Cesar Quilodran, Imperial College London

Fast data assimilation for improved ocean weather forecasting over the North Brazil Current

Coauthors
Ralf Toumi

Abstract:

Atmosphere and ocean coupled models are characterised by their high dimensionality and nonlinearity. The application of data assimilation techniques over these models, in order to improve their forecasting, can be tempting but, unfortunately, very computationally expensive. This research utilises a procedure based in (Frolov et al., 2009) where a Kalman Filter is applied onto a reduced dimensionality space, i.e. a projection of the full space coupled model. The area of study is the North Brazil Current. Daily inputs are obtained from a 10 km resolution hindcast ROMS model (full space FS), during the period of 2001-2010. The procedure consists of two stages: offline and online. During the offline stage, an Empirical Orthogonal Function (EOF) analysis is applied on FS to reduce its high dimensionality. Since this is a rather memory consuming task, the grid is divided into sub grids where the EOF analysis is performed individually. The compound of these sub problems creates the reduced space (RS). An adequate amount of EOFs are retained. Here, the first 50 EOFs hold ~85% of the variability when sea surface temperature (SST), sea surface height (SSH), and U and V current components are considered. The retained variability ranges from 79 to 99.8% when using different combinations of these four variables. In order to create short-range forecasts, the RS is trained using an autoregressive artificial neural network (ANN). The sensibility of the ANN will strongly depend, but not exclusively, on the selected combination of the four variables. An Ensemble Kalman Filter is utilised during the online stage, where the ANN provides short-range forecasts from the ensembles. Observational data is obtained from GHRSST, AVISO, and OSCAR to assimilate and validate SST, SSH and currents, respectively. In general, it is possible to assimilate 30 days in less than 30 minutes, without the use of a “supercomputer”. Forecasting by using this methodology is yet to be assessed, but promising

File available here