
The Library
Improving neural networks for geospatial applications with geographic context embeddings
Tools
Klemmer, Konstantin (2022) Improving neural networks for geospatial applications with geographic context embeddings. PhD thesis, University of Warwick.
|
PDF
WRAP_Theses_Klemmer_2022.pdf - Submitted Version - Requires a PDF viewer. Download (17Mb) | Preview |
Official URL: http://webcat.warwick.ac.uk/record=b3821873
Abstract
Geospatial data sits at the core of many data-driven application domains, from urban analytics to spatial epidemiology and climate science. Over recent years, ever-growing streams of data have allowed us to quantify more and more aspects of our lives and to deploy machine learning techniques to improve public and private services. But while modern neural network methods offer a flexible and scalable toolkit for high-dimensional data analysis, they can struggle with the complexities and dependencies of real-world geographic data. The particular challenges of geographic data are the subject of the geographic information sciences (GIS). This discipline has compiled a myriad of metrics and measures to quantify spatial effects and to improve modeling in the presence of spatial dependencies. In this dissertation, we deploy metrics of spatial interactions as embeddings to enrich neural network methods for geographic data. We utilize both, functional embeddings (such as measures of spatial autocorrelation) and parametric neural-network embeddings (such as semantic vector embeddings). The embeddings are then integrated into neural network methods using four different approaches: (1) model selection, (2) auxiliary task learning, (3) feature learning, and (4) embedding loss functions. Throughout the dissertation, we use experiments with various real-world datasets to highlight performance improvements of our geographically-explicit neural network methods over naive baselines. We focus specifically on generative and predictive modeling tasks. The dissertation highlights how geographic domain-expertise together with powerful neural network backbones can provide tailored, scalable modeling solutions for the era of real-time Earth observation and urban analytics.
Item Type: | Thesis (PhD) | ||||
---|---|---|---|---|---|
Subjects: | G Geography. Anthropology. Recreation > G Geography (General) Q Science > QA Mathematics Q Science > QA Mathematics > QA76 Electronic computers. Computer science. Computer software |
||||
Library of Congress Subject Headings (LCSH): | Geospatial data, Neural networks (Computer science), Geographic information systems, Embeddings (Mathematics) | ||||
Official Date: | January 2022 | ||||
Dates: |
|
||||
Institution: | University of Warwick | ||||
Theses Department: | Department of Computer Science | ||||
Thesis Type: | PhD | ||||
Publication Status: | Unpublished | ||||
Supervisor(s)/Advisor: | Neill, Daniel ; Wen, Hongkai ; Jarvis, Stephen | ||||
Sponsors: | Engineering and Physical Sciences Research Council ; University of Warwick. Centre for Doctoral Training in Urban Science and Progress ; Alan Turing Institute | ||||
Format of File: | |||||
Extent: | xviii, 161 leaves : colour illustrations, charts | ||||
Language: | eng |
Request changes or add full text files to a record
Repository staff actions (login required)
![]() |
View Item |
Downloads
Downloads per month over past year