Suspended sediment load and bedload flux from the Glacier d’Otemma
proglacial forefield (summers 2020 and 2021)
D. Mancini1, M. Dietze2,3, M. Jenkin1, F. Miesen1 and S.N. Lane1
1
Institute of Earth Surface Dynamics (IDYST), Université de Lausanne, Lausanne,
Switzerland.
2
Institute of Geography, Georg-August-University Göttingen, Göttingen, Germany.
3
German Research Center for Geosciences (GFZ), Potsdam, Germany.
1. Introduction
The Glacier d’Otemma proglacial margin, located in the Swiss Alps at an altitude of about 2450
m a.s.l. (45.93423 N, 7.41160 E), is characterized by a ca. 1 km long by 200 m wide active
braided forefield (Figure 1). In this setting we installed two gauging stations for the monitoring
of both suspended sediment and bedload transport within the proglacial margin: GS1 at about
350 m from the glacier terminus and GS2 at the forefield outlet.
Figure 1: Glacier d’Otemma proglacial forefield with location of gauging station (GS) 1 and 2. Yellow dots refers to the location
of geophones, while orange squares to turbidity and water stage probes. Source of background images: Swisstopo.
Monitoring stations were equipped with water pressure sensors (CS451 from Cambell
Scientific), turbidity probes (OBS300+ from Cambell Scientific) and geophones (3-components
PE-6/B from Sensor Nederland connected to a DiGOS DATA-CUBE type 2 logger). Water
discharge were determined following modalities described in Müller and Miesen (2022).
Suspended loads were quantified using a conventional turbidity-suspended sediment
concentration relationship, while bedload transport was derived seismically using the
geophysical Fluvial model inversion (FMI) algorithm developed in Dietze et al. (2018). The
dataset covers two melt seasons characterized by different climatic conditions: summer 2020
from June 25th (Julian Day [JD] 177) to August 29th (JD 242), which was warmer and drier; and
1
summer 2021 from June 11th (JD 162) to August 21st (JD 233), which was colder and wetter
especially in the first part of the season. Here a brief description of the methodology used to
obtain sediment transport quantifications is given, but further details are available in Mancini
et al. (2023).
2. Suspended sediment load transport
Turbidity records were automatically collected at a sampling frequency of 2 minutes. In 2020
we installed single probes at both stations, while in 2021 we had two probes per location (one
installed above the other) to deal with burial problem. For each year, and for each gauging
station, we developed a specific calibration curve describing the relation between manually
collected suspended sediment concentration through the USDH-48 technique [g/l] and
recorded turbidity [NTU]. These were fitted using robust regression using the iteratively-
reweighted least squares (IRLS) method after Beaton and Tukey (1974) (also known as the
bi-square method). For both stations in 2020 and 2021 we observed the relationship between
turbidity and concentration to be linear as expected for the range of concentrations sampled
given the probe specification, thus we fitted first order polynomials (y = ax+b, where y is the
concentration [g/l] and x is the turbidity [NTU]). The obtained R2 values between the two
variables are generally very high. The results are shown in Table S1.
Table 1: Summary of suspended sediment calibration and analysis results. GS2 data in 2021 are divided into two
periods because of burial problems and sensor failure: the higher probe was used until the 6 th August and the lower
probe from the 6th August.
Date 2020 2021
GS2
Site GS1 GS2 GS1
Period 1; Period 2
Start date 25th June 26th June 10th June 10th June
End date 14th September 14th September 8th September 8th September
Analysis
26th June 26th June 10th June 10th June
period after 31st August 31st August 21st August 21st August
cleaning
Percentage of
record
common to 89.6% 90.0%
both stations
after cleaning
Manual
suspended
66 67 45 28; 30
sediment
samples
a 0.0040 0.0032 0.0038 0.0072; 0.0043
b 0.467 -0.007 0.180 -0.107; 0.026
R2 85.3 % 91.7 % 93.5% 85.8%; 77.1%
The calibration curves were then combined with the cleaned turbidity data to estimate time-
series of concentration (Ci, where i is the time stamp). For each fit, confidence intervals were
used to determine the uncertainty in concentration for each estimation (𝜎𝐶,𝑖 ). The
instantaneous load was calculated for each time period as QiCi, where Qi is the instantaneous
discharge rate measured in Müller and Miesen (2022). We applied a product rule (Taylor, 1997)
to estimate the associated uncertainty (𝜎𝑄𝐶,𝑖 ) under the assumption that the errors are
Gaussian and independent:
2
𝜎 2 𝜎 2 0.5
𝜎𝑄𝐶,𝑖 = ±𝑄𝑖 𝐶𝑖 [( 𝑄𝑄,𝑖 ) + ( 𝐶𝐶,𝑖) ] (Eq. 1)
𝑖 𝑖
3. Bedload transport
All the three components of seismic signals (two horizontal and one vertical) were collected.
Each seismic station was composed of three geophones installed in a triangular array with two
of them placed at about ca. 5 m from the river and the other one (the middle one) spaced out
ca. 70 m towards the valley sidewalls, and following best practices described in Dietze et al.
(2019). Data issued from geophones A were used to infer bedload flux, while data collected by
the other geophones were used for cleaning the obtained timeseries (see below). Loggers
were set at a sampling frequency of 400 Hz with an amplifier gain of 32.
The post-processing of collected seismic data to quantify bedload transport rates used the
geophysical inversion model of Dietze et al. (2019) in the open source R package eseis
(version 0.5.0) (Dietze, 2018). The principle consists of comparing the measured seismic
spectra to a set of pre-calculated simulations of physical models predicting the spectra of fluvial
turbulence and bed sediment transport. The FMI takes into account the two physical models
to describe the seismic signals generated by water turbulence (Gimbert et al., 2014) and the
motion of bedload particles on the riverbed (Tsai et al., 2012). These rely on a set of 15
constant site-specific parameters describing the sedimentological proprieties of bedload
particles and of the fluid, the topography of the studied reach, the computational needs for the
inversion and the seismic ground proprieties of the surveyed area (Table 2).
Table 2: Input parameters describing the FMI and calibrated values for both seismic stations.
Parameter Symbol GS1 GS2
D50 bedload grain diameter [m] ds 0.08 0.04
Sedimentological Grain diameter standard deviation [log m] ss 1.1 1.4
and fluid 3
proprieties Sediment density [kg/m ] rs 2650 2650
Fluid density [kg/m3] rw 1040 1040
Average channel width [m] ww 10 8
Topography Channel slope [rad] aw 0.03 0.02
Distance river center to sensor [m] r0 10 11
Reference frequency [Hz] f0 1 1
Computational
Model frequency range [Hz] f 20-70 20-90
Material quality factor at f0 [-] q0 32.84 32.14
Seismic ground Rayleigh wave phase velocity at f0 [m/s] v0 197 4550.9
proprieties Variation coefficient for v0 [-] p0 0.43 0.95
Q increase with frequency [-] e0 0.02 0.004
Bedload flux [kg/ms] qs 0.001 - 10 0.001 - 10
Variables
Water depth [m] hw 0.3 - 1 0.3 - 1
These parameter values were quantified directly in the field, or extracted from previous studies,
as was the case for both sediment (rs) and fluid (rw) densities which were retrieved from Dietze
et al. (2019). The three parameters describing the topography of the seismically monitored
stream reaches (ww, aw and r0) were quantified using dGPS measurements, while the
computational ones are set according to conventional seismic rules (f0) and quality of the
collected data (f). This latter parameter was determined looking at the shape of the empiric
spectra recorded in periods having different hydraulic conditions in terms of water stage and
bedload transport rates. The median diameter of moving bedload particles (ds) and their
3
standard deviation (ss) were quantified using a log-raised cosine distribution function following
Tsai et al. (2012) to take into account the disproportionately large effect of larger grain sizes
that have on the signal compared to smaller ones because of the higher vertical impact velocity
when they are moving (hopping) in contact with the riverbed.
The seismic ground properties are used in the Green’s function (Eq. S2; Aki and Richards,
2002; Burtin et al., 2016; Bakker et al., 2020) and they describe the way in which seismic
waves propagate through the ground (v0 and p0), and the attenuation imposed by the medium
on the carried seismic magnitude (q0 and e0).
𝑢(𝑓, 𝑥) = 2𝜋𝑓𝐹(𝑓, 𝑥0 )𝐺(𝑓, 𝑥, 𝑥0 ) (Eq. 2)
Where:
u(f,x) = seismic signal ground velocity at frequency f and at a distance x from the source
F(f, x0) = Fourier transform of the magnitude at a given temporal period t
G(f,x,x0) = Greens function converting force to ground velocity and describing the
frequency-dependent wave attenuation related to proprieties of the medium through
which the seismic waves have passed. Under the assumption that seismic forces
impact orthogonal against the seismic device and that Rayleigh surface waves are the
main excited waves, G can be expressed as (e.g. Sanchez-Sesma et al., 2011):
𝑘 2
𝐺(𝑓, 𝑥, 𝑥0 ) = √ 𝑒 −𝜋𝑓𝑟 /(𝑣𝑔 𝑄) (Eq. 3)
8𝑝𝑠 𝑣𝑝 𝑣𝑔 𝜋𝑘𝑟
Where:
ps = volumetric mass density of the medium [kg/m3]
vp = seismic wave phase velocity at f0 [m/s]
vg = seismic wave group velocity [m/s]
k = angular wavenumber (k = 2f/vc)
r = distance seismic source-receiver (|x-x0|) [m]
Q = quality factor [-]
We determined seismic ground parameter values for the study area combining active tests
(e.g. Bakker et al., 2020) with a Generalized Likelihood Uncertainty Estimation (GLUE)
approach allowing to calibrate these unknown parameter values statistically.
Once the FMI was calibrated, the inversion process to convert seismic date into bedload flux
and water stage was applied following Dietze (2018). First, the raw vertical component of the
seismic records for the entire period of investigation collected by geophones A was de-
convoluted, de-trended and finally clipped according to a pre-defined time step ti (3 minutes in
our case) in order to be converted into the frequency domain to produce a series of observed
seismic spectra. Second, the two physical models describing the seismic activity related to
flow turbulence (Gimbert et al., 2014) and bedload transport (Tsai et al., 2012) were used into
a Monte Carlo simulation to produce a series of synthetic spectra describing different potential
seismic conditions occurring in the monitored river reach that serve as a lookup table during
the inversion process. This was achieved keeping all parameters values constant, but allowing
for random variation of the two unknown parameters of interest (i.e. bedload flux and water
depth) within pre-defined possible ranges (Table 2). As bedload transport is an unsteady
variable over time, to enhance the representation of all possible seismic conditions that may
have occurred, the FMI allows production of synthetic spectra with and without bedload
transport. The inversion is then computed by comparing the observed and the synthetic
spectra: for each spectrum measured at time ti, the synthetic one having the lowest RMS error
is kept and the associated instantaneous bedload flux and water depth values are retrieved
from the lookup table. Inverted and measured water stage records were compared to evaluate
both quality and reliability of model parametrization (Mancini et al., 2023).
4
Inferred bedload transport rates were cleaned from potential interferences with unwanted
seismic sources such as anthropogenic noise (e.g. hikers and operator footsteps, as well as
seismic noise generated by other scientific activities taking place in the forefield) or due to
other geomorphic processes (e.g. mass movements) occurring on adjacent hillslopes. We
investigated the source of every single transport peak to be sure that they were effectively
related to riverine processes. We extracted the temporal occurrence of these events and we
localized the seismic source by applying the signal_migrate function (also available in the eseis
R-package; Dietze, 2018) thanks to our triangular array of geophones used at each monitoring
station. This allowed us to sort the signal into two origins: if the estimated source was located
into or close to the nearest channel we kept the inversion result, in the opposite case we
discarded it. Then, we also discarded results matching with the occurrence of strong seismic
activates exciting the entire range of considered frequencies in power spectral density plots as
due to activities occurring in close proximity to geophones (e.g. wild animals) re-interpolating
bedload flux timeseries using a moving mean operator (window size of 20 elements). Finally,
bedload flux estimations in kg/ms are converted into kg/s multiplying them by the average
channel width (10 m in GS1 and 8 m in GS2, Table 2), and interpolated at the same frequency
of suspended sediment records (i.e. 2 minutes).
As the main parameters controlling FMI performance are those describing seismic ground
proprieties, bedload flux uncertainty was assessed using Latin Hypercube Sampling (LHS;
McKay, 1992) applied to q0, v0, p0, e0 and ss parameters (Table 2). This method consists of
randomly sampling within a hypercube having the same dimensions of the investigated
variables (in this case 5), in our case four, in a way that every combination of parameter value
can be sampled only once. Ground seismic parameter ranges are based on both mean and
standard deviation of the best 100 simulations used to define final parameter values, while that
for ss is defined as the standard deviation obtained from the probabilistic log-raised cosine
function. We generated a total of 1000 possible combinations and we classified ground seismic
parameter sets according to a normal distribution in 5 different classes of standardized lengths
for each parameter space. A total of 20 possible combinations (5 per parameter, one per each
class associated to a specific ss value) are selected and implemented into the FMI to produce
several time series of bedload flux. Each set of inversion took about two weeks for a whole
melt season on a high performance computer, and this justifies our choice to limit the total
number of inversions. To assess whether or not this was sufficient, we looked at the ranges of
predictions for these 20 inversions. In all cases the inversion with the optimized parameter set
fell with the range of the 20 uncertainty inversions. However, these included some parameter
sets with poor fit to the reference spectrogram. Thus, we used the RMSE fit to the reference
spectrogram to identify the best 5 inversions and then we took for each time period the highest
and lowest inverted bedload transport rate. These optimized transport rate fell within the range
defined by the maxima and minima for GS1 in 2020 and 2021 and GS2 in 2020, and thus they
were deemed as behavioral and providing a basic uncertainty estimation. For GS2 in 2021 the
optimized transport rate was higher than the maximum and so for that time series we
considered the best 6 inversions.
4. Data details
The datasets are composed by:
Raw hourly 3-components collected seismic data in SAC format. Folder structure is
composed by year, Julian day, hourly data in UTC. File names are denoted as station
ID (OTM1 for GS1 and OTM2 for GS2), Julian day, hour, minute, second, channel and
file name format.
Text files of discharge, suspended sediment loads and bedload rates used in this study.
Data are organized in the following way:
Column 1 = Time in Julian day
Column 2 = Suspended sediment flux in GS1 [kg/s]
5
Column 3 = Bedload flux in GS1 [kg/s]
Column 4 = Suspended sediment flux in GS2 [kg/s]
Column 5 = Bedload flux in GS2 [kg/s]
5. References
Aki, K. & Richards, P. G. (2002). Quantitative Seismology. University Science Book.
Bakker, M., Gimbert, F., Geay, T., Misset, C., Zanker, S. & Recking, A. (2020). Field
Application and Validation of a Seismic Bedload Transport Model. Journal of
Geophysical Research: Earth Surface, 125(5). [Link]
Beaton, A.E. & Tukey, J.W. (1974). The fitting of power series, meaning polynomials, illustrated
on band-spectroscopic data. Technometrics, 16, 147-85.
[Link]
Burtin, A., Hovius, N. & Turowski, J. M. (2016). Seismic monitoring of torrential and fluvial
processes. Earth Surface Dynamics, 285–307. [Link]
2016
Dietze, M. (2018). The R package “eseis” – a software toolbox for environmental seismology.
Earth Surface Dynamics, 6(3), 669–686. [Link]
Dietze, M., Gimbert, F., Turowski, J. M., Stark, K. A., Cadol, D. & Laronne, J. B. (2019). The
seismic view on sediment laden ephemeral flows—Modelling of ground motion data for
fluid and bedload dynamics in the Arroyo de los Pinos [Computer software manual].
Retrieved from [Link]
[Link]/pages/publications/other/Dietze_et_al_2019b.pdf (Paper to SEDHYD
conference)
Gimbert, F., Tsai, V. C. & Lamb, M. P. (2014). A physical model for seismic noise generation
by turbulent flow in rivers. Journal of Geophysical Research: Earth Surface, 119(10),
2209–2238. [Link]
Mancini, D., Dietze, M., Müller, T., Jenkin, M., Miesen, F., Roncoroni, M., Nicholas, A. & Lane,
S.N. (2023). Filtering of the signal of sediment export from a glacier by its proglacial
forefield. Geophysical Research Letters.
McKay, M.D. (1992). Latin hypercube sampling as a tool in uncertainty analysis of computer
models. Proceedings of the 24th conference on winter simulation, 557-564.
[Link] 10.1145/167293.167637
Müller, T. & Miesen, F. (2022). Stream discharge, stage, electrical conductivity & temperature
dataset from Otemma glacier forefield, Switzerland (from July 2019 to October 2021)
[Data set]. Zenodo. [Link]
Sanchez-Sesma, F.J., Weaver. R.L., Kawase, H., Matsushima, S., Luzon, F. & Campillo, M.
(2011). Energy partition among elastic waves for dynamic surface loads in a semi-
infinite solid. Bulletin of the Seismological Society of America, 101, 1704-1709.
[Link]
Taylor. J.R. (1997). An Introduction to Error Analysis: The Study of Uncertainties in Physical
Measurements. University Science Books: Sausalito.
Tsai, V. C., Minchew, B., Lamb, M. P. & Ampuero, J.-P. (2012). A physical model for seismic
noise generation from sediment transport in rivers. Geophysical Research Letters,
39(2). [Link]
