Climate-Driven Linear and Non-Linear Sediment Dynamics: A Machine-Learning Approach

Climate variability affects sediment dynamics; however, the relative contributions of individual climate parameters and the nature of these relationships (linear versus non-linear) remain unexplored. This knowledge gap hinders the development of effective, climate change-adaptive water quality management strategies. This study develops a multi-model framework to identify the complex relationships between climate parameters and sedimentary parameters (Turbidity and Total Dissolved Solids). Two modelling approaches, multivariate linear regression (MLR) as a baseline, and Random Forest (RF), were compared to capture both linear and nonlinear sediment dynamics. The SHAP (Shapley Additive exPlanations) method is used to quantify the contributions of climate parameters in variations of turbidity and total dissolved solids concentration. The model has developed a relationship between five climate variables (precipitation, average temperature, wind speed, solar radiation), reservoir operations (reservoir level), and sedimentary parameters. SHAP feature importance was quantified through an evidence-based evaluation for both models, providing a methodology and interpretation for both linear and non-linear pathways.

The results indicate that random forest substantially outperformed linear regression (R² = 0.65 versus 0.47, representing 38% improvement), with RMSE reduced by 19% and MAE by 36%, indicating significant non-linear climate-turbidity dynamics. Whereas the total dissolved solids model suggests an improved R² of 0.30 compared to 0.04 for linear regression. Furthermore, SHAP analysis revealed a divergence in precipitation importance between random forest and linear models, which attributed only 12.8% of the linear contribution. While non-linear models identified precipitation as the dominant driver, accounting for 50.8% of the contribution, a 38 percentage-point divergence was observed. SHAP dependence analysis identified a 10 mm/day as a critical precipitation threshold, below which the impacts of turbidity remain minimal. The precipitation contributions increase exponentially, reaching +10 to +13 SHAP units at precipitation levels exceeding 100 mm/day. The SHAP dependence result suggests that air temperature interactions amplify precipitation effects, with high-temperature periods generating 30-40% larger turbidity events to equivalent precipitation. In contrast, other climate parameters show consistent SHAP values across models (solar radiation: 28.8% MLR versus 16.1% RF), indicating predominantly linear relationships that were adequately explained by simple regression. Additionally, the reservoir level is a major contributor to total dissolved solids, with 35.7% non-linear contribution compared to 28.2% linear contribution, followed by precipitation and solar radiation. The reservoir level, ranging from 255 to 265 m, provides buffering capacity to absorb precipitation-driven sediment loads without significant fluctuations in turbidity and total dissolved solids.

The identified thresholds enable the development of climate-informed, tiered operational protocols: standard operations below 10 mm/day precipitation, enhanced operations at 10-100 mm, and advanced operations above 100 mm with different treatment dosages. Instead of the proportionate responses predicted by linear extrapolation. The non-linear dynamics for climate adaptation planning suggest that anticipated 20-30% increases in monsoon precipitation intensity could lead to 50-80% increases in peak turbidity events. This multi-model SHAP system provides a modelling approach for determining operational thresholds, measuring parameter contributions, and assessing the complexity of climate-water quality interactions to inform practical management strategies.