Abstract:

The skill of dynamical atmospheric forecasts is intimately connected to the accuracy and coverage of observations. Observational resources are limited. Optimizing the use of observations is key to improving atmospheric forecasts of the future. The development and testing of algorithms that can be used to optimize the design of observational networks is the subject of this thesis.

Observations can generally be broken down into two categories; Adaptive and Routine. Adaptive (Routine) observations consist of measurements from instruments whose locations can (can not) be willingly altered at each observing time. Chapter 2 develops a novel Estimation Theory based framework for developing solutions to the Adaptive Observations problem. A series of computationally efficient approximations of the general theory are then discussed. The theoretical relationships of these methods to the general framework are pointed out. In Chapter 3, a series of Observing System Simulation Experiments are performed to investigate a number of key issues concerning Estimation Theory based Adaptive Observation algorithms. Fundamental problems associated with computationally efficient Adaptive Observing schemes are examined. The potential benefits of the most general Adaptive Observation methodology in Chapter 2 are also examined. The implications of these results for operational Adaptive Observations is discussed.

In Chapter 4, a framework for defining a cost function appropriate for designing optimal configurations of routine observing networks is developed. The cost function can be evaluated using a retrospective analysis of output data from an ensemble filter based prediction system. Evaluation of the cost function does not require repeated integrations of the prediction model, and therefore offers significant computational advantages over traditional approaches such as Observing System Simulation Experiments. For a wide variety of observing system design problems, the statistical and dynamical significance of the algorithm (called the Retrospective Design Algorithm) is demonstrated using the nonlinear Lorenz 1996 model (Chapter 4) and an Atmospheric General Circulation model (Chapter 5). The implications of the results in Chapters 4 and 5 for the general problem of routine observing network design in very high dimensional systems is explored.