Loading Events

« All Events

  • This event has passed.

ECE PhD Proposal Review: Yumin Liu

November 13, 2020 @ 2:00 pm - 3:00 pm

PhD Proposal Review: Learning from Spatio-Temporal Data with Applications in Climate Science

Yumin Liu

Location: Zoom Link

Abstract:Climate change is one of the major challenges to human beings and many other species in our time. In the recent decade, the number of disasters related to climate change such as wildfires, storms, floods and droughts are increasing, and the casualty and economic losses caused by them are larger compared to those of decades ago. This calls for better and efficient ways to predict climate change in order to better prepare and reduce losses. Predicting climate change involves using historical observational data and model simulated data, both of which usually involve multiple locations and timestamps and are spatio-temporal. With the rapid development and progress of machine learning, these methods have achieved several impactful contributions in many domains; we would like to translate its impact to climate science.
In this thesis we addressseveral problems in climate science. This challenging complex domain enable us to develop, innovate, adapt, and advance machine learning in the following ways. 1) We develop a multi-task learning method to estimate relationships between tasks and learn the basis tasks in different locations especially for nearby locations which may share similar climate patterns. This method assumes that the weights of an observed task is a linear combination of several latent basis tasks and that the task relationships can be learnt by imposing a regularized precision matrix. 2) We propose a nonparameteric mixture of sparse linear regression models to cluster and identify important climate models for prediction. This model incorporates Dirichlet Process (DP) to automatically determine the number of clusters, imposes Markov Random Field (MRF) constraints to guarantee spatio-temporal smoothness, and selects a subset of global climate models (GCMs) that are useful for prediction within each spatio-temporal cluster with a spike-and-slab prior. We derive an effective Gibbs sampling method for this model. 3) We adapt image super resolution method to climate downscaling — increasing spatial resolution for climate variables for local impact analysis. The proposed method is called YNet which is a novel deep convolutional neural network (CNN) with skip connections and fusion capabilities to perform downscaling for climate variables on multiple GCMs directly rather than on reanalysis data.