Unveiling Sediment Deposition Thresholds in Discontinuous Emergent Vegetation Patches Using a SHAP-Interpretable Ensemble Machine Learning

Predicting sediment deposition in vegetated channels is challenging due to nonlinear flow-vegetation-sediment interactions that are poorly captured by empirical models. While physics-based equations (e.g., Saint-Venant, advection-diffusion) govern these processes, their application to discontinuous vegetation patches faces parameterization difficulties and computational demands. This study introduces a novel interpretable ensemble machine learning (ML) framework as a complementary approach, integrating novel feature engineering that defines spatial positioning relative to vegetation patches and vegetation-induced hydrodynamic effects. Five ensemble algorithms, including random forest (RF), gradient boosting (GB), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM) and categorical boosting (CatBoost), were evaluated using 498 observations from 11 laboratory flume experiments under varying hydraulic and vegetation conditions. Bayesian hyperparameter optimization with 5-fold cross-validation ensured robust hyperparameter tuning. A multi-criteria ranking for selecting top performers, with sensitivity analysis showed CatBoost and LightGBM consistently ranking highest. Instead of focusing solely on prediction, the discovery of a quantitative physical threshold is targeted. Through Shapley additive explanations (SHAP) analysis, a previously undocumented threshold between deposition-promoting and deposition-inhibiting regimes is identified and physically interpreted. CatBoost achieves the highest generalization (unseen holdout: R² = 0.953). SHAP reveals a crisp threshold at ≈ 0.3 m "cumulative vegetation exposure" , which aligns with the theoretical adjustment length scale 1/( C D ·a v ) for emergent canopies. "Local flow velocity" shows consistently weak influence (mean |SHAP| = 0.049 to 0.063), indicating that it is not a primary control within dense patches. This study demonstrates that interpretable ML can generate testable physical hypotheses from existing experimental data.