Towards Data-Driven Large Scale Scientific Visualization and Exploration

Abstract

Technological advances have enabled us to acquire extremely large datasets but it remains a challenge to store, process, and extract information from them. This dissertation builds upon recent advances in machine learning, visualization, and user interactions to facilitate exploration of large-scale scientific datasets. First, we use data-driven approaches to computationally identify regions of interest in the datasets. Second, we use visual presentation for effective user comprehension. Third, we provide interactions for human users to integrate domain knowledge and semantic information into this exploration process. Our research shows how to extract, visualize, and explore informative regions on very large 2D landscape images, 3D volumetric datasets, high-dimensional volumetric mouse brain datasets with thousands of spatially-mapped gene expression profiles, and geospatial trajectories that evolve over time. The contribution of this dissertation include: (1) We introduce a sliding-window saliency model that discovers regions of user interest in very large images; (2) We develop visual segmentation of intensity-gradient histograms to identify meaningful components from volumetric datasets; (3) We extract boundary surfaces from a wealth of volumetric gene expression mouse brain profiles to personalize the reference brain atlas; (4) We show how to efficiently cluster geospatial trajectories by mapping each sequence of locations to a high-dimensional point with the kernel distance framework. We aim to discover patterns, relationships, and anomalies that would lead to new scientific, engineering, and medical advances. This work represents one of the first steps toward better visual understanding of large-scale scientific data by combining machine learning and human intelligence.