Abstract
AbstractA probabilistic tracking model is introduced that identifies storm tracks from feature vectors that are extracted from meteorological analysis data. The model assumes that the genesis and lysis times of each track are unknown and estimates their values along with the track’s position and storm intensity over time. A hidden-state dynamics model (Kalman filter) characterizes the temporal evolution of the storms.The model uses a Bayesian methodology for estimating the unknown lifetimes (genesis–lysis pairs) and tracks of the storms. Prior distributions are placed over the unknown parameters and their posterior distributions are estimated using a Markov Chain Monte Carlo (MCMC) sampling algorithm. The posterior distributions are used to identify and report the most likely storm tracks in the data. This approach provides a unified probabilistic framework that accounts for uncertainty in storm timing (genesis and lysis), storm location and intensity, and the feature detection process. Thus, issues such as missing observations can be accommodated in a statistical manner without human intervention.The model is applied to the field of relative vorticity at the 975-hPa level of analysis from the National Centers for Environmental Prediction Global Forecast System during May–October 2000–02, in the tropical east Pacific. Storm tracks in the National Hurricane Center best-track data (HURDAT) for the same period are used to assess the performance of the storm identification and tracking model.
Highlights
By training the dynamics model on Geostationary Operational Environmental Satellite (GOES) data and applying it to National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS) data, we demonstrate that the trained dynamics model is transferrable to different datasets
We have presented a probabilistic model for automatically extracting storm tracks from a set of domainspecific feature detections
We developed an Markov Chain Monte Carlo (MCMC)-based inference method for fitting the model parameters to observed data and a practical methodology for extracting multiple tracks from a single season of data
Summary
Prior nonprobabilistic approaches to solving the initiation and association problems utilize gating functions and hard constraints on certain track properties, such as storm velocity and track smoothness, in order to prune away a large amount of the search space (Hodges 1994; Gauvrit et al 1997) These methods can produce excellent tracking results within a particular domain, but may lack flexibility because the track constraints and heuristics are often an integral part of the method. Oh et al (2004) propose a similar model to ours, along with a sampling strategy to explore the posterior space of data associations; their model does not place prior distributions over the genesis and lysis times Their sampling strategy tends to be less effective when the feature vectors are spatially sparse relative to the scale of the storm dynamics, which is the case in our application. In contrast to this MHT-based method, our approach can leverage both past and future data when making inferences about states and parameters for a particular time
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
