Assessing the Limits of Improving Subseasonal Predictability Indices

The goal of this doctoral thesis is to identify new sources of subseasonal predictability for temperature over the United States. Historically, distinct components of large-scale climate variability were identified using a variety of methods, and their impact on predictability was later derived using knowledge of that variation. By focusing on identifying components of variability first, have we overlooked some sources of predictability? In the first part of this thesis, we identify predictability in temperature by calculating the lagged correlations in temperature fields over the United States using Canonical Correlation Analysis (CCA). This lets us identify the predictability in temperature without knowing what causes this predictability. Then, when we know the predictable temperature signal we can investigate the source of the predictability using the same widely used methods in predictability studies. We examine predictability at weeks 1-2 and, separately, at weeks 3-4. Because the El Nino Southern Oscillation (ENSO) has a strong affect on subseasonal predictability, if it is not removed CCA will focus on the temperature response to ENSO, obscuring other predictable signals. Therefore, the ENSO signal is removed from the temperature data by removing the seasonal mean prior to any analysis. We identify several modes of predictability at weeks 1-2 and weeks 3-4 for all seasons. Several of these modes are independent of known sources of predictability. The sources of these new modes are investigated. The CFSv2 reforecasts are analyzed to see if they can capture the identified predictable patterns; in many cases it is able to, but in some cases it cannot. This thesis also introduces a practical advance in CCA, particularly for verifying canonical correlations in independent data. Since the predictability identified in the first part of the thesis is independent of ENSO, in the second part of this thesis we return to ENSO. A commonly used index of ENSO is the Nino 3.4 index, which is the area average surface temperature over a particular region of the tropical Pacific. However, this index was defined in the 1990s based on available observations and was never intended to be the optimal predictor for subseasonal climate. Accordingly, in the second part of this thesis, we attempt to find more useful indices of subseasonal predictability. To do this, machine learning algorithms were trained on observed SSTs, but the resulting predictions were worse than a simple prediction based on the Nino 3.4 index. To make a more skillful model, the machine learning algorithms were trained on the SST of long climate simulations and verified on observations. Ultimately, the skill of the best machine learning models are only modestly better than ordinary least squares based on the Nino 3.4 index. These results provide a cautionary tale about the potential of machine learning to discover new sources of predictability. In the first place, machine learning algorithms were not able to produce better predictions than simple linear regression using the Nino 3.4 index. Second, when the predictors are correlated, very different regression coefficients can produce virtually identical predictions, making interpretation difficult or misleading. The models based on machine learning were also compared to predictions from the CFSv2 model, a fully coupled dynamical model. Even though the CFSv2 includes ocean, land, and atmospheric components, its skill in predicting the ENSO-forced pattern is comparable to the machine learning models. Although the best predictions come from machine learning models trained on long climate simulations, the skill is only modestly better than predictions based on the Nino 3.4 index alone.

Author

• Buchmann, Paul

• Canonical Correlation Analysis
• CCA
• ML
• subseasonal predictions
• week 3-4

Identifier

hdl:1920/14087

OAI identifier oai:identifier

oai:MARS:1920/14087