Geospatial machine learning

Stewart, Adam J.

Permalink

https://hdl.handle.net/2142/121934 Copy

Description

Title

Geospatial machine learning

Author(s)

Stewart, Adam J.

Issue Date

2023-08-22

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

Banerjee, Arindam

Doctoral Committee Chair(s)

Banerjee, Arindam

Committee Member(s)

• Lazebnik, Svetlana
• Hoiem, Derek W
• Robinson, Caleb

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• machine learning
• remote sensing
• PyTorch
• geospatial
• deep learning
• software library
• dynamic topography
• digital agriculture
• self-supervised learning

Abstract

Geospatial datasets provide a rich source of unique problems of interest to the machine learning community, including multispectral and hyperspectral imagery, the time series nature of most datasets, and large scale applications for semantic segmentation and change detection. At the same time, machine learning is capable of making important advancements in a myriad of geospatial applications, including climate change, natural disaster monitoring, precision agriculture, and geodynamics. However, due to the complexities of working with geospatial data, this research domain is often underexplored. In this dissertation, we discuss progress made at the intersection of machine learning and remote sensing. In particular, we highlight the design and development of TorchGeo, a PyTorch domain library for working with geospatial data. TorchGeo is the first machine learning library to provide foundation models pre-trained on multispectral imagery. We also introduce SSL4EO-L, the first Landsat dataset designed for self-supervised learning and the largest Landsat dataset in history. We describe the process by which we create the first foundation models pre-trained on the Landsat family of satellites and benchmark their transfer learning capabilities on novel land cover mapping datasets. We also demonstrate the success of machine learning algorithms applied to a number of geospatial applications. The first application centers around detecting center pivot irrigation in the Midwestern United States using satellite imagery. Using various time-series modeling approaches, we develop a U-Net-based approach capable of detecting irrigated farmland with an overall accuracy of 86%. The second application involves plate tectonic modeling for the purpose of seafloor bathymetry estimation. By removing the topographic signal predicted by crustal properties, we introduce an objective method for estimating dynamic topography.

Graduation Semester

2023-12

Type of Resource

Thesis

Copyright and License Information

Copyright 2023 Adam Stewart

Owning Collections