Headwater streams represent a substantial proportion of river systems and many of them have intermittent flows due to their upstream position in the network. These intermittent rivers and ephemeral streams have recently seen a marked increase in interest, especially to assess the impact of drying on aquatic ecosystems. The objective of this paper is to quantify how discrete (in space and time) field observations of flow intermittence help to extrapolate over time the daily probability of drying (defined at the regional scale). Two empirical models based on linear or logistic regressions have been developed to predict the daily probability of intermittence at the regional scale across France. Explanatory variables were derived from available daily discharge and groundwater-level data of a dense gauging/piezometer network, and models were calibrated using discrete series of field observations of flow intermittence. The robustness of the models was tested using an independent, dense regional dataset of intermittence observations and observations of the year 2017 excluded from the calibration. The resulting models were used to extrapolate the daily regional probability of drying in France: (i) over the period 2011–2017 to identify the regions most affected by flow intermittence; (ii) over the period 1989–2017, using a reduced input dataset, to analyse temporal variability of flow intermittence at the national level. The two empirical regression models performed equally well between 2011 and 2017. The accuracy of predictions depended on the number of continuous gauging/piezometer stations and intermittence observations available to calibrate the regressions. Regions with the highest performance were located in sedimentary plains, where the monitoring network was dense and where the regional probability of drying was the highest. Conversely, the worst performances were obtained in mountainous regions. Finally, temporal projections (1989–2016) suggested the highest probabilities of intermittence (> 35 %) in 1989–1991, 2003 and 2005. A high density of intermittence observations improved the information provided by gauging stations and piezometers to extrapolate the temporal variability of intermittent rivers and ephemeral streams.
Headwater streams represent a substantial proportion of river systems (Leopold et al., 1964; Nadeau and Rains, 2007; Benstead and Leigh, 2012). From an ecological point of view, headwater catchments are at the interface between terrestrial and aquatic ecosystems and they often harbour a unique biodiversity with a very high spatial turnover (Meyer et al., 2007; Clarke et al., 2008; Finn et al., 2011). Their contribution to the functioning of hydrographic networks is essential: sediment flows, inputs of particulate organic matter and nutrients, refugia/colonisation, sources of aquatic organisms (Meyer et al., 2007; Finn et al., 2011).
Headwater streams are generally naturally prone to flow intermittence, i.e. streams which stop flowing or dry up at some point in time and space, mainly due to their upstream position in the network and their high reactivity to natural or human disturbances (Benda et al., 2005; Datry et al., 2014b). These waterways which cease flow and/or dry are referred to as intermittent rivers and ephemeral streams (IRES). The geographic extent of IRES is poorly documented due to mapping limitations (digital elevation models, satellite images, aerial photos) and because of their size and their location (Leopold et al., 1994; Nadeau and Rains, 2007; Benstead and Leigh, 2012; Fritz et al., 2013). However, the proportion of IRES in hydrological networks can be very large: for example, they represent 60 % of the length of rivers in the United States (Nadeau and Rains, 2007) and are considered to represent probably more than 50 % of the global hydrological network (Larned et al., 2010; Datry et al., 2014b). Considering only gauging stations with continuous records may lead to severe underestimation of their regional extent (Snelder et al., 2013; De Girolamo et al., 2015; Eng et al., 2016).
Recently, IRES have seen a marked increase in interest stimulated by the challenges of water management facing the global change context (water scarcity issues, climate change impact, etc.) (Acuña et al., 2014; Datry et al., 2016b). Studies have characterised the hydrological functioning of IRES (Gallart et al., 2012; Costigan et al., 2016; Sarremejane et al., 2017) to assess the effects of flow intermittence on aquatic ecosystems (Larned et al., 2010; Datry et al., 2016b; Leigh et al., 2016; Leigh and Datry, 2017). IRES have been altered due to human actions (abstraction, hill dams, low-water support, pollution, etc.) despite their high and unique biodiversity (Datry et al., 2014a; Garcia et al., 2017a). In addition, some perennial streams are becoming intermittent due to global change, water abstraction or river damming (Skoulikidis, 2009), and the extent of IRES may increase in the future (Döll and Schmied, 2012; Jaeger et al., 2014; Pumo et al., 2016; Garcia et al., 2017b; De Girolamo et al., 2017a).
A better hydrological understanding of IRES is now essential and improved management requires knowledge of both the spatial extent and arrangement of IRES within the river network (Boulton, 2014; Acuña et al., 2017). Efforts have been made to estimate the spatial distribution of IRES at the catchment scale (Skoulikidis et al., 2011; Datry et al., 2016a), at the regional scale (Gómez et al., 2005) and at the national scale (Snelder et al., 2013). In France, Snelder et al. (2013) suggested a classification of IRES regimes and spatialised their distribution. Based on an analysis of the continuous gauging network, they showed that the proportion of IRES accounted for 20 to 39 % of the hydrographic network. The accuracy of the obtained map is highly dependent on the density of the flow monitoring network. The installation of additional gauging stations is expensive and headwater systems may be difficult to monitor due to active geomorphology processes or to difficult access.
As a promising tool to advance the mapping of IRES, citizen science creates
opportunities to overcome the lack of hydrological data, contributes to
densifying the flow-state observation network (Turner and Richter, 2011;
Buytaert et al., 2014; Datry et al., 2016b) and could be used for
hydrological model calibration (van Meerveld et al., 2017). In France, Datry
et al. (2016a) used such data to describe the spatio-temporal dynamics of
aquatic and terrestrial habitats within five river catchments located in the
western part of France. They showed that processes resulting in flow
intermittence were complex at a fine scale and could vary substantially among
nearby catchments. However, these data were only available in a few
catchments, limiting any attempt to map large-scale patterns of flow
intermittence in river networks. Since this first attempt, new sources of
observational data have become available in France thanks to the ONDE network
(Observatoire National des Etiages,
However, discrete observations of intermittence do not provide any information on the persistence of dry conditions between two consecutive dates of observation. The rewetting–drying events could have significant impacts on communities whose survival is conditioned by the duration/frequency of drying. The duration of drying is of importance for ecologists, as one key driver for the composition and persistence of aquatic species (Vardakas et al., 2017; Kelso and Entrekin, 2018; Vadher et al., 2018). Temporal extrapolations of river flow regimes are thus necessary to summarise the different facets of flow intermittence at various timescales, from daily to inter-annual.
The main objective of this paper is to use discrete (in space and time) field observations of flow intermittence to extrapolate over time the daily probability of drying (averaged at the regional scale). We first carried out a quantitative analysis of the ONDE network data in order to characterise the information that they contribute in comparison with the data resulting from the conventional hydrological monitoring. Then, we developed two empirical models based on linear or logistic regressions to convert discontinuous series of flow intermittence observation from ONDE into continuous daily probability of drying, defined at the regional scale across France. Explanatory variables were derived from available continuous daily discharge and groundwater-level data of a dense gauging/piezometer network, and models were calibrated using the ONDE discrete observations. The robustness of the models was tested using (1) an independent, dense regional dataset of intermittence observations and (2) observations of the year 2017 excluded from the calibration. Finally, the resulting models were used to extrapolate the regional probability of drying in France: (i) over the period 2012–2017 to identify the regions most affected by flow intermittence; (ii) over the period 1989–2017, using a reduced input dataset, to analyse temporal variability of flow intermittence at the national level.
The study area is continental France and Corsica (550 000 km
We defined regions as combinations of “level-2 Hydro-EcoRegions” (HER2) and
classes of hydrological regimes (HR). Hydro-EcoRegion (HER) corresponds to a
typology developed for river management in accordance with the European Water
Framework Directive. The Hydro-EcoRegion classification includes 22 “level-1
Hydro-EcoRegions” (HER1) based on geology, topography and climate, and
considered the primary determinants of the functioning of water ecosystems
(Wasson et al., 2002). HER2 correspond to a finer classification accounting
for stream size. HER2 have a mean drainage area of 5000 km
The ONDE network was set up in 2012 by the French Biodiversity Agency (AFB, formerly ONEMA) with the aim of constituting a perennial network recording summer low-flow levels and used to anticipate and manage water crisis during severe drought events (Nowak and Durozoi, 2012).
There are 3300 ONDE sites distributed throughout France (Fig. 1). ONDE sites
are located on headwater streams with a Strahler order of strictly less than
5 and balanced across HER2 regions to take into account the
representativeness of the hydrological contexts (Nowak and Durozoi, 2012).
The ONDE network is stable over time. Observations have been made monthly
(around the 25th) by trained AFB staff, between April and September, every
year since 2012. One of the statuses is assigned at each observation among
“visible flow”, “no visible flow” and “dried out”. Here, we consider
two intermittency statuses: “Flowing” when there is visible flow across the
channel (“visible flow”) and “Drying” when the channel is entirely devoid
of surface water (“dried out”) or when there is still water in the river
bed but without visible flow (disconnected pools, lentic systems) (“no
visible flow”). The proportion of drying sites determined on the basis of
the ONDE network for each HER2–HR combination is considered a good estimate
of the daily regional probability of drying (RPoD
Location of the 3300 ONDE sites and partition into HER2.
Distribution of the 3300 ONDE sites and of the 1600 gauging stations
available in the HYDRO database against
Figure 2 illustrates the complementary nature of the ONDE network to the
already existing HYDRO (
A spatially denser citizen science dataset of flow-state observations in
western France (Poitou-Charente region)
(
Two discharge datasets (continuous daily time series) were used as
explanatory variables of discrete intermittence observations, with the
objective of extrapolating the intermittence frequency over time. The two
datasets included time series of daily discharge extracted from the French
river flow monitoring network (HYDRO database,
The 2011–2017 dataset was composed of 1600 gauging stations distributed
across France. Each stream where a HYDRO gauging station is located has been
defined as IRES or perennial. Several definitions of IRES can be found in the
literature (Huxter and van Meerveld, 2012; Eng et al., 2016; Reynolds et al.,
2015). In this study, we considered stations to be intermittent when five
consecutive days with discharge less than 1 L s
The 1989–2017 dataset consisted of 630 gauging stations selected with less than 5 % of missing data (continuous or not) during the period 1989–2017. This dataset was thereafter used to estimate the regional probability of drying before the creation of the ONDE network.
Because groundwater resources influence stream intermittence, we used
available time series of the daily groundwater level available in the ADES
database (
The 2011–2017 dataset was composed of 750 piezometers with daily groundwater-level data with less than 5 % of missing data (continuous or not). The selection of the 1989–2017 dataset was not easy because few groundwater-level measurements were available in the database before 2000. For example, only five piezometers met the tolerance limit on missing values considered for the 1989–2017 discharge dataset. In order to extend the dataset and because groundwater levels were less variable than stream discharges, the proportion of permitted gaps was fixed to 20 % between 1989 and 2017. This led us to select 150 piezometers. Thereafter, when the missing data period was less than 10 days, groundwater levels were reconstructed by linear interpolation in order to reduce the proportion of missing values to less than 5 % for the 150 piezometers selected.
Strategy of parametric modelling (steps 1–4) developed to predict (step 5) the regional probability of drying (RPoD) by HER2–HR combination in France.
The parametric modelling strategy was based on five main steps (Fig. 3). The
first step consisted in selecting all ONDE sites, gauging stations
and piezometers located in the same HER2–HR combination. When the total
number of gauging stations and piezometers was less than five for a HER2–HR
combination, we merged the HER2–HR combination with a neighbouring one
located in the same HER1. This was done for 20 of the 280 regions. The second
step consisted in calculating the RPoD A truncated logarithmic linear regression (LLR), with two parameters A logistic regression (LR) with two parameters
Finally, in a fifth step, a daily regional probability of drying (RPoD) could be predicted for each HER2–HR combination with both models following analytical formulas (Eqs. 3 and 5).
We used (1) the POC independent data and (2) the 2017 ONDE year to test the robustness of the LLR and LR model to predict the intermittence frequency (1) in space and (2) over time. Note that when predicting on the POC datasets, a new model was calibrated using only ONDE sites located out of POC streams.
For both datasets (POC and ONDE, 2017), the relative performance of the LLR
and LR models was compared in multiple ways using both the 2011–2017 and
1989–2017 datasets. The performance of each model was evaluated by the
Nash–Sutcliffe efficiency criterion (NSE) (Nash and Sutcliffe, 1970):
Both models have been calibrated over the period 2012–2016 and were then applied in a fifth step to predict the daily RPoD in France (Fig. 3). The RPoD was firstly predicted over the period 2012–2016 in order to identify the most affected regions by flow intermittence using the 2011–2017 datasets. The second application concerned the extrapolation of RPoD in France over a longer period using the 1989–2017 dataset to analyse the temporal variability of flow intermittence at the national level. It should be noted that model predictions only concern streams with a Strahler order of lower than 5 due to the ONDE site location.
A total of 1127 ONDE sites have recorded at least one drying event during the period 2012–2016 representing 35 % of the 3300 ONDE sites. From the ONDE database the probability of drying at the country scale was computed as the total number of drying statuses over France divided by the total number of ONDE observations available during statuses the same year (Fig. 4a). Between 2012 and 2016, the most critical year is 2012, with 15 % of drying statuses followed by 2016 (14 %) and 2015 (14 %) (Fig. 4a). The years 2013 and 2014 are less affected with only 6 % of drying statuses observed (Fig. 4a).
Drying events mainly occur between July and September, but the evolution of the month's proportion of drying can differ between years (Fig. 4b). In more detail, water levels in 2012 decrease in August when the proportion of drying is 27 %, and the situation lasts until the end of September with 25 % of drying (Fig. 4b). In 2013, the proportion of drying is lower than in 2012, but follows the same pattern with an increase at the end of July (3 %), reaching 9 % in August and in September. In 2014, the first peak of drying (5 %) is reached early in June. Then, the proportion of drying decreases in July (3 %) and increases slightly in August (4 %), reaching 7 % in September. In 2015, the critical period occurs at the end of July with 19 % of drying statuses, and the proportion of drying decreases slightly at the end of August (17 %) until it reaches 9 % in September. Finally, in 2016, the situation gradually deteriorates every month, reaching 20 % of drying statuses in August, and 28 % in September.
Distribution of the percentages of drying observed at ONDE sites for
the years
Between 2012 and 2016, a proportion of drying higher than 50 % is recorded on 93 ONDE sites and their spatial distribution is very patchy at the scale of France (black and dark grey dots, Fig. 5a). There are only 158 ONDE sites with at least one drying event every year and a variability of drying locations can be observed across years. The south-east of France is heavily affected by rivers drying, where the proportion of drying can exceed 75 % annually (black dots, Fig. 5b–f). The north-western part of France is less affected, although many ONDE sites show a proportion of drying observed above 50 % in 2014 and 2016 (Fig. 5d and f). North-eastern France is rather affected in 2012, 2014 and 2015, where several ONDE sites have more than 75 % of drying statuses (Fig. 5b, d and e). South-western France is particularly affected in 2012 and 2015 (Fig. 5b and e).
The HYDRO dataset includes 90 gauging stations located on streams considered IRES, which represents only 5.6 % of the 1600 gauging stations against 35 % for ONDE sites. At the national scale, the number of IRES seems underrepresented in the south-western, central, and north-eastern parts of France and Corsica in comparison with sites experiencing drying in the ONDE network (Fig. 6).
Map of ONDE sites and HYDRO gauging stations with at least one drying.
The number of gauging stations with at least one drying event (discharge
< 1 L s
Annual statistics on flow intermittence calculated on HYDRO gauging stations between 1 May and 30 September.
Map of Nash–Sutcliffe criteria (NSE) obtained for each HER2–HR
combination between 2012 and 2016 with the 2011–2017 and 1989–2017 datasets
according to
The LLR and LR models, calibrated over the period 2012–2016, perform well with the 2011–2017 dataset, with a mean NSE of 0.8 with the LR model against 0.7 with the LLR model (Fig. 7a and b). With the LR model, 50 % of the HER2–HR combinations obtain a NSE greater than 0.8, representing a coverage of 65 % of the French territory, while 33 % of HER2–HR combinations display a NSE higher than 0.8 (50 % of coverage of France) with the LLR model. Regions with the highest performances are located in sedimentary plains, in the south-east of France and in the Pyrenees Mountains. Conversely, the worst performances are obtained in the mountainous regions of the Alps as well as in the Massif Central. In these regions the size of the HER2 is rather small and the number of ONDE sites, gauging stations and piezometers per HER2–HR combination are certainly too few to derive reliable relations. Despite pooling, estimating RPoD remains impossible for nine HER2–HR combinations (4.5 % of coverage of France) because the number of ONDE sites, gauging stations and piezometer sites is insufficient (less than five) to perform the regression analysis.
The performance level is lower when the 1989–2017 dataset is used in models: the mean NSE with the LR and LLR models is 0.7 and 0.6, respectively (Fig. 7c and d).
The LR and LLR models lead to a similar performance range. However, the LR model outperforms the LLR model in terms of number of HER2–HR combinations, with NSE greater than 0.8 (Fig. 7c and d). The performance is sensitive to the dataset. As expected, the best results are obtained with the denser network. A decrease in NSE by more than 0.2 is identified for 5 % of the French territory when the 1989–2017 dataset is used (black areas; Fig. 7e and f). The regions with the most degraded values of NSE are small HER2–HR combinations located in eastern France (Fig. 7e and f).
NSE calculated for each HER2–HR combination between 2012 and 2016
with the 1989–2017 dataset as a function of NSE calculated with the
2011–2017 dataset with, respectively,
The decrease in performance is mainly due to the difference in the number of gauging stations and piezometers between the two datasets (Fig. 8). The most degraded NSEs correspond to HER2–HR combinations where the number of gauging stations and piezometers considered in regressions is the most reduced, i.e. with a loss higher than 50 % of the stations (black and dark grey dots; Fig. 8a and b). However, the decrease in performance remains low (the difference in NSE is below 0.1 for 75 and 64 % of HER2–HR combinations with the LLR and LR models, respectively).
Comparison between observed proportion of drying RPoD
The observed proportion of drying RPoD
When the 1989–2017 dataset is used for explanatory variables, the simulations of RPoD are weakly degraded with both models (Fig. 9d, e, f). However, the simulated pattern is similar to the observed one. The LLR model outperforms the LR model during the 3 years of validation with the 1989–2017 dataset (black curve; Fig. 9d, e, f).
NSE criteria obtained between 2012 and 2017 with the LLR and LR models calibrated over the period 2012–2016.
During the calibration period, the LLR and LR models tend to better simulate
the RPoD during dry years 2012 and 2016 (NSE
During the validation year of 2017, both models obtain a similar performance
over the year independent of datasets (NSE
Scatter plot of the predicted RPoD (
Monthly NSEs in 2017 follow the same trend as monthly NSEs of the calibration
period, with lower NSEs in May (NSEs < 0.4) and June
(NSEs
Both models have been applied using the 2011–2017 dataset. Figure 11
displays the maximum number of consecutive days (D
The LR model tends to simulate shorter periods of drying, particularly in HER2–HR combinations located in south-eastern France in 2013 and 2014 (Fig. 11). However, there is an overall agreement between RPoD simulated by both models in terms of spatial and temporal extent of dry streams.
Maximum duration of consecutive days with RPoD higher than 20 % simulated with the LLR and LR models.
RPoD simulated between 1989 and 2016 with the 1989–2017 dataset
with
The trend temporal patterns of RPoD predicted by the two models, considering
the 1989–2017 dataset, look similar between 1989 and 2016, and the simulated
RPoD fit well to RPoD
The proportion of drying is highly variable over the total simulation period, with alternating dry (1989 to 1991, 2003 to 2006, 2009 to 2012) and wet (1994 to 1995, 2000 to 2002, 2013 to 2014) phases. In spite of inter-annual variability, peaks of RPoD occur regularly between August and September, whether in dry years or wet years. This finding is consistent with the preeminence of rainfall-fed river flow regimes with low flows in summer, in France.
The highest values of RPoDs (above 35 % over France) are observed in 1989, 1990, 1991, 2003 and 2005 (black curve, Fig. 12a and b). The RPoDs simulated during these dry years are out of the range of the observed values over the calibration period (2012–2016). Estimations are thus uncertain. However, the high values of RPoD are consistent with observations reported in previous studies (e.g. Larue and Giret, 2004; Snelder et al., 2013; Caillouet et al., 2017). Conversely, the years less affected by drying are simulated in 1994, 2001 and 2014, with an average RPoD below 15 % throughout the year (black curve, Fig. 12a and b).
Results obtained with the LLR model are more contrasted in terms of extreme values than those obtained with the LR model (Fig. 12b).
The analysis of the ONDE observations shows that the proportion of rivers undergoing drying is significantly higher (35 %) than that observed with the conventional monitoring (HYDRO database, 8 %). This proportion, although related to a short period of records 2012 and 2016, is consistent with the percentage of 39 % of river segments classified as intermittent by Snelder et al. (2013). This analysis confirms the under-representation of IRES in the French HYDRO database, and probably others in other countries (flows are often uncontrolled in IRES). Without gauging stations located on headwaters, Snelder et al. (2013) were unable to predict IRES in eastern France (see Fig. 9, p. 2694). The high density of ONDE sites makes it possible to improve the detection of drying events and lead to better understand the spatial distribution of IRES located at the upstream extent of the hydrographic network. The ONDE network encompasses various hydrological conditions, which provides a more accurate assessment of inter-annual variability, differentiating between dry years (2012, 2015 and 2016) and wet years (2013, 2014) with clearly few drying occurrences.
The validation of the LR and LLR models against the spatially dense POC database also demonstrates the spatial representativeness of the ONDE network. Thanks to the qualitative information provided and to models such as the statistical models developed here, it is now possible to capture drying events at the regional scale.
The ONDE sites are located on small headwater streams which can be very reactive to external disturbances (rainfall deficit, change in air temperature, increase in water withdrawals, etc.) and by nature are more likely to be IRES. The gauging stations available in the HYDRO database are located on larger streams and their hydrologic response to changes in external factors (environmental or human) is slower and drying occurred with greater inertia under temperate climate. Their uneven distribution across France does not allow us to accurately characterise the inter-annual variability of drying development. Overall, the ONDE network provides very complementary information to conventional flow monitoring, leading to a better understanding of the processes of drying in upstream catchments.
The performance obtained with the LR and LLR models is slightly better with
the 2011–2017 dataset (mean NSE
Comparison of NSE obtained with regression including only the
discharge variable as a function of NSE obtained with discharge and
groundwater-level variables in the 2011–2017 dataset with
We have chosen to average the non-exceedance frequencies of flows and groundwater levels in order to increase the monitoring network. If models had been calibrated using only gauging stations, performance will have been globally similar, or slightly better, in some HER2–HR combinations (Fig. 13). Therefore, we could not validate the real gain of using groundwater-level data in addition to discharge data. This is certainly due to the dominant proportion of the gauging stations compared to the piezometers. Indeed, in the 2011–2017 dataset, the proportion of gauging stations is greater than 75 % for more than 70 % of HER2–HR combinations, whereas the proportion of piezometers exceeds 70 % in only 5 % of HER2–HR combinations. Groundwater-level data thus have a small weight in regressions for this dataset. However, in the 1989–2017 dataset, the proportion of piezometers is greater than 70 % in more than 30 % of HER2–HR combinations. The presence of piezometers increases the density of the monitoring network in HER2–HR combinations with few available gauging stations. Thanks to groundwater-level data, RPoD can be predicted on more HER2–HR combinations.
Spatio-temporal simulation of the probability of drying is crucial for
advancing our understanding of IRES ecology and management. Some aquatic
species can persist in a dry reach for a few days, weeks or months, while
some are highly sensitive to desiccation (Datry, 2012; Storey and Quinn,
2013; Stubbington and Datry, 2013). Estimation of the total duration of days
with drying at the reach scale is therefore needed to understand biological
patterns in river networks (Kelso and Entrekin, 2018). To our knowledge, no
study has proposed to reconstruct daily flow-state time series of headwater
streams at the country scale such as France (> 500 000 km
Regional probability of drying simulated with
This study provides a first regional approach to use discrete data obtained
from regular observations. The average non-exceedance frequency is a global
hydrological statistic that only captures the hydrological conditions at the
regional scale in modelling the RPoD. For rainfall-driven river flow regimes,
the effect of rainfall events on flow intermittence at the HER2–HR scale is
probably indirectly reflected by the daily discharge and groundwater levels
used to calculate the average non-exceedance frequency. However, when more
observation data are available, it is likely that including more detailed
descriptors of rainfall events and local geology could improve our approach.
In France, based on the 2011–2017 dataset, both models suggest the highest
values of RPoD along the Mediterranean coast
(D
The second application aimed at reconstructing historical RPoD over the period 1989–2016. Both models suggest the highest values of mean RPoD (> 35 %) in 1989–1991, 2003 and 2005. During these dry years, predicted values of RPoD result from extrapolation but are consistent with published studies (Mérillon and Chaperon, 1990; Moreau, 2004). For example, Mérillon (1992) estimated that for the whole of France, 11 000 km of rivers were dried at the end of the summers of 1989 and 1990. Caillouet et al. (2017) found that the low-flow event observed in 1989–1990 was particularly severe in terms of duration and affected 95 % of France. Snelder et al. (2013) showed from 628 gauging stations that the years 1989–1991, 2003 and 2005 witnessed particularly high values of duration and frequency of drying events. They found that regions with the highest probability of drying were located along the Mediterranean and Atlantic coasts, which is consistent with ONDE observations and with our results.
Both models suggest the same sequence of dry and wet years. However, the application of the LLR model leads to less contrasted RPoD than the LR model (Fig. 12).
To illustrate these differences, the RPoD has been simulated by both models
with an extreme
This paper investigates the spatial and temporal dynamics of the regional probability of drying (RPoD) of headwater streams by taking benefit from qualitative and discontinuous data provided by the ONDE network. Two models based on linear or logistic regressions have been developed and succeeded in reconstructing the temporal dynamics of RPoD. They are based on a strong relationship between the non-exceedance frequencies of discharges and groundwater levels as a function of the proportion of drying statuses observed at ONDE sites per HER2–HR combination. LLR and LR models show similar performance and perform well between 2011 and 2017. The accuracy of predictions is dependent on the number of gauging stations, ONDE sites and piezometers available to calibrate the regressions. Regions with the highest performance are located in the sedimentary plains, where the monitoring network is dense and where the RPoD is the highest. Conversely, the worst performances are obtained in the mountainous regions. Finally, both models have been used to reconstruct historical RPoD between 1989 and 2016 and suggest the highest values of mean RPoD (> 35 %) in 1989–1991, 2003 and 2005. This is consistent with other published studies, but the high density of ONDE sites makes it possible to improve the detection of drying events and lead to better capturing of the spatial distribution of IRES located at the upstream extent of the hydrographic network. Moreover, the duration of drying is of importance for ecologists and the prediction of a daily RPoD provides one key driver for the composition and persistence of aquatic species.
From a methodological point of view, our method relating discrete drying observation obtained by citizen science networks to continuous daily gauging data seems robust across the highly diverse (climate and topography) regions of France, and provides good predictions in an independent region excluded from the calibration process (PoC). These two results suggest a potential application of our approach in other countries. Citizen science creates opportunities to overcome the lack of hydrological data, contributes to densifying the flow-state observation network (Turner and Richter, 2011; Buytaert et al., 2014) and remains less expensive than the installation of additional gauging stations to survey flow intermittence. The next step will be to use this regional approach to simulate the RPoD in future periods by taking into account effects of climate change through predicted discharge and groundwater-level data. This would allow quantification of the evolution of the probability of drying between the current period and the different climate projections provided by the latest IPCC Report (IPCC 2014) and would assist decision makers in defining protocols for restoring flows with appropriate measures to preserve aquatic ecosystems (Woelfle-Erskine, 2017).
Secondly, further work is needed to develop an approach capable of reconstructing the drying dynamics locally by differentiating each stream. Our approach remains spatially valid for estimating RPoDs at the scale of HER2–HR combinations, but does not allow characterisation of the variability of drying occurrence between nearby streams within these regions. From a methodological point of view, statistical tools such as neural networks (Breiman, 2001) have shown good ability to assess both the occurrence and extent of perennial and temporary segments (González-Ferreras and Barquín, 2017) and could be investigated as an alternative method to reconstruct locally the temporal variability of drying.
Data used in this study are freely available on the
Eaufrance data portal (
AB, NL and ES developed the main ideas and designed the experiments. AB implemented the algorithms and analysed the results. AB prepared the manuscript with contributions from all co-authors.
The authors declare that they have no conflict of interest.
The authors wish to thank Anne van Loon and Catherine Sefton for their
valuable comments, suggestions and positive feedback on the manuscript. The
research project was partly funded by the French Agency for Biodiversity
(AFB, formerly ONEMA). This study is based upon works from COST Action
CA15113 (SMIRES, Science and Management of Intermittent Rivers and Ephemeral
Streams,