William Hsieh

• Machine learning methods and their applications to the environmental sciences
• Seasonal climate and extreme weather prediction
• Atmosphere-ocean climate dynamics

Machine learning (ML), a major branch of artificial intelligence, has a huge impact on our everyday lives through its ability to recognize complicated, nonlinear signals in large datasets. When we post a letter, the post office uses ML technology to understand our handwriting. Online vendors such as Amazon and Netflix suggest books and movies of interest using ML. Self-driving automobiles promise a revolution in transportation. While most environmental scientists are not familiar with machine learning, ML has been the fastest growing field in the last twenty years, fueled by Apple, Alphabet/Google and Microsoft, the three largest companies in the world by market capitalization according to Wikipedia.

The question that has intrigued me for over two decades has been: How would machine learning impact the environmental sciences? This single question has been the main driving force of my research program.

A prominent example of climate variability is the famous El Niño-La Niña phenomenon, an irregular fluctuation of the climate system which produces anomalous warming in the equatorial Pacific during El Niño and cooling during La Niña, with notable influence on the Canada winter climate. El Niño/La Niña episodes can be forecast with reasonable accuracy 3-12 months in advance. Our group has built models for El Niño/La Niña prediction using artificial neural networks and other ML methods. Our forecasts are updated monthly on our web site http://www.ocgy.ubc.ca/projects/clim.pred/ and are included in the IRI ensemble forecasts.

We have developed neural network models for nonlinear principal component analysis, nonlinear canonical correlation analysis, and nonlinear singular spectrum analysis (our codes are freely downloadable and have users from over 60 countries). We have identified nonlinear atmospheric teleconnection patterns associated with the El Niño/La Niña, the Arctic Oscillation, the quasi-biennial oscillation and the Madden-Julian oscillation.

In the last few years, our research efforts have been directed to the following areas:

• With general circulation models having spatial resolution too coarse to reveal climate variability at local scales, ML methods have been developed to nonlinearly downscale the model output to finer spatial scales, especially for precipitation and streamflow.
• Machine learning methods are ideal for extracting information from satellite data. Crop yield prediction models have been developed by applying ML methods to vegetation indices derived from satellite data.
• ML methods have been used to improve forecasts of air quality over Canadian cities.
• ML methods have been used in data fusion, i.e. combining various gridded products, to improve estimates of snow depth (i.e. snow water equivalent) over British Columbia.
• While ML methods such as artificial neural networks are able to extract nonlinear signals missed by linear statistical methods, they are computationally much more expensive. We have been developing new ML models which are several orders of magnitude faster than the standard ML models, especially when the models need to be updated frequently as new data arrive continually (i.e. online learning).

My graduate-level book "Machine Learning Methods in the Environmental Sciences: Neural Networks and Kernels" was published by Cambridge Univ. Press in 2009.

***** Note: I am no longer accepting new graduate students/post-docs, as I have moved to live in Victoria on Vancouver Island after becoming professor emeritus *****

Selected Publications