An integrated SWJ-LSTM-ETS modeling strategy for investigating upper morphology dynamics of a stochastic laboratory delta with environmental changes

Abstract

Accurately modeling delta morphological evolution remains challenging because of the complexity of interacting processes (e.g., runoff, wave and tidal flow). This study develops an integrated framework combining a shallow water jet (SWJ) based model, long short-term memory (LSTM) neural networks, and exponential smoothing (ETS) time series models to predict the cross-sectional topographic dynamics of a stochastic experimental delta's upper plain. The integrated approach enhances explanatory power compared to individual models by assimilating mechanistic and empirical techniques. Hindcasting and forecasting under vegetation and intermittent discharge impacts yield insights aligning with validated findings on fluvio-morphological responses. Notably, new interactions are revealed like discharge inducing central deposition while vegetation drives bi-lateral erosion near the inlet. Farther downstream, discharge transitions result in alternating and pronounced patterns of erosion and deposition, with abrupt changes in discharge having a greater impact than gradual ones, particularly considering the mitigating effect of vegetation on these dynamics. This demonstrates an effective strategy widely applicable to integrating physics-based and data-driven models for analogous environmental systems involving both deterministic and random elements. The key contributions are developing a versatile integrated modeling methodology through synergizing physical simulators, statistical learners, and stochastic representations. This provides enhanced capabilities for elucidating complex delta dynamics.