Machine learning models on geographic spatial-temporal data predictions

Li, Yanye

Permalink

https://hdl.handle.net/2142/115632 Copy

Description

Title

Machine learning models on geographic spatial-temporal data predictions

Author(s)

Li, Yanye

Issue Date

2022-04-29

Director of Research (if dissertation) or Advisor (if thesis)

Brunner, Robert J

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

Machine Learning, Time Series, Spatial-temporal Data

Abstract

Geographic data was not a primary area for early machine learning research. But just as computers rapidly became important tools in radiology, financial trading, and other fields that require fast, highly accurate prediction-based work, machine learning is also showing its ability to push the limits in geospatial data prediction in a very short period. Furthermore, many geographic data analysis include the time dimension to accommodate the temporal dependencies of observations since they often desire to quantify certain changes in environments or landscapes. This added dimension often makes machine learning predictions much harder. In general, it is common for scientists to migrate models used in speech processing, such as recurrent neural networks, to geographic spatial-temporal datasets because the knowledge about temporal dependencies can relatively easily be applied in a similar manner. This thesis first introduces simple regression models and discusses the special considerations required for the three-dimensional data, and this thesis also introduces a state-of-the-art deep learning method, spatial-temporal neural network (STNN), together with its variations. STNN is a specialized recurrent neural network that aims to learn from a series of observations that share both spatial and temporal interactions. We implement these models and compare their performances on experimental results from two different geographic spatial-temporal datasets. Both of the datasets are representative of predictions works in geographic information science, although they differ in some characteristics such as size, timescale, and reversibility. In the end, the comparison leads to a discussion on different strategies of learning and potential improvement.

Graduation Semester

2022-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/115632

Copyright and License Information

Copyright 2022 Yanye Li

Owning Collections