Blume, Christian (2012) Statistical Learning to Model Stratospheric Variability. (PhD/ Doctoral thesis), Freie Universität Berlin, Berlin, Germany, 145 pp.

Preview

Text dissertation_online.pdf - Published Version Download (5MB) | Preview

Abstract

The Variability of the stratosphere is crucial for the evolution of the Earth-climate system as a whole. Stratospheric variability on various time scales is influenced by a number of forcings, such as the Quasi-Biennial Oscillation, the El Nino- Southern Oscillation, the 11-yr solar cycle, or volcanic eruptions, that interact to create a complex system. This link is particularly nonlinear during winter when planetary waves can propagate upward to interact with the stratospheric mean flow. Most commonly, sophisticated chemistry-climate models simulate stratospheric variability, driven by the interactions between dynamics, radiation, and chemistry. However, climate models are computationally expensive and quantifying the importance of forcing factors is difficult. In contrast, statistical methods are mathematically simpler, computationally less expensive, and weight forcing factors according to their importance. Statistical methods learn variability patterns from historical data and can potentially forecast these patterns into the future.
For the first time, a wide class of statistical methods is used in this work to model stratospheric variability in data from observations, reanalyses, and model simulations. The statistical methods are partly nonlinear and nonstationary making them appropriate to cope with the complex feedbacks that govern the stratosphere. These advanced methods, along with a standard linear method, are compared with respect to their ability to model stratospheric variables on different temporal and spatial domains. The considered methods are linear discriminant analysis (LDA), a cluster method based on finite elements (FEM-VARX), a neural network, namely the multi-layer perceptron (MLP), and the support vector machine (SVM). It is shown how an optimal, method-specific set of tuning parameters is estimated using information criteria along with cross-validation.
A prominent example of dynamical wave-mean flow interactions during winter are sudden strato- spheric warmings (SSWs). SSWs are dramatic extreme events characterized by a great temperature increase on daily time scales and a breakdown of the polar vortex. While the resulting anomalies can descend downward and provide predictive skill for tropospheric weather conditions, forecasting SSWs themselves remains a difficult task. It is shown in this work that polar stratospheric variability can be modeled and forecasted using nonlinear and nonstationary statistical methods while incorporating all significant forcing factors. Moreover, an approach based on a nonlinear neural network is presented that can classify SSWs in major, minor, and final warmings for the recent climate. The statistical importance of the forcing factors and their nonlinear interrelationships are estimated. In addition, global stratospheric temperature and ozone are statistically modeled due to their specific importance for indicating changes in dynamics and composition. The four statistical methods are used to quantify the natural variability inherent in the stratosphere so that the impact of anthropogenic forcings can be attributed appropriately. Considering various data sets along with the different independent statistical methods makes it feasible to estimate robust uncertainties. Using the statistical methods, variability in temperature and ozone is successfully forecasted up to the year 2100. It is shown in this work that the standard linear method leads to robust results on the monthly scale but is clearly outperformed by the advanced methods on the daily scale.

Actions (login required)

View Item