Version of Record: https://www.sciencedirect.

com/science/article/pii/S1001627916301536
Manuscript_3ad95cfb4d34956a59242ed7b98e7514

A SEM-based method to determine the mineralogical composition and the

particle size distribution of suspended sediment

Sylvain Pineta,*, Bruno Lartigesa, Jean-Michel Martineza,b, Sylvain Ouillonc

a
University of Toulouse (Paul Sabatier), Geosciences Environment Toulouse, CNRS (Centre National

de la Recherche Scientifique), IRD (Institut de Recherche pour le Développement), 14 av. E. Belin,

31400 Toulouse, France.

b
Instituto de Geosciências, Universidade de Brasilia, Campus Universitario Darcy Ribeiro, ICC

Centro, 70910-900 Brasilia, Brazil.

c
University of Toulouse, LEGOS (Laboratoire d'Etudes en Géophysique et Océanographie Spatiales),

14 av. E. Belin, 31400 Toulouse, France.

* Corresponding author. E-mail address: [email protected] (S. Pinet)

© 2018 published by Elsevier. This manuscript is made available under the Elsevier user license
https://www.elsevier.com/open-access/userlicense/1.0/
 A SEM-based method to determine the mineralogical composition and the

particle size distribution of suspended sediment

Keywords: Mineralogy; Scanning Electron Microscopy; Suspended sediment; Machine

learning; Particle size distribution

Abstract

A robust method for characterizing the mineralogy of suspended sediment in continental

rivers is introduced. It encompasses 3 steps: the filtration of a few milliliters of water,

measurements of X-ray energy dispersive spectra using Scanning Electron Microscopy

(SEM), and robust machine learning tools of classification. The method is applied to

suspended particles collected from various Amazonian rivers. A total of more than 204,000

particles were analyzed by SEM-EDXS (Energy Dispersive X-ray Spectroscopy), i.e. about

15,700 particles per sampling station, which lead to the identification of 15 distinct groups of

mineralogical phases. The size distribution of particles collected on the filters was derived

from the SEM micrographs taken in the backscattered electron imaging mode and analyzed

with ImageJ freeware. The determination of the main mineralogical groups composing the

bulk sediment associated with physical parameters such as particle size distribution or aspect

ratio allows a precise characterization of the load of the terrigenous particles in rivers or lakes.

In the case of the Amazonian rivers investigated, the results show that the identified

mineralogies are consistent with previous studies as well as between the different samples

collected. The method enabled the evolution of grain size distribution from fine to coarse

material to be described in the water column. Implications about hydrodynamic sorting of

mineral particles in the water column are also briefly discussed. The proposed method appears

well suited for intensive routine monitoring of suspended sediment in river systems.

1
1. Introduction

Understanding and quantifying the processes controlling the production, transportation, and

deposition of sediment in watersheds is a major issue for both river geomorphology and river

management. Considering that sediment, and especially clay minerals, are key indicators of

various environmental processes (Moriarty, 1977; Naidu & Mowatt, 1983; Petschick et al.,

1996; Roddaz et al., 2005; Washner et al., 1999), knowledge of suspended sediment

composition as well as its temporal and spatial variability in the river, provides significant

insight about the evolution of watersheds. Conventional analytical procedures for water

analysis such as spectrophotometry, chromatography, or spectrometry as Inductively Coupled

Plasma - Atomic Emission Spectroscopy (ICP-AES) and Inductively Coupled Plasma - Mass

Spectroscopy (ICP-MS) involve a heavy protocol of sample preparations and data analysis.

The development of simple and efficient tools to characterize the bulk of suspended sediment

in rivers is, thus, particularly needed by water agencies and the scientific community to

identify any change in the distribution and in the nature of the material transported within the

catchment during the hydrological cycle.

Such a need is particularly relevant in large river basins, such as the Amazon River basin,

where the amount and the nature of suspended sediment may impact the environment at

various scales, from global (Gibbs, 1977; Ludwig & Probst, 1998; Milliman & Meade, 1983;

Syvitski et al., 2003) to local (Armijos et al., 2013; Guyot et al., 2005; Meade et al., 1985).

Most often, high frequency monitoring is associated with a rough characterization of

suspended sediment. The solid discharge of the Amazon River is now systematically

monitored every 10 days at various sampling stations by the Observation Service program

HYBAM, providing both estimates of Andean erosion rates and of the total amount of

sediment transported through the Amazon River basin to the Atlantic Ocean (Filizola et al.,

1999; Filizola, 2003; Guyot et al., 1988; Laraque et al., 2004; Martinez et al., 2009; Molinier
2
et al., 1996; Safran et al., 2005). On the other hand, there is a significant body of literature

reporting detailed chemical characterizations of the particulate or dissolved load in the

Amazon basin on a one time basis only (Bouchez et al., 2011; Gaillardet et al., 1997; Guyot et

al., 1996; Guyot et al., 2007; Irion, 1983; Johnsson & Meade, 1990; Martinelli et al., 1993;

Moquet et al., 2011; Sondag et al., 2010; Stallard & Edmond, 1983; Tardy et al., 2005), and it

is now envisaged to perform a large analysis of their variations amongst the different sub-

basins, using a unique, simple, and robust method.

Scanning Electron Microscopy coupled with automated Energy Dispersive X-ray

Spectroscopy (SEM-EDXS) and image analysis, simultaneously provide both the size

characteristics and the chemical analysis of thousands of particles. Such a method has

previously been used to characterize the occurrence of heavy metal bearing mineral phases in

polluted soils and sediments (de Boer & Crosby 1995; El Samrani et al., 2004) or to support

the analysis of light-scattering features of particles in water bodies (Peng et al., 2009; Pinet et

al., 2017).

The aim of this paper is to introduce a method that determines the mineralogical composition

of suspended sediment in continental waters. This method involves a simple sampling

protocol combined with the robustness of powerful statistical techniques, making it an ideal

tool for processing a significant amount of water samples collected across a watershed. The

main purpose is to take advantage of the expertise in mineralogical determination, then to

reinforce it by the ability of machine learning tools to identify statistical patterns in large

datasets, and finally to supply an efficient tool to characterize suspended sediment in lakes or

rivers. In the current paper, the method is fully illustrated for suspended sediment sampled in

the Amazon basin, where many one-time analyses are already available in the literature.

2. Materials and methods

2.1. Sampling
3
Three sampling campaigns were done between March 2013 and December 2014 by the

Service National d’Observation (SNO) HYBAM network and its national partners. Water

samples were taken from the surface [0-30 cm] and at various depths in rivers belonging to

the Amazon basin. Eight liters of water were collected at each sampling depth, using a bucket

for the surface samples and a specific depth sampler (Callède, 1994) in the water column.

Turbidity was measured with a turbidity meter (Hach 2100Q) giving for each sample a proxy

of the sediment load. Depending on the turbidity values, from 1 ml of water (for the samples

with the highest turbidity) to 6 ml (for the samples with the lowest turbidity) were collected

with a micropipette (Eppendorf), and filtered through 0.4 µm pore sized polycarbonate

membrane filters (GE Osmonics). The filters were then dried and kept in PetriSlides

(Millipore) until further processing in the laboratory.

2.2. Scanning electron microscopy (SEM)

After preliminary examination with a binocular microscope, a piece of membrane filter was

selected, mounted on an aluminum stub using carbon double-sided conductive tape, and

coated with carbon. SEM observations were done with a Jeol JSM 6363LV microscope

equipped with a silicon drift detector (SDD) PGT operating at 20 kV for chemical analysis.

The backscattered electron imaging mode was used to identify the mineral particles from the

membrane filter (Fig. 1). A series of SEM images was randomly collected at 400x

magnification, taking care not to superimpose successive images. Each identified particle was

analyzed by Energy Dispersive X-ray Spectroscopy (EDXS) to determine its elemental

composition. To obtain the mineralogical composition of classified groups, the stoichiometric

ratios were calculated from the atomic percentages given by the EDXS analysis, and were

then compared with known mineralogical compositions.

SEM images were also used to determine the size distribution of the suspended solids using

the ImageJ freeware (National Institutes of Health, Bethesda, MD, U.S.) in order to compare
4
the size distribution of the samples. Projected area, perimeter, and shape descriptors were

automatically determined for each particle with ImageJ, while excluding any particle touching

the edge of the image. The equivalent diameter, Deq, of each particle was calculated according

to Saleh and Guigon (2009) (after Allen, 1975) as:

Deq = (S / π) ½ (1)

where S is the projected area of the particle in µm2. The particle size distribution (PSD) was

then determined both for each sampling station and for each mineral phase identified, using

Junge’s power-law distribution (Junge, 1963):

N(D) = K x D – J (2)

where N(D) represents the number of particles of diameter D, K is the concentration of

particles and the exponent J is the slope of the distribution. The size bins range from 0.14 to

550 µm. However, because of the low magnification used to acquire the SEM images, only

the size classes greater than 1 µm were taken into account for the calculation of the exponent

J. An aspect ratio (AR) was also obtained from the ratio between the major axis and the minor

axis of the best-fit ellipse around each particle. An AR equal to 1 is a circle, whereas an AR

greater than 1 indicates elongated particles.

2.3. Machine learning

The particles were partitioned into homogeneous groups based on their elemental composition

using the Partitioning Around Medoids (PAM) unsupervised classification (Kaufman &

Rousseeuw, 1987). The PAM method is based on a simple to use k-medoids algorithm that

does robust classifications (Breitkreutz & Casey, 2008). The elemental content of particles

representing 99% of the total of elements detected by the EDXS analysis was considered to

derive the mineralogical groups. These elements are sodium (Na, 2.2%), magnesium (Mg,

2.7%), aluminum (Al, 25.7%), silicon (Si, 57.7%), potassium (K, 2.8%), calcium (Ca, 0.7%),

titanium (Ti, 1.2%), and iron (Fe, 6.9%).

5
The randomForest supervised learning classification algorithm (Breiman, 2001) was applied

on the same training dataset to test the robustness of the groups and to be compared with the

classification resulting from the PAM method. The randomForest classification was then used

on all the remaining sampling stations to extend the mineralogical determination. The

randomForest classification computes hundreds of classification trees, selecting randomly m

variables (i.e. elements) that lead to the split of the dataset at each tree node. A ‘vote’ is

assigned to each tree built during the classification process, and then the majority vote over

the set of trees built in the process designates the most consistent and recurrent classification

as the final result. In the current study, the package used was “randomForest” ported in R by

Liaw and Wiener (2002) from the original Fortran program by Breiman and Cutler (available

at http://www.stat.berkeley.edu/users/breiman/). The sampling stations used for the training

and the validation of the two classifications are described in Table 1 by their dates, locations,

depth, and the corresponding numbers of SEM images and particles analyzed.

Building only two mineralogical groups would determine a minimum error rate in the

classification, since if one particle does not fit in a group, it necessarily belongs to the other

group. However, in that case the consistency within each group is likely to be quite poor, e.g.,

the intra-group variability of the particle types will be much too high to identify the resultant

dominant mineralogy. In contrast, a very high number of groups would correctly classify the

particles according to their chemical analysis, but the similarity between those groups would

be too close, and the mineralogy would be difficult to assess and misclassification would be

frequent. Therefore, the optimal number of groups achieved by the classification was

investigated in order to obtain the best consistency (i.e. a single mineralogy) for each group

while keeping a reasonable number of groups and minimizing the possible randomForest

misclassifications. Such tests required an expertise in earth sciences in support of the

statistical analysis of the classification results that was accomplished in collaboration with

mineralogists. In order to reach the best agreement between PAM and randomForest
6
classifications, various parameterizations for the randomForest classification were tested on

the training dataset.

Fig. 1. Imaging procedure and data processing from SEM micrographs of suspended particles:

(left) image used in grain size analysis and (right) the same image with the particles analyzed

by EDXS and manually tagged using green numbers.

2.4. Study area

The Amazon River basin is the most important watershed in the world, both in terms of

drainage area (6.1 x 106 km², Goulding et al., 2003) spreading from the Andes to the Atlantic

Ocean through a major part of South America, and in terms of water discharge (on average

208 x 103 m3 s-1, Callède et al., 2010) (Fig. 2). The region is characterized by a humid tropical

climate (average rainfall = 2,015 ± 112 mm yr-1, Espinoza Villar et al., 2009) with contrasted

seasonal rainfall patterns depending on the latitude. The Amazon River main stream is known

as the Solimões River from the border between Peru and Brazil until downstream of the Rio

Negro confluence near Manaus. The Madeira River is the third main tributary of the Amazon

River, 200 km downstream from the confluence between the Negro River and Solimões

River, and drains the southeast part of the Amazon River basin (Fig. 2).

Fig. 2. Location of the sampling stations in the Amazon River basin (source Google Earth).

Most sediment transported by the Amazon River originates from the Andean mountains

which are composed of highly erodible rocks (Dunne et al., 1998; Filizola & Guyot, 2009).

Such hydro-morphological characteristics induce massive erosion processes at a large scale in

Ecuador, Peru, and Bolivia (Aalto et al., 2006; Guyot, 1993; Masek et al., 1994); as a result,
7
the Amazonian basin is one of the most important contributors in the world in terms of

sediment input to the ocean (average = 800 x 106 t yr-1, Martinez et al., 2009). The Solimões,

Madeira, and Amazon rivers are classified as white rivers (Gibbs, 1967; Sioli, 1957) because

of their high sediment load, originating mainly from the erosion of the Andes, but also coming

from tributaries, river bank erosion, and remobilization of the older sediments trapped in the

floodplains (Dunne et al., 1998). These sediments correspond mainly to quartz and

assemblages of various clays.

3. Results

3.1. Unsupervised PAM classification

The training of the classifications was done using samples from the Solimões, Madeira and

Amazon rivers during the March 2013 sampling campaign (Table 1, Fig. 2), representing

various water types and various terrigenous origins of suspended particles.

Table 1. Description of the sampling stations used in this study - Ni stands for the number of

SEM images, and Np for the number of particles analyzed by the machine learning process -

Samples used for the PAM classification and the randomForest training are marked with *.

Date Station Depth (m) Ni Np

March 9, 2013 Manacapuru 0 13 1,141*
- Foz de Madeira 0 10 1,971*
- Amazonas 0 14 1,525*
- Manacapuru 10 21 1,729*
April 3, 2014 Belém 0 13 1,481
April 6, 2014 Fonte Boa 0 16 1,154
- - 10 12 1,478
- - 20 8 1,603
- - 30 14 1,236
April 11, 2014 Itapeua 0 14 1,390
- - 25 12 1,787
- - 35 8 1,720
April 13, 2014 Manacapuru 0 16 1,874
- - 30 14 1,902

8
 December 15, 2014 Careiro 0 16 2,291

The unsupervised PAM and the supervised randomForest were both built from the chemical

analyses of 6,366 particles sampled during this campaign. The resulting randomForest

classification was then used on the stations of the Solimões River sampled during the other

field sampling trips (Table 1). The sampling campaigns corresponding to the period of April

and December 2014 were not used in the PAM classification procedure. The classification

process involved a trial-and-error phase in the choice of the number of groups. The

consistency of each group was evaluated through the dispersion of the elementary contents (in

percent). The average standard deviation of the total classification was, thus, calculated and

compared with classifications built with a number of groups ranging from 2 to 20. A value of

less than 3% expressing a great consistency within the different groups built by the PAM

classification was first reached for a minimum of 15 groups. Table 2 lists the mean

elementary composition of the 15 groups obtained after the PAM classification based on the

6,366 particles sampled in March 2013.

Table 2. Mean values and standard deviations of the elementary contents (in %) of each

group representing 99% of the bulk particle composition and built from the unsupervised

PAM classification of 6,366 suspended particles.

Na Mg Al Si K Ca Ti Fe
Mean 0.04 0.01 0.06 99.79 0.00 0.00 0.01 0.07
Group 1
St. dev. 0.33 0.14 0.35 0.59 0.03 0.05 0.11 0.30
Mean 1.34 11.21 26.35 35.00 1.08 0.59 0.47 21.48
Group 2
St. dev. 2.32 7.21 5.75 6.31 1.85 1.23 2.35 7.47
Mean 1.60 2.30 36.02 49.23 8.01 0.06 0.09 2.64
Group 3
St. dev. 1.80 1.29 2.49 2.67 2.04 0.23 0.47 1.41
Mean 1.38 3.34 30.34 54.39 3.22 0.98 0.18 5.47
Group 4
St. dev. 2.27 1.91 2.85 2.24 2.11 1.48 0.62 2.09
Mean 2.39 0.45 20.90 60.18 14.16 0.09 0.04 0.76
Group 5
St. dev. 3.10 1.09 3.42 3.33 4.40 0.40 0.22 1.75
Group 6 Mean 18.94 0.36 21.45 56.35 0.75 1.38 0.03 0.54

9
 St. dev. 4.02 0.85 3.14 3.33 1.60 2.13 0.26 1.10
Mean 0.82 2.26 25.39 63.65 1.73 0.62 0.16 4.75
Group 7
St. dev. 1.68 2.55 4.40 4.28 1.96 1.26 1.02 4.10
Mean 2.06 5.36 29.71 47.13 2.62 1.39 0.90 9.73
Group 8
St. dev. 2.40 3.45 3.66 3.55 2.11 3.90 3.86 3.87
Mean 1.31 118 40.90 50.22 0.96 0.17 0.13 2.74
Group 9
St. dev. 2.27 1.40 6.04 6.66 1.36 0.47 0.70 1.91
Mean 2.08 14.88 10.00 43.35 0.52 11.47 2.03 8.92
Group 10
St. dev. 2.63 11.57 7.46 11.24 1.00 13.53 6.16 7.34
Mean 1.16 1.40 15.20 76.31 1.61 0.28 0.47 2.96
Group 11
St. dev. 2.32 1.54 4.85 4.23 1.65 1.89 3.21 3.77
Mean 1.30 2.28 11.34 18.22 0.77 1.27 2.26 57.64
Group 12
St. dev. 3.19 4.41 6.59 9.66 2.02 2.55 5.85 15.27
Mean 0.64 0.65 8.11 87.61 0.83 0.07 0.17 1.71
Group 13
St. dev. 1.25 1.04 2.17 2.72 0.80 0.47 1.07 1.84
Mean 0.48 0.97 9.61 13.32 0.33 0.21 69.75 4.28
Group 14
St. dev. 1.58 1.79 5.22 7.57 0.65 0.69 16.59 5.96
Mean 0.19 0.05 4.02 94.93 0.16 0.00 0.04 0.55
Group 15
St. dev. 0.69 0.28 1.20 1.79 0.41 0.07 0.26 0.80

This statistical consistency of the classification was confirmed by the expertise of

mineralogists and enabled the association of a unique mineralogy to each group. Two

different groups may, however, be attributed to the same mineralogy as natural particles such

as clays can occur in a range of elemental compositions. These 15 groups represent the

reference from which the remainder of the statistical procedure was applied (i.e. the training

of the randomForest classification). The classification reveals an increase in the Al/Si ratio

from group 1 (defined as pure quartz with Al/Si close to 0) to the most developed and mature

sediments of group 9 (defined as kaolinite with Al/Si close to 1). The classification also

detected two poorly represented groups that were very distinct from the others, with major

contents in iron or titanium (respectively, iron oxides, and titanium dioxides). Except for these

two groups representing less than 3% of the dataset, the classification distinguished 8 groups

of clays (represented by chlorite, smectite, illite, and kaolinite) corresponding to 65% of the

classified particles, 3 groups of quartz (22% of the dataset) and 2 groups of feldspars

(orthoclase and albite, 3% and 7% of the dataset, respectively). The same mineralogies were

10
established for different groups expressing either a chemical variability within a group of

minerals (e.g., the 3 groups of smectite), or the mixing of different particles signatures by the

SEM-EDXS measurements (e.g., the 3 groups of silica particles).

3.2. Supervised randomForest classification

By varying the number of trees, Ntree, built during the classification and the number of

variables, Nvar, sampled at each split, the minimum error rate was obtained for Ntree = 500 and

Nvar = 2. The most representative tree of the forest having the majority ‘votes’ and

corresponding to the best split of the dataset is shown in Fig. 3. This tree highlights the

importance of Si, Al, K, Fe, and Na contents in the classification process. On the other hand,

the 3 remaining elements (Mg, Ti, and Ca) did not significantly contribute to the repartition of

the chemical analyses into distinct groups. These results were confirmed by the internal

calculations of variables importance of randomForest called MeanDecreaseAccuracy and

MeanDecreaseGini (Breiman, 2002), which assign the lowest values (i.e. the minimum

contribution) to Mg, Ti, and Ca, and the highest values to the Si content.

Fig. 3. The classification tree resulting from the randomForest process and the splitting

criterion used to separate the dataset at each node of the tree. Bold numbers on the terminal

nodes represent the corresponding classification group.

3.3. Comparison of the classifications

The resulting tree and the corresponding supervised randomForest classification were

compared to the PAM partitioning. Table 3 shows a confusion matrix where the

misclassification between PAM and randomForest are detailed and quantified via the

classification error rate. The entire process was repeated because of the random aspect of the

variables used during the classification. All the confusion matrices thus obtained showed
11
similar results, and the classification error rate between both classification methods was

always lower than 4% using the combined (Ntree = 500 and Nvar = 2) parametrization. The

randomForest classification was used on the remaining dataset, which included stations

sampled during the campaigns of April and December 2014 (i.e., 17,916 particles). The

mineralogical content of each Solimões River station were then determined from the number

percentage of particles associated with the different groups (Fig. 4).

Table 3. Example of a confusion matrix among the different groups classified by the PAM

and the randomForest machine learning.

Supervised randomForest classification Class

G1 G2 G3 G4 G5 G6 G7 G8 G9 G10 G11 G12 G13 G14 G15 error
G1 662 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0.00150
G2 0 399 1 0 0 0 1 12 3 2 0 2 0 0 0 0.05000
G3 0 0 596 17 0 0 0 8 9 0 0 0 0 0 0 0.05396
G4 0 0 5 1120 0 5 11 13 7 0 0 0 0 0 0 0.03531
Unsupervised PAM classification

G5 0 0 0 1 174 1 3 0 0 3 1 0 0 0 0 0.04918
G6 0 0 0 0 2 436 1 0 0 0 1 0 0 0 0 0.00909
G7 0 0 0 20 0 0 400 1 0 0 4 0 0 0 0 0.05882
G8 0 4 4 16 0 0 0 588 7 1 0 0 0 0 0 0.05161
G9 0 0 6 6 0 0 0 7 467 0 0 0 0 0 0 0.03909
G10 0 4 0 0 0 1 0 4 0 109 1 0 0 0 0 0.08403
G11 0 0 0 0 1 3 8 0 0 4 234 0 3 0 0 0.07509
G12 0 3 0 0 0 0 0 0 0 0 0 163 0 0 0 0.01807
G13 0 0 0 0 0 0 0 0 0 1 7 0 301 0 4 0.03833
G14 0 1 0 0 0 0 0 0 0 0 0 1 0 60 0 0.03225
G15 1 0 0 0 0 0 0 0 0 0 0 0 9 0 415 0.02352

As illustrated in Fig. 4, the Solimões River samples exhibit a very similar pattern in

mineralogical composition, with a maximum standard deviation of σ = 2.5%. The comparison

between Solimões River and Madeira River stations exhibits contrasting dissimilarities,

highlighting the differences in mineralogical composition of the suspended sediments

collected during the March 2013 sampling trip and used for the training of the randomForest

classification. For all the Solimões River stations, the smectite content (sum of groups 3, 4,

12
and 10) represents the main component of the mineralogical assemblage (33.01% on average).

Group 4 is especially dominant representing on average 19.12% of all particles analyzed by

SEM. The remainder of the particles occurs as other clay minerals (illite, chlorite, and

kaolinite, on average 19.74, 7.38, and 6.78%, respectively), quartz (on average 19.17%) and

feldspars (orthoclase and albite, on average 3.03 and 6.87%, respectively). Finally, the less

frequent types of particles are iron oxides and titanium dioxides (on average 2.90 and 1.12%,

respectively).

Fig. 4. Distribution of the mineralogical groups from 1 to 15 at 3 different stations on the

Solimões River sampled in a) March 2013, c) April 2014, and d) December 2014, and a

Madeira River station b) also sampled in March 2013.

The average Pearson’s correlation coefficients between Solimões River sampling stations

taken two by two reached 0.90, but reached only 0.73 when comparing Madeira River and

Solimões River stations. After combining the groups showing analogous mineralogies (e.g.,

groups 1, 13, and 15 representing quartz), the correlation coefficients are even higher,

increasing to 0.97 on average for the Solimões River stations. The mineralogical similarities

evidenced in Fig. 4 between Solimões River sampling stations were confirmed by variance

analysis, which showed no statistical difference between the randomForest classifications of

the Solimões River stations (mean adjusted p-value = 0.96), but established a statistical

distinction between the Madeira and the Solimões sampling stations (mean adjusted p-value =

0.48).

3.4. Particle Size Distribution

A total of more than 204,000 particles were scanned by SEM to analyze their size, which

represents more than 15,700 particles per station on average. Contrasted types of particles
13
were examined with projected areas ranging between 0.062 and 25,204 µm² (5.78 µm² on

average) and perimeters ranging between 0.70 and 1,874 µm (6.02 µm on average). The

resulting equivalent diameters were calculated and ranged between 0.14 and 69.57 µm (0.76

µm on average) (Fig. 5a), which led to the calculation of the grain size distribution of the

particles for each station (Fig. 5b), following the least-squares fitting method. The smallest

value of the slope exponent was obtained for the suspended particles taken at a 30 m depth at

Manacapuru (J = 2.54) and the highest value was for the surface sample of Fonte Boa station

(J = 3.36). As expected, a decrease of J was observed with sampling depth in the water

column, thus, showing an increase of the coarser fraction of the particles (Fig. 5c). As an

example, at the Itapeua station, J is 3.14 at the surface level, 2.96 at a 25 m depth, and is only

2.62 at a 35 m depth. On average, the J value for surface samples is about 3.15, but only

reached 2.84 for samples collected at depth. The particle size distribution of each

mineralogical group was examined. J values ranging from 2.48 for group 14 to 3.43 for group

4 were obtained. Similar J values (3.31) characterized the smectite groups (including groups

3, 4, and 11), the lowest J values were obtained for iron oxides (2.60, group 12) and with

titanium dioxides (2.48, group 14), whereas the other groups (quartz, chlorite, illite,

orthoclase, albite, and kaolinite) exhibited a Junge’s exponent close to 3 (3.01 on average).

The majority of particles showed aspect ratios in the [1 – 2] range, which reveals that most of

the particles were significantly rounded. Such a result was confirmed by the calculation of the

circularity coefficient, C (included in the [0 – 1] range) that corresponds to the ratio between

the projected area S and the squared value of perimeter P of each particle (C = 4π x S / P²).

More than 45% of the particles showed a C value equal to 1, thus, indicating a spherical

particle, and it increased to more than 59% for C values higher than 0.9. Those quasi-

spherical particles showing an AR value of 1 were characterized by small dimensions with a

maximum area of 2 µm² (0.11 µm² on average) and a maximum perimeter of 4.83 µm (0.90

µm on average).
14
Fig. 5. Grain size distribution of the particles analyzed by SEM: a) histogram of the calculated

equivalent diameters; b) particle number/size distribution at each sampling station (for D > 1

µm); c) corresponding Junge’s distribution exponent, J, values at various depths (black and

white triangles represent Manacapuru station sampled in 2013 and 2014, respectively. Black

and white circles represent Fonte Boa station and Itapeua station, respectively. Black and

white squares represent Belém station and Careiro station, respectively).

4. Discussion

With more than 96% of particles getting a correct assignment (Table 3), the supervised

randomForest classification ensures great consistency with the groups built from the

unsupervised PAM method. The relative importance of the mineralogical groups, thus,

obtained compares favorably with those inferred by X-ray diffraction in Guyot et al. (2007),

at similar locations in the Amazon basin (including the Solimões River). These authors used a

semi-quantitative method to manually identify the clay minerals present in each sample

collected from the banks of the different rivers. In the Brazilian floodplain, Guyot et al.

(2007) found approximatively 25% by weight (wt.) for the illite/chlorite assemblage, 10% wt.

for kaolinite, and 65% wt. for the smectite/mixed layers assemblage for the Solimões River.

In the current case, excluding quartz from the analysis, the composition of the same

assemblages are on average 33.55, 8.39, and 58.06% (percent in number), respectively. The

clay content in the Solimões River retrieved by the two methods are very close, the smectite

content being predominant, and the amount of kaolinite, which indicates the presence of

mature sediments, being significant. Despite the analysis of more than 200,000 particles, the

initial sampling protocol and the magnification used for the SEM observations preclude the

consideration of large particles in the size distribution. Indeed, very few particles larger than
15
10 µm² and almost none larger than 20 µm² were collected, thus, leading to an overestimation

of the J exponent of Junge’s law which describes the particle size distribution. The absence of

the coarse size classes may smooth the differences in the PSD between samples collected at

different depths within the water column. Nevertheless, such underestimation of large

particles does not conceal the decrease in Junge’s parameter, J, with increasing depth at a

given sampling site, either considering every size class greater than 1 µm, or just taking into

account size classes between 1 and 10 µm. In other words, an increasing proportion of coarse

sedimentary materials is observed with increasing depth, as illustrated in Fig. 6.

Fig. 6. Two SEM images (x400 magnification) corresponding to the Manacapuru station

sampled at the same scale in April 2014, a) at the water surface and b) at 30 m deep.

On the other hand, the analysis of the PSD of the various mineralogical groups showed more

contrast. Indeed, iron oxides and titanium dioxides are characterized by coarse material,

whereas smectites, and especially those associated with group 4, are mainly composed of

small particles. Such results are consistent with Chipera and Bish (2001), who showed that,

among various types of clays, smectite represented the finest size fraction. However, there

was no significant difference between the other groups in terms of particle size distribution.

Slight variations were observed, but the general pattern led to a J value of approximately 3.

Regarding the aspect ratio analysis, Li et al. (2016) have revealed that particles from deep

rivers are generally characterized by a regular shape, with slight roughness. The current

results in terms of sediments shape, with more than 59% of particles characterized by a

roundness coefficient typical of a quasi-spherical shape, are consistent with that study. It is

worth noting that the sphericity index of particles is inferred from projected surfaces and not

from direct volume measurements. Thus, taking into account that clays occur as flake-shaped
16
particles, it is expected that clay minerals present a less spherical shape compared with that of

other silicates, such as quartz or feldspar. These differences in mineral shapes are likely to be

magnified for coarser materials (Krinsley & Smalley, 1973). It will induce de facto an

increasing dissociability of the number of particles and associated volume between clays and

quartz sediments when considering deeper samples in the water column. It then becomes

obvious that the absolute content in Si will increase with depth even if the relative number

percentage of quartz particles remains unchanged. On the contrary, the volume percent

occupied by clays will decrease with depth, considering that quartz minerals occupy more

volume than clay minerals. Thus, using the Al/Si ratio, Bouchez et al. (2011) showed profiles

describing a direct relation between depth and mineralogical composition, and suggested that

coarser materials are Si enriched. In contrast, the current results indicate that, considering the

composition of single particles instead of the bulk composition of suspended sediment, there

is no significant pattern between Al/Si ratio and depth (Fig. 7a). In addition, no specific trend

allows the description of a split between the different mineralogies based on sampling depth

(Figs. 7a and 7b).

Fig. 7. Mineralogical composition of suspended sediments of the Solimões River stations at

various depths: a) the Al/Si ratio, illite, chorite, kaolinite and orthoclase contents; b) the

smectite, quartz, albite, iron oxide and titanium oxide contents.

The Si content is then truly higher and the Al content is indeed lower as the sampling depth

increases, but this should not be related to an increased proportion of quartz or feldspar

particles compared with clay minerals. Obviously, the latter result is somehow related to the

range of particle size explored in the current study by SEM, and should be confirmed by

extending the investigation to both finer and coarser materials. Nevertheless, it should be
17
pointed out that the levels of turbulence (and, thus, of vertical mixing) were likely different in

stations sampled by Bouchez et al. (2011) and in the current study. Such a parameter may be

estimated in future sampling to support the analysis of vertical profiles (e.g., Lefebvre et al.,

2012).

5. Conclusions

The new PAM-randomForest procedure presented in this paper shows robust results and

offers an easy method to characterize suspended sediments. The determination technique of

the mineralogical assemblages in rivers is highly representative of the total sediment, even

though just a few milliliters of water were sampled. It remains a powerful procedure even in

the presence of a large dataset, which is essential in a spirit of maximum representativeness of

the measurements. It also allows a more precise characterization of the suspended sediment

sampled compared with an estimate of the total mass of the bulk particles. This method is

particularly interesting for intensive sampling programs, where a series of samplings are made

and where it is easier to keep filters instead of several litters of water or kilograms of

sediments. Complementary analyses have to be made to confirm the distribution of the

different mineralogies within the water column for larger grain sizes. However, as part of a

long term monitoring program, this technique could be routinely used and would provide

important information on the sediment content and its variability in large river basins.

Acknowledgements

The authors are grateful to the French Centre National d’Etudes Spatiales (CNES) and the

French company NOVELTIS for co-funding the first author’s Ph.D. This work was also

supported by the Brazilian Agência Nacional de Águas (ANA) for funding the field

campaigns. The authors thank also Thierry Aigouy of the French Laboratory Géosciences

Environnement Toulouse (UMR 5563), Sophie Gouy from the French Centre de Micro
18
Caractérisation Raimond Castaing (UMS 3623), and Jacques Yvon from the French

laboratory GeoRessources (UMR 7359) for technical and scientific support.

References

Aalto, R., Dunne, T., & Guyot, J. L. (2006). Geomorphic controls on Andean denudation

rates. The Journal of Geology, 114(1), 85-99.

Allen, T. (1975). Particle size measurement. Bradford: Springer.

Armijos, E., Laraque, A., Barba, S., Bourrel, L., Ceron, C., Lagane, C., …Guyot, J. L. (2013).

Yields of suspended sediment and dissolved solids from the Andean basins of

Ecuador. Hydrological Sciences Journal, 58(7), 1478-1494.

Bouchez, J., Gaillardet, J., France‐Lanord, C., Maurice, L., & Dutra‐Maia, P. (2011). Grain

size control of river suspended sediment geochemistry: Clues from Amazon River

depth profiles. Geochemistry, Geophysics, Geosystems, 12(3).

Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5-32.

Breiman, L. (2002). Manual on setting up, using, and understanding random forests v3.1.

Statistics Department, University of California at Berkeley, U.S.

Breitkreutz, D., & Casey, K. (2008). Clusterers: A comparison of partitioning and density-

based algorithms and a discussion of optimisations. Technical Report 11999, James

Cook University, Townsville, Australia, 14 p.

Callède, J. (1994). Coletor de amostra d'água. Publication HiBAm, Brasília, 16 p. (In

Portuguese)

Callède, J., Cochonneau, G., Alves, F. V., Guyot, J. L., Guimarães, V. S., & De Oliveira, E.

(2010). Les apports en eau de l'Amazone à l'Océan Atlantique. Revue des Sciences de

l'Eau/Journal of Water Science, 23(3), 247-273. (In French)

Chipera, S. J., & Bish, D. L. (2001). Baseline studies of the clay minerals society source

clays: Powder X-ray diffraction analyses. Clays and Clay Minerals, 49(5), 398-409.
19
de Boer, D.H., & Crosby, G. (1995). Evaluating the potential of SEM/EDS analysis for

fingerprinting suspended sediment derived from two contrasting topsoils. Catena, 24,

243–258.

Dunne, T., Mertes, L. A., Meade, R. H., Richey, J. E., & Forsberg, B. R. (1998). Exchanges

of sediment between the flood plain and channel of the Amazon River in Brazil.

Geological Society of America Bulletin, 110(4), 450-467.

El Samrani, A. G., Lartiges, B. S., Ghanbaja, J., Yvon, J., & Kohler, A. (2004). Trace element

carriers in combined sewer during dry and wet weather: An electron microscope

investigation. Water Research, 38, 2063-2076.

Espinoza Villar, J. C., Ronchail, J., Guyot, J. L., Cochonneau, G., Filizola, N., Lavado, W.,

…Vauchel, P. (2009). Spatio-temporal rainfall variability in the Amazon basin

countries (Brazil, Peru, Bolivia, Colombia, and Ecuador). International Journal of

Climatology, 29, 1574–1594. doi:10.1002/joc.1791

Filizola, N. (2003). Transfert sédimentaire actuel par les fleuves amazoniens. (Ph.D.

dissertation), Université Paul-Sabatier, Toulouse III, France. (In French)

Filizola, N., & Guyot, J. L. (2009). Suspended sediment yields in the Amazon basin: An

assessment using the Brazilian national data set. Hydrological Processes, 23(22),

3207-3215.

Filizola, N., Guyot, J. L., & Boaventura, G. (1999). Fluxo de sedimentos em suspensão na

Amazônia– a análise a partir da base de dados da ANEEL. Proceedings of the

Hydrological and Geochemical Processes in Large Scale River Basins, Manaus, 99.

(In Portuguese)

Gaillardet, J., Dupré, B., Allègre, C. J., & Négrel, P. (1997). Chemical and physical

denudation in the Amazon River basin. Chemical Geology, 142(3), 141-173.

Gibbs, R. J. (1967). Amazon rivers: Environmental factors that control its dissolved and

suspended load. Science, 156, 1734-1737.

20
Gibbs, R. J. (1977). Clay mineral segregation in the marine environment. Journal of

Sedimentary Research, 47(1), 237-243.

Goulding, M., Barthem, R., & Ferreira, E. (2003). The Smithsonian atlas of the Amazon.

Washington, D.C.: Smithsonian Books.

Guyot, J. L. (1993). Hydrogéochimie des fleuves de l'Amazonie bolivienne. (Ph.D.

dissertation), Editions de l’ORSTOM, Collection Etudes et Thèses, Paris, France.

Guyot, J. L., Bourges, J., Hoorelbecke, R., Roche, M. A., Calle, H., Cortes, J., & Guzman, M.

C. B. (1988). Exportation de matières en suspension des Andes vers l'Amazonie par

le Rio Béni, Bolivie. In: M. P. Bordas & D. E. Walling (Eds.), Porto Alegre

Symposium, 1988, Sediment Budgets (pp. 443-451), IAHS Publication 174. (In

French)

Guyot, J. L., Filizola, N., & Laraque, A. (2005). Régime et bilan du flux sédimentaire de

l’Amazone à Óbidos (Pará, Brésil) de 1995 à 2003. In: D. E. Walling & A. J.

Horowitz (Eds.), Sediments Budgets 1 (pp. 347-354) IAHS Publication 291. (In

French)

Guyot, J. L., Filizola, N., Quintanilla, J., & Cortez, J. (1996). Dissolved solids and suspended

sediment yields in the Rio Madeira basin, from the Bolivian Andes to the Amazon.

In: D. E. Walling, & B.W. Webb (Eds.), Exeter symposium, 1996, Erosion and

sediment yield: Global and regional perspectives (pp. 55-63), IAHS Publication 236.

Guyot, J. L., Jouanneau, J. M., Soares, L., Boaventura, G. R., Maillet, N., & Lagane, C.

(2007). Clay mineral composition of river sediments in the Amazon Basin. Catena,

71(2), 340-356.

Irion, G. (1983). Clay mineralogy of the suspended load of the Amazon and of rivers in the

Papua-New Guinea mainland. Transport of carbon and minerals in major world

rivers, 2, 482-504.

21
Johnsson, M. J., & Meade, R. H. (1990). Chemical weathering of fluvial sediments during

alluvial storage: The Macuapanim Island point bar, Solimoes River, Brazil. Journal

of Sedimentary Research, 60(6), 827-842.

Junge, C. E. (1963). Air chemistry and radioactivity. New York: Academic Press.

Kaufman, L., & Rousseeuw, P. (1987). Clustering by means of medoids. In: Y. Dodge (Ed.),

Statistical data analysis based on the L1 norm and related methods (pp. 405-416)

Amsterdam: North-Holland.

Krinsley, D. T., & Smalley, I. J. (1973). Shape and nature of small sedimentary quartz

particles. Science, 180(4092), 1277-1279.

Laraque, A., Ceron, C., Armijos, E., Pombosa, R., Magat, P., & Guyot, J. L. (2004). Sediment

yields and erosion rates in the Napo River basin: An Ecuadorian Andean Amazon

tributary. In: V. Golosov, V. Belyaev, & D. E. Walling (Eds.), Sediment transfer

through the fluvial system, (pp. 220- 225), IAHS Publication 288.

Lefebvre, J. P., Ouillon S., Vinh, V. D., Arfi, R., Panche, J. Y., Mari, X., …Torréton, J. P.,

(2012). Seasonal variability of cohesive sediment aggregation in the Bach Dang-Cam

Estuary, Haiphong (Vietnam). Geo-Marine Letters, 32(2), 103-121.

Li, D., Li, Y., Wang, Z., Wang, X., & Li, Y. (2016). Quantitative, SEM-based shape analysis

of sediment particles in the Yellow River. International Journal of Sediment

Research, 31(4), 341-350.

Liaw, A., & Wiener, M. (2002). Classification and regression by randomForest. R news, 2,

18–22.

Ludwig, W., & Probst, J. L. (1998). River sediment discharge to the oceans; present-day

controls and global budgets. American Journal of Science, 298(4), 265-295.

Martinelli, L. A., Victoria, R. L., Dematte, J. L. I., Richey, J. E., & Devol, A. H. (1993).

Chemical and mineralogical composition of Amazon River floodplain sediments,

Brazil. Applied Geochemistry, 8(4), 391-402.

22
Martinez, J. M., Guyot, J. L., Filizola, N., & Sondag, F. (2009). Increase in suspended

sediment discharge of the Amazon River assessed by monitoring network and

satellite data. Catena, 79(3), 257-264.

Masek, J. G., Isacks, B. L., Gubbels, T. L., & Fielding, E. J. (1994). Erosion and tectonics at

the margins of continental plateaus. Journal of Geophysical Research, 99, 13941-

13956.

Meade, R. H., Dunne, T., Richey, J. E., Santos, U. D. M., & Salati, E. (1985). Storage and

remobilization of suspended sediment in the lower Amazon River of Brazil. Science,

228(4698), 488-490.

Milliman, J. D., & Meade, R. H. (1983). World-wide delivery of river sediment to the oceans.

The Journal of Geology, 9(1), 1-21.

Molinier, M., Guyot, J. L., De Oliveira, E., & Guimarães, V. (1996). Les régimes

hydrologiques de l'Amazone et de ses affluents. In: P. Chevallier, & B. Pouyaud

(Eds.), Tropical hydrology: A geoscience and a tool for sustainability, dedicated to

the memory of Jean Rodier, Paris, 1995, (pp. 209-222), IAHS Publication 238. (In

French)

Moquet, J. S., Crave, A., Viers, J., Seyler, P., Armijos, E., Bourrel, L., …Guyot, J. L. (2011).

Chemical weathering and atmospheric/soil CO2 uptake in the Andean and Foreland

Amazon basins. Chemical Geology, 287(1), 1-26.

Moriarty, K. C. (1977). Clay minerals in Southeast Indian Ocean sediments, transport

mechanisms and depositional environments. Marine Geology, 25(1-3), 149-174.

Naidu, A. S., & Mowatt, T. C. (1983). Sources and dispersal patterns of clay minerals in

surface sediments from the continental-shelf areas off Alaska. Geological Society of

America Bulletin, 94(7), 841-854.

23
Peng, F., Effler, S. W., Pierson, D. C., & Smith, D. G. (2009). Light-scattering features of

turbidity causing particles in interconnected reservoir basins and a connecting

stream. Water Research, 43, 2280-292.

Petschick, R., Kuhn, G., & Gingele, F. (1996). Clay mineral distribution in surface sediments

of the South Atlantic: Sources, transport, and relation to oceanography. Marine

Geology, 130(3), 203-229.

Pinet, S., Martinez, J. M., Ouillon, S. Lartiges, B., & Villar, R. E. (2017). Variability of

apparent and inherent optical properties of sediment-laden waters in large river

basins - Lessons from in situ measurements and bio-optical modeling. Optics

Express, 25(8), A283-A310.

Roddaz, M., Viers, J., Brusset, S., Baby, P., & Hérail, G. (2005). Sediment provenances and

drainage evolution of the Neogene Amazonian foreland basin. Earth and Planetary

Science Letters, 239(1), 57-78.

Safran, E. B., Bierman, P. R., Aalto, R., Dunne, T., Whipple, K. X., & Caffee, M. (2005).

Erosion rates driven by channel network incision in the Bolivian Andes. Earth

Surface Processes and Landforms, 30(8), 1007-1024.

Saleh, K., & Guigon, P. (2009). Caractérisation et analyse des poudres: Propriétés physiques

des solides divisés. Techniques de l'ingénieur. Génie des Procédés, (J2251). (In

French)

Sioli, H. (1957). Sedimentation im Amazonasgebiet. Geologische Rundschau, 45(3), 608-633.

Sondag, F., Guyot, J. L., Moquet, J. S., Laraque, A., Adele, G., Cochonneau, G., …Vauchel,

P. (2010). Suspended sediment and dissolved load budgets of two Amazonian rivers

from the Guiana Shield: Maroni River at Langa Tabiki and Oyapock River at Saut

Maripa (French Guiana). Hydrological Processes, 24(11), 1433-1445.

24
Stallard, R. F., & Edmond, J. M. (1983). Geochemistry of the Amazon: 2. The influence of

geology and weathering environment on the dissolved load. Journal of Geophysical

Research: Oceans, 88(C14), 9671-9688.

Syvitski, J. P., Peckham, S. D., Hilberman, R., & Mulder, T. (2003). Predicting the terrestrial

flux of sediment to the global ocean: A planetary perspective. Sedimentary Geology,

162(1), 5-24.

Tardy, Y., Bustillo, V., Roquin, C., Mortatti, J., & Victoria, R. (2005). The Amazon. Bio-

geochemistry applied to river basin management: Part I. Hydro-climatology,

hydrograph separation, mass transfer balances, stable isotopes, and modelling.

Applied Geochemistry, 20(9), 1746-1829.

Washner, M., Müller, C., Stein, R., Ivanov, G., Levitan, M., Shelekhova, E., & Tarasov, G.

(1999). Clay‐mineral distribution in surface sediments of the Eurasian Arctic Ocean

and continental margin as indicator for source areas and transport pathways - a

synthesis. Boreas, 28(1), 215-233.

25