2026-05-24T01:13:56Zhttps://keep.lib.asu.edu/oai/request

oai:keep.lib.asu.edu:node-1953442024-12-23T18:01:48Zoai_pmh:alloai_pmh:repo_items

195344 https://hdl.handle.net/2286/R.2.N.195344 http://rightsstatements.org/vocab/InC/1.0/ All Rights Reserved 2024 136 pages Doctoral Dissertation Academic theses Text eng Chowdhury, Kanchan Sarwat, Mohamed Zou, Jia Davulcu, Hasan Bryan, Chris Arizona State University Partial requirement for: Ph.D., Arizona State University, 2024 Field of study: Computer Science Geospatial machine learning (ML) models and their applications have recently gained significant attention due to the rising availability of raster and spatiotemporal datasets. Three important limitations in ML for the geospatial domain are the following. Firstly, real-world geospatial datasets are often too large, and many geospatial ML algorithms represent the geographical region in terms of a grid. If the granularity of the grid is too fine, it results in a large number of grid cells, leading to long training time and high memory consumption issues during the model training. Secondly, current machine learning systems are mainly designed for text, image, audio, and video data, and they often fall short of adequately supporting geospatial datasets. This is because machine learning and data preprocessing techniques in this domain fail to capture the spatial autocorrelation property, a key characteristic available in geospatial datasets. Thirdly, many real-world inference workflows in this domain involve preprocessing steps that join data from multiple data silos to assemble feature vectors. Often, these geospatial joins are expensive and become bottlenecks in the inference process. In this dissertation, I will discuss novel solutions to these three major concerns of spatiotemporal machine learning. In particular, the dissertation includes three main research components. The first one presents a machine learning-aware technique for re-partitioning geospatial data to shorten the training duration of spatial machine learning models; the second component introduces an end-to-end framework for deep learning and data preprocessing with spatiotemporal vector and raster dataset; the third solution presents a strategy to co-optimize a preprocessing and inference pipeline consisting of costly join queries and model inferencing. Additionally, I will present experimental evaluation results using a variety of real-world datasets to demonstrate the effectiveness of all three solutions. Computer Science Geographic Information Science and Geodesy Database Systems Deep learning Geospatial data Neural Networks Spatial Machine Learning Spatiotemporal Data Exploration of Location-Aware Machine Learning for Spatiotemporal Vector and Raster Datasets