Sensitivity analysis, ocean state estimation and diagnostics in the California current

The effects of sharply different wind forcing patterns on the upwelling system and upwelling source waters over the California Current System (CCS) are investigated using adjoint-based sensitivity analyses in the Regional Ocean Modeling System (ROMS). The wind stress curl field appears to control the locations of the equatorward flow and cross -shore gradient of isopycnal. A deeper upwelling cell and more remote source waters for the upwelling are found when the wind stress curl field changes sharply cross-shore. In contrast, a gradual change of wind stress curl causes a shallower upwelling and local source waters for the upwelling. Data assimilation (DA) combines numerical models and data to determine the best possible estimate of the state of a dynamic system. Data-assimilated ocean states are prepared using the ROMS four-dimensional variational data assimilation (4D-VAR) system with satellite and in situ data during four separate upwelling seasons. They are used for the diagnosis of observed phenomena such as an abrupt change in the Pacific sardine egg distributions. The ROMS 4D-VAR system adjusts the initial conditions and surface forcing for one-month time periods and successfully reduced the statistical differences from the observations. Analysis using optimally estimated ocean states shows stronger offshore transport during the April 2002 La Niña conditions than during the weak 2003 El Niño. This partially causes the extension of preferred spawning habitat for the Pacific sardine but distributes eggs over a broad area, resulting in the lower sampled egg concentration. The adjoint model runs with passive tracer reveal that the nutrient richness of the source waters also contributes to the sardine egg distributions. This dissertation also suggests two new data assimilation approaches to improve the ensemble representativeness of the true states in the ensemble Kalman filter (EnKF). These two approaches, an adaptive EnKF (AEnKF) and a four-dimensional AEnKF (4D-AEnKF), estimate the ensemble statistics better by including new members in the ensemble. The AEnKF creates new members at the current analysis time step, and the 4D-AEnKF creates new members in the past analysis time step, with the aid of the adjoint model to enrich the ensemble. The numerical experiments show that these two new methods improve the filter's performance significantl