Speaker: Bruno Sanso
Abstract: We consider the problem of fitting a statistical model to historical records of sea surface temperatures collected sparsely in space and time. The records span the whole of the last century and include the Atlantic Ocean north of the Equator. The purpose of the model is to produce and atlas of sea surface temperatures. This consists of climatological mean fields, estimates of historical trends and a spatio-temporal reconstruction of the anomalies, i.e., the transient deviations from the climatological mean. Our model improves upon the current tools used by oceanographers in that we account for all estimation uncertainties, include parameters associated with spatial anisotropy and non-stationarity, transient and long-term trends, and location-dependent seasonal curves. Additionally, since the data set is composed of four types of measurements, our model also includes four different observational variances. The model is based on discrete process convolutions and Markov random fields. Particular attention is given to the problem of handling a massive data set. This is achieved by considering compact support kernels that allow an efficient parallelization of the Markov chain Monte Carlo method used in the estimation of the model parameters. The model is based on a hierarchical structure that is physically sound, it is easily parallelizable and provides information about the quantities that are relevant to the oceanographers together with uncertainty bounds. The data set is sufficiently large and the area sufficiently complex and extended to serve as a good testbed for global applications.
This seminar is co-hosted by Statistical Sciences (CCS-6) and the UCSC/LANL Institute for Scalable Scientific Data Management, an Information Science and Technology Institute (ISTI) collaborative university partnership.