A geospatial deep learning framework for scalable hydrographic mapping

Jaroenchai, Nattapon

Permalink

https://hdl.handle.net/2142/132676 Copy

Description

Title

A geospatial deep learning framework for scalable hydrographic mapping

Author(s)

Jaroenchai, Nattapon

Issue Date

2025-12-04

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

Wang, Shaowen

Doctoral Committee Chair(s)

Wang, Shaowen

Committee Member(s)

• Banerjee , Arindam
• Diao , Chunyuan
• Wang , Jida

Department of Study

Geography & GIS

Discipline

Geography

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Hydrography
• Deep Learning
• Geospatial AI
• Transfer Learning
• Meta-Learning
• Multimodal Learning
• Multitask Learning
• Streamline Delineation
• Digital Elevation Models
• CyberGIS

Abstract

Accurate and up-to-date hydrographic mapping is essential for managing water resources, mitigating flood risk, and supporting disaster management and environmental sustainability. In the United States, the National Hydrography Dataset (NHD) provides the authoritative spatial framework for surface water representation, yet its reliance on manual interpretation and rule-based workflows limits the timeliness and consistency of updates across diverse landscapes. Traditional hydrologic mapping methods based on Digital Elevation Models (DEMs) are constrained by data quality, subjective parameterization, and limited scalability. Recent advances in deep learning and geospatial artificial intelligence (AI) present new opportunities to automate hydrography extraction by learning spatial and physical patterns directly from data. However, challenges remain in achieving spatial transferability, adaptability to new environments, and scalability to national applications. This dissertation develops a geospatial deep learning framework for scalable hydrographic mapping that addresses these three interrelated challenges through a progressive, multi-stage design. The first study applies transfer learning to improve model generalization across regions by fine-tuning pretrained convolutional neural networks for hydrographic feature extraction. The second study employs model-agnostic meta-learning to enhance adaptability, enabling rapid fine-tuning of segmentation models for previously unseen watersheds with limited labeled data. The third study integrates multimodal and multitask learning to achieve scalability and geospatial realism by fusing heterogeneous geospatial data—such as DEM, optical, thermal, and radar data—with auxiliary hydrologic tasks including flow-direction prediction. Across these studies, the framework demonstrates improved spatial generalization, accelerated adaptation to new regions, and enhanced model robustness under diverse physiographic conditions. The integration of cyberGIS-enabled high-performance computing resources, including the Taylor Geospatial Institute Regional AI Learning System (TGI RAILS) and Virtual ROGER, ensures scalable training, reproducibility, and transparent experiment tracking. Collectively, this dissertation advances the automation of hydrographic mapping by uniting transferable, adaptive, and scalable AI paradigms within a geospatial informed design. The outcomes contribute to the modernization of the NHD and establish a foundation for future large-scale, data-driven hydrography modeling in support of sustainable water resource management and disaster management.

Graduation Semester

2025-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/132676

Copyright and License Information

© 2025 Nattapon Jaroenchai

Owning Collections