1 Introduction

The area-specific sediment yield (SSY), expressed in t km−2 year−1, is the amount of sediment that is exported out of a given watershed. The SSY of a large diversity of watersheds has been studied worldwide since the nineteenth century and measured with increasing accuracy due to recent technological improvements (Asselman 2000; Verstraeten and Poesen 2001; Ward 2008; Métadier and Bertrand-Krajewski 2012). These studies have been conducted in a wide range of watershed scales, with different lithologies and climates (Moatar et al. 2006; Dumas 2007; Lefrançois et al. 2007; Picouet et al. 2009; Furuichi et al. 2009; Mano et al. 2009; Marttila and Kløve 2010; Oeurng et al. 2010; Gao and Josefson 2012; Araujo et al. 2012; Fortesa et al. 2021). They often showed a link between SSY and land cover, amount, and intensity of rainfall as well as the availability of fine sediment in the watersheds. For instance, the effects on SSY of heavy summer precipitation have been studied on the Mediterranean rim (Seeger et al. 2004; Nadal-Romero et al. 2008; Tena et al. 2011), the temporal distribution of soil erosion examined in high relief energy areas (Jansson 1996; Sadeghi et al. 2008; Navratil et al. 2011), and the variability of sediment transport analysed at spatial scale (Steegen et al. 1998; Lenzi and Marchi 2000; Meybeck et al. 2003; Vanmaercke et al. 2012b; Gericke and Venohr 2012). Impacts of traditional soil conservation practices were observed (Estrany et al. 2009), in terms of reducing soil erosion by an order of magnitude. Also widely analysed were the effect of wildfires on the increase of soil erosion (Lane et al. 2006; Warrick et al. 2015). Many sedimentary studies also showed the ecological impacts of sediment accumulation on stream habitats (Parkinson et al. 1999; Tramblay et al. 2010; Collins et al. 2011); some authors associated it to a decrease of biodiversity in silt-clogged river beds (Vaessen et al. 2021), while others to human impacts on sediment yield resulting in the consolidation of riverbanks or an increase of flooding risks (Dumas 2007; Vanmaercke et al. 2015).

In Europe, the analysis performed by Vanmaercke et al. (2011, 2012a) combining data from 1794 watersheds in Europe (area: 0.01–1,360,000 km2) showed that the average erosion rate observed in Europe was 341 t km−2 year−1, for a median rate of 92 t km−2 year−1, with six orders of magnitude separating the minimum values (0.3 t km−2 year−1) from the maximum values (30,000 t km−2 year−1). These differences were attributed to a combination of factors, such as differences in climate, topography, lithology, and land use (Vanmaercke et al. 2011).

This present study is intended to compensate for the lack of published quantified data on suspended sediment transport in rivers of the southern part of Belgium (Wallonia). Because of the influx of fine sediments into the waterways surrounding loamy soils, public river managers needed to know, for a large number of stations, the quantity of transported sediments, and their geographical origin, identified using reference stations in other regions with different lithology, land cover, and physical characteristics.

Indeed, suspended sediment concentrations (SSC) and SSY estimations have mainly been carried out, in Belgium, on watersheds of medium size, from agronomic researches at plot scale, up to watershed synthetic analyses (Steegen et al. 1998; Verstraeten and Poesen 2001; Pineux 2018). Vanmaercke et al. (2011) have highlighted the lack of sediment transport studies over long periods in large watersheds in Belgium (often < 100 km2 and mainly located in the Scheldt basin in loamy environment). For instance, the Dyle watershed showed SSY of 30 t km−2 year−1 in 1959–1960 at Leuven (742 km2), 70 t km−2 year−1 in 1985–1986 at Bertem (730 km2), and 210 t km−2 year−1 in 1998–2000 downstream of Leuven (820 km2) (Boardman and Poesen 2006). Other measurement campaigns were also carried out since the end of the nineteenth century in the Meuse and Scheldt watersheds (Spring and Prost 1884; Close-Lecocq et al. 1982; Lemin 1984; Lemin et al. 1987; Ward 2008) or some of their tributaries (Sine and Agneessens 1978; Petit 1985; Lemin et al. 1987; Lamalle et al. 1989; Perpinien 1998; Parkinson et al. 1999; Hombrouckx 2002; Monseur 2005). However, changes in land cover and agricultural practices are known to have great effect upon soil erosion, and therefore, updated SSY values encompassing rivers in different geographical regions were of interest for managers. A fairly rapid and inexpensive methodology was developed during two successive research projects (between 2006 and 2011) on watersheds ranging in size from 7 to 3600 km2 to estimate SSY in rivers located in the southern part of Belgium.

2 Materials and methods

2.1 Study location: geographical and geological features

The 72 study sites correspond to gauging stations which are located in Wallonia, the southern region of Belgium. A total of 65 of them belong to the Meuse basin, 6 to the Scheldt basin, and 1 to the Moselle basin (see Table 1; Fig. 1). This area experiences a warm-temperate and oceanic climate without a dry season (Cfb code in the updated Köppen-Geiger classification) and encounters annual rainfall ranging from 725 mm in westernmost Wallonia to 1400 mm in the easternmost part of the region, mainly in relation to the elevation gradient (Erpicum et al. 2018). The selection of sites where SSY have been quantified through SSC measurements has been guided by the presence of flow gauging stations. After preliminary results were acquired on major rivers, the sampling network has been extended to watersheds of lesser area and/or in other geographical regions while the set of installed gauging stations was growing in Wallonia under the aegis of the public service. The 72 stations are distributed as shown in Fig. 1 in the different geographical regions. The regional classification of stations depends on their location and on the sedimentary setting and the hydrological dynamics of the upstream area. The regional affiliation of each station (Van Campenhout et al. 2020) is shown in Table 2 as well as their geological substratum and loess availability. The land cover proportion in each watershed is derived from Copernicus Land Service and the Corine Land Cover maps (Panagos et al. 2015; CORINE Land Cover 2018). These 100-m resolution maps have been reclassified to give the spatial proportion of forest, grassland, cultivated area, and impervious area at watershed scale with the raw data available for the years 2000, 2006, 2012, and 2018 (Table 9 in Supplementary Material). The average percentage of cultivated area greatly varies with the region where the studied watersheds are located: 19% in Ardenne, 30% in Entre-Vesdre-et-Meuse, 37% in Fagne-Famenne, 28% in Lorraine, 54% in the Condroz, 50% in the Brabant Plateau, 49% in the Haine basin, and 75% in Hesbaye, based on the Corine Land Cover map of 2006. Conversely, the average forested area ranges from 2% in Hesbaye to 55% in Ardenne.

Table 1 Characteristics of the studied watersheds
Full size table
Fig. 1
figure 1

Location of sampling sites and simplified geological map of Wallonia (according to de Bethune (1954) and Dejonghe (2007), modified). Station numbering refers to Table 1. The sub-watershed area of each river studied is shown with its appurtenance to the main basin (Scheldt, Meuse, and Moselle)

Full size image
Table 2 Regional affiliation of each station in terms of lithology and loess availability
Full size table

2.2 Field sampling

2.2.1 Sampling methods

To estimate the concentration of wash load and suspended sediment, a manual sampling methodology was set up to allow a large number of sampling sites during flood periods, with the sampling of a bucket of 5 l of river water from a bridge in the centre of the river. This method was chosen for its speed of implementation (Lemin et al. 1987) and its efficiency in obtaining reproducible results with different operators. Between 1994 and 2002, 250 samples were taken during fieldwork for academic research master theses (Perpinien 1998; Hombrouckx 2002; Monseur 2005). The representativeness of the results obtained with this sampling method has been validated across the cross section of several rivers in Wallonia (Pironet 1995; Monseur 2005). After this first phase of samplings, around 1300 samples were taken between 2003 and 2010. Then 450 samples were collected in the 2010–2016 intervals. The average number of samples per study site was 40, with a coefficient of variation of 1.76. Over the whole dataset, 25% were taken in rising limbs, 24% in recession limbs, with different and independent events, and the other 51% were sampled during almost constant flow rates between floods. Recent reproducibility tests (2011–2014) were performed in the watersheds of the Gette, Senne, and Dyle rivers in order to validate the method of systematic subsurface sampling at the point where the flow is the fastest, according to the morphology of the stream bed (Van Campenhout et al. 2013).

2.2.2 Hydrologic series availability

Computing median sediment yield over years needs discrete or continuous water concentration sampling and the recording of discharge data at the same gauging station. With the aim of comparing sediment transport from different geographical regions, and due to the variability of SSY, computations of median SSY at gauging stations have to be compared over the same period of time in order to have consistent results. Discharge data from the dataset do not cover the same period at each location because their installation date varies. Figure 2 gives information about the mean number of days per year where discharge is above bankfull discharge (Qb), which has been observed in the field or computed from partial series with the methodology of Van Campenhout et al. (2020). Figure 2 also shows the number of stations with at least 90% of hourly discharge data available for a given year. Median SSY will be computed from annual SSY interpolated data in the 1996 to 2018 intervals. This time span maximizes the availability of discharge data over most of the monitored stations, and is consistent with the sampling period.

Fig. 2
figure 2

Average number of days above bankfull discharge (Qb) for studied gauging stations, availability, and representativeness of hydrologic data

Full size image

2.3 Laboratory analysis of the suspended sediment particles

The concentration of suspended solids is measured by vacuum-filtering the samples with 110-mm diameter Whatman GF/C glass microfibers filters — mesh = 1.2 µm (Smith 2007) with 98% retention efficiency (Zimmermann et al. 2012). These filters allow working with the huge concentrations that can be attained in loamy rivers (up to 20 g l−1) and are compatible with Loss On Fire procedure. The methodology is based on the NF EN 872–2005 standard (Marttila and Kløve 2010). The accuracy of the measurements depends on the mass of the sediment-filled filter. Tests performed on precisely known mass and volume samples indicate an error of the order of 10% for samples < 100 mg l−1, of the order of 5% for samples of 500 mg l−1, and the order of < 2% for samples of 2000 mg l−1. Concentration values ≤ 10 mg l−1 were rejected due to the uncertainty, representing ~ 8% of samples.

2.4 Sediment yield computations

2.4.1 Rating curves and logarithmic corrections

Water sampling strategies and sediment load calculation are crucial to provide accuracy and reliability in results. Interannual variability of sediment load complicates the mid- and long-term estimations of the sediment yield (Syvitski et al. 2000). The sampling strategy used in our study was set up as a first-line approach to obtain sediment transport data for gauging stations that had never been the subject of suspended sediment studies in Wallonia. While many fluvial systems show a non-linearity behaviour in the relationship between discharge and sediment concentration and high uncertainty related with non-continuous monitoring (Webb et al. 1997; Araujo et al. 2012), sampling campaigns were first made to obtain flood water samples related to real-time water level alerts. These flood water samplings were augmented by other samples taken during recessions and low water periods. The sampling regimen was, at most, one sample per station per event in order to prevent intra-event correlation. The campaigns lasted for at least 5 years in order to cover as thoroughly as possible the observed discharges, from low water to above bankfull discharge.

The instantaneous concentration of suspended solids (Cs) in a river with well homogenised flow is commonly related to the instantaneous flow rate (Q) by a power function (Eq. 1), better known as the sediment rating curve (Campbell and Bauder 1940), where a and b are two empirical constants (Phillips et al. 1999; Meybeck et al. 2003; Li et al. 2005; Rovira and Batalla 2006; Doomen et al. 2008; Delmas et al. 2011).

$${C}_{s}=a{Q}^{b}$$
(1)

Cs is generally expressed in mg l−1 and Q in m3 s−1. Parameters a and b are computed by least squares regression in logarithmic space formed by log Cs/log Q from measured concentration and discharge values.

A key factor of the sediment rating curve is its statistical significance. The coefficient of determination (R2) of the curve, the standardized root-mean-squared error (RMSE), referred to as RSR (RMSE-observations standard deviation ratio, Eq. 2), the Nash–Sutcliffe efficiency coefficient (NSE, Eq. 3), and the percent of bias (PBIAS, Eq. 4) have been computed following the equations presented by Jung et al. (2020) based on Moriasi et al. (2007) and shown in Table 3.

Table 3 Validation criteria of the sediment rating curves,
Full size table
$$\text{RSR } = \frac{\text{RMSE}}{{\sigma }_{obs}}=[\frac{\sqrt{\sum\limits_{i=1}^{n}{({C}_{i}^{obs}-{C}_{i}^{est})}^{2}}}{\sqrt{\sum\limits_{i=1}^{n}{({C}_{i}^{obs}-{C}_{i}^{mean})}^{2}}}]$$
(2)
$${\text{NSE}}=1-[\frac{\sum\limits_{i=1}^{n}{({C}_{i}^{obs}-{C}_{i}^{est})}^{2}}{\sum\limits_{i=1}^{n}{({C}_{i}^{obs}-{C}_{i}^{mean})}^{2}}]$$
(3)
$${\text{PBIAS}}=[\frac{\sum\limits_{i=1}^{n}({C}_{i}^{obs}-{C}_{i}^{est})*100}{\sum\limits_{i=1}^{n}({C}_{i}^{obs})}]$$
(4)

Equations 2, 3, and 4 give the RSR, NSE, and PBIAS calculations, respectively, where RMSE is the root-mean-squared error, \({\sigma }_{obs}\) is the standard deviation of the observed concentrations, \({C}_{i}^{obs}\) is the observed concentration, \({C}_{i}^{est}\) is the estimated concentration for the same index i, and \({C}_{i}^{mean}\) is the mean SSC observed concentration from n observations. The values of the NSE can range from − ∞ to 1 (optimal value), where value between 0 and 1 are acceptable, and those values smaller than 0 are not considered as usable. The PBIAS measures the tendency of the estimated concentrations to be higher or lower than the observed data (Jung et al. 2020). The value of RSR varies from the optimal value of 0, which indicates zero RMSE or residual variation and therefore perfect model simulation, to a large positive value. The lower the RSR, the lower the RMSE, the better the model simulation performance (Moriasi et al. 2007). The validation criteria are shown in Table 3, with modifications compared to Jung et al. (2020) because they use an additional constant term p to Eq. 1 in order to take into account the non-linearity of the suspended sediment rating-curve. This would have needed a greater number of samples to be computed.

2.4.2 Spatial and temporal sediment yield calculation

At watershed scale, the specific annual sediment yield was computed as the sum, for each time interval, of the product of the instantaneous flow and the concentration of suspended solids estimated via Eq. 1 and discrete samples (Syvitski et al. 2000; Cerdan et al. 2012). This method of estimation is based on the hypothesis of a unique relationship between Q and Cs and on the assumption that an instantaneous Cs depends only on Q at any given time without hysteresis phenomena (Ritchie 2007; Cerdan et al. 2012). Not enough samples were collected in this study to differentiate flow conditions and obtain flood- and recession-sediment curves at each gauging station The method proposed by Ferguson (1986, 1987) has been used to correct the bias due to the logarithmic de-transformation on both axes, which is only effective when the residuals of Cs follow a log-normal distribution and Cs is a power function of Q (Asselman 2000). A corrective factor is applied to the total suspended and wash load estimations (Phillips et al. 1999), as shown by Eq. 5 to give the corrected daily mass of suspended and wash loads (fdaily corr.) from Qh, the hourly discharge; Ch the estimate suspended sediment concentration, and with s, the standard error of the estimation of the least squares regression in log10 units.

$${f}_{daily\text{corr.}}=\sum_{\text{h}}^{24}{3600}{Q}_{\text{h}}{C}_{\text{h}}\mathrm{exp}(2.651{s}^{2})$$
(5)

2.5 Physical characteristics of the studied watersheds

The physical data of the watersheds have influence on soil erosion and sediment transport efficiency (Syvitski et al. 2000). These parameters have been extracted from the global 1-arcsecond (30-m) Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (DEM) downloaded from the United States Geological Survey’s EarthExplorer site (http://earthexplorer.usgs.gov). However, the mean slope of the watersheds may not be ideal to describe the runoff concentration time of the watershed. The Roche’s slope index Ip, also called the index of runoff susceptibility and based on the compactness coefficient of Gravelius (1914), was computed for each watershed (Roche 1963). Additionally, the hypsometric curve (Davis 1899) and the hypsometric integral were computed for each (Table 8 in Supplementary Materials) watershed (Strahler 1952; Demoulin 2012).

2.6 Multi-criteria analysis on physical watershed variables

Correlation matrices will be used in the first approach in order to isolate the variables with the greater influence on the integrated value of sediment transport at the outlet. The physical variables that will be used are the watershed area, the mean elevation and slope, the Roche’s slope index, and the hypsometric value, whilst the variables related to the land use will be the relative proportion of forests, grassland, and cultivated areas. The Corine Land Cover map that has been used is the reference year 2006 because a majority of samples were taken around 2006 (CORINE Land Cover 2018). Land cover maps of 2000, 2012, and 2018 will also be taken into account in the correlation analysis. In the second approach, weighted coefficients of these variables will be computed as well as the coefficient of multiple determination.

3 Results

3.1 Sediment rating curves analysis

The results of this study are presented in Table 4, which gives all the parameters that were computed from the water samplings and the computed SSY. The parameters ‘a’ (Fig. 3) and ‘b’ (Fig. 4) of the sediment rating curves were plotted in relationship with the watershed area and the regional affiliation of the stations. Even if lower ‘a’ values are observed in the Ardenne rivers and higher ‘a’ values in the Hesbaye rivers, no other obvious regionalization effect appears. The parameter ‘b’ does not show any relationship with the watershed area among the studied sites. However, when dealing with several stations on the same river, the parameter ‘b’ tends to increase with the drained area, while the parameter ‘a’ decreases with the watershed area.

Table 4 Sedimentary parameters of the stations
Full size table
Table 5 Sampling representativeness
Full size table
Fig. 3
figure 3

Relationship between the parameter ‘a’ of the rating curve equation Cs = a Qb and the watershed area with a regionalized differentiation. The values between brackets stand for the sediment rating curves considered as unsatisfactory

Full size image
Fig. 4
figure 4

Relationship between the parameter ‘b’ of the rating curve equation Cs = a Qb and the watershed area with a regionalized differentiation. The values between brackets stand for the sediment rating curves considered as unsatisfactory

Full size image

Based on the validity criterion (Table 3), the unsatisfactory sediment rating curves are presented in Table 5. In addition to the validity criterion, a small number of stations hugely overpredict suspended sediment concentration based on maximum recorded discharge: #61 Grande Gette River at Sainte-Marie-Geest (~ 59 g l−1) and #42 Bocq River at Yvoir (~ 35 g l−1). All the other stations show peak concentrations below 17 g l−1. For comparison, in a 134-km2 watershed in the loess belt region (#59), automatic sampling far below the water surface during the rupture of a storm dam on 29 June 2011 led to measured concentrations around 25 g l−1.

3.2 Spatial and temporal specific sediment yield variability analysis

Through the computation of sediment rating curves and the hydrological series, the annual sediment transport at the stations was calculated. Due to the significant temporal variability of annual sediment yield, which is directly linked to the variability of rainfall and river flows, the average and the median annual sediment yields during the 1996–2018 period were computed (Table 4) and the median SSY values were mapped (Fig. 5). The overall weighted-area median SSY was 32.7 t km−2 year−1, taking into account the 58 valid stations. The median sediment yield for the period 1996–2018 reveals clear regional variability. Considering the computed data for valid sediment rating curves (therefore without the unsatisfactory stations), the median SSY reaches, on average, 19.2 t km−2 year−1 in Lorraine, 24.9 t km−2 year−1 in Ardenne, 26.9 t km−2 year−1 in the Haine basin, 28.4 t km−2 year−1 in Fagne-Famenne, 41.4 t km−2 year−1 in Hesbaye, 49.2 t km−2 year−1 in Entre-Vesdre-et-Meuse, and 119.0 t km−2 year−1 in the Brabant Plateau. The annual sediment yield is very dependent upon the annual runoff and the intensity of the floods. Regional differences were observed in relationship to the proportion of agricultural areas, the availability of fine sediment, and to a less extent, the slope of the watershed in a runoff concentration time point of view.

Fig. 5
figure 5

Sediment yield median values within studied watersheds (period 1996–2018). Median SSY related to sediment rating curves that are considered as unsatisfactory are shown between brackets

Full size image

The region with the most marked median SSY (Hesbaye and Brabant Plateau) also shows the most marked annual variation. In Hesbaye, the studied watersheds belong to two large watersheds: the Meuse basin and the Scheldt basin. The median SSY of the rivers belonging to the Scheldt basin (Petite Gette and Grande Gette rivers) is much higher than the values observed in the Meuse basin (Geer and Mehaigne watersheds). In the Brabant Plateau region and in the Senne watershed, SSY values tend to be even more important than in the Hesbaye region. The Entre-Vesdre-et-Meuse rivers tend to present large annual variations. The other studied regions show less marked annual and spatial variations. Rivers from the Haine basin (16 to 35 t km−2 year−1) and the Fagne region (18 to 37 t km−2 year−1) show lower median SSY values. These values in Fagne are understandable given the grassland which is the predominant land cover and the low slope of the watersheds. However, in the case of the tributaries of the Haine River, it seems that the type of flow — a predominant base flow while Ardennian rivers often show a more flashy discharge regime — tends to explain the lower SSY value. In Famenne, the interannual variability is lower for the studied stations due to a smaller proportion of agricultural areas, preventing high sediment concentrations after summer storms, compared to the large plots in Hesbaye. The Lesse and the Lhomme rivers are subject to karstic losses at the interface of the Ardenne and the Famenne regions. Hart and Schurger (2005) showed that karstic zones can also behave like sedimentary sources, when floods re-suspend old deposits trapped within these cavities.

The median SSY of the rivers of the Ardenne region show, at first sight, a certain variability. However, the order of magnitude of these values agrees with the measurements made by estimating the sedimentary volume trapped by the Bütgenbach dam on the Warche River (A = 72 km2), in a similar lithological and land use context than the upper Amblève River watershed where samples were taken (#3 Amblève River in Targnon, 803 km2). Rivers from the southern part of the Ardenne region and the Lorraine region show lower SSY, due to the forestry and the grassland land cover respectively.

Due to the very high interannual variability of SSY, we analysed 2 years, i.e., 2002, representing a humid year (Fig. 7 in Supplementary Materials), and 2018, representing a dry year (Fig. 8 in Supplementary Materials). Table 5 gives the average SSY by region for the humid and dry years in terms of days > Qb, taking into account valid sediment rating curves. In the case of the humid year and even if the number of flood events and their intensity play a role, the Lorraine and Ardenne regions present quite low sediment transportation in comparison to the other regions. Entre-Vesdre-et-Meuse and Fagne-Famenne regions present intermediate results (~ 100 t km−2 year−1). The larger SSY are observed in the Haine Basin, and above all in the Brabant Plateau and Hesbaye regions. These watersheds from the loess belt have greater availability of fine sediment to transport. In the case of long period of intense floods, SSY can exceed 300 to 600 t km−2 year−1. The Senne River showed a SSY value of 1032 t km−2 year−1 but is still considered as a valid value in terms of maximum extrapolated concentration (9.3 g l−1), contrary to the station #61, the Grande Gette River. Years with heavy runoff are those that contribute the most to sediment transport but uncertainties in SSY calculations are more important in the case of above-bankfull discharge and especially when extreme floods occur.

The lowest annual discharge of the studied time interval was reached in 2018 in a vast majority of watersheds, although some significant flooding occurred in January and June. Figure 8 shows the average SSY values of 2018 taking into account only valid sediment rating curves and Table 6 presents the average values computed by region. In comparison to the year 2002, the year 2018 and its cumulative drought over a period of 3 years display great differences in terms of regional SSY values. The Haine basin presents the lowest SSY values, while this region was showing one of the highest SSY in a humid year. Ardenne and Lorraine regions present values pretty close to 2002 values. Fagne-Famenne region, for its part, has up to 5 times less SSY. Hesbaye and Brabant Plateau regions show around 10 to 20 times less SSY during a drought period. It should be noted that the Berwinne and Bolland watersheds, in Entre-Vesdre-et-Meuse, show very high SSY values, 131 t km−2 year−1 and 116 t km−2 year−1, respectively. This is due to an intense thunderstorm accompanied by rainfall of 89 mm per day on 1 June 2018 in the town of Battice (headwaters of the Berwinne and the Bolland rivers). The Berwinne River experienced flood discharge of 62 m3 s−1 (recurrence ~ 120 years), while the Bolland River reached a flow rate of 12.4 m3 s−1 (recurrence > 175 year). (Table 7)

Table 6 Average SSY by region for 2002 (humid year) and 2018 (dry year)
Full size table
Table 7 Correlation between median SSY (1996–2018) and physical variables of the watersheds
Full size table

3.3 Effects of physical characteristics of the watersheds on sediment yield

Table 1 gives the proportion of each type of land cover in the watersheds. Moreover, the sediment concentration at bankfull discharge has been used in order to compare rivers. This reference discharge has been selected because it is the most suitable discharge value to compare stations with each other. The recurrence of the bankfull discharge (Q0.625 in partial series according to Van Campenhout et al. (2020) in the same geographical area) is usually in the range of sampled discharges. Comparison with greater floods (Q2, Q5, or Q10) may lead to uncertainty due to the extrapolation of the sediment rating curve. 

As a preamble to the analysis of regional differences in suspended sediment yield, some discrepancies in the relation between bankfull discharge and watershed area appears. The rivers located in the Haine Basin and those from the Hesbaye region clearly show Qb values lower than the other rivers for a given value of watershed area (Petit and Pauquet 1997; Petit et al. 2007). In this case, it is more an influence of the hydrological regime, with a general weakness of the flows, rather than of the morphology of the bed and the size of the bed material (Petit et al. 2007). In the opposite case, with their soft-shale substratum that tends to increase the depth of the bed incision, Fagne and Famenne rivers show a higher bankfull discharge compared to the other rivers. 

Ardennian rivers show less possibility of the accumulation of large alluvial plains, regarding their less important SSC at Qb, unlike Brabant Plateau and Hesbaye rivers, with a large availability of loess. However, the rarity of overflowing floods prevents having a precise estimate of the concentrations of suspended sediment above Qb. Some rivers in the southern part of the Lorraine region show local particularities, such as natural levees on the Ton River or an artificial dam in the Vire River, can induce, in the case of the Vire River, a significantly greater bankfull discharge value and thus, a greater sediment concentration at Qb.

Figure 6 shows the estimated concentration at Qb (see Table 4 for values). The variance of SSC is partially explained by the variance of Qb (R2 = 0.60), given by Eq. 6 (where SSC at Qb is the suspended sediment concentration at bankfull discharge in mg l−1, and Aa% is the proportion of agricultural area in the watershed expressed in percent).

Fig. 6
figure 6

Estimated suspended sediment concentrations at bankfull discharge and Corine Land Cover simplified map for 2006 (100-m resolution). Stations #41 and #42 were not shown because of their unrealistic SSC value at bankfull discharge

Full size image
$$SSC\;at\;Q_b=38.72A_{a\%}^{5.29}$$
(6)

Another important physical parameter is the mean slope of the watersheds: 4.2% in the Haine basin and in Hesbaye, 5.5% in the Brabant Plateau, 7.7% in Fagne-Famenne, 8.0% in Entre-Vesdre-et-Meuse, 8.9% in the Condroz, 9.8% in Lorraine, and 10.3% in Ardenne. It can be observed that slopes in central and south-western Ardenne are less steep than in the north-eastern Ardenne. The watersheds with the greatest average slopes are more likely to be forested. Inversely, the cultivated watersheds are prone to present the least steep slopes. However, local lower slopes in agricultural areas may lead to mudflows and flash-floods that contribute to higher sediment yield at the watershed integration scale (Evrard et al. 2007; Van Campenhout et al. 2015). The Roche’s slope index is given in Table 8 (see Supplementary Materials) and ranges from 0.045 to 0.197. It is a better representation of the overall slope of the watershed. The higher values of Roche’s slope index (> 0.100) are observed in the watershed with an area that is lower than 200 km2.

Correlation matrices have been used to detect the explanatory variables with respect to the median 1996–2018 specific sediment yields (Table 7). Despite our efforts to determine physical explanatory variables, only weak correlations between median SSY and land cover percentages are visible. The percentage of cultivated areas tends to be correlated with median SSY (R = 0.42), while the percentage of forested areas (R = −0.45) is inversely correlated with sediment yields. Physical variables that are linked to the relief energy show less meaningful correlations with median SSY. The mean slope of the watershed (R = −0.28) presents an inverse trend in comparison to the median SSY. The hypsometric value of the watershed showed a very weak correlation (R = 0.21) with median SSY. The mean elevation of the watershed is inversely linked with the median SSY (R = −0.43). However, the mean elevation is also related to the location of the agricultural areas, and thus with the availability of suspended sediment runoff in the watershed. No correlation was found between the median SSY and the watershed area. For the duration of the water samplings, the land cover has not been drastically modified. However, some trends were detected in the studied watersheds (see Table 9 in Supplementary materials). From 2000 to 2018, the forest land cover has increased in all regions — and especially in Lorraine (+ 0.9%) and Ardenne (+ 0.4%) — except in Hesbaye and Entre-Vesdre-et-Meuse. Grassland tends to decrease in every region (up to −2.6% in the Haine Basin), except in Brabant Plateau and Entre-Vesdre-et-Meuse. Cultivated areas increase in the Haine Basin (+ 1.8%) and in the Condroz (+ 0.61%); they decrease especially in the Brabant Plateau region (−1.8%) and in Hesbaye (− 0.9%). This decrease is related to the urbanization of the watershed (+ 1% in the loess belt). The Ardenne region is less marked by recent urbanization (+ 0.2%). Multi-criteria analyses have been undertaken to extract the coefficients of the most correlated or inversely correlated variables with SSY (Eq. 7).

$$\begin{aligned}\text{Median SSY } [\text{t km}^{-2}{\text{year}}^{-1}] &= 3.10 \text{ Mean slope} [\%] \\&- 0.527\text{ Forest cover }[\%] \\&+ 0.396\text{ Cultivated cover }[\%] + 11.2\end{aligned}$$
(7)

The coefficient of multiple determination (R2) reaches a value of 0.48, with a standard error of 26.1 t km−2 year−1. The multi-criteria analysis hardly explains the variance of median SSY with confidence, even if physical parameters of the watersheds play a role.

4 Discussion

The sediment rating-curve method tends to underpredict high and overpredict low SSC value (Horowitz 2003). Ideally, the calibration measurements should cover a full range of flow conditions from droughts to floods to ensure that extrapolation of the curve beyond the range of supporting measurements is minimized. However, achieving water sampling over a full range of flow rates is difficult where the sampling site is remote from the lab or where the watershed is small so that the river has a quick response to infrequent storms. For this reason, some degree of extrapolation of the rating curve is often necessary (McKerchar 2003). In case of flash-floods or pluri-centennial floods, Keaton (2019) estimates the threshold between normal streamflow to hyperconcentrated sediment flow at a sediment concentration by volume of about 0.05 to 0.1 (50 to 100 g l−1), while mud flood and mud flow are classified above 0.2 and 0.45 respectively. These reference values were used to ban the over-extrapolated values of the stations #42 and #61 (with about 35 and 59 g l−1 respectively); they are considered as unrealistic for their stream environment. In the end, the #62 station could be considered as invalid due to a probable underestimation of the sediment concentration (max. 986 mg l−1) even if the NSE and PBIAS give a “very good” state of validity. There is a possibility that sediment accumulation occurred upstream of the sampling station according to Hombrouckx (2002). In any case, the manual sampling of over-bankfull discharges will remain hard due to the rarity of these events. Only automatic sampling stations would have achieved the gathering of data for very infrequent floods to study the within-event sediment variability and the hysteresis (Oeurng et al. 2010; Rodríguez-Blanco et al. 2010). In practice, this real-time monitoring and water sampling are rather difficult to set up in a large number of stations and requires high frequency field handlings and laboratory analyses.

Minimizing the sum of the squares of the logarithmic deviations results in the underestimation of the calculated concentrations (Wilson et al. 1990; Grasso and Jakob 2003; Hallot 2010). The bias increases with the degree of scatter about the regression (Ferguson 1986), despite the common use of a correction factor (Phillips et al. 1999; Ndomba et al. 2008), and the sediment yield underestimation may exceed 50% in some cases (Thomas 1985; Jansson 1985; Ferguson 1986, 1987; Koch and Smillie 1986; Cohn et al. 1989; Lamalle et al. 1989; Grasso and Jakob 2003). We have observed that the correction factor led to an increased value of around 30% when the R2 of the sediment rating curve is around 0.5. In the case of a R2 above 0.8, the increased value of the corrected SSY remains below 10%. The quality of the sediment rating curve has therefore a significant impact on the computed results, much higher than the exhaustiveness of the range of sampled discharges. However, the slope of the logarithmic rating curve could change significantly in case of new samples, because this type of mathematical relation is driven by extreme values. In addition, the values of the parameters ‘a’ and ‘b’ of the sediment rating curve may change with time: hydrologic changes may come from human-caused alterations of the global climate system or river restoration plans (Warrick 2015). Authors showed that variation in flow discharge can lead to a general decrease in ‘b’ and an increase in ‘a’ during the period with more flood events (i.e., an increase in sediment transport) and an increase in ‘b’ and a decrease in ‘a’ during droughts (Higgins et al. 2013).

The relationship between the SSY and the area of the watershed is rarely significant. In a large study of 60 watersheds in Spain, Verstraeten et al. (2003) could not explain more than 17% of the variability of the SSY with the watershed area. Multiple regression models, based on climatic, topographic, and land use properties, often are insufficient to model the sedimentary behaviour of the watersheds (Verstraeten et al. 2003). The same observations were made in Italy by de Vente et al. (2006). In the presence of clastic materials such as badlands, gullies, or landslides, the prediction of SSY seems to be easier, based on the drainage density (de Vente et al. 2006; Grauso et al. 2008). In Wallonia, the prediction seems to be more difficult due to the large range in the pattern of types of lithologies and soils over a rather small area. The high variability of lithologies prevents drawing a clear link between drainage density and sediment yield. The land cover is more likely to be used to explain a part of the spatial variability of sediment yields as it can be presented in correlation and multicriteria analyses. For instance, predominant agricultural land cover conjugated to higher watershed slope in the vicinity of the watercourses in the loess belt (Senne, Dyle and Gette rivers) is likely to explain higher SSY values.

From the samplings of this study, it was shown that, as a general rule, the concentrations of suspended sediment observed in the summer period are higher than the concentrations observed for an identical discharge in the winter period, even when the sample is taken during the rising phase of the flood in the studied stations. In addition, it was observed that this difference between the concentration of samples in winter and in summer is maximum in the small watersheds located in silty areas. Large-area watersheds are less subject to this phenomenon of temporal differentiation because the impact of runoff due to summer storms is integrated over the entire surface.

Despite the high variability of factors controlling the sediment yields (Zabaleta et al. 2016), the physical characteristics of the watersheds play a role in sediment yield. In addition to the scale effect of their area, it has been demonstrated that parameters such as slope, land cover, lithology, and soil erosivity have an effect on sediment budget (Van Oost et al. 2000; Yan et al. 2013; de Vente et al. 2013; Fang et al. 2016; Messina and Biggs 2016). Moreover, the link between higher drainage density and higher sediment yield is only present in homogeneous bedrock geology (Dragičević et al. 2018). In addition, the connectivity of the river networks also plays a significant role. The presence of weirs and dams impacts the fine sediment transport (Lajczak, 1996). In urban area, the rectified reaches also affect its transport by the narrowing due to bridges and hydraulic obstacles. Even if the land cover data are not sufficient to realistically represent all the sedimentary processes that take place from soil erosion to the outlet and the accumulation processes, the more complex models are generally very difficult to calibrate for large watersheds. The use of Corine Land Cover maps with a resolution of 100 m was the only available technique to compare different years with thematic consistency over the last decades. Due to the international environment of the studied watersheds, pan-European data are needed to deal with the part of them that are in boundary countries and regions.

Compilation studies often highlight the influence of runoff in the variability of SSY. In Wallonia, Perpinien (1998) observed very low denudation rates at the level of the Mehaigne, at its confluence with the Meuse (14 t km−2 year−1 at Wanze in 1997), while Sine and Agneessens (1978) observed a SSY of barely 10 t km−2 year−1 in the upper Mehaigne (20.4 km2). These differences highlight the great interannual variability of SSY which makes it difficult to compare the values based on different analysis periods, especially when they include particularly dry years, such as the 1973–1977 period studied by Sine and Agneessens (Lamalle et al. 1989). The results of Vanmaercke et al. (2012b) with a worldwide dataset remain inconclusive about the potential impact of land use on the inter-annual variability of SSY, but indicate a weak correlation. Costa et al. (2017) only managed to observe a link with land cover modifications over a 40-year period of data with continuous monitoring in an Alpine environment.

This study has demonstrated that nowadays the rivers that were suspected of clogging the downstream waterways, such as the Trouille River and the Samme River, did not represent a massive input of sediment each year. However, taking into account the estimated value by the public managers of 1.3 million tons of accumulated sediment in the Condé-Pommeroeul canal since 1818 (connected to the Trouille River via the Haine River) and the data from the station #40 (Trouille River at Hyon, with a watershed area of 224 km2 and a median SSY of 28.4 t km−2 year−1), the simple extrapolation by multiplication of our median SSY results over the period 1818–2010 reaches 1.22 million tons of potentially accumulated suspended sediment for this 192-year period. The computation by other means of the sediment yield was in accordance with the results found in our study. For instance and despite the difference in watershed area, the SSY was estimated to be between 28 and 33 t km−2 year−1 with an accumulation in the Bütgenbach dam from 1932 to 2004 (Hallot et al. 2012) compared to 30 t km−2 year−1 of median 1996–2018 SSY value with suspended sediment samplings.

Several types of potential uncertainty were addressed in the computation of the SSY from sediment rating curves. The first type is due to the sampling method and the representativeness of the sample across the water column and the water section. Perpinien (1998) observed, thanks to measurements conducted by Lamalle (1987) in the Burdinale River and Pironet (1995) in the Magne River, a low variability in suspended sediment concentration of samples taken at different depths on cross-sections. This variability is different with average concentrations. The coefficient of variation is 8 to 12% for average concentrations below 30 mg l−1, between 1.3 and 6% for average concentrations of 150 to 600 mg l−1 and 8.4% for a concentration of 1800 mg l−1 on the Burdinale River. The second type of uncertainty relates to the quality of the sediment rating curve, linked to the trap efficiency of the filtration method and the ability for modelling suspended sediment concentration across a wide range of discharge. Jung et al. (2020) summarize a computing method of several validation criteria to describe the quality of the sediment rating curve. Another source of uncertainty in the SSY computation is the model of interpolation of annual or pluri-annual sediment yield. A great variety of types of calculations exist to interpolate from infrequent samples (Phillips et al. 1999; Delmas et al. 2011). The summation of modelled concentrations on an hourly discharge basis was used in this study. However, each type of computation model may give very different SSY results, in relation to the number of samples, the coefficient of determination of the sediment rating curve, and the distribution of the samples over the range of flow discharge values. Estimating sediment yield over large watersheds is a difficult task, taking into account the great spatial and temporal variability of physical parameters of the watersheds, and the difficulty of sampling with representativeness the suspended sediment at a wide range of discharge values.

The frequency of suspended sediment sampling also has a non-negligible impact on the annual sediment yield calculated from the flow series (Horowitz 2002, 2003; Vanmaercke et al. 2012b; Skarbøvik et al. 2012). Walling (1977) showed significant differences in the interpolation of SSY with different sampling timings such as daily, monthly or seasonal intervals. Underestimates of up to 70% are reported in the literature when the sampling frequency is weekly (Li et al. 2006). A fairly small number of samples, for example 12 samples taken on a hydrological basis rather than a calendar basis, may allow a first estimate of the annual sediment transport to be obtained while minimizing the trips required to cover a large network of stations (Skarbøvik et al. 2012). The representativeness of the sediment rating curves obtained is correct for the majority of the study sites. However, the number of samples or the range of sampled discharges was too small in 10 cases, leading to unsatisfactory results. This compromise between the precision of the estimate of the annual sediment yield and the ability of sampling a large number of measurement sites is dependent upon each watershed, because the bias due to a low sampling frequency can vary greatly with the area and physical characteristics of the studied watersheds (Moatar et al. 2006).

In the end, the source data of land cover maps, their reference year, and their resolution may be another issue affecting the multicriteria analyses. With a variable sampling period for the studied sites, the land cover map at different epochs has been used and integrated to the searching for correlations between land cover and the SSY at the outlet of the watershed. The Corine Land Cover map for the year 2006 has shown better correlation values compared to ProbaV 2015 maps (Buchhorn et al. 2019), previously tested. The current trend of urbanizing the areas in the agricultural area such as the Hesbaye and the Brabant Plateau region could lead to an increase of the sediment transport by raising the connectivity of the sediment source and the rivers downstream through the impervious area.

5 Conclusions

Since the end of the 1990s, campaigns of suspended sediment samplings have been carried out on 72 study sites in the Southern part of Belgium in order to acquire sediment yields at the outlet of watersheds located in different geographical regions. The statistical validity and representativeness of the sediment curves has been described using several descriptors. This type of study aiming at accumulating many suspended sediment samples from a great number of gauging stations confronted us with the difficulty of obtaining representative results at different time and space scales. The classical method of interpolation of the suspended sediment concentration with the Ferguson correction required by the bi-logarithmic space was the methodology used to compute the SSY over a representative period with the available series of hourly discharges. However, the high variability of SSY related to the seasonality of sediment transport in fluvial regimes would, due to logistics and costs, tend to result in only a small number of stations, representative of geographic regions or watersheds, being equipped with continuous measurement devices to acquire data during single events. The methodology described here made it possible to obtain an order of magnitude of the sediment transport in suspension for a large number of stations, which was the initial request of the river managers.

Belonging to seven very different geographical regions, with their specific geologic substratum, type of soils, and land cover, differences in SSY have been observed. Considering the computed data for valid sediment rating curves, median SSY reaches, on average, 20.1 t km−2 year−1 in Lorraine, 28.7 t km−2 year−1 in Ardenne, 52.9 t km−2 year−1 in the Haine basin, 30.7 t km−2 year−1 in Fagne-Famenne, 96.1 t km−2 year−1 in Hesbaye, 63.2 t km−2 year−1 in Entre-Vesdre-et-Meuse, and 192.2 t km−2 year−1 in the Brabant Plateau region. The estimated SSC value at bankfull discharge was related to the proportion of agricultural areas in the watershed (R2 = 0.60). Indeed, rivers from Ardenne, Fagne, and Famenne regions and the northern part of the Lorraine region depict low SSC at Qb. Conversely, Brabant Plateau and Hesbaye rivers (and to a lesser extent, those located in the southern part of Lorraine) show the larger sediment concentrations at bankfull discharge, correlated to a predominantly agricultural land cover. The percentage of cultivated areas (in 2006) is weakly correlated with median SSY of the 1996–2018 period (R = 0.42), while the percentage of forested areas (R = −0.45) is inversely correlated with sediment yields. The mean slope of the watershed (R = −0.28) and the average elevation (R = −0.43) present an inverse trend in comparison to the median SSY, but the mean elevation is linked to the proportion of agricultural areas in terms of climatic environment and the availability of arable lands in Wallonia.

Compared to other sediment transport analyses, the order of magnitude of median SSY (33 t km−2 year−1) was consistent with other studies in the same climatic context and for the same range of watershed areas. This study also confirms the great temporal and spatial variability of SSY. In terms of clogging of the waterways and dams, the computations made from sediment rating curves and hydrological data matched with the estimates of accumulated sediment over different time intervals.

The uncertainties that are linked to the sampling methods, the quality of the sediment rating curves in terms of representativeness, and the soundness of the choice of the available computation methods still make complex today the study of sediment transport in rivers. Further analyses that are based upon high-frequency water samplings and long-term data gathering would be necessary to define more precisely the intrinsic and complex sedimentary processes that take place in watersheds.