Using a GLM with spatial random effects to model fractures in antarctic iceshelves

Add to your list(s) Download to your calendar using vCal

• Sergio Bacallado, University of Cambridge
• Tuesday 02 June 2020, 11:00-12:30
• https://ukri.zoom.us/j/92946956029.

If you have a question about this talk, please contact Jonathan Rosser.

Emetc et al. (The Cryosphere, 12, 3187–3213, 2018) recently gathered a dataset of >100,000 locations on antarctic iceshelves classified according to whether or not a satellite image shows a fracture. The purpose of their study is to model the statistical relationship between the outputs of a fluid model for the iceshelf and the probability of fractures within it. There is great interest in this type of analysis, as physics-based simulations of iceshelves with different types of climate forcing could potentially be used to predict the effect of climate change on the risk of iceshelf collapse. Emetc et al. propose a logistic regression model with variable selection, which is applied to different regions within the iceshelves separately, from which they derive consensus estimates of regression coefficients.

In this talk, I will discuss several issues with the analysis of Emetc et al. and propose an alternative binomial regression model with spatially correlated random effects. Approximate Bayesian inference is done using a technique proposed by Hensman et al. (NIPS proceedings, 2015) which applies MCMC to sample a sparse gaussian process variational approximation of the posterior. The technique is scalable to datasets with hundreds of thousands of points. I will show an example with iceshelf fracture data.

Joint work with Marko Closs and Nick Barrand at the University of Birmingham.

This talk is part of the AI4ER Seminar Series series.

This talk is included in these lists:

• AI4ER Seminar Series
• Cambridge Centre for Data-Driven Discovery (C2D3)
• Cambridge talks
• Chris Davis' list
• Interested Talks
• Trust & Technology Initiative - interesting events
• bld31
• https://ukri.zoom.us/j/92946956029
• ndk22's list
• ob366-ai4er
• rp587

Note that ex-directory lists are not shown.