Rainfall was measured at nine locations using tipping bucket rainfall recorders that allow the recording of every 0.1 or 0.2 mm of rain (cone surface of 1000 or 2000 cm2). Table 3 lists all the rain gauges, the type of instrument and logger, the coordinates and elevations, and the period of data acquisition. All the rain gauges except three were equipped with an insulated chamber, and a heating system was installed only for the Laval rain gauge in 2007. Data are stored in a logger system that records tipping times. All the rain gauges except two are also equipped with a container that collects water under the rain gauge. Every 2 weeks, data are retrieved from the logger, the rain gauge is cleaned if needed (leaves, needles), and the volume stored in the container is measured to check that there is no bias or drift in the instrument.

The raw data series is processed as follows: first, when the data are retrieved from the logger, we check for consistency between cumulated recorded rainfall and water volume stored in the container below the rain gauge. If these two values differ by more than 2 % and less than 10 %, the rain gauge is recalibrated, and the data for the antecedent rainfall period are corrected such that the cumulated rain over the period corresponds to the real water volume that is collected. Secondly, data from all the rain gauges are compared. Due to the small spatial extent of the site, most rainfall events reach all the rain gauges. Snow events are detectable in the raw data set by the difference in the cumulated rainfall recorded by heated and non-heated rain gauges. Cumulated rainfall recorded by non-heated rain gauges increases slowly over the days that follow the precipitation event as snow melts, especially during the daytime, whereas cumulated rainfall recorded by the heated rain gauge increases more quickly over the duration of the precipitation event. Moreover, the site is also equipped with a disdrometer (data not presented here) which measures drop size and velocity distribution and allows us to distinguish snow from rain events. When a snow event is detected, the data from the non-heated rain gauges are kept as they are with a lower quality code.

Following Table 2, quality code 0 is attributed to all missing data. Quality code 2 is attributed to correct data. When there is a misfit smaller than 10 % between the recorded rainfall and the container, the data that are stored in the database are the corrected data with a quality code 2. Quality code 3 is attributed to snow melt data over a short period, and quality code 4 is attributed to snow melt data when the snow and melting last several days. Quality codes 4 or 5 are attributed when the bucket is found to be partially obstructed by leaves or needles or for cumulated records that differ by more than 10 % from the container. Quality code 1 generally corresponds to the oldest data that have not been qualified. A quantitative uncertainty estimation for rainfall data with quality codes 2 and 3 is therefore, respectively, ±2% and ±10%. Note that the rainfall data from the rain gauges of Laval and Pompe are very similar; therefore, we recommend using Pompe data when Laval data are missing.

Detailed records of rainfall allow for the computation of rainfall intensity at a short timescale. Figure 5 shows an example of a flood from 3 June 2013. Rainfall intensity computed over 1 min reaches 84 mm h-1 at the Pompe rain gauge and 126 mm h-1 at the Pépin rain gauge.

Rainfall records are also cumulated and analyzed at the daily, monthly and yearly scale. Figure 6 shows the monthly repartition of the rainfall over the year for the Laval rain gauge. For a total annual rainfall of 920 mm, the rainiest periods are spring and autumn.

A full climatic station was installed in 2000 on the ridge between the Laval and Moulin stations (L93 coordinates are 968 755 and 6 343 638; elevation is 862 m; see location in Fig. 4). It includes temperature, air humidity, wind direction and speed, and radiation (short wavelength). The station is equipped with a Campbell Scientific CR10X data logger, solar panels and batteries.

Table 4 summarizes the types of instruments, measurements and associated acquisition periods. A second climatic station was installed in 2013 in the Brusquet watershed, but the data are not validated yet; therefore, they are not presented here.

Meteorological data are checked for consistency but do not generally require processing before storage in the database unless failures are detected in the instruments or logger. In such cases, the data are indicated as missing data until the device is repaired. The uncertainty associated with this data is therefore the measurement precision given in Table 4.

Hydrological stations are installed at the outlet of each catchment, and their location is visible in Fig. 4. All the stations have the same design, with a data logger (Campbell Scientific or Serosi) that controls sensor acquisition and stores the data. Some stations are connected to a power supply, while others are equipped with solar panels and batteries. Except for the Bouinenc, the gauging system is either a Parshall flume or a V-notch weir, and it is associated with one device or several devices for water level measurement. A summary of the gauging devices and water level instruments for each station is presented in Table 5. Photos of all the gauging stations are presented in Fig. 7.

For the Laval station, the gauging system is a Parshall flume (size 2 ft – throat size 0.61 m). The Parshall standard relation between water depth and discharge is used until the flume is flooded, which corresponds to a discharge of 1.9 m3 s-1. Above this value, the flow is constrained inside a rectangular weir, and a weir equation is used to relate water depth and discharge (all gauging relations are detailed in Table 6). The water level is currently measured with an ultrasonic sensor (Paratronic US6) and two nilometers (Serosi – vertical arrays of electrodes that are progressively connected as water level rises) for low and high water stages (sizes 960 and 1920 mm). Note that, in the past, different devices were used to measure the water level. A bubbler sensor was mainly used before 1997, and since then, several sensors including ultrasonic (US) sensors, radar sensors, a homemade sensor called ELLAN (Olivier and Pebay-Peroula, 1995) and nilometers have been used. Water level measurements are recorded with a Serosi data logger with a frequency that is controlled by water level changes. Water level is measured every 10 s in summer and every minute in winter, and the data are stored only when they differ from the previous record by more than a given threshold (5 mm). The resulting record therefore has a varying time step depending on the flood dynamics. For summer floods generated by storms, a time period of 10 s was chosen to best capture the flood dynamics, whereas for the slower and longer events of autumn and winter, a time period of 1 min was chosen to save some storage capacity. At every field visit (every 2 to 4 d for the Laval, Moulin and Roubine and every 2 weeks for the Brusquet), the discharge is manually measured with a calibrated bucket and chronometer at low flow. A standard water level scale is also installed in the Parshall flume, and the water level is visually recorded during every field visit. Specific conditions such as ice on the sensors or in the flume are also recorded during each field visit.

For the Moulin station, two gauging systems are available: a trapezoidal flume and a Parshall flume of size 1.5 ft. The trapezoidal flume is located a few meters upstream of the Parshall flume. The Parshall standard relation is used until the flume is flooded, which corresponds to a discharge of 0.75 m3 s-1. The depth–discharge relation of the trapezoidal flume is then used when the Parshall flume is flooded. Water level is measured with a nilometer and a radar (Vegapulse WLS61) in the Parshall flume. Water level was recorded by a Serosi data logger using the same procedure as for the Laval until 2015. It is now recorded on a Campbell CR1000 logger using a simpler procedure where one point is stored every minute.

For the Roubine station, the gauging system is a V-notch weir. The water level is measured with a nilometer and recorded on a Serosi logger. It is measured every 10 s in summer (every 1 min in autumn and winter), and the data are recorded only if the change in the water depth since the last measurement exceeds a given threshold (same procedure as the Laval). The discharge is computed from the water depth using the weir equation listed in Table 6.

For the Francon station, the gauging system is a Parshall flume, and the water level is measured with an ultrasonic sensor (Paratronic US10) and recorded on a Serosi logger following the same type of procedure as that described for the Laval.

For the Brusquet station, the gauging system is a Parshall flume, and the water level is measured with a nilometer (1920 mm) located in a well. The acquisition frequency is one point per minute. A V-notch weir with a nilometer was installed in 2010 to better measure low flows, but it was damaged by a flood in 2018.

The Bouinenc station is different from the others since it is located in a natural river section next to a bridge in the village of Draix. Until 2016, the water level was measured with an ultrasonic sensor (Paratronic US10) attached to the bridge. A gauging curve was built to compute the discharge based on manual measurements of flow velocity using the salt dilution method. In June 2013, the section was greatly modified by a series of floods that moved bars around, and a new bar appeared just below the US sensor. The gauging relation is therefore out of date. As a result, no discharge data are available after this date.

When available, the nilometer is the reference sensor for the water depth. It is completed or corrected by data from other sensors only when needed. The data-processing scheme for the water depth record is the following:

Remove human artifacts from the record. In particular, cleaning operations in the sediment deposit area upstream of the gauging station and on the grid that dams this deposit area generally generate short peaks of water depth in the data record which do not correspond to actual flows. These points are therefore removed from the record.

Regular field visits are a great advantage for data quality since they allow manual measurements of low discharges and scale reading, and field observations, which are useful to interpret, criticize and possibly correct the sensor data. Several issues may affect the quality of the water depth record. When sediment or ice are deposited next to the measurement area, when the gauging system is flooded, or when the deposit trap is full of sediment, a lower quality is associated with the water depth data (quality code 3). This is also the case when the sensor record is very noisy or differs significantly from the scale reading. When none of these issues have been noted, the data are considered to be of good quality (code 2), and the uncertainty is mostly related to the noise from ripples and raindrops. Old data were not given any quality code. The uncertainty in the gauging relation is also accounted for in the quality code attributed to the discharge data. In particular, at the Laval station, when the Parshall flume is flooded, a rectangular-weir gauging relation calibrated only from numerical modeling is used to estimate discharge (Table 6). For this reason, quality code 3 is attributed to all discharge data above 1.89 m3 s-1. Current work based on flow surface image analysis is ongoing to improve the accuracy of high-discharge estimates. Another source of possible uncertainty for the highest flood peaks could be a buffer effect due to the sediment trap located upstream of the station. Based on our current knowledge, quantitative estimates of the discharge uncertainty are therefore 10 % and 30 %, respectively, for data with quality codes 2 and 3.

An example of hydrographs obtained for all the catchments on 3 June 2013 is given in Fig. 5. The response time for this summer event is small, ranging from 6 to 10 min for the mostly denuded badland catchments (Roubine, Moulin and Laval), and is slightly higher for the larger Bouinenc catchment (15 min) and much larger for the reforested Brusquet catchment (1 h). Peak discharge reaches 35 m3 s-1 km-2 for the smallest Roubine catchment.

Suspended-sediment concentration (SSC) is measured at all the hydrosedimentary stations using a combination of turbidity sensors and automatic samplers (see Table 7). All the stations are equipped with an ISCO automatic sampler, even two for the Laval. Each sampler is able to collect up to 24 samples of 1 L each. The samplers are controlled by the data logger to collect water (including suspended matter) when water depth exceeds a threshold (e.g. 20 cm at Laval), remains higher than this threshold for more than half an hour or changes faster than a threshold (e.g. 10 cm min-1 at Laval). The samplers are able to collect up to one sample every minute. The bottles are retrieved after each flood and brought to the laboratory for decantation, oven drying (24 h at 105 ∘C) and weighing. SSC is then calculated as the ratio of the dry weight to the sample volume. As an example for the Laval station, more than 1000 samples were collected over the period 2017–2019, representing an average of 11 samples per flood.

Moreover, the stations are equipped with sensors for a continuous measurement of turbidity. The Laval, Moulin and Francon (until 2018 for Francon) stations are equipped with an optical fiber sensor that was specifically designed for the high concentrations that are encountered in these catchments (several hundreds of grams per liter (g L-1); Bergougnoux et al., 1999). This sensor is able to measure concentrations in the range of 10–800 g L-1 with an accuracy of ±5 g L-1. The Bouinenc and the Roubine are equipped with WTW turbidimeters (VISolid 700IQ). For the Laval, Moulin and Francon, the measurement from the optical fiber sensor is recorded only if it differs by more than a given threshold from the previous one; it is recorded with an algorithm similar to what is used for water depth recording. The acquisition frequency for the turbidimeter is one point per minute at the Roubine and one point every 10 min at the Bouinenc station. For low flows, the water depth is not sufficient to reach the turbidimeter or the suction strainer of the sampler pipe. The concentration measurement is therefore only available for flood events.

Following the 2013 flood at the Bouinenc station, the turbidimeter and the suction pipe for the sampler have been buried in a gravel bar; therefore, the concentration is not measured anymore at this station.

Common practice for concentration measurements consists of using the discrete concentration values obtained from the sampler to calibrate a turbidity–concentration relation (or voltage–concentration) and then converting the full continuous turbidity time series in a concentration time series based on this relation. In Draix, however, the relation between turbidity (from the WTW turbidimeter) or voltage (from the optical fiber sensor) on one hand and concentration on the other hand is often noisy and varies a lot from one flood to the other. In particular, a hysteresis is frequently observed between the water discharge and the concentration on one hand and between the tension and the concentration on the other hand. Moreover, the optical fiber sensor is not reliable at low concentrations (i.e. less than 10 g L-1). On the contrary, the concentration measured from the samples is a reliable source of information for which the uncertainty remains small, especially for high values of concentration (1 % to 5 % for high and low concentrations, respectively). Therefore, we have defined a specific procedure to reconstruct the concentration time series based on the discrete concentration values measured in the samples. This procedure is meant to give more credit to sample concentration data than turbidimeter data and to ensure that the reconstructed concentration is equal to the sample concentration where samples are available, all while preserving the hysteretic pattern between discharge and concentration. For each flood, the concentration time series is reconstructed, starting from the sample values followed by combining several procedures, as illustrated in Fig. 8.

First, an event-specific concentration–voltage relation is built for the event if at least four samples are available. If not, the yearly or interannual relation is used.

Finally, a flux of suspended matter is computed by multiplying the discharge and the concentration, and a total mass exported for each flood is computed by integrating this flux over the duration of the flood. This mass can be transformed into a volume of eroded rock using the marl density value of 2750 kg m-3.

A full quantitative assessment of the uncertainty associated with the SSC is currently in progress but requires detailed investigations and modeling that are out of the scope of this paper. However, we provide below some elements to estimate this uncertainty and to help the reader use this data. First, the measurement uncertainty associated with the concentration measured in the samples ranges from 1 % to 5 %. Secondly, the interannual calibration curve between turbidity and concentration measured in samples at the Laval station over the period 2012–2019 yields a mean absolute error of, respectively, 60 % and 13 % for points of concentration lower than 50 g L-1 and higher than 50 g L-1. Note that, at the Laval station, 80 % of the annual sediment flux is exported at concentrations higher than 50 g L-1 (Seve, 2021). As explained in Bergougnoux et al. (1999), part of this error is related to the time needed to pump the water sample, which is longer than the duration of the turbidity measurement, and another part is linked to the variation in the sediment grain size during one flood and from one flood to the other. Thirdly, a preliminary analysis over a few event-scale turbidity calibration curves suggests that the mean absolute error when estimating SSC from turbidity using these curves is approximately ±10%.

When no samples are available for a flood, the annual or interannual turbidity calibration is used, and the resulting SSC time series is given quality code 3. When at least four samples are available and are well distributed over the flood duration and over the range of discharges, an event-scale calibration is used and combined with the sample data, and the resulting SSC time series is given quality code 2. If the samples are less than four and/or are not well distributed over the flood, a quality code 3 is given. Based on our current knowledge, a rough estimate of the uncertainty for SSC data is therefore ±10% for data with quality code 2 and ±13% to ±60% for data with quality code 3 with concentrations that are, respectively, lower or higher than 50 g L-1.

Note that the quality of the sediment concentration data has been assessed only recently; therefore, most historical data in the data base have no quality code attributed to them. For this data, we therefore recommend considering an uncertainty of ±60% for concentrations lower than 50 g L-1 and ±13% for concentrations higher than 50 g L-1, although we hope to refine and reduce these values in the future by means of a more detailed assessment of uncertainty, which is in progress.

An example of a suspended-sediment flux time series obtained for the flood of 3 June 2013 is given in Fig. 5. The concentration reaches 594 g L-1 at the Laval station and 30 g L-1 at the Brusquet station. Summing up with the volume exported as bedload, the total sediment yield for this flood is 7840 t for the Laval, corresponding to a surface ablation of 5 mm.

The highest concentrations are observed at the Laval station, and instantaneous concentrations higher than 600 g L-1 have been observed 11 times over the last 30 years (see Fig. 9). For such values, the fluid becomes non-Newtonian, which might result in an overestimation of the discharge when combined with a low water depth, i.e. in the early moments of floods (Le Bouteiller et al., 2021). This effect has not been accounted for in the depth–discharge relations and might affect the discharge estimation in the earliest moments of highly concentrated floods. However, we tested this effect for the flood of 6 July 2006 at the Laval station and found that the resulting error at the flood scale was less than 1 % for the total water volume and less than 3 % for the total sediment volume.

Between flood events, the water depth is not sufficient to perform concentration measurements. However, for the small catchments of Laval, Moulin, Brusquet, Francon and Roubine, base flow is low or null; therefore, most of the discharge passes during floods, and the sediment flux between the events should be negligible. This was checked for the Laval: we compared the yearly sediment export obtained by summing only the contributions of the floods or by adding a contribution of the interflood periods assuming an average concentration of 0.1 and 1 g L-1 (which is a high value of concentration for low flows) for these periods. An interflood contribution of concentration 0.1 g L-1 (1 g L-1) was found to increase the yearly sediment flux by only 0.12 % (1.2 %). It is therefore reasonable to sum the flood contributions to estimate the yearly sediment export from such catchments.

Sediment transport does not occur only through suspended matter but also as bedload. Apart for the Bouinenc, all the stations are equipped with a sediment trap located upstream of the gauging station. An example of the Laval sediment trap is presented in Fig. 10. Some traps are dug into the marl bedrock (Laval, Francon), while others have a concrete bottom (Moulin, Roubine). The storage capacities of these traps are summarized in Table 8. The downstream side of the trapping area is closed by a grid. Coarse sediment transported as bedload is deposited in the traps due to the reduction of the transport capacity (flow divergence and slope reduction) and/or is stopped by the grid. After each flood or after a few consecutive flood events, the topography of the deposit area is measured and compared with the previous topography to estimate the volume of sediment that has been deposited by the flood. For the Laval, Moulin, Francon and Brusquet, the topography is measured using a tacheometer pointing at orange spray dots painted at the surface of the deposit. Approximately 120 dots are used for the Laval deposit trap, with one transect every 4 m, and 70 dots are used for the Moulin deposit trap, with one transect every 2 m. For the Roubine, the topography is measured with a rule across the deposit, with a resolution of 60 measurements for 17.7 m2. The traps need to be emptied regularly, a few times a year. This is done with a power shovel for the largest traps and manually for the Roubine, and the total volume excavated is also recorded at that time (number of buckets or number of trucks filled with sediment). Note that some fine sediments (<20%) are stored together with coarser material in the trap, as shown by Liébault et al. (2016), but they only represent a small fraction of the total suspended-sediment volume at the event scale.

For tacheometer measurements, the x, y and z coordinates of the painted dots are interpolated into a gridded surface with a grid size of 50 cm for the Laval and 16 cm for the Moulin. Subtraction between two consecutive surfaces over the extent of the sediment trap yields a volume of sediment which is attributed to the flood event. In the case where it is not possible to measure the topography between one flood and the next one, the volume is attributed to the batch of floods that occurred between the two consecutive measurements of the topography. The volume is then transformed into a mass of sediments using a bulk density of 1700 kg m-3. This value was obtained from in situ measurements (Mathys, 2006) on samples collected in the trap and accounts for the porosity and water content of the deposit. The trap is emptied regularly, and at every cleaning, the cumulated volume obtained by summing up the contributions of all floods since the previous cleaning is compared to the volume that is removed from the trap.

The uncertainty of bedload volume measurement results from the uncertainty in measuring the x, y and z coordinates of the points at the surface and interpolating these points into a surface model. Moreover, subsidence of the bedload volume is possible as the material dries after being deposited by a flood. The uncertainty for the bedload volumes can therefore be estimated at ±10%.

Summing up the suspended-load mass and the bedload mass for each event provides a series of event-scale total sediment yields. Figure 11 shows an example of the cumulated suspended-load and bedload yield for 2013 at the Laval station. The total yield for 2013 is 20 900 t, which corresponds to a distributed ablation of 13.4 mm of marl across the catchment (computed for a fresh bedrock of density 2750 kg m-3). Note that 2013 was one of the most erosive years recorded for the observatory. The contributions of the suspended load and bedload are, respectively, 72 % and 28 %. The largest contribution for this year is the flood of 3 June, which was presented in Fig. 5. Averaging over the whole period of 2003–2022, the respective contributions of suspended load and bedload for the Laval station are 32 % and 68 %.