Random forest modeling and Shapley values reveal controls of sediment transport and hysteresis patterns in a low-gradient watershed

Understanding and modeling the mechanisms of watershed-scale sediment transport is crucial for stream management and ecosystem protection. The increased availability of high-frequency, in situ water quality data has facilitated the use of Machine Learning (ML) algorithms to predict sediment transport rates in streams and rivers. While many ML algorithms outperform classical process-based modeling approaches with respect to sediment load prediction, these algorithms have been criticized for having limited potential to reveal underlying sediment processes, which is largely attributed to the black box nature of such algorithms. This study explores the potential of using Shapley values coupled with Random Forest (RF) predictions to investigate the controls of sediment transport and hysteresis patterns in a low-gradient watershed in Kentucky, USA. Specifically, the Shapley values were used to identify hydrometeorological variables controlling both sediment load and hysteresis patterns. We characterized hydrological forcings, sediment loads, and hysteresis patterns for 71 hydrologic events between August 2017 and July 2019 using high-frequency discharge, specific conductance, turbidity, and precipitation data collected in the Upper South Elkhorn watershed (62-km2). Preliminary results indicate that the fraction of flow contributed from baseflow pathways, peak discharge and total event precipitation predicted sediment load and hysteresis patterns well, which suggests that the provenance and timing of sediment delivery is closely linked with hydrologic connectivity during storm events. Furthermore, our results show that the Shapley value approach is a promising technique for identifying non-linear relationships among explanatory variables that may remain concealed in principal component analyses and multiple linear regression algorithms. This study demonstrates the utility of widely available flow and water quality data to elucidate sediment processes at the watershed scale, which has important implications for watershed managers and land management decisions.