Determining Change Points in Balloon Based Measurements of the Atmosphere

08/01/2013 - 3:45am to 4:05am
ML - Main Seminar Room
Joseph Usset


Radiosonde measurements form the backbone of our knowledge of the vertical structure of the atmosphere. Besides their use in weather prediction, longer records can also add to our knowledge of climate. These data are gathered from a heterogenous and global network of measurement sites but can contain artificial discontinuities, also known as change points, due to new technology, measurement error, or changes in observation location and time. Because a change point can influence the quality of a record it is important to detect these features, and the large size of the radiosonde archive makes automated methodologies attractive.  For this project, several  automated methods are compared for the identification of change-points. From a statistical perspective, these methods attempt to identify points where there exists an underlying distributional change in in a sequence of random variables. The first method, developed by Lazante (used by the US National Climate Data Center), is a rank-based technique that seeks to be robust to outliers. The second method, known as binary segmentation, is widely used outside atmospheric science, and follows a traditional penalized likelihood framework. We also propose a hybrid method that applies the binary segmentation algorithm to the ranked data points used by Lazante. All methods are evaluated via simulation, and their viability is demonstrated with numerous applications to real data.