Doctoral Dissertations
Date of Award
5-2025
Degree Type
Dissertation
Degree Name
Doctor of Philosophy
Major
Data Science and Engineering
Major Professor
Jitendra Kumar
Committee Members
Jitendra Kumar, Forrest M. Hoffman, Monica Papeş, Qiusheng Wu
Abstract
The annual melting of river ice in Arctic regions presents a significant hazard, leading to ice jam flooding (IJF), ecosystem disruption, property damage, and risks to local communities. Accurately forecasting river ice breakup is essential for mitigating these risks, ensuring safe transportation, and deepening our understanding of Arctic river ecosystems.
This research first develops a deep learning-based approach to predict river ice breakup using meteorological and hydrological data. A long short-term memory (LSTM) model, trained on Daymet meteorological inputs and static hydrological data from the pan-Arctic catchment database (ARCADE), predicts breakup timing across 33 locations with an average error of 5.40 days. Additionally, seasonal forecast ensembles from the Copernicus Climate Data Store (CCDS) are incorporated to assess the model's capability in predicting future events. Shapley value analysis is employed to interpret the time-varying input contributions.
Building upon these findings, the study expands to 93 locations across Alaska and Canada and leverages CMIP6 and NEX-GDDP Earth system model (ESM) outputs to serve as inputs, allowing for future projections under four socio-economic pathway scenarios (SSPs). Results indicate that under more extreme warming scenarios, river ice breakup occurs earlier in the season, with the possibility that Arctic rivers may fail to freeze entirely in the distant future due to rising temperatures and increasing uncertainty in the region.
The role of atmospheric rivers (ARs) in river ice breakup is also investigated. ARs, capable of transporting vast amounts of moisture, have a notable impact on the Alaskan climate. Analysis across 26 locations reveals that ARs elevate local air temperatures for over a week, contribute to 40% of annual precipitation, and account for a majority of extreme precipitation events. A physics-based analysis shows that winter precipitation delays ice breakup, whereas precipitation near the breakup date has little effect.
This research contributes to improving river ice breakup forecasting and understanding its climate-driven variability. Future work will refine predictive models and explore additional climate drivers affecting Arctic hydrology.
Recommended Citation
Limber, Russ, "Deep Learning-Based Time Series Methods for Predictive Understanding of River Ice Breakup in North American Cold Regions. " PhD diss., University of Tennessee, 2025.
https://trace.tennessee.edu/utk_graddiss/12385
Included in
Applied Statistics Commons, Data Science Commons, Environmental Indicators and Impact Assessment Commons, Glaciology Commons, Longitudinal Data Analysis and Time Series Commons, Other Earth Sciences Commons