Data Assimilation

Data assimilation (DA) is a method that is employed to merge the solution from a numerical model, with available observational data to produce a 3-D time varying fields. The output from DA is considered to the “best estimate” of the system state, where model uncertainty is reduced, and sparse observation data is dynamically interpolated using the numerical model. Within the context of eReefs, DA has been applied to both the marine hydrodynamic and biogeochemical (BGC) models. It has been used for both parameter estimation in the hydrodynamic model, and also state updating in the hydrodynamics and BGC.

Observations of temperature from remote sensing (Sea surface Temperature), moorings and gliders, have been assimilated into the hydrodynamic model. Two approaches were used.  The first assimilation approach was to constrain three parameters relating to the transmission, attenuation and bottom reflectance of short wave radiation. These parameters vary according to water clarity, bottom type and errors in surface heat fluxes.  By including these variables in the data assimilation system, spatially varying fields were produced that were subsequently used in the production run of the hydrodynamics. By estimating the spatially varying short wave radiation parameters, we reduced the bias of the hydrodynamic model by up to 3 degrees C, and reduced the RMS error to less than 1 degree C when compared to a withheld dataset.

The second approach used an efficient Ensemble Optimal Interpolation system for sequential updating of the model state (temperature) to assimilate observation from January 2011 – December 2014 in the hydrodynamic model, to create the first version of a high resolution reanalysis product for the GBR region. The average forecast error for temperature ranged between 0.4 and 0.9 degrees C, depending on observation density, typically related to cloud cover.

The assimilation system was also applied to the biogeochemistry, with observations from satellite remote sensing products being the only assimilated data. A simple ensemble optical interpolation system was applied and the forecast error from this system was approximately half that of the control run. By applying a more advance ensemble Kalman filter system, there were further reduction in forecast error, with the domain wide error reducing by a further 5-10% when compared to the simple data assimilation system. It should be noted that recent BGC model developments and further improvements to the BGC DA scheme has led to the mean absolute prediction error dropping from 75% in earlier versions, to 25% in the latest system.