Machine learning seminar series - Probabilistic downscaling to detect regional present and future climate hazards

11:30 BST | April 28, 2020

Host

Peter Dueben (ECMWF)

Speakers

Sherman Lo holds a PhD in statistics and is currently working, at the University of Warwick, on precipitation forecasting of the UK using climate models. He has obtained degrees in Physics and Computational Statistics & Machine Learning at University College London. He is interested in applied and computational statistics and has previously worked on x-ray imaging for 3D printing, fusion energy and chromosome imaging.

Ritabrata Dutta is an Assistant Professor in the Department of Statistics, University of Warwick, broadly working on the applications of statistical methodology in the domain of meteorology, population genetics and complex bio-medical processes. He has developed methods for Approximate Bayesian computation (ABC), collaborated with Swiss National Supercomputing Centre (CSCS) to develop a Python package for ABC utilizing HPC infrastructure optimally and applied ABC methodology to learn expensive mechanistic models explaining epidemic processes on network, molecular dynamics of water, arterial flow of blood, volcanic eruption and passenger dynamics in airports. Presently Dr Dutta is supervising 6 PhD students and 1 Postdoc, working broadly on methodological developments of statistics and machine learning, applications in the domain sciences. He is currently PI of a project on Quantifying effects of climate change on extreme weather events via distributional downscaling funded by the Alan Turing Institute (ATI) as well as a special project of ECMWF on Data-driven calibration of stochastic parametrization of IFS using ABC.

Abstract

We present a statistical methodology to do forecasting of precipitation at 0.1° resolution using model fields from computer simulations (air temperature, geopotential, specific humidity, total column water vapour, wind velocity) at a lower resolution (5/9° and 5/6° lon/lat). Observed precipitation at the 0.1° resolution for the past 4 decades was used to fit our model onto it in order to do forecasting with quantifiable uncertainty.

Our model is the compound Poisson distribution (Revfeim, K.J.A., 1984, Dunn, P.K., 2004) which can model both occurrence and quantity of precipitation as a random variable. We impose time autocorrelation by introducing auto-regressive and moving average terms into the model. Spatial dependencies were introduced by putting a Gaussian process prior on the parameters.

The model fields were interpolated, with uncertainty bars, from low resolution to high resolution using a Gaussian process. The entire Bayesian inference (or model fitting) was done using a Gibbs sampling scheme consisting of Metropolis-Hastings, slice sampling and elliptical slice sampling.