Graph-Guided Regularization for Improved Forecasting of Southwestern US Winter Precipitation
Abstract
Understanding the factors that determine regional climate variability is a challenge with important political, economic, and environmental implications. Unprecedented quantities of high-resolution observations and climate model simulations provide an opportunity to discover previously unknown teleconnections between large-scale modes of variability and regional hydroclimate using tools from machine learning and statistics. In this work, we seek to improve seasonal forecasting of winter precipitation over the Southwestern US by regressing on summer sea-surface temperatures (SST) over the entire Pacific basin. However, due to the small sample size of the observational record and highly correlated SSTs, off-the-shelf methods can yield misleading results when applied directly to such data. Our method explicitly accounts for spatiotemporally correlated predictors via a regularization approach based on an underlying correlation graph. The nodes of the correlation graph are SST observations at different locations and time lags, and edge weights correspond to pairwise correlations between predictors. This graph is used to define a "graph total variation" regularizer that promotes similar weights for highly correlated features. In order to overcome statistical issues introduced by estimating a high-dimensional covariance graph using the limited samples from the observational record, we use the CESM1 Large Ensemble (LENS) output, which consists of 40 trajectories of SSTs with which we can estimate correlation strengths. We demonstrate that using this covariance graph to add structure to coefficients in the regression scheme provides stronger predictive performance and more interpretable model selection than competing methods.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFMGC43D1352S
- Keywords:
-
- 0555 Neural networks;
- fuzzy logic;
- machine learning;
- COMPUTATIONAL GEOPHYSICS;
- 1626 Global climate models;
- GLOBAL CHANGE;
- 1942 Machine learning;
- INFORMATICS;
- 4313 Extreme events;
- NATURAL HAZARDS