Date of Award
2025
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Atmospheric and Earth Science
Committee Chair
Sundar Christopher
Committee Member
Sean Freeman
Research Advisor
Christopher Hain
Subject(s)
Soil moisture--Measurement--Computer simulation, Deep learning (Machine learning), Neural networks (Computer science)
Abstract
This work examines the ability of deep learning time series generative models to accurately and efficiently emulate the hourly temporal dynamics of the Noah Land Surface Model (Noah-LSM) out to a 2 week forecast horizon, given atmospheric forcings and static parameterization provided by the second phase North American Land Data Assimilation System (NLDAS-2) framework. Results from multiple neural network architectures are compared alongside variations in prediction target, loss function characteristics, and model properties. The most performant model types are subsequently evaluated with respect to forecast distance, annual seasonality, and against a variety of regional scenarios, including several event case studies. Ultimately, we present a software system and suite of evaluation techniques for developing and testing neural networks that use time-varying and static data to estimate temporal dynamics, with the goal of providing a foundation for similar data-driven modeling techniques to be implemented within the upcoming third phase of the NLDAS data record.
Recommended Citation
Dodson, Mitchell T., "Deep learning time series prediction strategies for efficiently emulating Noah Land Surface Model soil moisture dynamics" (2025). Theses. 773.
https://louis.uah.edu/uah-theses/773