This work assesses the estimation of surface volumetric soil
moisture (VSM) using the global navigation satellite system interferometric reflectometry (GNSS-IR) technique. Year-round observations were acquired from
a grassland site in southwestern France using an antenna consecutively placed
at two contrasting heights above the ground surface (3.3 and 29.4 m). The VSM
retrievals are compared with two independent reference datasets: in situ
observations of soil moisture, and numerical simulations of soil moisture and
vegetation biomass from the ISBA (Interactions between Soil, Biosphere and
Atmosphere) land surface model. Scaled VSM estimates can be retrieved
throughout the year removing vegetation effects by the separation of growth
and senescence periods and by the filtering of the GNSS-IR observations that
are most affected by vegetation. Antenna height has no significant impact on
the quality of VSM estimates. Comparisons between the VSM GNSS-IR retrievals
and the in situ VSM observations at a depth of 5 cm show good agreement
(
Soil moisture is a key component in the hydrological cycle and in the soil–plant–atmosphere continuum. It is also important for irrigation management and flood prediction (Rodriguez-Iturbe and Porporato, 2007). However, in situ observations of soil moisture are very sparse and with small sampling volumes. On the other hand, L-band satellite-derived products, for example, from the Soil Moisture Active Passive (SMAP) mission or the Soil Moisture and Ocean Salinity (SMOS) mission, have a coarse resolution of tens of kilometers (Chan et al., 2016; Kerr et al., 2001). These products consist of surface volumetric soil moisture (VSM) and concern the top soil layer (from the soil surface to a depth of 1 to 5 cm). There is a need to monitor VSM at the local scale in order to validate model simulations, and satellite-derived products. The International Soil Moisture Network (Dorigo et al., 2013) has been collecting such in situ observations. The Committee on Earth Observation Satellites (CEOS) Land Product Validation group has recommended expanding the soil moisture networks (Morisette et al., 2006). In particular, the development of new automatic monitoring techniques to measure VSM is needed.
The global navigation satellite system interferometric reflectometry (GNSS-IR) technique has demonstrated strong potential to monitor VSM using ground-based receivers (Chew et al., 2014). GNSS antennas measure the signal directly emitted by the GNSS satellites, together with the signal reflected by the surface surrounding the antenna. The GNSS-IR technique allows relating the reflected signal to the characteristics of the reflecting surface and to retrieve geophysical variables. Over land, variables such as soil moisture, snow depth and vegetation parameters can be observed using this technique (Larson et al., 2008; Small et al., 2010; Larson and Nievinski, 2013; Wan et al., 2015; Larson, 2016; Roussel et al., 2016; Zhang et al., 2017). GNSS satellites emit active L-band microwave signals (between 1.2 and 1.6 GHz). The L-band signal is less affected by vegetation effects than shorter wavelengths, which is an asset for retrieving surface soil moisture (Kerr et al., 2001). The GNSS-IR footprint can cover up to thousands of square meters, depending on the antenna height and on the satellite elevation angle (Larson et al., 2010; Vey et al., 2016).
In addition to an antenna specially designed to receive the reflected GNSS
signal from the land surface (Zavorotny et al., 2014), classical
geodetic-quality GNSS antennas can be used to estimate VSM (Larson et al.,
2008). Such antennas have an antenna gain pattern optimized for right hand
circular polarization (RHCP) and minimized for left hand circular
polarization (LHCP). A GNSS network called Plate Boundary Observatory (PBO)
H
The present-day Block II R-M (Replenishment Modernized) and Block II F (Follow-on) GPS satellites now transmit a L2C (1227.60 Hz) civilian signal. Power and precision of the L2C signal are higher than for the L1 C/A signal (1575.42 Hz) transmitted by all GPS satellites. Several previous studies, such as Larson et al. (2008, 2010), Chew et al. (2014), Chew et al. (2016) and Small et al. (2016) exclusively analyzed the SNR data from the GPS L2C signal to retrieve soil moisture. The Block II F satellites also transmit the latest L5 signal (1176.45 Hz) as well, which features even higher power, greater bandwidth and an advanced signal design. There are now seven Block II R-M satellites (pseudo-random noise, PRN, numbers 5, 7, 12, 15, 17, 29 and 31, identifying each satellite) and 12 Block II F satellites (PRN 1, 3, 6, 8, 9, 10, 24, 25, 26, 27, 30 and 32).
Due to the motion of the satellites, the direct and reflected signals cause
an interference pattern in SNR data. The SNR oscillations depend on known
attributes such as the satellite elevation angle, signal wavelength and
antenna height. The SNR amplitude and phase can be solved by using the least
squares estimation (LSE) method (Larson et al., 2008; Chew et al., 2016).
Larson et al. (2008, 2010) empirically showed that phase correlates with
near-surface soil moisture, with values of the coefficient of determination
(
Zhang et al. (2017) used the GNSS-IR technique for a wheat field throughout
the growth and senescence period in 2015. The L1 C/A signal was acquired over
a wheat field during a period of about 7 months using a Leica GR25 receiver
and a Leica AR10 antenna at a constant height of 2.5 m above the soil
surface. They showed that VSM could not be retrieved when the vegetation
canopy is too dense, i.e., plant height and simulated dry aboveground biomass
larger than one wavelength (
The objectives of this study are to (1) investigate VSM estimation over a
meadow, contrasting conditions of plant phenology (growth, senescence,
after and before cutting), (2) compare the use of L2C and L5 signals
(3) assess the impact of a major change in the height of the receiving
antenna above the soil surface, in relation to the SNR sampling interval.
Investigating the impact of the sampling interval on VSM retrievals is needed
due to the fact that small sampling intervals (e.g., 1 s) generate a large
amount of data (
A key difference between this study and Zhang et al. (2017) is related to the type of observed vegetation canopy. The meadow considered in this study and the wheat field considered by Zhang et al. (2017) present contrasting characteristics. The meadow is cut once a year and consists of a multi-species permanent grassland incorporating a litter composed of dead leaves. On the other hand, the wheat crop in Zhang et al. (2017) consisted of a single plant species with no litter.
Past microwave remote sensing studies (e.g., Saleh et al., 2007) have shown that permanent grasslands behave differently from crops. Because permanent grasslands incorporate a litter composed of dead leaves, they can intercept precipitation considerably more than annual crops. The short growing cycle of annual crops does not allow for the accumulation of large amounts of litter material. This property of permanent grasslands can have a major effect on the microwave signal and can perturb the retrieval of VSM, even at GPS L-band (Saleh et al., 2007). Also, the structure of grass canopies differs from the structure of crops such as wheat and this has an impact on the attenuation of the microwave signal by vegetation (Wigneron et al., 2002).
GPS SNR data from both L2C and L5 signals are obtained using a geodetic-quality GNSS antenna. SNR analysis using the GNSS-IR technique is used to retrieve VSM over a field covered with grass using the normalization method based on the newly established scaled wetness index proposed by Zhang et al. (2017). Another point to underline is the impact of the antenna height (here two levels: 3.3 and 29.4 m above the soil surface) on the VSM retrieval. Moreover, the VSM retrievals from two kinds of GPS signal wavelengths (24.45 and 25.40 cm for L2C and L5, respectively) are compared with field observations. We analyze the vegetation effects on VSM retrieval accuracy. Another important topic addressed is the influence of the sampling interval on the VSM estimates. As the SNR period changes depending on the antenna height, satellite elevation angle, elevation angle change rate and GNSS signal wavelength, the sampling interval has to be adjusted accordingly in order to maintain the VSM retrieval accuracy.
Data are described in Sect. 2 and methods in Sect. 3. The obtained soil moisture retrievals are presented in Sect. 4 and compared with independent VSM estimates. Results are discussed in Sect. 5, and the main conclusions are summarized together with prospects for further research in Sect. 6.
The study site is located at the premises of Meteo-France in Toulouse,
France, over an experimental field covered with grass
(43
The grass height did not exceed 0.3 m during the experiment time period.
This is much lower than maximum height of the wheat crop (
Experimental site of Meteopole-Flux. The specular reflection points and first Fresnel zone (FFZ) areas from the selected satellite tracks are shown in orange for a 29.4 m GNSS antenna (“H” red dot). The specular reflection points and FFZ areas for a 3.3 m GNSS antenna (“L” red dot) are shown in blue. The red star indicates the location of in situ soil moisture observations. Background geographic information is from Google Earth.
Mean in situ VSM observations at 5 and 1 cm depths were performed using precise Delta-T ML2x ThetaProbes and low-cost Decagon EC-5 VSM sensors, respectively. Three ThetaProbes measured VSM at a depth of 5 cm and were located within a few meters of each other (red star in Fig. 1). The mean value was derived from these probes to represent the in situ VSM observations at 5 cm. Only one EC-5 sensor was used to measure VSM at 1 cm. Precipitation measurements were made in the experimental field by one rain gauge close to the in situ soil moisture sensors. A small fraction of the precipitation time series was missing. Missing data were replaced by the precipitation data obtained from the SAFRAN atmospheric analysis (Durand et al., 1993, 1999). Additionally, scaled VSM observations at a depth of 1 cm and scaled VSM simulations for the top 1 cm thick soil layer were used as independent benchmarks for validation.
VSM simulations for the top 1 cm were produced using the ISBA (Interactions
between Soil, Biosphere, and Atmosphere) land surface model within the SURFEX
(version 8.0) modeling platform (Masson et al., 2013). In addition to VSM,
simulations included the soil iced water content and the vegetation
aboveground dry biomass. The ISBA model used the atmospheric forcing data
produced by the SAFRAN atmospheric analysis of Météo-France. The
model version used in this study was designed for generic country-scale
simulations over France at a spatial resolution of 8 km
In this study, GNSS SNR data were acquired using a Leica GR25
multi-constellation and multi-band geodetic receiver equipped with an AR10
antenna for more than 1 year. Two measurement configurations were
explored (Fig. 1). First, from 1 August 2015 to 5 June 2016, the antenna was
placed at the top of a building close to the studied grassland, at a height
of 29.4 m above the soil surface (43
In this study, both L2C and L5 SNR data from the GPS Block II R-M and Block II F satellites were used. The ascending and descending parts of the same satellite were processed separately and were considered as independent satellite tracks (Roussel et al., 2015, 2016).
Timeline of experiment.
Characteristics of the selected satellite tracks from the GNSS
antenna at a 29.4 m height and at a 3.3 m height (north is 0
Soil moisture scores for four time segments from the comparison
between scaled VSM validation data (in situ VSM observations at 1
and 5 cm and ISBA VSM simulations at 1 cm) and scaled GNSS VSM retrievals (both
L2C and L5). Scores in m
The valid SNR segment for each ascending or descending satellite track was
limited based on the available satellite elevation angle range (90
Measurements from the antenna at a height of 29.4 m were affected by
surrounding obstructions (buildings and impervious areas like car park,
roads, etc.) and by an under-sampling issue at a sampling interval of 10 s
(see Sect. 4.2). In order to cope with these problems, only six satellite
tracks were used to retrieve VSM from L2C SNR data (GPS PRN 03, 07, 08, 17,
25 and 26), and four satellites tracks from L5 SNR data (GPS PRN 03, 08, 25 and
26). Satellite track characteristics and instantaneous FFZ areas are given in
Table 1. The selection of satellite tracks and elevation angles was performed
by comparing VSM retrievals with the in situ VSM observations described in
Sect. 2.1. It must be noted that this limitation only affected measurements
at a height of 29.4 m and was caused by the more complex experimental
constraints in this configuration (e.g., possible parasitic signal reflection
on buildings). For the low antenna configuration (3.3 m), this additional
data sorting was not needed and all available satellite tracks with a
complete elevation angle range (between 7 and 30
The modulation of the SNR by the multipath frequency can be expressed as
(Larson et al., 2008, 2010; Chew et al., 2016):
Due to the good linear relationship between
In this study, the method proposed by Zhang et al. (2017) is used.
Normalizing
Moreover,
SNR amplitude (
Another step is to select relevant satellite tracks under significant
vegetation effects. This is particularly challenging in dense vegetation
conditions. Even in conditions presenting significant vegetation effects,
some satellite tracks can be selected to retrieve VSM. This occurs during
TS3, corresponding to low
Figure 2 presents the VSM estimates derived from both the L2C and L5 SNR data using the normalized SNR phase method (see Sect. 3.1) and the vegetation correction method (see Sect. 3.2). Results are shown for the whole experiment period from 1 August 2015 to 6 October 2016, and for all the experimental configurations of antenna height, sampling interval and grass cutting (time segments).
The first grass cutting event occurs during TS1 but has no effect on
The scaled wetness indexes (
Scatter plot of daily mean in situ VSM observations (
Figure 2 shows that the GNSS VSM retrievals are more sensitive to light rainfall events than in situ VSM observations at 5 cm depth. Such events occur during the summer and autumn of 2016. It can be observed that while GNSS VSM estimates peak at the same time as light rain, the diffusion of water in the soil does not reach the probes at 5 cm depth. This is why the GNSS VSM tends to be larger than in situ VSM. This difference reduces the correlation and increases the errors and can be attributed to a GNSS-IR sensing depth less than 5 cm (Chew et al., 2014; Shellito et al., 2016), in relation to vegetation litter effects (see Sect. 5.3).
In the following subsections, more detailed comparisons are presented for antenna heights of 29.4 and 3.3 m.
Median of the daily VSM retrievals (
In most previous studies, VSM was retrieved from GNSS antennas at about 2 or
3 m above the soil surface. Increasing the antenna height can significantly
expand the size of the observed areas. In this study, the impact of using a
29.4 m antenna on VSM retrievals is assessed using TS1 and TS2 data. The
whole observation area for each track is about 900 m
Scatter plots of daily mean in situ VSM observations at a depth of
5 cm vs. GNSS VSM retrievals (
Soil moisture scores between daily mean in situ VSM observations at
a depth of 5 cm and GNSS VSM retrievals (either L2C or L5) during TS1 (SNR
data from the 29.4 m antenna with 10 s sampling interval from 1 August 2015
to 18 March 2016). MAE, RMSE, SDD and
Figure 5 and Table 3 show that VSM retrievals using L5 SNR data are very
close to those derived from L2C SNR data. The retrieval accuracies from L2C
and L5 SNR data are similar (Table 3), showing that both L2C and L5 SNR data
can be used to retrieve VSM. In Table 2, L2C and L5 SNR data are combined.
Results for TS1 in Table 2 show slightly improved scores with respect to
those in Table 3. This can be explained by the larger number of available
satellite tracks per day. It is interesting to note that results very similar
to those presented in Fig. 5 can be obtained by multiplying the
Overall, the scores obtained during TS1, at a height of 29.4 m and a
sampling interval of 10 s are comparable to those obtained in other time
segments, including TS2 with a sampling interval of 1 s. The scores
(Table 2) in TS2 are similar to the scores in TS1. This does not mean that
there is no effect from the sampling interval because vegetation conditions
are different in TS1 and TS2. TS2 corresponds to a vegetation growing period.
Vegetation growth impacts the reflecting surface and has an impact on the SNR
data as illustrated by the fast decrease of
Substantial vegetation effects are observed during TS3, at the end of the
growing season of 2016. This is evidenced by
Median of the daily VSM retrievals (red lines) with
In order to remove vegetation effects, the SNR data before and after cutting are considered as distinct datasets (see Sect. 3.1 and 3.2). SNR data are used, time segment by time segment, to obtain soil wetness index and then VSM estimates. The observed soil moisture minimum and maximum values are derived for each time segment. For L2C (L5), 10 (6) satellite tracks out of 36 (21) are selected for use during TS3. Figure 6a shows the VSM retrievals for each time segment TS3 and TS4 for L2C SNR data after removing vegetation effects by applying the Zhang et al. (2017) method. The corresponding scores are listed in Table 4. Similar results are obtained for L5 and both L2C and L5 SNR data (Table 4). Results obtained by applying the Zhang et al. (2017) method to the merged time segments (TS3 and TS4) for L2C SNR data are also shown in Fig. 6 and in Table 4. In this case, SNR-derived VSM are too dry before the cutting and too wet after the cutting (Fig. 6b).
Soil moisture scores between daily mean in situ VSM observations at
a depth of 5 cm and GNSS VSM retrievals (either L2C or L5 or both) during
TS3 and TS4 (SNR data from the 3.3 m antenna with 1 s sampling interval
from 8 June to 6 October 2016). The Zhang et al. (2017) method is used for
separated time segments, and also for merged time segments. MAE, RMSE, SDD and
While VSM could not be retrieved by Zhang et al. (2017) after wheat
tillering, i.e., for plant heights larger than 0.2 m, we could retrieve scaled
VSM values throughout time segments of the grass growing and senescence
phases. However, retrieving VSM values in m
Scatter plots of daily mean in situ VSM observations (
Section 4.3 showed that the VSM retrieval from SNR data during TS3 is of lower quality than during TS4, i.e., after cutting the vegetation. Not all satellite tracks can be used (Table 1) and skill scores are systematically worse (Table 2). Moreover, Fig. 6 shows that a specific calibration (see Sect. 3.2) of the retrieval method is needed for TS3. Because the retrieval method is based on the minimum phase which is related to the vegetation height and density, the lack of a priori information about this factor is likely to trigger marked discrepancies.
Based on Eq. (1), SNR amplitude
Figure 7 illustrates the improvement associated with the vegetation correction.
The systematic bias caused by the mismatch in
As a consequence, monitoring VSM using a GNSS network could be difficult when
vegetation effects are noticeable. However, we show that one may use the
information from
L2C SNR VSM retrieval time series using GPS PRN 10 ascending tracks
with different sampling intervals:
In this study, we used independent VSM in situ observations to harmonize the VSM time series across TS3 and TS4. Since in situ observations are not extensively available, this technique is not readily applicable at other sites. In practice, one could possibly use a data assimilation framework able to integrate the VSM retrievals into model VSM simulations such as those produced by the ISBA land surface model (Albergel et al., 2017). In such land data assimilation systems (LDASs), a complex seasonal rescaling of VSM observations is needed (Reichle and Koster, 2004; Draper and Reichle, 2015), especially when the observations are not properly decontaminated from vegetation effects (Stoffelen et al., 2017). Our results show that using this rescaling technique would be feasible since the ISBA simulations of VSM correlate well with the retrieved VSM (Fig. 8). The main reason for this result is that ISBA is forced by the SAFRAN atmospheric analysis, incorporating a large number of in situ rain gauge observations (Sect. 2.1). This is another way of using ancillary in situ observations.
The effects of vegetation on GNSS SNR data are threefold: from plant height, aboveground biomass and litter. At the end of the growing season, plant height and aboveground biomass values can be much larger for annual crops than for grass. On the other hand, while litter is usually missing during the growing phase of annual crops, litter is characteristic of grasslands (Quested and Eriksson, 2016).
Over our grassland site, the measured grass height at the end of the growing
period is 30 cm on 22 June 2016. The grass height is then only slightly
larger than one GNSS wavelength (
Zhang et al. (2017) showed that over a wheat field the vegetation gradually
replaces the soil as the dominant reflecting surface when plant height
becomes comparable to, or larger than, one wavelength, even at relatively low
values of the aboveground biomass (an estimate of 0.08 kg m
This study shows that VSM retrieval above these biomass and plant height thresholds are feasible for grass. However, a limited number of suitable tracks, less affected by vegetation, have to be selected using the grass cutting event (see Sect. 3.2). In real practical applications, such tracks are not a priori known and retrieving VSM would be challenging when vegetation effects are significant.
In order to analyze the possible impact of litter on the differences between GNSS VSM and either in situ VSM or ISBA VSM, in situ VSM observations at 5 cm, in situ VSM observations at 1 cm and ISBA VSM simulations at 1 cm are compared with the GNSS VSM retrievals. The GNSS VSM is retrieved applying the Zhang et al. (2017) method to both L2C and L5 SNR data, and the vegetation effects are removed from the retrievals. For ensuring the comparability of these various soil moisture estimates, GNSS retrievals, ISBA 1 cm simulations, in situ 1 cm observations and in situ 5 cm observations are scaled to dimensionless values.
Figure 8 shows a comparison between the four scaled VSM time series during
TS3 and TS4. Soil moisture values tend to increase drastically during
precipitation events. Most of the VSM peaks observed in 1 cm in situ
observations are also found in 5 cm observations, except for 5–7 July and
5 August 2016. On the other hand, GNSS VSM peaks can occur while in situ VSM
observations do not display any response to rain, e.g., on 8–14, 25 and
30 June, 30–31 July, and 29 August 2016. A contrasting result is found
comparing GNSS and ISBA VSM estimates, which peak, more often than not, at
the same time. As a consequence, the GNSS VSM estimates correlate much better
with ISBA VSM (
The scores resulting from the comparison between scaled VSM validation data
and GNSS VSM estimates are separately recorded in Table 2 for each time
segment. The highest correlations are with ISBA simulations at 1 cm, for all
time segments. The scores based on in situ VSM observations at 1 cm are
similar to those based on in situ VSM observations at 5 cm. For TS4, the
correlation with in situ VSM observations at 1 cm is much higher than with
those at 5 cm. The main difference between observations at 1 cm and at
5 cm is that the former respond to rainfall events more rapidly. This is
illustrated by Fig. 8 for events occurring after 9 July 2016 (TS4). The
differences observed between GNSS VSM estimates and in situ VSM observations
at 1 cm can be explained by the interception of light rain by the litter.
Water contained in the litter tends to directly reflect the GNSS signal and
to prevent the GNSS signal from further penetrating into the soil. This
difference is not observed with ISBA simulations because the litter is not
implemented in this version of the ISBA model. The good correspondence
between ISBA and GNSS VSM estimates can be considered as an artifact: ISBA
simulates a VSM peak which does not exist, and the GNSS SNR data are
sensitive to a sudden increase in the litter water content and/or to the rain
intercepted by the litter or by the leaves. Another demonstration of the
impact of the litter effects can be made, removing rainy days from TS4. The
When the antenna height increases, the size of the observing areas is extended. But at the same time the period of the SNR data decreases (Eq. 1), and a smaller sampling interval is needed to ensure the usability of the SNR data for VSM retrieval. On the other hand, because the SNR period from a high antenna is much smaller, it is possible to use smaller elevation angle ranges and shorter observing time periods per track. The number of complete SNR waveforms is much larger than using a low antenna. We investigate the impact of under-sampling for the 3.3 m antenna and for the 29.4 m antenna. It should be noted that in the examples illustrated by Figs. 9, 10, and 11 the SNR frequency is always lower than the Nyquist frequency.
Two examples of L2C SNR data sets (from the GPS PRN 10 ascending
tracks) acquired by the 3.3 m antenna at two contiguous
dates:
First, an example of the impact of the sampling interval for the 3.3 m
antenna is shown in Fig. 9. L2C SNR observations (
Soil moisture scores from the comparison between daily mean in situ VSM observations at a depth of 5 cm and GNSS VSM retrievals during TS4 (after grass cutting, from 9 July to 6 October 2016). The L2C SNR data from GPS PRN 10 ascending tracks were used, which were acquired by the 3.3 m antenna. MAE is the mean absolute error, RMSE is the root mean square error and SDD is the standard deviation of difference.
Two examples of L2C SNR data sets (from the GPS PRN 25 ascending
tracks) acquired by the 29.4 m antenna at two contiguous
dates:
SNR amplitudes are also affected by the sampling interval in TS4. For
29 July 2016, the estimated SNR amplitude is 26 V V
For the 29.4 m antenna, the sensitivity to the sampling interval is more critical. Figure 11 shows the SNR oscillations for the GPS PRN 25 ascending track. The SNR period is only about 38 s. With 10 s sampling interval, three or four samples are available for a complete waveform. This is about the same situation as for the 100 s sampling interval for the 3.3 m antenna. Figure 11a shows that pit and peak information is missing on 18 March 2016 with respect to the 1 s sampling interval data on the next day in Fig. 11b. Nevertheless, Table 2 shows that the 10 s under-sampling had a limited impact on VSM retrievals during TS1 since the best scores are observed during this segment. This paradoxical result can be explained by the prior use of the in situ VSM data to select the satellite tracks and the satellite elevation angles (see Sect. 2.2).
GPS L2C and L5 SNR data were obtained at a grassland site
in southwestern France during a period of 15 months. A dimensionless scaled
wetness index was derived from the SNR observations based on the GNSS-IR
technique, using indiscriminately L2C or L5 signals. Surface soil moisture
was derived from this scaled wetness index. We show that accurately
estimating soil moisture in units of m
The data used in this work are available for research from the corresponding author.
The authors declare that they have no conflict of interest.
The work of Sibo Zhang was supported by the STAE (Sciences et Technologies pour l'Aéronautique et l'Espace) foundation, in the framework of the PRISM (Potentialités de la Réflectométrie GNSS In-Situ et Mobile) project. The authors would also like to thank Eric Moulin and Joel Barrié (CNRM) for their technical support during the field campaign and Anne Belleudy and Diane Tzanos (CNRM) for performing biomass observations. Edited by: Miriam Coenders-Gerrits Reviewed by: two anonymous referees