Interpretable Machine Learning to Improve Predictions of Fluvial Sediment Transport

Abstract

Fluvial sediment-transport contributes to many environmental concerns including flooding, nutrient loading, aquatic habitat degradation, reservoir sedimentation impacting dam operation, filling in of navigable waterways, and degrading of streams requiring costly restorations. However, there is a general lack of fluvial sediment-transport predictive power as it is difficult to comprehensively understand due to its complex process controlled by many factors such as hydrology, geology, land-use, and sediment supply. Field data can be used to better understand fluvial sediment-transport but is often limited due to it being expensive, technical, and labor intensive to collect. Minnesota provides a unique opportunity to study sediment-transport due to a complex glacial history that produced diverse landforms which combine with varying land uses and land covers to yield surface water conditions. Fortunately, Minnesota has a large sediment-transport dataset available with which to build and test a statewide predictive model in order to increase the knowledge of sediment-transport and help solve environmental concerns. XGBoost machine learning models were developed and trained to predict suspended-sediment concentration (SSC) and bedload transport rates at unsampled rivers and streams by using SSC samples collected from 56 sites and bedload samples collected from 43 sites by the U.S. Geological Survey from 2007-2019. Basin (full upstream area), catchment (nearby landscape), near-channel, and in-channel feature variables were compiled from available state and national datasets (NHDPlusV2, StreamCat, U.S. Stream Classification System, and StreamStats). The 2-year recurrence interval statistic for each site was used to normalize streamflow. The slope of the dimensionless hydrograph was calculated to teach the model the rate of rising or falling streamflow conditions before and after each sample was collected. Models for both bedload and SSC transport explained roughly 70% of the variance in the dataset. Shapley additive explanation values (SHAP) facilitated model interpretation and connected important model features to their roles with the sediment-transport process. Cumulative suspended sediment loads were calculated from model output and compared to in-situ surrogate loads from four sites in the study area to show model utility, and test model improvements. Results show that these models can inform sediment loads and stream-restoration activities across Minnesota by providing estimates of suspended-sediment concentrations and bedload rates where samples have not been collected.