Authors: Dan Ames and Sujan Mondol – Brigham Young University
Title: Forecasting River Ice Breakup in the Northeast RFC Domain using Machine Learning
Presentation Type:
Abstract: River ice dynamics are a fundamental component of hydrologic processes in cold regions, with major implications for streamflow, fluvial morphology, aquatic ecosystems, water resources, and infrastructure. In particular, spring river ice breakup, which typically occurs in April or May depending on latitude and elevation, can increase river discharge and trigger erosion, ice jam formation, and flooding. Despite its importance, river ice is not represented in the current National Water Model (NWM). This gap largely stems from the complexity of river ice processes, which involve highly nonlinear interactions among hydrodynamics, thermodynamics, channel morphology, and atmospheric forcing across multiple spatial and temporal scales. As a result, physics-based modeling of river ice
dynamics remains computationally demanding, and difficult to implement operationally at large scales. These challenges are further compounded by the scarcity of continuous in situ river ice observations needed for model development, calibration, and validation. While physics-based models provide valuable process understanding and can be used to predict breakup events, machine learning offers a complementary data-driven framework that can learn complex nonlinear relationships from historical hydroclimatic and remote sensing data, making it especially promising for capturing the spatial and temporal variability of breakup processes when physical parameters are uncertain or unavailable. To address these challenges, we develop a machine learning framework for predicting river ice breakup from historical hydroclimatic, hydrologic, and geospatial data. The methodology leverages spatiotemporal predictors describing antecedent meteorological conditions, snowpack evolution, thermal forcing, watershed attributes, and river states extracted from NWM retrospective data at monitored locations to train data-driven models of breakup timing and occurrence. By learning from multi-year and multi-site observations, the framework captures the nonlinear controls and regional variability of breakup processes without requiring full physical parameterization of river ice dynamics. This enables prediction in data-sparse environments where continuous in situ ice observations are limited and physics-based modeling is difficult to apply operationally. The proposed approach is designed to provide a scalable complement to existing hydrologic forecasting systems and to support improved streamflow prediction and river ice hazard assessment in cold regions.