Deep Learning Synthetic Likelihood Approximations for Spatial Extremes

Probability that the GEV location parameter for annual streamflow maxima has been increaasing over the last 50 years.

This is work done jointly with Brian Reich (NCSU) and Ben Shaby (CSU). We developed a process mixture model (PMM) which interpolates between a max stable process and a Gaussian process. The likelihood is intractable, so we develop a computational approach which uses a Vecchia approximation to simplifiy the likelihood and a simulation based approach to learn the approximate likelihood. This can then be used in a hierarchical Bayesian framework for posterior inference. In our first work we analyzed changes in streamflow across the US over the last 50 years.

The Vecchia approximated density regression workflow

In follow-up work, we concentrated on the storm-prone central US. We used precipitation as a predictor to developed a nonstationary PMM for the region. Using climate projections, we also obtained near-future forecasts of extreme streamflow for the region.

RCP 4.5 projections for the CUS

Reetam Majumder
Reetam Majumder
Postdoctoral Fellow