimilation

“…A radically different approach consists in optimizing the internal cogs of DA schemes by finding an efficient latent space representation for either the ensemble Kalman filter [91][92][93][94] or the variational schemes [95]. One can also try to learn the solvers of, for instance, variational schemes since these are critical to the DA analysis and its numerical efficiency and cost [96,97]. Beyond supplementing ML techniques to DA schemes as was originally proposed, one may further replace key components of the DA scheme with advanced NNs such as transformers and multi-headed attention [98].…”