William Wei Hsieh
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
William Wei Hsieh has a total of 23 co-authors. Here are the top co-authors by number of shared publications:
Aiming Wu
Department of Earth and Ocean Sciences University of British Columbia Vancouver, British Columbia Canada
5 shared publicationsAmir Shabbar
Climate Research Division , Environment Canada , Toronto, Ontario
3 shared publicationsLawrence A. Mysak
McGill University
2 shared publicationsSanjay S. P. Rattan
Department of Earth and Ocean Sciences University of British Columbia Vancouver, British Columbia Canada
2 shared publicationsShuyong Li
Department of Earth and Ocean Sciences University of British Columbia Vancouver, British Columbia Canada
2 shared publicationsA. Shabbar
Climate Research Division, Environment Canada, Toronto, Ontario, Canada
1 shared publication
Latest Publications
- Open Access
- 1 January 2023
- Published by International Society for Environmental Information Science (ISEIS) in Journal of Environmental Informatics
- Vol. 41 (2) , 75-87-87
Improving Predictions by Nonlinear Regression Models from Outlying Input Data
When applying machine learning/statistical methods to the environmental sciences, nonlinear regression (NLR) models often perform only slightly better and occasionally worse than linear regression (LR). The proposed reason for this conundrum is that NLR models can give predictions much worse than LR when given input data which lie outside the domain used in model training. Continuous unbounded variables are widely used in environmental sciences, whence not uncommon for new input data to lie far outside the training domain. For six environmental datasets, inputs in the test data were classified as “outliers” and “non-outliers” based on the Mahalanobis distance from the training input data. The prediction scores (mean absolute error, Spearman correlation) showed NLR to outperform LR for the non-outliers, but often underperform LR for the outliers. An approach based on Occam’s Razor (OR) was proposed, where linear extrapolation was used instead of nonlinear extrapolation for the outliers. The linear extrapolation to the outlier domain was based on the NLR model within the non-outlier domain. This NLROR approach reduced occurrences of very poor extrapolation by NLR, and it tended to outperform NLR and LR for the outliers. In conclusion, input test data should be screened for outliers. For outliers, the unreliable NLR predictions can be replaced by NLROR or LR predictions, or by issuing a “no reliable prediction” warning.
- 1 January 2023
- Published by International Society for Environmental Information Science (ISEIS) in Journal of Environmental Informatics
- Vol. 41 (2) , 75-87-87
Open AccessImproving Predictions by Nonlinear Regression Models from Outlying Input Data
When applying machine learning/statistical methods to the environmental sciences, nonlinear regression (NLR) models often perform only slightly better and occasionally worse than linear regression (LR). The proposed reason for this conundrum is that NLR models can give predictions much worse than LR when given input data which lie outside the domain used in model training. Continuous unbounded variables are widely used in environmental sciences, whence not uncommon for new input data to lie far outside the training domain. For six environmental datasets, inputs in the test data were classified as “outliers” and “non-outliers” based on the Mahalanobis distance from the training input data. The prediction scores (mean absolute error, Spearman correlation) showed NLR to outperform LR for the non-outliers, but often underperform LR for the outliers. An approach based on Occam’s Razor (OR) was proposed, where linear extrapolation was used instead of nonlinear extrapolation for the outliers. The linear extrapolation to the outlier domain was based on the NLR model within the non-outlier domain. This NLROR approach reduced occurrences of very poor extrapolation by NLR, and it tended to outperform NLR and LR for the outliers. In conclusion, input test data should be screened for outliers. For outliers, the unreliable NLR predictions can be replaced by NLROR or LR predictions, or by issuing a “no reliable prediction” warning.
- Open Access
- 13 April 2022
- Published by Cambridge University Press (CUP) in Environmental Data Science
- Vol. 1
Evolution of machine learning in environmental science—A perspective
The growth of machine learning (ML) in environmental science can be divided into a slow phase lasting till the mid-2010s and a fast phase thereafter. The rapid transition was brought about by the emergence of powerful new ML methods, allowing ML to successfully tackle many problems where numerical models and statistical models have been hampered. Deep convolutional neural network models greatly advanced the use of ML on 2D or 3D data. Transfer learning has allowed ML to progress in climate science, where data records are generally short for ML. ML and physics are also merging in new areas, for example: (a) using ML for general circulation model parametrization, (b) adding physics constraints in ML models, and (c) using ML in data assimilation.
- 13 April 2022
- Published by Cambridge University Press (CUP) in Environmental Data Science
- Vol. 1
Open AccessEvolution of machine learning in environmental science—A perspective
The growth of machine learning (ML) in environmental science can be divided into a slow phase lasting till the mid-2010s and a fast phase thereafter. The rapid transition was brought about by the emergence of powerful new ML methods, allowing ML to successfully tackle many problems where numerical models and statistical models have been hampered. Deep convolutional neural network models greatly advanced the use of ML on 2D or 3D data. Transfer learning has allowed ML to progress in climate science, where data records are generally short for ML. ML and physics are also merging in new areas, for example: (a) using ML for general circulation model parametrization, (b) adding physics constraints in ML models, and (c) using ML in data assimilation.
- Open Access
Improving predictions by nonlinear regression models from outlying input data
When applying machine learning/statistical methods to the environmental sciences, nonlinear regression (NLR) models often perform only slightly better and occasionally worse than linear regression (LR). The proposed reason for this conundrum is that NLR models can give predictions much worse than LR when given input data which lie outside the domain used in model training. Continuous unbounded variables are widely used in environmental sciences, whence not uncommon for new input data to lie far outside the training domain. For six environmental datasets, inputs in the test data were classified as "outliers" and "non-outliers" based on the Mahalanobis distance from the training input data. The prediction scores (mean absolute error, Spearman correlation) showed NLR to outperform LR for the non-outliers, but often underperform LR for the outliers. An approach based on Occam's Razor (OR) was proposed, where linear extrapolation was used instead of nonlinear extrapolation for the outliers. The linear extrapolation to the outlier domain was based on the NLR model within the non-outlier domain. This NLR$_{\mathrm{OR}}$ approach reduced occurrences of very poor extrapolation by NLR, and it tended to outperform NLR and LR for the outliers. In conclusion, input test data should be screened for outliers. For outliers, the unreliable NLR predictions can be replaced by NLR$_{\mathrm{OR}}$ or LR predictions, or by issuing a "no reliable prediction" warning.
Open AccessImproving predictions by nonlinear regression models from outlying input data
When applying machine learning/statistical methods to the environmental sciences, nonlinear regression (NLR) models often perform only slightly better and occasionally worse than linear regression (LR). The proposed reason for this conundrum is that NLR models can give predictions much worse than LR when given input data which lie outside the domain used in model training. Continuous unbounded variables are widely used in environmental sciences, whence not uncommon for new input data to lie far outside the training domain. For six environmental datasets, inputs in the test data were classified as "outliers" and "non-outliers" based on the Mahalanobis distance from the training input data. The prediction scores (mean absolute error, Spearman correlation) showed NLR to outperform LR for the non-outliers, but often underperform LR for the outliers. An approach based on Occam's Razor (OR) was proposed, where linear extrapolation was used instead of nonlinear extrapolation for the outliers. The linear extrapolation to the outlier domain was based on the NLR model within the non-outlier domain. This NLR$_{\mathrm{OR}}$ approach reduced occurrences of very poor extrapolation by NLR, and it tended to outperform NLR and LR for the outliers. In conclusion, input test data should be screened for outliers. For outliers, the unreliable NLR predictions can be replaced by NLR$_{\mathrm{OR}}$ or LR predictions, or by issuing a "no reliable prediction" warning.
- 14 July 2016
- Published by Springer Nature in Air Quality, Atmosphere & Health
- Vol. 10 (2) , 195-211
Evaluating hourly air quality forecasting in Canada with nonlinear updatable machine learning methods
- 14 July 2016
- Published by Springer Nature in Air Quality, Atmosphere & Health
- Vol. 10 (2) , 195-211
Evaluating hourly air quality forecasting in Canada with nonlinear updatable machine learning methods
- Open Access
- 15 February 2011
- Published by Taylor & Francis in Atmosphere-Ocean
- Vol. 49 (1) , 22-31
Surface Wind Speed Prediction in the Canadian Arctic using Non-Linear Machine Learning Methods
Two non-linear, machine-learning/statistical methods, i.e., Bayesian neural network (BNN) and support vector regression (SVR), plus multiple linear regression (MLR), were used to forecast surface wind speeds at lead times of 12, 24, 48 and 72 h. Three different schemes, a statistical downscaling model (Scheme 1) using daily reforecast data from the National Centers for Environmental Prediction (NCEP) Global Forecasting System (GFS), an autoregressive model (Scheme 2) based on past wind observations, and a full model (Scheme 3) combining the two, were investigated in this study for the October–March winds from two meteorological stations in the Canadian Arctic (Clyde River and Paulatuk). At very short lead times, Scheme 2 provides better wind speed prediction than Scheme 1, but its forecast scores decrease rapidly with lead time. Scheme 3 generally performs best, especially at shorter lead times. All the linear and non-linear downscaling methods have significantly higher forecast scores at the two stations than the GFS reforecast. The non-linear methods tended to have slightly better forecast scores than linear methods (MLR and the linear version of SVR). There is particular interest in high-wind events, defined as having wind speeds over 22 knots (11.3 m s−1). After rescaling, the continuous wind predictions from Scheme 3 were classified into two types — high-wind event or non-event. For high-wind event forecasting, the non-linear methods have marginally better binary forecast scores than the linear methods for Clyde River but not for Paulatuk. The alternative approach of using support vector classification (SVC) did not perform better, but weighting the high-wind events more heavily than the non-events during model training improved the binary forecast scores.
- 15 February 2011
- Published by Taylor & Francis in Atmosphere-Ocean
- Vol. 49 (1) , 22-31
Open AccessSurface Wind Speed Prediction in the Canadian Arctic using Non-Linear Machine Learning Methods
Two non-linear, machine-learning/statistical methods, i.e., Bayesian neural network (BNN) and support vector regression (SVR), plus multiple linear regression (MLR), were used to forecast surface wind speeds at lead times of 12, 24, 48 and 72 h. Three different schemes, a statistical downscaling model (Scheme 1) using daily reforecast data from the National Centers for Environmental Prediction (NCEP) Global Forecasting System (GFS), an autoregressive model (Scheme 2) based on past wind observations, and a full model (Scheme 3) combining the two, were investigated in this study for the October–March winds from two meteorological stations in the Canadian Arctic (Clyde River and Paulatuk). At very short lead times, Scheme 2 provides better wind speed prediction than Scheme 1, but its forecast scores decrease rapidly with lead time. Scheme 3 generally performs best, especially at shorter lead times. All the linear and non-linear downscaling methods have significantly higher forecast scores at the two stations than the GFS reforecast. The non-linear methods tended to have slightly better forecast scores than linear methods (MLR and the linear version of SVR). There is particular interest in high-wind events, defined as having wind speeds over 22 knots (11.3 m s−1). After rescaling, the continuous wind predictions from Scheme 3 were classified into two types — high-wind event or non-event. For high-wind event forecasting, the non-linear methods have marginally better binary forecast scores than the linear methods for Clyde River but not for Paulatuk. The alternative approach of using support vector classification (SVC) did not perform better, but weighting the high-wind events more heavily than the non-events during model training improved the binary forecast scores.
- Open Access
- 6 January 2011
- Published by Copernicus GmbH in Hydrology and Earth System Sciences
- Vol. 15 (1) , 65-74
Seasonal prediction of winter extreme precipitation over Canada by support vector regression
For forecasting the maximum 5-day accumulated precipitation over the winter season at lead times of 3, 6, 9 and 12 months over Canada from 1950 to 2007, two nonlinear and two linear regression models were used, where the models were support vector regression (SVR) (nonlinear and linear versions), nonlinear Bayesian neural network (BNN) and multiple linear regression (MLR). The 118 stations were grouped into six geographic regions by K-means clustering. For each region, the leading principal components of the winter maximum 5-d accumulated precipitation anomalies were the predictands. Potential predictors included quasi-global sea surface temperature anomalies and 500 hPa geopotential height anomalies over the Northern Hemisphere, as well as six climate indices (the Niño-3.4 region sea surface temperature, the North Atlantic Oscillation, the Pacific-North American teleconnection, the Pacific Decadal Oscillation, the Scandinavia pattern, and the East Atlantic pattern). The results showed that in general the two robust SVR models tended to have better forecast skills than the two non-robust models (MLR and BNN), and the nonlinear SVR model tended to forecast slightly better than the linear SVR model. Among the six regions, the Prairies region displayed the highest forecast skills, and the Arctic region the second highest. The strongest nonlinearity was manifested over the Prairies and the weakest nonlinearity over the Arctic.
- 6 January 2011
- Published by Copernicus GmbH in Hydrology and Earth System Sciences
- Vol. 15 (1) , 65-74
Open AccessSeasonal prediction of winter extreme precipitation over Canada by support vector regression
For forecasting the maximum 5-day accumulated precipitation over the winter season at lead times of 3, 6, 9 and 12 months over Canada from 1950 to 2007, two nonlinear and two linear regression models were used, where the models were support vector regression (SVR) (nonlinear and linear versions), nonlinear Bayesian neural network (BNN) and multiple linear regression (MLR). The 118 stations were grouped into six geographic regions by K-means clustering. For each region, the leading principal components of the winter maximum 5-d accumulated precipitation anomalies were the predictands. Potential predictors included quasi-global sea surface temperature anomalies and 500 hPa geopotential height anomalies over the Northern Hemisphere, as well as six climate indices (the Niño-3.4 region sea surface temperature, the North Atlantic Oscillation, the Pacific-North American teleconnection, the Pacific Decadal Oscillation, the Scandinavia pattern, and the East Atlantic pattern). The results showed that in general the two robust SVR models tended to have better forecast skills than the two non-robust models (MLR and BNN), and the nonlinear SVR model tended to forecast slightly better than the linear SVR model. Among the six regions, the Prairies region displayed the highest forecast skills, and the Arctic region the second highest. The strongest nonlinearity was manifested over the Prairies and the weakest nonlinearity over the Arctic.
- Open Access
- 24 December 2009
- Published by Wiley in International Journal of Oceanography
- Vol. 2009, 1-13
Forecasts of Tropical Pacific Sea Surface Temperatures by Neural Networks and Support Vector Regression
Two nonlinear regression methods, Bayesian neural network (BNN) and support vector regression (SVR), and linear regression (LR), were used to forecast the tropical Pacific sea surface temperature (SST) anomalies at lead times ranging from 3 to 15 months, using sea level pressure (SLP) and SST as predictors. Datasets for 1950–2005 and 1980–2005 were studied, with the latter period having the warm water volume (WWV) above the isotherm integrated across the equatorial Pacific available as an extra predictor. The forecasts indicated that the nonlinear structure is mainly present in the second PCA (principal component analysis) mode of the SST field. Overall, improvements in forecast skills by the nonlinear models over LR were modest. Although SVR has two structural advantages over neural network models, namely (a) no multiple minima in the optimization process and (b) an error norm robust to outliers in the data, it did not give better overall forecasts than BNN. Addition of WWV as an extra predictor generally increased the forecast skills slightly; however, the influence of WWV on SST anomalies in the tropical Pacific appears to be linear.- 24 December 2009
- Published by Wiley in International Journal of Oceanography
- Vol. 2009, 1-13
Open AccessForecasts of Tropical Pacific Sea Surface Temperatures by Neural Networks and Support Vector Regression
Two nonlinear regression methods, Bayesian neural network (BNN) and support vector regression (SVR), and linear regression (LR), were used to forecast the tropical Pacific sea surface temperature (SST) anomalies at lead times ranging from 3 to 15 months, using sea level pressure (SLP) and SST as predictors. Datasets for 1950–2005 and 1980–2005 were studied, with the latter period having the warm water volume (WWV) above the isotherm integrated across the equatorial Pacific available as an extra predictor. The forecasts indicated that the nonlinear structure is mainly present in the second PCA (principal component analysis) mode of the SST field. Overall, improvements in forecast skills by the nonlinear models over LR were modest. Although SVR has two structural advantages over neural network models, namely (a) no multiple minima in the optimization process and (b) an error norm robust to outliers in the data, it did not give better overall forecasts than BNN. Addition of WWV as an extra predictor generally increased the forecast skills slightly; however, the influence of WWV on SST anomalies in the tropical Pacific appears to be linear. - 1 May 2007
- Published by Elsevier in Neural Networks
- Vol. 20 (4) , 434-443
Nonlinear principal component analysis of noisy data
- 1 May 2007
- Published by Elsevier in Neural Networks
- Vol. 20 (4) , 434-443
Nonlinear principal component analysis of noisy data
- Open Access
- 12 April 2006
- Published by American Geophysical Union (AGU) in Geophysical Research Letters
- Vol. 33 (7)
Nonlinear atmospheric teleconnections
Neural network models are used to reveal the nonlinear winter atmospheric teleconnection patterns associated with the El Niño‐Southern Oscillation (ENSO) and with the Arctic Oscillation (AO) over the N. Hemisphere. The nonlinear teleconnections (for surface air temperature, precipitation, sea level pressure and 500 hPa geopotential height) are found to relate quadratically to the ENSO and AO indices. Relative to linear teleconnections, nonlinear teleconections appear to propagate perturbations farther, into regions where classical linear teleconnections are insignificant.
- 12 April 2006
- Published by American Geophysical Union (AGU) in Geophysical Research Letters
- Vol. 33 (7)
Open AccessNonlinear atmospheric teleconnections
Neural network models are used to reveal the nonlinear winter atmospheric teleconnection patterns associated with the El Niño‐Southern Oscillation (ENSO) and with the Arctic Oscillation (AO) over the N. Hemisphere. The nonlinear teleconnections (for surface air temperature, precipitation, sea level pressure and 500 hPa geopotential height) are found to relate quadratically to the ENSO and AO indices. Relative to linear teleconnections, nonlinear teleconections appear to propagate perturbations farther, into regions where classical linear teleconnections are insignificant.