Lagrangian and Coupled Data Assimilation enhanced by Machine Learning to improve Operational Ocean Prediction
PI: Carton, James (formerly Penny, Stephen) (University of Maryland)
Co-PI(s): Abarbanel, Henry (University of California San Diego), Cornuelle, Bruce (University of California San Diego), Hogan, Patrick (NRL-SSC)
Start Year: 2019 | Duration: 3 years
Partners: UCSD, NRL-SSC
We aim to improve high resolution (1/25º) ocean state estimation at the sea surface, including the positioning of fronts and eddies, the corresponding upper ocean currents, and air-sea fluxes. To achieve this goal, two essential guiding principles are to: (1) utilize new observational data types that are typically ignored for operational ocean data assimilation, and (2) reduce the computational costs of cycled ocean data assimilation. New observing platforms are producing ever more data that need to be assimilated, including altimetry (at upcoming SWOT resolutions), high-resolution SST, ocean color images from VIIRS, and SAR images. These can and should be incorporated into the data assimilation process, with reasonable computational costs. We use the Navy’s NCODA file conventions to permit use with both the global HYCOM and regional coupled COAMPS modeling systems.
We begin by using a combination of in situ and satellite-based data that track the near surface flow fields, while simultaneously assimilating conventional ocean observation data. The PI has developed a Lagrangian data assimilation (LaDA) approach compatible with the Local Ensemble Transform Kalman Filter (LETKF) that extends the operational ocean data assimilation system developed for NCEP (Penny et al., 2015; Penny 2017). Sun and Penny (2019) assimilated the position data of surface drifters and found that the estimated forecast error covariance between drifter positions and prognostic model state variables such as temperature and salinity were sufficiently accurate to reduce errors in surface ocean currents and also the temperature, salinity, and kinetic energy down to 1000m depth. By using image processing algorithms, satellite imagery of the ocean surface can be classified and mapped with key features that track similarly to surface drifters (e.g. using ocean color). These points can be verified against the model using a virtual ‘grid’ of simulated drifters corresponding to features that move with the surface flow, allowing LaDA to be applied.
Sluka et al. (2016) showed that by using strongly coupled data assimilation (SCDA), atmospheric observations can be used to constrain ocean state estimates at all depths, while also improving surface fluxes. New multiscale DA methods are required to isolate temporal and spatial scales in the analysis to enhance predictive skill in the initialized coupled forecasts. We plan to explore the impacts and limitations of SCDA at the air-sea interface at high resolution by first implementing the SCDA version of the LETKF with the Navy’s regional COAMPS.
We attempt to develop low-cost proxy models using machine learning (ML) methods in order to make high resolution ensembles viable. We will explore the utilization of long offline model runs and reforecasts to reduce the computational burden in online state estimation. A key element in this investigation is to determine whether the ML models display similar dynamical error characteristics so that they can adequately serve as a proxy for the high-resolution numerical model used in the DA cycle. We will then investigate the use of ML to generate low-cost large ensemble forecasts to complement the high-resolution deterministic model integration. We go further to apply ML methods throughout the DA cycle, including: (1) the ensemble forecast, (2) model bias correction, (3) the observation operator, and (4) the analysis method.
Anticipated Outcome & Impact
We anticipate a working prototype ocean/coupled data assimilation system that functions similarly to the SCDA-LETKF system but at reduced computational cost due to use of ML-based ensembles, with the additional capability to assimilate position information from drifters and tracking of images of the ocean surface. We expect this project to accelerate the capabilities of high resolution operational ocean data assimilation beyond any methods currently available today.
BAA Topic: Topic 6: New Approaches for Data Assimilation to Improve Operational Ocean Prediction