Dr. Poterjoy presented a seminar on “Probabilistic Storm-Scale Analysis and Prediction Using a Non-Parametric Ensemble Filter: Implications for Tropical Cyclone Forecasting”.
Initial condition uncertainty remains a major factor contributing to forecast errors for tropical cyclones and mesoscale convective systems. Data assimilation strategies based on linear/Gaussian methods, such as ensemble Kalman filters and ensemble-variational hybrids, currently provide the best means of using observations to reduce this source of error. Nevertheless, the underlying assumptions of these methods are inappropriate for most mesoscale weather applications. For example, the dynamical processes governing the evolution of tropical cyclones are nonlinear and poorly observed, leading to forecast errors that can be highly non-Gaussian. Remote sensing platforms onboard satellites and aircrafts (e.g., radars and radiometers) may provide valuable measurements of this system, but the nonlinear mapping between model space and measurements also challenges the linear assumptions of Gaussian techniques. Suboptimal state estimates from these methods often manifest themselves as imbalanced initial conditions, which can lead to degraded forecasts. The spin down of tropical cyclones immediately after model initialization presents one example of this issue. In this presentation, a brief overview of data assimilation strategies used routinely for numerical weather prediction will be provided. An idealized convective-scale weather application will then be used to demonstrate several limitations of these techniques and introduce a new strategy for avoiding them. This talk will also summarize recent work towards the development and testing of a non-parametric filter for data assimilation in weather models, and discuss future implications for tropical cyclone prediction.
A recording of the presentation is available on the anonymous ftp site: