A Bayesian framework for storm tracking using a hidden-state representation

The problem of tracking is of fundamental interest in the atmospheric sciences where it is natural to study objects of interest -- storms, hurricanes and other coherent phenomena -- by tracking their evolution over time. We propose a Bayesian tracking methodology which is built on a hidden state representation of an object's temporal behavior that obeys a Linear Dynamic Model (LDM). The dynamic model may either be constructed “by hand” from consideration of an underlying physical model, or trained from a representative set of storm tracks. The data association process is formulated to be independent of any specific dynamic model which allows different models to be plugged into the same tracking framework. The use of a hidden state model automatically produces smoothed tracks from noisy or grid-aligned feature detections. Our method provides numerical estimates of the statistical strength of track segments, as well as rankings of tracks based on their likelihood under the chosen LDM.

We develop an efficient solution to the data association problem using Markov Chain Monte Carlo sampling to extract tracks from a set of feature detections and marginalize over the possible associations to obtain posterior distributions over the genesis and lysis times of storms. This Bayesian methodology is applied to a multi-year set of vorticity fields from the NCEP Final Analysis and ERA40 data sets to generate tracks of Westward Propagating Disturbances (WPDs) in the tropical Pacific region.