Comparison of Sediment Rating Curve Developed by Using KNN, SVM and Random Forest Method for Gandaki River Basin, Nepal

Abstract

Accurate estimation of suspended sediment load (SSL) is crucial for effective water resource management in Nepal's monsoon-influenced Himalayan River systems, where traditional sediment rating curves (SRCs) often fail to capture complex, non-linear discharge-sediment relationships. This study developed season-specific sediment rating curves for Station 430.5 on the Seti-Gandaki River using three machine learning algorithms: K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Random Forest (RF) to address the limitations of conventional power-law approaches across Nepal's four distinct hydrological seasons. Daily discharge and suspended sediment concentration data from 2004-2017 were analyzed to characterize seasonal variability, with monsoon periods exhibiting peak flows exceeding 1,400 m³/s and sediment concentrations reaching over 20,000 ppm, contrasting sharply with winter baseflow conditions below 35 m³/s and sediment concentrations under 250 ppm. Model performance was evaluated using coefficient of determination (R²) and Mean Absolute Percentage Error (MAPE), revealing optimal algorithm selection varied significantly across seasons: KNN demonstrated superior accuracy for monsoon (R² = 0.22) and post-monsoon (R² = 0.97) periods, while SVM showed better performance during winter (R² = 0.81) conditions, though all models struggled with pre-monsoon transitional dynamics (R² < 0.25). The resulting power-law equations—S = 0.85 × Q^1.81 (pre-monsoon), S = 333.66 × Q^0.31 (monsoon), S = 26.55 × Q^0.62 (post-monsoon), and S = 17.61 × Q^0.72 (winter) captured distinct seasonal sediment transport mechanisms, with monsoon conditions presenting the greatest modeling challenges (MAPE = 90.67%) due to complex hysteresis effects, variable sediment availability, and extreme non-linearity that conventional machine learning approaches cannot fully represent. While machine learning models demonstrated substantial improvements over traditional SRCs, particularly during stable post-monsoon and winter periods, the persistent high prediction errors during monsoon conditions highlight the need for advanced deep learning architectures capable of capturing temporal dependencies and multi-factorial sediment transport controls. These findings provide essential foundations for improving sediment load estimation in data-scarce Himalayan basins and inform sustainable water resource management, hydropower planning, and climate adaptation strategies in rapidly changing mountain river systems.