2013 Academic Year Seminars
A new probabilistic data assimilation method for short-term rainfall forecasting from radar observations is introduced. The spatial model relies on a decomposition of the observed rainfall field into precipitation areas ('cells'). The characteristics of the 'cells', that is the parameters of the model, are given relevant priors and their estimation is performed within a Bayesian framework. The cell parameters are estimated using an approximate variational Bayesian method, in which the Kullback Leibler divergence between the approximate posterior (known up to its parameters) and the true posterior (given by Bayes rule, known up to a constant) is minimised. Variational Bayesian methods could be thought of as a probabilistic version of 3D VAR assimilation. The cells advection is represented as a smooth vector Gaussian process, for which the posterior distribution is estimated in a Kalman filter-like fashion, using the cells' displacements over time as pseudo-observations. The model is tested on real radar data from the UK Met. Office, both in a convective and a frontal setting. The performance of the model is assessed using standard and probabilistic validation methods. Overall, the model shows very good assimilation skill and reasonable forecasting skill given the simplistic forecasting method used. Extensions are discussed.
Remi Barillec is a research fellow working for the MUCM (Managing Uncertainty in Complex Models) project at Aston University, under supervision of Dr Dan Cornford. He completed his MSc in Computer Science and Mathematics at the ENSIIE in France before joining the Neural Computing Research Group at Aston University in 2002 for an MSc by research in Machine Learning. He recently obtained his PhD in data assimilation and Bayesian rainfall nowcasting. His interests lie in dynamic modelling, emulation of complex computer models and probabilistic inference.