Dissertation
Bridging hydrological modeling and AI: advancing streamflow forecasting with multi-task learning and spatio-temporal data improvements
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Autumn 2025
DOI: 10.25820/etd.008197
Abstract
Floods remain among the most destructive natural disasters worldwide, intensified by climate change and rapid environmental variability. Reliable streamflow forecasting is essential for early warning, reservoir management, and water allocation, yet achieving accurate and transferable predictions remains a central challenge in hydrology. This dissertation advances data- and AI-driven approaches for hydrological prediction by improving data fidelity, enhancing model architectures, and moving toward unified learning frameworks that represent multiple hydrological processes within a single system.
The first phase of the research focuses on improving the spatial and temporal quality of environmental datasets through deep learning–based super-resolution methods. By enhancing Digital Elevation Models and radar rainfall datasets, these studies provide higher-fidelity inputs that strengthen the accuracy of hydrological modeling. The second phase of the research investigates advanced deep learning architectures—such as Transformers and state-space models—for streamflow forecasting across multiple basins and prediction horizons. These models demonstrate improved efficiency, generalization, and interpretability compared with traditional recurrent networks. Building upon these advancements, the final research phase focuses on developing a unified framework that leverages shared representations across tasks such as streamflow forecasting, soil-moisture estimation, and rainfall reconstruction.
Collectively, this work demonstrates how data enhancement and shared-learning frameworks can improve the fidelity, generalizability, and interpretability of hydrological prediction. The unified architecture represents a first step toward comprehensive, multi-process hydrological modeling, contributing to the development of scalable and physically consistent AI systems for streamflow forecasting and flood-risk management.
Details
- Title: Subtitle
- Bridging hydrological modeling and AI: advancing streamflow forecasting with multi-task learning and spatio-temporal data improvements
- Creators
- Bekir Z Demiray
- Contributors
- Ibrahim Demir (Advisor)Susan Meerdink (Committee Member)Bijaya Adhikari (Committee Member)Marian Muste (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Informatics
- Date degree season
- Autumn 2025
- DOI
- 10.25820/etd.008197
- Publisher
- University of Iowa
- Number of pages
- xiv, 289 pages
- Copyright
- Copyright 2025 Bekir Z Demiray
- Language
- English
- Date submitted
- 12/03/2025
- Description illustrations
- illustrations, tables, graphs, maps
- Description bibliographic
- Includes bibliographical references (pages 231-284).
- Public Abstract (ETD)
- Floods are among the most damaging natural disasters, threatening lives, infrastructure, and ecosystems around the world. Predicting river flow and flood potential is essential for protecting communities and managing water resources. This dissertation combines artificial intelligence (AI) and hydrology to improve how we forecast streamflow and understand the water cycle. The research focuses on three key areas. First, it improves the quality of environmental data by applying deep learning techniques to increase the resolution of maps and rainfall measurements. Second, it evaluates modern AI models to predict river discharge more accurately and efficiently across different regions. Third, it explores a new approach called multi-task learning, where one model learns several related tasks, such as predicting both streamflow and soil moisture, at the same time. Together, these studies demonstrate how AI can help scientists create more reliable, transparent, and transferable hydrological models. The improved datasets, predictive tools, and unified modeling frameworks developed in this work have the potential to enhance flood forecasting systems and support data-driven decision-making for climate adaptation and water management.
- Academic Unit
- IDGP in Informatics
- Record Identifier
- 9985135147302771
Metrics
1 Record Views