2025 CSDMS meeting-027
From CSDMS
(if you haven't already)
Log in (or create account for non-CSDMS members)
Forgot username? Search or email:CSDMSweb@colorado.edu
Browse abstracts
Accelerate Bed-Change Simulations Using Spatiotemporal Deep Learning Framework
Mohamed Fathi Said,
Florida Gulf Coast University Fort Myers Florida, United States. m.fathi.said0@gmail.com
Virginia Smith, Villanova University Villanova Pennsylvania, United States. virginia.smith@villanova.edu
Zihan Liu, Villanova University Villanova Pennsylvania, United States. zliu5@villanova.edu
Anjali M. Fernandes, Denison University Granville Ohio, United States. fernandesa@denison.edu
Michael T. Hren, University of Connecticut Storrs Connecticut, United States. michael.hren@uconn.edu
Dennis O. Terry, Temple University Philadelphia Pennsylvania, United States. dennis.terry@temple.edu
C. Nataraj, Villanova University Villanova Pennsylvania, United States. c.nataraj@villanova.edu
Traditional sediment transport models are constrained by their spatial and/or temporal resolution, requiring high computational power and consuming extensive processing time. This limitation poses a substantial challenge in modeling two-dimensional landscape evolution over long-term scales and in predicting the impacts of system dynamics, such as climate change. To address this shortcoming, we introduce a novel Deep Learning (DL) framework that integrates Convolutional Neural Networks (CNN) with Long Short-Term Memory (LSTM) to capture the spatiotemporal morphodynamics of rivers in response to flooding events. This framework combines three models to simulate essential hydrodynamic and morphodynamic features: water depth, flow velocity, and bed change, which operate in a continuous loop to ensure dynamic updates to bed topography. The proposed framework was trained and evaluated using a dataset generated by HEC-RAS, a physics-based model, for a 22 km segment of Ninnescah River in Kansas. The hydrodynamic results demonstrate proficiency in capturing the flood dynamics, where the average Root Mean Square Error (RMSE) across the testing dataset is 0.19 m and 0.04 m/s for water depth and flow velocity, respectively. These hydrodynamic features are essential for the bed-change model, which exhibits high accuracy with a normalized RMSE and an R2 of 27% and 0.93, respectively, at the end of the testing dataset. Furthermore, the trained framework can generate predictions 4700 times faster than HEC-RAS. This work signifies a paradigm shift in the long-term simulation of river evolution and sets the stage for exploring new frontiers in fluvial morphodynamic modeling.
