Machine Learning Approach to Modeling Sediment Transport

Inaccuracies of sediment transport models largely originate from our limitation to describe the process in precise mathematical terms. Machine learning (ML) is an alternative approach to reduce the inaccuracies of sedimentation models. It utilizes available domain knowledge for selecting the input and output variables for the ML models and uses modern regression techniques to fit the measured data. Two ML methods, artificial neural networks and model trees, are adopted to model bed-load and total-load transport using the measured data. The bed-load transport models are compared with the models due to Bagnold, Einstein, Parker et al., and van Rijn. The total-load transport models are compared with the models due to Ackers and White, Bagnold, Engelund and Hansen, and van Rijn. With the chosen data sets on bed-load and total-load transport the ML models provided better accuracy than the existing ones.     

Modeling Bed-Load Rates from Fine Grain-Size Patches during Small Floods in a Gravel-Bed River

This study investigates the applicability of five bed-load-transport formulas (the Meyer-Peter and Müller, Schoklitsch, Bagnold, Smart and Jaeggi, and Rickenmann equations) to predict bed-load transport rates of frequent, low-magnitude flood events (maximal bankfull discharge) for a mountainous, poorly sorted gravel-bed river characterized by a bimodal sediment-size distribution and spatially distributed patches. For model parametrization, special emphasis was placed on the spatial composition of the grain-size distribution (GSD) to evaluate the impact of preferential removal of sediments from patches with finer sediments on bed-load transport. Three parametrization approaches to the choice of an appropriate sediment size that considered the apparent bimodality of the GSD to varying degrees were tested. The modeling study demonstrated that the incorporation of spatial structure of GSD and its bimodal character has an important impact on model performance—a unimodal parametrization failed to reproduce measured bed-load rates for all tested bed-load formulas; a threshold parametrization approach that considered only finer sediments from the small patches as bed-load source material in combination with the Schoklitsch, Smart and Jaeggi, and Rickenmann equations yielded the best results, whereas the Meyer-Peter and Müller and the Bagnold equations failed to predict bed-load rates for all parametrization approaches. The modeling study thus showed that bed-load formulas are sensitive to the spatial structure of the GSD, which should not be treated as a continuum of sediment size fractions but rather as composition of finer sediment patches to enable an adequate reproduction of measured bed-load data from low-magnitude floods in gravel-bed rivers.     

Effect of Time Scale on the Performance of Different Sediment Transport Formulas in a Semiarid Region

This note compares the performance of various sediment transport formulas for a semiarid region of southern Italy (Puglia region). The sediment transport that occurs in the main streams of the region has been monitored for about 30 years (1958–1986) and the measured data are available at monthly intervals. The field data were compared with those predicted from four formulas (Ackers-White, Engelund-Hansen, Yang, and Van Rijn) selected on the basis of the sediment transport mode, the theoretical aspects, and the sensitivity to basic physical properties. The study examines the ability of the formulas to adapt to local conditions, the influence of time scale on predictions, and the accuracy of the prediction varying the discharge. The Engelund-Hansen and Van Rijn formulas (suspended sediment transport) estimate well the sediment load and are largely unaffected by variations in time scale. For low sediment loads, these formulas were less reliable.     

One-Dimensional Modeling of Dam-Break Flow over Movable Beds

A one-dimensional model has been established to simulate the fluvial processes under dam-break flow over movable beds. The hydrodynamic model adopts the generalized shallow water equations, which consider the effects of sediment transport and bed change on the flow. The sediment model computes the nonequilibrium transport of bed load and suspended load. The effects of sediment concentration on sediment settling and entrainment are considered in determining the sediment settling velocity and transport capacity. In particular, a correction factor is proposed to modify the Van Rijn formulas of equilibrium bed-load transport rate and near-bed suspended-load concentration for the simulation of sediment transport under high-shear flow conditions. The governing equations are solved by an explicit finite-volume method with the first-order upwind scheme for intercell fluxes. The model has been tested in two experimental cases, with fairly good agreement between simulations and measurements. The sensitivities of the model results to parameters such as the sediment nonequilibrium adaptation length, Manning’s roughness coefficient and the proposed correction factor have been verified. The proposed model has also been compared to an existing model and the results indicate the new model is more reliable.     

Bed-Load Transport Equation for Sheet Flow

When open-channel flows become sufficiently powerful, the mode of bed-load transport changes from saltation to sheet flow. Where there is no suspended sediment, sheet flow consists of a layer of colliding grains whose basal concentration approaches that of the stationary bed. These collisions give rise to a dispersive stress that acts normal to the bed and supports the bed load. An equation for predicting the rate of bed-load transport in sheet flow is developed from an analysis of 55 flume and closed conduit experiments. The equation is ib = ω where ib = immersed bed-load transport rate; and ω = flow power. That ib = ω implies that eb = tan?α = ub/u, where eb = Bagnold’s bed-load transport efficiency; ub = mean grain velocity in the sheet-flow layer; and tan?α = dynamic internal friction coefficient. Given that tan?α ≈ 0.6 for natural sand, ub ≈ 0.6u, and eb ≈ 0.6. This finding is confirmed by an independent analysis of the experimental data. The value of 0.60 for eb is much larger than the value of 0.12 calculated by Bagnold, indicating that sheet flow is a much more efficient mode of bed-load transport than previously thought.     

Assessment of Methods Used in 1D Models for Computing Bed-Load Transport in a Large River: The Danube River in Slovakia

Comprehensive measurements of bed-load sediment transport through a section of the Danube River, located approximately 70?km downstream from Bratislava, Slovakia, are used to assess the accuracy of bed-load formulas implemented in 1D modeling. Depending on water discharge and water level, significant variations in the distribution of bed load across the section were observed. It appeared that, whatever the water discharge, the bed shear stress τ is always close to the estimated critical bed shear stress for the initiation of sediment transport τcr. The discussion focuses on the methods used in 1D models for estimating bed-load transport. Though usually done, the evaluation of bed-load transport using the mean cross-sectional bed shear stress yields unsatisfactory results. It is necessary to use an additional model to distribute the bed shear stress across the section and calculate bed load locally. Bed-load predictors also need to be accurate for τ close to τcr. From that point of view, bed-load formulas based on an exponential decrease of bed-load transport close to τcr appear to be more appropriate than models based on excess bed shear stress. A discussion on the bed-load formula capability to reproduce grain sorting is also provided.     

Applicability of Sediment Transport Capacity Formulas to Dam-Break Flows over Movable Beds

In this paper, we investigate the extent to which well-known sediment transport capacity formulas can be used in one-dimensional (1D) numerical modeling of dam-break waves over movable beds. The 1D model considered here is a one-layer model based on the shallow-water equations, a bed update (Exner) equation, a space-lag equation for the nonequilibrium sediment transport and an empirical formula calculating the sediment transport capacity of the flow. The model incorporates a variety of sediment transport capacity formulas proposed by Meyer-Peter and Müller, Bagnold, Engelund and Hansen, Ackers and White, Smart and Jaeggi, van Rijn, Rickenmann, Cheng, Abrahams and Camenen, and Larson. We examine the performance of each formula by simulating four idealized laboratory cases on dam-break waves over sandy beds. Comparisons between numerical results and measurements show that for each case better predictions are obtained using a particular formula, but overall, formulas proposed by Meyer-Peter and Müller (with the factor 8 being replaced by 12), Smart and J?ggi, Cheng, Abrahams and Camenen, and Larson rank as the best predictors for the entire range of conditions studied here. Moreover, results show that in the cases where a bed step exists, implementing a mass failure mechanism in the numerical modeling plays an important role in reproducing the bed and water profiles.     

Unified View of Sediment Transport by Currents and Waves. II: Suspended Transport

The problem of suspended sediment transport in river and coastal flows is addressed. High-quality field data of river and coastal flows have been selected and clustered into four particle size classes (60–100, 100–200, 200–400, and 400–600?μm). The suspended sand transport is found to be strongly dependent on particle size and on current velocity. The suspended sand transport in the coastal zone is found to be strongly dependent on the relative wave height (Hs/h), particularly for current velocities in the range 0.2–0.5?m/s. The time-averaged (over the wave period) advection–diffusion equation is applied to compute the time-averaged sand concentration profile for combined current and wave conditions. Flocculation, hindered settling, and stratification effects are included by fairly simple expressions. The bed-shear stress is based on a new bed roughness predictor. The reference concentration function has been recalibrated using laboratory and field data for combined steady and oscillatory flow. The computed transport rates show reasonably good agreement (within a factor of 2) with measured values for velocities in the range of 0.6–1.8?m/s and sediments in the range of 60–600?μm. The proposed method underpredicts in the low-velocity range (<0.6?m/s). A new simplified transport formula is presented, which can be used to obtain a quick estimate of suspended transport. The modeling of wash load transport in river flow based on the energy concept of Bagnold shows that an extremely large amount of very fine sediment (clay and very fine silt) can be transported by the flow.     

Depth-Averaged Two-Dimensional Numerical Modeling of Unsteady Flow and Nonuniform Sediment Transport in Open Channels

A depth-averaged two-dimensional (2D) numerical model for unsteady flow and nonuniform sediment transport in open channels is established using the finite volume method on a nonstaggered, curvilinear grid. The 2D shallow water equations are solved by the SIMPLE(C) algorithms with the Rhie and Chow’s momentum interpolation technique. The proposed sediment transport model adopts a nonequilibrium approach for nonuniform total-load sediment transport. The bed load and suspended load are calculated separately or jointly according to sediment transport mode. The sediment transport capacity is determined by four formulas which are capable of accounting for the hiding and exposure effects among different size classes. An empirical formula is proposed to consider the effects of the gravity on the sediment transport capacity and the bed-load movement direction in channels with steep slopes. Flow and sediment transport are simulated in a decoupled manner, but the sediment module adopts a coupling procedure for the computations of sediment transport, bed change, and bed material sorting. The model has been tested against several experimental and field cases, showing good agreement between the simulated results and measured data.     

Unified View of Sediment Transport by Currents and Waves. I: Initiation of Motion, Bed Roughness, and Bed-Load Transport

Attention is given to the properties of sediment beds over the full range of conditions (silts to gravel), in particular the effect of fine silt on the bed composition and on initiation of motion (critical conditions) is discussed. High-quality bed-load transport data sets are identified and analyzed, showing that the bed-load transport in the sand range is related to velocity to power 2.5. The bed-load transport is not much affected by particle size. The prediction of bed roughness is addressed and the prediction of bed-load transport in steady river flow is extended to coastal flow applying an intrawave approach. Simplified bed-load transport formulas are presented, which can be used to obtain a quick estimate of bed-load transport in river and coastal flows. It is shown that the sediment transport of fine silts to coarse sand can be described in a unified model framework using fairly simple expressions. The proposed model is fully predictive in the sense that only the basic hydrodynamic parameters (depth, current velocity, wave height, wave period, etc.) and the basic sediment characteristics (d10, d50, d90, water temperature, and salinity) need to be known. The prediction of the effective bed roughness is an integral part of the model.     

Effect of Coarse Surface Layer on Bed-Load Transport

Existing bed-load transport formulas may overestimate the transport rate in mountain rivers by two orders of magnitude or more. Recently published field data sets provide an opportunity to take a fresh look at the bed-load transport relationship and it is hypothesized that the overestimate is due to a failure to account for the effect of a coarse surface layer of bed material inhibiting the release of fine subsurface material. Bed-load transport is determined as gs = aρ(q?qc) where q=water discharge per unit width; qc=critical value for initiation of bed material movement; ρ=water density; and a=coefficient. The gs/q relationship is typically piecewise linear, characterized by two transport phases with, respectively, low and high rates of change. Twenty-one flume and 25 field data sets were used to quantify the relationship for Phase 2. The flume data confirm the dependence of a on S1.5, where S=channel slope, in agreement with earlier studies. The field data additionally show that a varies inversely with the degree of bed armoring, given by the ratio of surface to subsurface bed material size. The finding is consistent with the hypothesis and suggests the need to account for the bed material supply limitation in the bed-load transport formula. However, the available data are not entirely sufficient to rule out an alternative dependency, or codependency, on flow resistance. The critical conditions for initiation of Phase 2 transport are also quantified as a function of bed material size and channel slope. The resulting set of equations allows a more accurate estimation of Phase 2 bed-load transport rates. However, the equations are empirical and should be restricted for use within the range of conditions used in their development, to determine mean rather than instantaneous transport rates and to determine bulk transport rates, not transport by size fraction.     

Modification of the Einstein Bed-Load Formula

The original derivation of the Einstein bed-load formula involves several oversimplified assumptions concerning step length of a particle, exchange time, and probability of being lifted into motion. A modified version of the Einstein formula that relaxes these assumptions and incorporates a nonuniform sediment model is proposed and tested with both field and laboratory data. The modified formula shows promise as being more reliable than the original version in estimating bed-load transport rates.     

Field Assessment of Alternative Bed-Load Transport Estimators

Measurement of near-bed sediment velocities with acoustic Doppler current profilers (ADCPs) is an emerging approach for quantifying bed-load sediment fluxes in rivers. Previous investigations of the technique have relied on conventional physical bed-load sampling to provide reference transport information with which to validate the ADCP measurements. However, physical samples are subject to substantial errors, especially under field conditions in which surrogate methods are most needed. Comparisons between ADCP bed velocity measurements with bed-load transport rates estimated from bed-form migration rates in the lower Missouri River show a strong correlation between the two surrogate measures over a wide range of mild to moderately intense sediment transporting conditions. The correlation between the ADCP measurements and physical bed-load samples is comparatively poor, suggesting that physical bed-load sampling is ineffective for ground-truthing alternative techniques in large sand-bed rivers. Bed velocities measured in this study became more variable with increasing bed-form wavelength at higher shear stresses. Under these conditions, bed-form dimensions greatly exceed the region of the bed ensonified by the ADCP, and the magnitude of the acoustic measurements depends on instrument location with respect to bed-form crests and troughs. Alternative algorithms for estimating bed-load transport from paired longitudinal profiles of bed topography were evaluated. An algorithm based on the routing of local erosion and deposition volumes that eliminates the need to identify individual bed forms was found to give results similar to those of more conventional dune-tracking methods. This method is particularly useful in cases where complex bed-form morphology makes delineation of individual bed forms difficult.     

Method for Estimating Sediment Transport in Ice-Covered Channels

A method is proposed for estimating rates of sediment transport in ice-covered alluvial channels. The method extends existing, open-water procedures for estimating rates of sediment transport to conditions of ice-covered flow. A key aspect of the method is the assessment of flow resistance attributable to bed-surface drag. That assessment is used to estimate rates of bed load and suspended load, and thereby total bed-sediment transport rate. Estimation of ice-covered suspended load additionally entails an approximation whereby open-water suspended load is scaled in proportion to the ratio of a reference sediment concentration for ice-covered flow relative to that for open-water flow. The reference concentration is calculated in terms of bed-load rate and shear velocity attributed to bed-surface drag. Flume data are used to develop the method and tentatively verify it. Field verification of the method presently is hampered by the absence of field data on bed sediment transport in ice-covered channels.     

Nonintrusive Method for Detecting Particle Movement Characteristics near Threshold Flow Conditions

Bed-load measurements comprise an important component in the development of reliable formulas, in an effort to obtain the necessary constitutive relations between the amount of transported material and flow parameters. The uncertainty of such measurements is rather well known, being much more pronounced at lower transport rates. This uncertainty stems from the multitude of factors affecting bed-load transport and the lack of available trustworthy measuring technologies. Predictions of the limiting case of nearly zero bed-load transport, typically reported in literature as threshold of motion or critical condition, are even more challenging. The purpose of this contribution is twofold. First, to examine the sensitivity of bed-load transport measurements at conditions moderately higher than critical, to the presence of a rather unobtrusive trap, designed through several iterations. Even under relatively simple laboratory flume channel and flow conditions, it proved difficult to measure the bed-load transport rate in a completely unbiased way. Second, to develop a methodology, together with the appropriate instrumentation, for determining the condition of incipient motion. The nonintrusive approach described here proved to be reliable in detecting even the slightest movements of a particle. At the same time, it demonstrates the complexity of the problem due to the highly fluctuating nature of turbulent flow.     

Evaluation and Improvement of Bed Load Discharge Formulas based on Helley–Smith Sampling in an Alpine Gravel Bed River

Bed load discharge formulas have been evaluated by analyzing them in relation to measured Helley–Smith data for the gravel-bedded armored Drau River, Austria. Comparison of calculations with measurements leads to ranking of the formulas that depends on the evaluation parameters. The choice of formula is made with respect to our specific aims: the investigation of individual floods requires a different approach from that of long-term budgets. Formula performance is consistently improved when conditions for the threshold of motion are modified according to data measured up on the initiation of motion. Formulas such as those reported by Parker in 1990, Zanke in 1999, and Sun and Donahue in 2000 are capable of coping with partial transport, which is commonly found in Alpine rivers. These formulas therefore provide encouraging results, particularly after the introduction of modifications. The augmentation of field measurements, even if limited in scope, considerably improves the performance of bed load discharge formulas.     

Surface-Based Fractional Transport Predictor: Deterministic or Stochastic

Stochastic bed load transport formulas for nonuniform sediment exist, but most of them do not account for the composition of surface material to predict fractional transport rate. This study transformed a surface-based bed-load transport predictor to a stochastic one by approximating the fluctuation of bed-shear stress with a standard log-normal distribution. The deterministic predictor underpredicts fractional transport rate at low values of bed-shear stress and Reynolds number. The modified stochastic predictor predicts fractional transport rate more accurately and converges to the deterministic one at high shear stresses.     

Hydraulic Interpretation of Cross-Stream Variations in Bed-Load Transport

The hydraulic control of bed-load transport rates in Nahal Yatir and Nahal Eshtemoa, two coarse-grained ephemeral channels in the semiarid northern Negev, Israel, provides a rare opportunity to infer the spanwise variation in bed-shear stress from an analysis of cross-stream variations in bed-load transport rate. Automatic sediment transport monitoring stations were used to obtain synchronous measurements of bed-load discharge at a number of locations across the widths of two straight channel reaches. In both streams, channel-average bed-load fluxes demonstrated a common and well-defined response to changing channel-average shear stress and approximated the transporting capacity of the flow over much of the range of monitored discharges. However, transport rates measured at the channel margins are only half those at the channel centerline, and, at high discharges, a marked asymmetry in the pattern of bed-load transport develops across the central section of the widest channel. This variation in bed-load discharge over the two channel cross sections is thought to reflect lateral variations in shear stress induced by sidewall drag and, more tentatively, the generation and disposition of cellular secondary currents. But no systematic relation is found for the ratios of sediment fluxes at off-center sampling locations and those recorded at the channel center, even though the off-center locations are thought to move into and out of the region affected by sidewall drag as aspect ratio of the flow decreases and increases with changing water-stage. The results suggest that it is difficult to generalize about the changing influence of the sidewall on local shear and bed load as aspect ratio changes during the course of a flood.     

Equilibrium Near-Bed Concentration of Suspended Sediment

A new approach is presented for calculating the equilibrium near-bed concentration of suspended sediment in an alluvial channel flow. It is formulated from the balance between bed sediment entrainment and suspended sediment deposition across the near-bed boundary. The entrainment flux is determined making use of a turbulent bursting outer-scale-based function and the flux of deposition by the product of near-bed concentration and hindered settling velocity of sediment. A number of flume data records in the literature are analyzed to calibrate and verify the present approach. The observed near-bed concentrations for the data records are obtained by first isolating the suspended load transport rate from the observed total load transport rate using Engelund and Fredsoe's bed-load formula and then equating the suspended load transport rate to the shape integration of Dyer and Soulsby. The present approach is shown to perform satisfactorily compared to the results of data analysis. It is found that the near-bed concentration is evidently dependent on sediment particle size in addition to the Shields parameter due to skin friction. This finding seems to challenge previous relationships that simply represent the near-bed concentration as empirical functions of the purely skin-friction-related Shields parameter.     

Performance of Bed-Load Transport Equations Relative to Geomorphic Significance: Predicting Effective Discharge and Its Transport Rate

Previous studies assessing the accuracy of bed-load transport equations have considered equation performance statistically based on paired observations of measured and predicted bed-load transport rates. However, transport measurements were typically taken during low flows, biasing the assessment of equation performance toward low discharges, and because equation performance can vary with discharge, it is unclear whether previous assessments of performance apply to higher, geomorphically significant flows (e.g., the bankfull or effective discharges). Nor is it clear whether these equations can predict the effective discharge, which depends on the accuracy of the bed-load transport equation across a range of flows. Prediction of the effective discharge is particularly important in stream restoration projects, as it is frequently used as an index value for scaling channel dimensions and for designing dynamically stable channels. In this study, we consider the geomorphic performance of five bed-load transport equations at 22 gravel-bed rivers in mountain basins of the western United States. Performance is assessed in terms of the accuracy with which the equations are able to predict the effective discharge and its bed-load transport rate. We find that the median error in predicting effective discharge is near zero for all equations, indicating that effective discharge predictions may not be particularly sensitive to one’s choice of bed-load transport equation. However, the standard deviation of the prediction error differs between equations (ranging from 10% to 60%), as does their ability to predict the transport rate at the effective discharge (median errors of less than 1 to almost 2.5 orders of magnitude). A framework is presented for standardizing the transport equations to explain observed differences in performance and to explore sensitivity of effective discharge predictions.