ORCID
0000-0002-8527-0231
Date of Award
2024
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Computer Science
First Advisor
Safwan Wshah
Abstract
The past decade has seen an explosion in the popularity of deep learning algorithms due to their uncontested performance in tasks including image classification, object segmentation, and natural language processing. However, their remarkable performance on these tasks has led researchers to narrowly focus their research towards applications for which deep learning is an easy fit. Much less research has been devoted to adapting deep learning models to new domains, such as geospatial research, due to challenges associated with incorporating geospatial reasoning directly into the architecture of models. This work will begin by providing a comprehensive survey of the most prominent geospatial domains researchers are currently adapting deep learning to. We observe that for many tasks, existing research adapts traditional computer vision models to these datasets without designing architectures specifically tailored to learn to reason about geospatial relationships. In this work, we outline a series of techniques for building learnable algorithms that incorporate geospatial relationships directly into their reasoning, in contrast to current approaches which develop fixed, hand-built algorithms to aggregate the output of computer vision models into a geospatial format.
This work will address three specific research problems. First, we will discuss traffic sign geo-localization, in which a key challenge is the development of an algorithm which re-identifies repeated occurrences of the same signs between images and geo-localizes their positions. Second, we discuss geospatial wetland segmentation, which involves predicting a spatial map indicating regions containing wetlands from drone imagery, LiDAR data, and a digital elevation model. Third, we discuss geographically-informed synthetic satellite image generation, which involves using a digital elevation model, infra-red bands, and weather data as conditioning information to generate a synthetic satellite imagery depicting the layer of a region.
Our methods result in models directly capable of reasoning about geospatial relationships in the dataset, resulting in improved performance compared to exclusively vision-based approaches. The core concepts we employ are not task-specific. Our modular approaches enable other researchers in similar fields to adapt our techniques to build deep learning models capable of learning to reason about geospatial relationships.
Language
en
Number of Pages
238 p.
Recommended Citation
Wilson, Daniel, "Geocentric Learning: Incorporating Learnable Geospatial Reasoning into Deep Learning Architectures" (2024). Graduate College Dissertations and Theses. 1997.
https://scholarworks.uvm.edu/graddis/1997