The California Current System (CCS) is undoubtedly one of the most well studied regions of the world ocean. The seasonal variability of equatorward

upwelling winds allow for the development of a complex set of dynamical features while also sustaining a rich diversity of marine organisms.

With use of the state-of-the-art Regional Ocean Modeling System (ROMS) 4-dimensional variational (4D-Var) data assimilation system,

the circulation of the CCS is explored in varying capacities over the time period 1980-2010.

Methods for improving the accuracy of circulation estimates provided during ROMS 4D-Var model integrations are explored through

quantification of a source of intrinsic model error that arrises from unavoidable numerical inaccuracies in the model and which is systematic to all oceanic and atmospheric forecast models.

Accounting for this source of model error in a model of the CCS leads to demonstrable improvements to the circulation estimates

compared to those computed without use of these techniques (Section 2).

Additionally, an in depth investigation of the low-frequency modulation of the circulation in a 31-year historical analysis of the CCS

computed using the ROMS 4D-Var system is presented and helps to elucidate the complex evolution of the 3-dimensional ocean state, as well as, assisting to

link local variations with basin scale climate modes; including the Pacific Decadal Oscillation, North Pacific Gyre Oscillation

and El Niño/Southern Oscillation (Section 3). Using a linear inverse model, estimates of the dynamical modes governing the temporal evolution

of the CCS circulation are computed. The dynamical modes capture the role ENSO plays in modulating the temperature, salinity and velocity structure in the CCS (Section 4) and the potential for decadal variability in the CCS to be dominated by forcing by basin scale climate modes (Section 5).

Data assimilation is a method to combine the numerical solutions calculated by circulation models with observations from different platforms in order to obtain an optimal estimate of the state of the system. The computational and numerical difficulties associated with processing the increasing number of observations necessitate the use of thinning techniques to reduce the number of data assimilated. In this thesis the impact of thinning two types of particularly voluminous data sets on the overall performance of an ocean data assimilation system were evaluated. In particular, an analysis of the ocean circulation along the U.S. West Coast with a 10 km resolution grid was performed, spanning a period of one year. Two different thinning methods were tested: an intelligent data thinning (IDT) algorithm to thin gridded satellite sea surface temperature (SST) and a simple thinning algorithm to reduce the volume of radial sea surface velocity measured by a network of coastal high frequency radars. The SST data were thinned by discarding data in regions with low spatial variability while retaining data in regions of high spatial variability. Conversely, the radar observations were averaged to create “super observations” consistent with the resolution of the model grid and prior assumptions about observation errors. The full and thinned data sets were assimilated using a 4-dimensional variational (4D-Var) data assimilation algorithm in the Regional Ocean Modeling System (ROMS). A statistical analysis of the diagnosed background and observation errors showed that the thinning experiments were well-behaved. Furthermore, the innovation and residual vectors (i.e. the difference between each observation and its prior and posterior model counterpart) in all cases generally satisfied the assumption of Gaussian distributions. Additionally, the topology of a the total error covariance matrix of the data assimilation system was explored via its eigen space. The thinning experiments amplified the eigen spectrum, modified the condition number, and in particular thinning SST changed the aspect ratio of the hyperellipse defined by the covariance matrix to change. Lastly the impact of each type of observation on the analyses was quantified for the different thinning methods, suggesting that the radial velocities thinning was perhaps too severe, while the thinning of SST levelled the impact different observations had on the DA analysis. Overall results showed that the thinning did not significantly degrade the analysis, hence the next step will be to test these algorithms in a near-real time forecasting system.