Use this URL to cite or link to this record in EThOS:
Title: Fast ocean data assimilation and forecasting using a neural-network reduced-space regional ocean model of the north Brazil current
Author: Quilodra´n Casas, Ce´sar
ISNI:       0000 0004 7658 4540
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Access from Institution:
Data assimilation is computationally demanding, typically many times slower than model forecasts. Fast and reliable ocean assimilation methods are attractive for multiple applications such as emergency situations, search and rescue, and oil spills. A novel framework which performs fast data assimilation with sufficient accuracy is proposed for the first time for the open ocean. Speed improvement is achieved by performing the data assimilation on a reduced-space rather than on a full-space. A surface 10km resolution hindcast of the North Brazil current from the Regional Ocean Modelling System (ROMS) serves as the full-space state. The target variables are sea surface height, sea surface temperature, and surface currents. A dimension reduction of the full-state is made by an Empirical Orthogonal Function analysis while retaining most of the explained variance. The dynamics are replicated by a state-of-the-art neural network trained on the truncated principal components of the full-state. An Ensemble Kalman filter assimilates the data in the reduced-space, where the trained neural network produces short-range forecasts from perturbed ensembles. The Ensemble Kalman filter of the reduced-space is successful in reducing the root mean squared error by ∼ 45% and increases the correlations between state variables and data. The performance is similar to other full-space data assimilation studies. However, the computations are three to four orders of magnitude faster than for other full-space data assimilation schemes. The forecast of ocean variables is a computationally demanding task in terms of speed and accuracy. This framework manages to create fast forecasts in ∼ 30 seconds, once data have been assimilated. The forecasts are obtained using the trained neural network. We performed additional experiments using data and forecasts from July 2015 and January 2016. The analysis and forecasts in our framework yield a higher skill score and high spatial correlation when compared to the operational dataset Global Ocean Physics Analysis and Forecast by the UK MetOffice. Forcing the neural network with 10 m surface winds in order to improve the total surface currents forecast was considered. There is no additional skill in the forecasts using wind forcing because of the low Ekman component compared to the dominant geostrophic currents. The reduced model approach could be a useful tool when full physics regional models are not available to make a forecast.
Supervisor: Toumi, Ralf Sponsor: Comision Nacional de Investigación Científica y Tecnologica (Chile)
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral