Boosting, Support Vector Machines, and Cloud Detection over Ice and Snow

University Of California-Berkeley, Berkeley CA

Investigators

Abstract

DMS-0306508 Bin Yu Boosting, support vector machines, and cloud detection over ice and snow Abstract Classification and regression are two fundamental statistical problems posed by the Information Technology (IT) age and at the same time aided by the computational tools it provides. Boosting and Support Vector Machines (SVMs) are two revolutionary methodologies from machine learning for regression and classification that meet the needs of massive data sets of our IT age. The investigator and her colleagues study when and why boosting and SVMs work to shed light on the design or tuning of Boosting and SVMs; in particular, understanding on the choice of the base learner in L2Boosting and the effect of the input density distribution on the Reproducing Kernel Hilbert space induced by the kernel in an SVM. Theoretical understandings on boosting and SVMs are then used, in collaboration with colleagues at Jet Propulsion Laboratory, in the cloud-detection problem over ice/snow based on Multi-angle Imaging SpectroRadiometer (MISR) data. The investigator and her colleagues also develop a novel cloud detection algorithm based on linear correlations using MISR data and compare this with the boosting and SVM based approaches. Information Technology (IT) is changing just about every facet of our lives. At a professional level, emerging innovations in IT areas represent tremendous opportunities for statistical research, providing both fundamental methodological challenges as well as applications with real-world impact. Advances in data collection and computing technologies have led to the proliferation of massive data sets such as those from remote sensing. Understanding the role of statistics in such data-rich applications forces us to reevaluate and revise traditional procedures and frameworks. The investigator and colleagues study Boosting and Support Vector Machines (SVMs), which are two revolutionary methodologies from machine learning for regression and classification that meet the needs of massive data sets of our IT age. In collaboration with colleagues at NASA's Jet Propulsion Laboratory, they apply the research results from these studies to the problem of cloud detection, which is a crucial step in any climate prediction or modeling including weather forecasting and global warming monitoring. Multi-angle Imaging SpectroRadiometer (MISR) was launched in 1999 by NASA to provide 9 angle (4 band) data to retrieve or estimate the cloud height and hence cloud detection. However, cloud detection even with MISR data has been proven very difficult over ice/snow. The investigator and her colleagues develop a novel cloud detection algorithm based on linear correlations using MISR data to work over ice/snow and compare it with boosting and SVM based approaches.