An ADCP-Based Data-Driven Framework for Proxy Sediment Transport Monitoring: From Controlled Flumes to Natural Rivers

Abstract. Acoustic Doppler Current Profilers (ADCPs) provide a rich yet underutilized source for monitoring hydrodynamics and sediment transport. Accurate prediction of sediment‐related variables is critical for river engineering, morphological studies, and environmental management. Among these, Bottom-Track Velocity (BT_Vel) serves as a robust proxy for near-bed sediment dynamics and bedload activity. This study develops a machine learning (ML) and deep learning (DL) framework to predict BT_Vel from ADCP-derived hydrodynamic and acoustic features, enabling proxy estimation of sediment transport processes in both controlled flume and natural riverine environments. Two datasets were analyzed: (i) a laboratory dataset of 22,650 ensemble samples obtained under controlled flow regimes, and (ii) a field dataset of 5,900 ensemble samples collected across seven campaigns at a fixed river cross-section. A consistent benchmarking strategy was applied across Random Forest, Gradient Boosting, LightGBM, CatBoost, XGBoost, LSTM, GRU, CNN, RNN, ANN, and a hybrid LSTM+CNN, with evaluation based on both an 80/20 split and a stratified 5-fold cross-validation (CV). SHAP analysis was conducted for model interpretability. In the laboratory, Random Forest (R² = 0.804 split / 0.783 CV) and Gradient Boosting (0.787 / 0.757) achieved the best generalization, while LSTM+CNN (0.770 / 0.730) and LSTM (0.775 / 0.718) remained competitive. In the field, Random Forest again delivered the strongest results (0.573 / 0.603), followed closely by CatBoost, LightGBM, and XGBoost. Notably, LSTM improved under cross-validation (0.468 → 0.529), suggesting fold-wise diversity stabilized training under noisy, heterogeneous river data. By contrast, the Stacking Regressor consistently showed the weakest generalization across both environments. SHAP revealed a shift in feature relevance: in the laboratory, Mean water velocity (Mean_Speed) dominated predictions, while in the field, Depth and signal-to-noise ratio (SNR) emerged as stronger drivers, reflecting the influence of stage variability and acoustic quality. Overall, the study demonstrates that ADCP-derived features, coupled with explainable ML/DL models, provide robust potential for proxy sediment transport modeling. Conversion to absolute transport rates requires paired sediment measurements, while future work should expand field campaigns and explore hybrid physics–data frameworks toward operational forecasting.