Professional Documents
Culture Documents
Estimating Sediment Settling Velocities From A Theoretically Guided Data-Driven Approach
Estimating Sediment Settling Velocities From A Theoretically Guided Data-Driven Approach
Estimating Sediment Settling Velocities From A Theoretically Guided Data-Driven Approach
Abstract: Sediment settling velocities are commonly estimated from analytical or process-based approaches. These approaches have
Downloaded from ascelibrary.org by University of Birmingham on 07/22/20. Copyright ASCE. For personal use only; all rights reserved.
theoretical constraints due to the incompletely resolved settling physics. A parametric data-driven approach was recently proposed without
theoretical constraints, but it is limited by its mathematical assumptions. To overcome these limitations, this study applies a machine learning
algorithm to an aggregated sediment settling experimental database and develops a nonparametric data-driven model to estimate the non-
cohesive sediment settling velocity in water. A cross-comparison against five process-based equations and a parametric data-driven equation
demonstrates the higher accuracy and better consistency of the new model in estimating sediment settling velocities under various physical
regimes. The new model also shows an easily implemented self-update capability by assimilating theoretical data derived from the process-
based equations. The updated model, incorporating experimental and theoretical data of sediment settling processes, further improves the
accuracy and reduces the uncertainty in estimating sediment settling velocities. This study demonstrates the capability of machine learning in
sediment transport study and illustrates an alternative framework for other hydraulic engineering challenges. DOI: 10.1061/(ASCE)
HY.1943-7900.0001798. © 2020 Published by American Society of Civil Engineers.
However, it is unclear how to select mathematical operators for GP, 2. Data are restricted to noncohesive sediment with D > 6.25 ×
which makes it challenging to reproduce the researchers’ work; 10−5 m (1,000 measurements).
furthermore, predefined operators may limit a model’s ability to 3. Data are classified into sand (D < 2 × 10−3 m) and gravel
fully capture the internal relationships in data. To overcome these (D > 2 × 10−3 m).
limitations, this study presents a nonparametric data-driven model 4. Only data in water are considered (ν < 2 × 10−5 m2 =s) for typ-
of sediment settling velocities based on an aggregated database. ical real-world applications.
The new model, built upon a machine learning (ML) algorithm After data standardization, the final database contains 756 ex-
called a random forest (RF), relaxes the theoretical or mathematical perimental measurements with the variable distributions shown
constraints of previous studies and describes the internal relation- in Fig. 1.
ships in the data from a fully data-driven approach. Background
knowledge of the RF and its applications to relevant research
New Data-Driven Approach Using Random Forest
are presented in the methodology section. The new model has the
capability to self-update with the addition of new theoretical data A RF (Breiman 2001) is a supervised machine learning algorithm
derived from process-based equations. Cross-comparisons between that learns intrinsic information within data, as demonstrated across
different approaches (process-based, parametric data-driven, and a variety of scientific fields and applications (Svetnik et al. 2003;
nonparametric data-driven) are also presented to evaluate the vari- Cutler et al. 2007; Shotton et al. 2013; Chen and Hu 2017; Kane
ety of sediment settling velocity models across multiple physical et al. 2014; Zhou et al. 2019; Chen et al. 2019). A brief overview
regimes. of a RF algorithm with simple examples is presented in Liaw and
The paper is organized as follows. First, data compilation, pre- Wiener (2002). Comparisons between RF and other ML algorithms
processing, and the ML model configuration are described in the (e.g., artificial neural network and support vector machine) demon-
methodology section. Then the ML model performance, cross- strate its easy implementation, competitively high accuracy, and
model comparison, and sensitivity tests are presented in the results transparency, as well as its ability to deal with small sample sizes
section. Differences between process-based and data-driven mod- and high-dimensional data (e.g., Svetnik et al. 2003; Liu et al.
els, including model sensitivity analysis and ML model update, are 2013; Smith et al. 2013; Biau and Scornet 2016; Tyralis and
provided in the discussion section. This is followed by a conclusion Papacharalampous 2017; Chen and Hu 2017; Chen et al. 2019).
section summarizing the value of this study and its illustration of Readers are referred to Verikas et al. (2011) for a review of the wide
ML capabilities for hydraulic engineering applications. applications of RF and its comparisons with other ML algorithms.
This study chooses a RF regression model because it (1) does not
require significant data preprocessing for discrete data, (2) guaran-
Methodology tees good performance with reduced sensitivity to outliers, which
are sometimes inevitable and hard to exclude in experimental data,
(3) is less susceptible to overfitting, (4) requires minimal parameter
Aggregated Database and Data Pre-Processing
tuning, (5) does not have linear/nonlinear assumptions, and (6) can
A multisource database of sediment settling measurements is com- provide the relative importance of each variable in the model, which
piled from Paphitis et al. (2002), Goldstein and Coco (2014), and is helpful in model sensitivity analysis (e.g., Segal 2004; Svetnik
Watts and Zarillo (2019). The database has 1,006 measurements et al. 2003; Raschka and Mirjalili 2017). Here, the open-source
(Table 1) with 4 independent variables: the nominal diameter of RF regression algorithm sklearn.ensemble.RandomForestRegressor
a particle (Dn ) (m), the particle submerged specific gravity (Δ) (Pedregosa et al. 2011) is applied in Python 3 (Summerfield 2010)
(dimensionless), the kinematic viscosity of fluid (ν) (m2 =s), and is applied in this study. Three parameters are fine-tuned in the
the corresponding particle settling velocity (w) (m=s). The specific model; the number of decision trees (ntrees), the depth of each tree
gravity Δ ¼ ρs =ρw − 1, where ρs and ρw are the density (kg=m3 ) of (max_depth), and the number of features (max_features) used to
sediment and water, respectively. Data standardization is carried determine tree splits. Each parameter is evaluated over a range of
out according to the following four steps: values, and the combination of values for best model performance
1. The nominal diameter (Dn ) is converted to sieve diameter (D) is determined using grid search (Lerman 1980). Fivefold cross-
by D ¼ Dn =1.1 (Raudkivi 1990) for consistency with paramet- validation (Kohavi 1995) is incorporated into the grid search to
ric settling equations that use D as an input variable. mitigate model overfitting (Raschka and Mirjalili 2017) during
(c) (d)
Fig. 1. Aggregated database sample distributions: (a) sediment sieve diameter; (b) fluid kinematic viscosity; (c) submerged specific gravity; and
(d) settling velocity.
training. The final RF model with optimized model parameters is accuracy: MAE describes the average error of the model directly,
selected from a trade-off between model simplicity and accuracy. and RMSE has the benefit of penalizing large errors in the model.
The database is split into training and test data sets based on D Among the nine optimized models (Table 2), the one with the best
because of the broad distribution of these values in the database performance on both training and test data is selected as the final
[Fig. 1(a)]. A stratified random split is applied to preserve a con- RF model:
stant sediment class ratio (0.35 for gravel and 0.65 for sand) in both
training and test data sets, with the test size ratio (test_size) varied ΣNi¼1
Tr
ðwiobs − wirfp Þ2
R2 ¼ 1 − ð1Þ
between 0.1 and 0.9 to produce nine parameter tuning scenarios. In ΣNi¼1
Tr
ðwiobs − wobs Þ2
each scenario, the RF decision trees are binarily split until the
depths reach max_depth. A grid search is performed for max_depth sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 N Tr i
from 5 to 20, max_features from 1 to 3, and ntrees from 100 to σ¼ Σ ðw − wobs Þ2 ð2Þ
400 at an interval of 50. A fivefold cross-validation assesses model N Tr i¼1 rfp
performance and the optimized model is selected based on both
the confidence of determination [R2 , Eq. (1)] and the standard 1 N Te i
MAE ¼ Σ jw − wirfp j ð3Þ
deviation [σ, Eq. (2)]. These two metrics are used to evaluate the N Te i¼1 obs
average model performance throughout the model cross-validation
process, i.e., higher R2 and lower σ indicate higher model accuracy qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
and stability. Each scenario generates one optimized model that is RMSE ¼ ΣNi¼1 Te
ðwiobs − wirfp Þ2 ð4Þ
N Te
verified by the test data across three metrics: R2 , mean absolute
error [MAE, Eq. (3)], and root-mean-square error [RMSE, Eq. (4)]. where N Tr and N Te = total number of samples in training and test
Both MAE and RMSE are the evaluation metrics for the model data, respectively; wiobs = observed settling velocity of ith sample;
(a) (b)
Fig. 2. (Color) Performance of RF estimator on the training and test data, respectively: (a) training data (605 measurements); and (b) test data (151
measurements).
wobs = averaged settling velocity; and wirfp = estimated settling Reynolds number, which will be further demonstrated and dis-
velocity of ith sample from RF model. cussed in the sensitivity analysis.
Cross-Model Comparison
Results
The RF estimator is compared against six selected algebraic para-
Model Performance metric sediment settling velocity equations, as listed in Table 3. The
selected equations include five process-based equations (VR89,
Model performance obtained during parameter tuning is shown SB97, CH97, WW06, and SH09) and one parametric data-driven
in Table 2. Each row denotes a scenario for the best combination equation (GC14). All these equations have been well validated by
of ntrees and max_depth. Model performance is the best when their experimental data and are reported to be applicable to different
max features ¼ 3. The fivefold cross-validation results (with flow regimes (Van Rijn 1989; Soulsby 1997; Cheng 1997; Wu and
all R2 values greater than 0.95) and the metric evaluations indicate Wang 2006; Sadat-Helbar et al. 2009; Goldstein and Coco 2014).
that no optimized model is overfit. Model structure (estimated Specifically, VR89 and CH97 solve the CD − R relationship sim-
by ntrees × max depth) and the model accuracy (MAE, RMSE, ilarly, but they derive different M, N, and n from different exper-
and R2 ) on both training and test data are compared to determine imental data; WW06 considers the sediment particle shape in
Scenario 2 as the final RF model, which is referred to as the RF solving the CD − R relationship; SB97 is one of the most simplified
estimator hereafter. process-based equations with optimization of the coefficients in a
The performance of the RF estimator on both training and test combined viscous plus bluff-body drag law; SH09 is a process-
data is shown in Fig. 2. The evaluation metrics in the training data based equation, but it is derived from the artificial data generated
(MAE ¼ 1.37 × 10−2 m=s, RMSE ¼ 2.35 × 10−2 m=s, and R2 ¼ by other equations; and GC14 is the first parametric data-driven
0.99) and the test data (MAE ¼ 2.42 × 10−2 m=s, RMSE ¼ 4.28 × equation using an aggregated experimental database. As shown
10−2 m=s, and R2 ¼ 0.97) indicate that an accurate model for in Fig. 3, each equation performs well in the test data, but none
sediment settling velocities has been developed without any phys- outperforms the RF estimator. Considering that GC14 is developed
ical or mathematical constraints. The one standard deviation using nearly the same database as used in this study, the model
(1-sigma) prediction interval shows that the RF estimator has rel- accuracy of GC14 and the RF estimator is further compared on
atively smaller prediction intervals for smaller settling velocities, the entire final database. The results shown in Fig. 4 indicate that
e.g., for w < 0.30 m=s, but the relative uncertainties as obtained the RF estimator has a higher accuracy and overall less bias with a
by normalizing the standard deviation with the velocity are smaller higher R2 and smaller MAE and RMSE.
for larger settling velocities (not shown here). The overall perform- Model performance across different sediment classes (gravel
ance of the RF estimator on the entire database is MAE ¼ 1.58 × and sand) is also evaluated. The results (Table 4) demonstrate that
10−2 m=s, RMSE ¼ 2.85 × 10−2 m=s, and R2 ¼ 0.98 (Fig. 4 red the RF estimator performance is comparably accurate on sand and
text and dots). Feature importance analysis identifies D as the most gravel classes, but the algebraic equations do not have consistent
important parameter for the RF estimator with an index of 0.89 (out accuracy across different sediment classes. For example, the WW05
of 1.0). The importance indices of ν and Δ are both quite small and is the best among the six equations on sand samples (MAE ¼ 0.96 ×
are less than 0.10 individually. This result indicates that the RF 10−2 m=s, RMSE ¼ 1.35 × 10−2 m=s, and R2 ¼ 0.94), but it
estimator is only sensitive to D and the corresponding particle performs poorly on gravel samples (MAE ¼ 10.34 × 10−2 m=s,
D
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 !n
Mν 4ND3 n
Wu and Wang (2006) 0.25 þ − 0.5 WW06a
ND 3M 2
8
3 0.963
> 0.033ν ΔgD
>
> if D ≤ 10
< D ν 2
Sadat-Helbar et al. (2009)
> SH09
>
> 0.51ν ΔgD3 0.535
: if D > 10
D ν2
37.8ΔDn þ 3,780ΔD2n
Goldstein and Coco (2014) GC14
0.383 þ 10,000Δν þ 100Δ2 Dn
1
a
D ¼ D½Δg =ν 2 3 is the particle effective diameter. The coefficients M, N, and n in WW06 consider the Corey shape factor Sf (Corey 1949),
M ¼ 53.5e−0.65Sf , N ¼ 5.65e−2.5Sf , and n ¼ 0.7 þ 0.9Sf , where Sf ¼ pcffiffiffiffi
ab
, with a, b, and c the lengths (m) of the longest, intermediate, and shortest axes
of the particle. In this study Sf ¼ 0.8.
sensitivity for coarse sand and pebbles, but less for fine sand.
GC14 overestimates w when the grain size is small [Fig. 5(a)],
indicated by the estimate range (red bar) not overlapping with
any other model results. This is consistent with the findings of
Goldstein and Coco (2014).
Sensitivity to Kinematic Viscosity and Submerged Specific
Gravity
Twenty-one different fluid kinematic viscosity ν values are selected
with constant intervals from the range of 10−6 m2 =s and 10−5 m2 =s.
Fig. 4. (Color) Comparison of RF estimator and GC14 on entire Constant nominal values of D ¼ 10−3 m and Δ ¼ 1.65 are defined
database (756 measurements). for each ν value. For the sensitivity test against Δ, 15 different Δ
values are selected from [1 3], with ν and D assigned to be
10−6 m2 =s and 10−3 m, respectively. Model sensitivity results to
ν and Δ are shown in Fig. 6. Fig. 6(a) shows that the RF estimator
RMSE ¼ 13.34 × 10−2 m=s, and R2 ¼ 0.65). This indicates that and GC14 are not sensitive to ν, but the other five equations are
the traditional process-based equations, derived from limited exper- strongly sensitive. Sensitivity results for the specific gravity Δ
imental data and not tested by new data, may fail to show their gen- are similar: the five process-based models are more sensitive to
eral reliability. Compared to WW05, the RF estimator improves the Δ than the RF estimator and GC14 [Fig. 6(b)]
accuracy by 9.4% for sand samples and 48.6% for gravel samples
according to the MAE values. Similarly, compared to GC14, which
performs the best on gravel samples but the worst on sand samples, Discussion
the RF estimator has an increased accuracy by 44.2% and 17.8% on
sand and gravel classes, respectively. In general, the RF estimator In contrast to the traditional process-based approach based on lim-
outperforms the other six algebraic equations and shows consistent ited experimental data and mathematical/physical laws, the data-
accuracy across different sediment classes. driven approach is inductive and relies on aggregated database to
develop insight, predictions, or relationships. The two distinctive
approaches have intrinsic differences. Here the differences are ad-
Model Sensitivity Testing dressed from two perspectives: (1) model sensitivity and (2) model
To evaluate the sensitivity of the RF estimator and the six equations scalability. The latter focuses on the improvement of our data-
to each input variable, three sensitivity scenarios are completed driven model with the addition of more data derived from sampling
with each scenario testing one feature. theoretical relationships.
(b)
(c)
Fig. 5. (Color) Model sensitivity to different grain sizes: (a) fine sand; (b) coarse sand; and (c) pebbles.
however, are much smaller. There are only 48 different ν values and experimental data sets in air, in which all the measurements share
70 different Δ values in the database. More specifically, ν and Δ the same air kinematic viscosity (1.5 × 10−5 m2 =s) and submerged
can be effectively clustered into single groups [Figs. 1(b and c)]: specific density (2,207). However, from that database they derived
646 out of the 756 samples (85.4%) have ν values around a linear equation of settling velocity in air with respect to only grain
10−6 m2 =s (log10 ν ∈ [−6.1 − 5.9]); more than 70.7% of the Δ val- size, which outperforms eight previous equations with more inde-
ues falls in between 1.45 and 1.85, with 295 values equal to 1.65. pendent variables (e.g., ν and Δ) and more complex formulations
This explains why the RF estimator (and GC14) is sensitive to D [Table 5 in Farrell and Sherman (2015)].
but not to either ν or Δ. The proposed model’s sensitivity testing demonstrates intrinsic
In fact, the database is largely composed of data on quartz par- differences between a data-driven model and a process-based
ticles falling in water. This is the most common engineering con- model. The key capability of a process-based model is its transpar-
dition in the real world: water kinematic viscosity is around ency in physics—each variable has a distinct role in determining
10−6 m2 =s, and most sediment particles are quartz or similar min- the physical process; for example, D determines the volume and
erals with a characteristic density of 2,650 kg=m3 ; water density is surface area of a particle, ν determines the drag forces on settling
around 1,000 kg=m3 ; and the specific gravity Δ ¼ 1.65 is the typ- particles, and Δ determines the relative gravity of particles in fluid.
ical value in sediment transport studies (Soulsby 1997). The data- These variables largely control the particle settling process, with
driven model’s sensitivity does not affect the model application in their contributions to the settling velocity represented as distinct
the real world. For example, Farrell and Sherman (2015) compiled terms in the equations (Table 3). The data-driven model, however,
a database from all five of the existing sediment settling velocity is an inductive approach that embeds theory and logic within the
(a)
(b)
Fig. 6. (Color) Model sensitivity results: (a) v ∈ ½10−6 ; 10−5 ; and (b) Δ ∈ ½1; 3.
model implicitly. The development of the data-driven model relies Under the inspiration of Sadat-Helbar et al. (2009), an ensemble
on the sufficiency and accuracy of the database, but it is not sus- learning method is introduced to generate the theoretical data
ceptible to overreductionism via conceptual assumptions. One con- from the five process-based equations in Table 3. This will
sequence is that the data-driven model might be applicable only incorporate theoretical knowledge of the settling process from
within the range of the data used to develop it. But this limitation the predicted data with reliable accuracy, and the ensemble learn-
also applies to the process-based models using empirical descrip- ing method will integrate five process-based equations and
tions for complex physical process, although it is rarely mentioned reduce the uncertainty of each individual estimate (e.g., Mendes-
(Goldstein et al. 2019). Moreira et al. 2012). The ensemble learning procedure is imple-
mented as follows:
1. Select 1,000 sediment particles with fixed density (ρs ¼
Improvement of RF Estimator with Theoretical Data 2,650 kg=m3 ) but varying grain size (D values that are evenly
The largest deficiency of the RF estimator is that it does not distributed in the range ½10−4 ; 10−2 m).
have theoretical knowledge of sediment settling processes as 2. Estimate the settling velocities of the particles in water (assume
do process-based equations. One solution is to have the RF es- ν ¼ 10−6 m2 =s and Δ ¼ 1.65) using each of the five process-
timator trained on data from theoretically derived predictions. based equations.
metric data-driven model has an easily implemented self-updating Brown, P. P., and D. F. Lawler. 2003. “Sphere drag and settling velocity
capability given more data. By integrating both experimental and revisited.” J. Environ. Eng. 129 (3): 222–231. https://doi.org/10.1061
theoretical knowledge of sediment settling process, the updated /(ASCE)0733-9372(2003)129:3(222).
BSI (British Standards Institution). 1967. Methods of testing soils for civil
model further enhances the model performance with improved esti-
engineering purposes. BS 1377:1967. London: BSI.
mate accuracy and consistency.
BSI (British Standards Institution). 1986. British standard specification for
There are many complex natural processes with incompletely test sieves. BS 410:1986. London: BSI.
resolved physics in the hydraulic engineering research literature, Camenen, B. 2007. “Simple and general formula for the settling velocity of
such as sediment transport, turbulent flows, and water resource particles.” J. Hydraul. Eng. 133 (2): 229–233. https://doi.org/10.1061
management. Deterministic mathematical frameworks for such /(ASCE)0733-9429(2007)133:2(229).
complex processes appear intractable from the perspective of current Chen, S., and C. Hu. 2017. “Estimating sea surface salinity in the northern
process-based approaches. Alternatively, with the increase in avail- Gulf of Mexico from satellite ocean color measurements.” Remote
able data and progress in ML techniques, data-driven approaches Sens. Environ. 201 (Nov): 115–132. https://doi.org/10.1016/j.rse
have been increasingly applied to complex hydraulic engineering re- .2017.09.004.
search, for example, in sediment transport [e.g., a review in Goldstein Chen, S., C. Hu, B. B. Barnes, R. Wanninkhof, W. J. Cai, L. Barbero, and
D. Pierrot. 2019. “A machine learning approach to estimate surface ocean
et al. (2019)], in turbulent flows (e.g., Mohan et al. 2019), and in
pCO2 from satellite measurements.” Remote Sens. Environ. 228 (Jul):
water resource research (e.g., Najafzadeh et al. 2017; Granata et al. 203–226. https://doi.org/10.1016/j.rse.2019.04.019.
2018). This study, using a nonparametric data-driven approach, not Cheng, N.-S. 1997. “Simplified settling velocity formula for sediment
only demonstrates the capability of ML models to better identify the particle.” J. Hydraul. Eng. 123 (2): 149–152. https://doi.org/10.1061
internal relationships present in data but also shows its potential in /(ASCE)0733-9429(1997)123:2(149).
leveraging theoretical knowledge for further model improvement. In Corey, A. T. 1949. “Influence of shape on the fall velocity of sand
addition, it provides an alternative framework that may be applied to grains.” M.S. thesis, Irrigation Engineering, Colorado Agricultural
other hydraulic engineering challenges. and Mechanical College.
Cutler, D. R., T. C. Edwards Jr., K. H. Beard, A. Cutler, K. T. Hess,
J. Gibson, and J. J. Lawler. 2007. “Random forests for classification
in ecology.” Ecology 88 (11): 2783–2792. https://doi.org/10.1890/07
Data Availability Statement -0539.1.
Dietrich, W. E. 1982. “Settling velocity of natural particles.” Water
The compiled data and python code for this study can be obtained Resour. Res. 18 (6): 1615–1626. https://doi.org/10.1029/WR018i006
by sending a written request to the corresponding author. p01615.
Donatelli, C., N. K. Ganju, S. Fagherazzi, and N. Leonardi. 2018.
“Seagrass impact on sediment exchange between tidal flats and salt
Acknowledgments marsh, and the sediment budget of shallow bays.” Geophys. Res. Lett.
45 (10): 4933–4943. https://doi.org/10.1029/2018GL078056.
The authors thank Drs. Goldstein and Coco for sharing data and for El-Nahhas, K., N. G. El-Hak, M. A. Rayan, and I. El-Sawaf. 2009. “Effect
helpful discussions motivating this study. The authors would also of particle size distribution on the hydraulic transport of settling slur-
ries.” In Proc., 13th Int. Water Technology Conf., IWTC13. Hurghada,
like to thank the editors and three anonymous reviewers for their
Egypt: International Water Technology Conference.
valuable comments that helped improve the manuscript. Funding Fagherazzi, S., P. L. Wiberg, S. Temmerman, E. Struyf, Y. Zhao, and P. A.
for this study was provided under the Los Alamos National Labora- Raymond. 2013. “Fluxes of water, sediments, and biogeochemical com-
tory Research and Development Directed Research project “Adaption pounds in salt marshes.” Ecol. Processes 2 (1): 3. https://doi.org/10
Science for Complex Natural-Engineered Systems” (20180033DR). .1186/2192-1709-2-3.
This publication has been supported by the Los Alamos Laboratory Farrell, E. J., and D. J. Sherman. 2015. “A new relationship between grain
Directed Research and Development project under LA-UR-20-22942. size and fall (settling) velocity in air.” Prog. Phys. Geogr. 39 (3):
361–387. https://doi.org/10.1177/0309133314562442.
Ferguson, R., and M. Church. 2004. “A simple universal equation for grain
settling velocity.” J. Sediment. Res. 74 (6): 933–937. https://doi.org/10
References .1306/051204740933.
Fredsoe, J., and R. Deigaard. 1992. “Advanced series on ocean engineer-
Ahrens, J. P. 2000. “A fall-velocity equation.” J. Waterway, Port, Coastal, ing.” In Vol. 3 of Mechanics of coastal sediment transport. Singapore:
Ocean Eng. 126 (2): 99–102. https://doi.org/10.1061/(ASCE)0733 World Scientific.
-950X(2000)126:2(99). Ganju, N. K., D. H. Schoellhamer, and B. E. Jaffe. 2009. “Hindcasting of
Alger, G. 1964. “Terminal fall velocity of particles of irregular shapes as decadal-timescale estuarine bathymetric change with a tidal-timescale
affected by surface area.” Ph.D. dissertation, Dept. of Civil Engineering, model.” J. Geophys. Res. Earth Surf. 114 (F4): F04019. https://doi
Colorado State Univ. .org/10.1029/2008JF001191.
Hallermeier, R. J. 1981. “Terminal settling velocity of commonly occurring Rubey, W. W. 1933. “Settling velocity of gravel, sand, and silt particles.”
sand grains.” Sedimentology 28 (6): 859–865. https://doi.org/10.1111/j Am. J. Sci. 25 (148): 325–338. https://doi.org/10.2475/ajs.s5-25.148
.1365-3091.1981.tb01948.x. .325.
Huijts, K. M. H., H. M. Schuttelaars, H. E. De Swart, and A. Valle- Sadat-Helbar, S. M., E. Amiri-Tokaldany, S. Darby, and A. Shafaie. 2009.
Levinson. 2006. “Lateral entrapment of sediment in tidal estuaries: An “Fall velocity of sediment particles.” In Proc., 4th IASME/WSEAS Int.
idealized model study.” J. Geophys. Res. Oceans 111 (C12): C12016. Conf. on Water Resources, Hydraulics and Hydrology, WHH’09.
https://doi.org/10.1029/2006JC003615. Cambridge, UK: WSEAS Press.
Imai, G. 1980. “Settling behavior of clay suspension.” Soils Found. 20 (2): Schulz, S. E., R. H. Wilde, and M. L. Albertson. 1954. Influence of shape
61–77. https://doi.org/10.3208/sandf1972.20.2_61. on the fall velocity of sedimentary particles. M.R.D Sediment Series
Jiménez, J. A., and O. S. Madsen. 2003. “A simple formula to estimate No. 5. Omaha, NE: USACE.
settling velocity of natural sediments.” J. Waterway, Port, Coastal, Segal, M. R. 2004. Machine learning benchmarks and random forest re-
Ocean Eng. 129 (2): 70–78. https://doi.org/10.1061/(ASCE)0733 gression. Technical Rep. San Francisco: Center for Bioinformatics &
-950X(2003)129:2(70). Molecular Biostatistics, Univ. of California.
Kane, M. J., N. Price, M. Scotch, and P. Rabinowitz. 2014. “Comparison of She, K., L. Trim, and D. Pope. 2005. “Fall velocities of natural sediment
ARIMA and random forest time series models for prediction of avian particles: A simple mathematical presentation of the fall velocity law.”
influenza H5N1 outbreaks.” BMC Bioinf. 15 (1): 276. https://doi.org/10 J. Hydraul. Res. 43 (2): 189–195. https://doi.org/10.1080/00221686
.1186/1471-2105-15-276. .2005.9641235.
Kohavi, R. 1995. “A study of cross-validation and bootstrap for accuracy Shotton, J., T. Sharp, A. Kipman, A. Fitzgibbon, M. Finocchio, A. Blake,
estimation and model selection.” IJCAI 14 (2): 1137–1145. M. Cook, and R. Moore. 2013. “Real-time human pose recognition in
Komar, P. D., and C. Reimers. 1978. “Grain shape effects on settling rates.” parts from single depth images.” Commun. ACM 56 (1): 116–124.
J. Geol. 86 (2): 193–209. https://doi.org/10.1086/649674. https://doi.org/10.1145/2398356.2398381.
Lerman, P. 1980. “Fitting segmented regression models by grid search.” Smith, D. A., and K. F. Cheung. 2003. “Settling characteristics of calca-
J. R. Stat. Soc. Ser. C Appl. Stat. 29 (1): 77–84. https://doi.org/10 reous sand.” J. Hydraul. Eng. 129 (6): 479–483. https://doi.org/10.1061
.2307/2346413. /(ASCE)0733-9429(2003)129:6(479).
Liaw, A., and M. Wiener. 2002. “Classification and regression by random Smith, P. F., S. Ganesh, and P. Liu. 2013. “A comparison of random forest
forest.” R News 2 (3): 18–22. regression and multiple linear regression for prediction in neuro-
Liu, M., M. Wang, J. Wang, and D. Li. 2013. “Comparison of random for- science.” J. Neurosci. Methods 220 (1): 85–91. https://doi.org/10
est, support vector machine and back propagation neural network for .1016/j.jneumeth.2013.08.024.
electronic tongue data classification: Application to the recognition of Soulsby, R. 1997. Dynamics of marine sands: A manual for practical
orange beverage and Chinese vinegar.” Sens. Actuators, B 177 (Feb): applications. London: Thomas Telford.
970–980. https://doi.org/10.1016/j.snb.2012.11.071. Stokes, G. G. 1851. Vol. 9 of On the effect of the internal friction of fluids
Mendes-Moreira, J., C. Soares, A. M. Jorge, and J. F. D. Sousa. 2012. on the motion of pendulums. Cambridge, UK: Pitt Press.
“Ensemble approaches for regression: A survey.” ACM Comput. Surv. Summerfield, M. 2010. Programming in Python 3: A complete introduction
45 (1): 1–40. https://doi.org/10.1145/2379776.2379786. to the Python language. Boston, MA: Addison-Wesley.
Mohan, A., D. Daniel, M. Chertkov, and D. Livescu. 2019. “Compressed Svetnik, V., A. Liaw, C. Tong, J. C. Culberson, R. P. Sheridan, and B. P.
convolutional LSTM: An efficient deep learning framework to model Feuston. 2003. “Random forest: A classification and regression tool for
high fidelity 3D turbulence.” Preprint, submitted February 28, 2019. compound classification and qsar modeling.” J. Chem. Inf. Comput. Sci.
https://arxiv.org/abs/1903.00033. 43 (6): 1947–1958. https://doi.org/10.1021/ci034160g.
Munson, B. R., D. F. Young, and T. H. Okiishi. 2006. Fundamentals of fluid Tyralis, H., and G. Papacharalampous. 2017. “Variable selection in time
mechanics. New York: Wiley. series forecasting using random forests.” Algorithms 10 (4): 114. https://
Najafzadeh, M., A. Tafarojnoruz, and S. Y. Lim. 2017. “Prediction of local doi.org/10.3390/a10040114.
scour depth downstream of sluice gates using data-driven models.” ISH US Inter-Agency Committee. 1957. Some fundamentals of particle size
J. Hydraul. Eng. 23 (2): 195–202. https://doi.org/10.1080/09715010 analysis: A study of methods used in measurement and analysis of sedi-
.2017.1286614. ment loads in streams. Rep. No. 12. Minneapolis: US Inter-Agency
Nasiha, H. J., and P. Shanmugam. 2018. “Estimation of settling velocity Committee on Water Resources.
of sediment particles in estuarine and coastal waters.” Estuarine Van Rijn, L. C. 1989. Handbook: Sediment transport by currents and
Coastal Shelf Sci. 203 (Apr): 59–71. https://doi.org/10.1016/j.ecss waves. Delft, Netherlands: Delft Hydraulics Laboratory.
.2018.02.001. Verikas, A., A. Gelzinis, and M. Bacauskiene. 2011. “Mining data with
Nielsen, P. 1986. “Suspended sediment concentrations under waves.” random forests: A survey and results of new tests.” Pattern Recognit.
Coastal Eng. 10 (1): 23–31. https://doi.org/10.1016/0378-3839(86) 44 (2): 330–349. https://doi.org/10.1016/j.patcog.2010.08.011.
90037-2. Warner, J. C., C. R. Sherwood, and W. R. Geyer. 2007. “Sensitivity of
Nielsen, P. 1993. “Turbulence effects on the settling of suspended par- estuarine turbidity maximum to settling velocity, tidal mixing, and sedi-
ticles.” J. Sediment. Res. 63 (5): 835–838. ment supply.” In Estuarine and coastal fine sediments dynamics, edited
Olabarrieta, M., W. R. Geyer, G. Coco, C. T. Friedrichs, and Z. Cao. 2018. by J. P.-Y. Maa, L. P. Sanford, and D. H. Schoellhamer, 355–376.
“Effects of density-driven flows on the long-term morphodynamic Amsterdam: Elsevier.