Models and monitoring designs for spatio-temporal climate data fields

• Main Author: Casquilho Resende, Camila Maria
• Language: English
• Published: University of British Columbia 2016
• Online Access: http://hdl.handle.net/2429/59293

In this thesis, we describe how appropriately modelling the spatio-temporal mean surface can help resolve complex patterns of nonstationarity and improve spatial prediction. Nonstationary fields are common in environmental science applications, and even more so in regions with complex terrain. Our analyses focus on the Pacific Northwest, a region where rapid changes and localized weather are very common, and where the terrain plays an important role in separating often radically different climate and weather regimes. To this end, we introduce two comparable strategies for spatial prediction. The first is based on a Bayesian spatial prediction method, where an exploratory analysis was performed in order to better understand the localized weather regimes. The other is based on tackling the anomalies of expected climate in the Pacific Northwest region, based on the average values of temperature computed over a 30-year range obtained through a climate analysis system. Secondly, we focus on one of the recent challenges in spatial statistics applications, the data fusion problem. There has been an increased need for combining information from multiple sources that may be on different spatial scales. Ensemble modelling is referred to as a statistical post-processing technique based on combining multiple computer model outputs in a statistical model with the goal of obtaining probabilistic forecasts. We give an overview of some ensemble modelling strategies, by combining observed temperature measurements with outputs from an ensemble of deterministic climate models. We also provide a comparison between the Bayesian model averaging approach and a dynamic Bayesian ensemble strategy for forecasting. Finally, we introduce a novel strategy for the design of monitoring network, where the goal is to select a high-quality yet diverse set of locations. The idea of spatial repulsion is brought to this context via the theory of determinantal point processes. Our design strategy is not only able to yield spatially-balanced designs, but it also has the ability to assess similarity between the potential locations should there be extra sources of information related to the underlying process of interest. We explore its relationship to existing design methods, such as the entropy-based and space-filling designs. === Science, Faculty of === Statistics, Department of === Graduate