University of British Columbia
Abstract: Having originated from the field of artificial intelligence, machine learning (ML) methods (e.g. artificial neural networks) have been increasing used in the environmental sciences, as they nonlinearly generalize the classical statistical methods such as regression, classification, principal component analysis and canonical correlation analysis. The advantages and limitations of ML methods in climate and weather applications are illustrated by several examples.
Our recent work has been on using linear regression (LR) and Bayesian neural network (BNN) methods to downscale global climate model output. For 10 weather stations located in Quebec and Ontario, daily values of maximum and minimum temperatures were obtained from downscaling for the period 1961-2000. Six annual indices of extreme weather (e.g. number of frost days, heat wave duration, etc.) were also computed from the downscaled temperature output. The results showed that while the nonlinear BNN models did not consistently outperform LR in downscaling daily temperature variability, they did in the annual climate indices for extreme weather. Using regional climate model (RCM) output used as .pseudo-observations. to verify the downscaled model data for future climate scenarios, we found that the BNN models were more accurate than LR in projecting future climate under the IPCC A2 scenario of global warming.
Bio: William W. Hsieh is Professor Emeritus in the Dept. of Earth and Ocean Sciences, at the University of British Columbia. He is internationally known for his pioneering work in developing and applying machine learning methods in the environmental sciences. He has over 90 peer-reviewed journal publications covering areas of climate variability, machine learning, oceanography, atmospheric science and hydrology. His graduate-level textbook Machine Learning Methods in the Environmental Sciences was published by Cambridge University Press in 2009. He is a Fellow of the Canadian Meteorological and Oceanographic Society.