Loading...
Enabling massive spatial data analysis
Blake, Lewis Rae
Blake, Lewis Rae
Citations
Altmetric:
Advisor
Editor
Date
Date Issued
2022
Date Submitted
Collections
Research Projects
Organizational Units
Journal Issue
Embargo Expires
Abstract
Modern spatial data sets generated by remote sensing instruments and numerical models are often noisy, exhibit complex dependence structures, and have a massive number of observations rendering them computationally infeasible for traditional methods. Even if computations were feasible, common modeling practices can hinder the scientific utility of these data by imposing unrealistic simplifying assumptions such as stationarity. In this thesis we address some key computational and methodological problems that arise in the analysis of massive spatial data sets on a global scale. The challenges addressed include (i) computing with data input sizes for which traditional techniques are computationally prohibitive, and (ii) developing nonstationary methods to accurately analyze and predict scientific processes on challenging global domains with precise uncertainty quantification. For the former, we develop and study a series of sophisticated algorithms for analyzing massive spatial data that adapt to available computational resources ranging from personal computers to state-of-the-art high-performance computing systems. We study our implementations' performance with massive satellite data with more than 47 million observations under various parallel computing settings. For the latter, this thesis develops methods to analyze highly nonstationary satellite sea surface temperature (SST) data collected by the Moderate Resolution Imaging Spectroradiometer (MODIS) with more than 43 million observations. We derive and implement a nonstationary covariance model and compare it to a stationary model, with the nonstationary model significantly outperforming the latter in terms of both point predictions and uncertainty quantification. Our approach provides a method for producing a high-resolution SST product in near real-time and can be easily adapted to other environmental fields of interest for applications in oceanography, atmospheric science, climate science, and beyond.
Associated Publications
Rights
Copyright of the original work is retained by the author.