A data-driven channel infilling rate forecasting framework with quantified uncertainty for Southwest Pass and Houston Ship Channel

Abstract

Dredging operations in navigational waterways are essential to maintain channel depth to ensure the safe passage of vessels. Existing shoaling forecasting tools often rely on heuristics or direct modeling of cumulative sedimentation, limiting lead time and predictive reliability. In this work, we propose an end-to-end data-driven forecasting framework that models the rate of change in sedimentation volume, referred to as Channel Infilling Rate (CIR), 30 days ahead, which is then integrated to recover cumulative forecasts. Our approach starts with the calculation of CIR using the Corps Shoaling Analysis Tool, developed by the U.S. Army Corps of Engineers. Following that, we employ variance decomposition-based global sensitivity analysis to identify the most influential upstream streamflow gauges from a large set of candidate gauges, which serve as inputs for CIR prediction. To model the complex relationship between upstream river discharge and the CIR of a region of interest, we explore four data-driven modeling approaches including multivariate linear regression, the long short-term memory (LSTM) networks, a hybrid model architecture that combines Convolutional Neural Network (CNN) with LSTM network (CNN-LSTM), and a Temporal Convolutional Network (TCN) model combined with LSTM network (TCN-LSTM). Finally, the trained models are employed to predict the CIR with uncertainty quantified to provide confidence intervals of the predictions. Unlike previous studies that model aggregate sedimentation volume across entire systems, we perform reach-level modeling to better capture spatial heterogeneity and hydrodynamic variability. Demonstrated on the Houston Ship Channel (HSC) and Southwest Pass (SWP), this framework represents a step toward data-adaptive shoaling forecasts that can support proactive sediment management in complex riverine environments. By providing rolling short-term forecasts with uncertainty bounds, the proposed framework can support risk-aware sediment management and proactive dredging decision-making. The comparison between HSC and SWP further demonstrates that model selection should be adapted to local survey-data availability, with simpler models favored under sparse survey conditions and LSTM-based models favored when dense survey records preserve nonlinear sedimentation dynamics.