29IOAS CYGNSS Observation Error Correlation in Data Assimilation Frameworks

Modern data assimilation systems that fuse a priori information about the atmosphere or oceans with observations rely on error correlation estimates of each. The quality of error covariance estimates provided by a background error matrix, B, and an observation error matrix, R, have a direct influence on the accuracy of the resulting analyses of atmosphere or ocean as initial conditions for prediction models. Much attention has been given by both research and operational environmental forecasting communities to estimating the error covariance of high-dimensional model state spaces, often referred to as the "background error covariance" matrix, or B. But observation error correlation in data assimilation frameworks is often treated more simplistically via heuristic methods. This typically involves heavily thinning satellite data to ensure sampling independence and applying average error figures rather than specific error estimates.

This research effort uses CYGNSS L2 retrieved wind speeds as a test case to demonstrate an approach to account for horizontal and temporal observation error correlation. The goal is to use all available CYGNSS data (i.e., no thinning), extracting the information content during data assimilation without overfitting or distortion. There is considerable error correlation because CYGNSS footprints on the ocean surface overlap significantly along each specular point track. Spatial and temporal clustering of global CYGNSS observations will be demonstrated. Each observation cluster is then assigned estimated errors, R, which are then used to compute contributions to the overall observation cost function, J_o. Sensitivity to differing degrees of observation error correlation to the aggregate observation cost function, J_o, will also be demonstrated.