Water restrictions under climate change : a Rhone-Mediterranean perspective combining ‘ bottom up ’ and ‘ top-down ’ approaches ”

Drought management plans (DMPs) require an overview of future climate conditions for ensuring long 13 term relevance of existing decision-making processes. To that end, impact studies are expected to best reproduce 14 decision-making needs linked with catchment intrinsic sensitivity to climate change. The objective of this study is 15 to apply a risk-based approach through sensitivity, exposure and performance assessments to identify where and 16 when, due to climate change, access to surface water constrained by legally-binding water restrictions may 17 question agricultural activities. After inspection of legally-binding water restrictions (WR) from the DMPs in the 18 Rhône-Méditerranée (RM) district, a framework to derive WR durations was developed based on harmonized low19 flow indicators. Whilst the framework could not perfectly reproduce all WR ordered by state services, as deviations 20 from socio-political factors could not be included, it enabled to identify most WRs under current baseline, and to 21 quantify the sensitivity of WR duration to a wide range of perturbed climates for 106 catchments. Four classes of 22 responses were found across the RM district. The information provided by the national system of compensation to 23 farmers during the 2011 drought was used to define a critical threshold of acceptable WR, related to the current 24 activities over the RM district. The study finally concluded that catchments in mountainous areas, highly sensitive 25 to temperature changes, are also the most predisposed to future restrictions under projected climate changes 26 considering current DMPs, whilst catchments around the Mediterranean Sea, were found mainly sensitive to 27 precipitation changes and irrigation use was less vulnerable to projected climatic changes. The tools developed 28

2 the text) and with daily discharges simulated by the rainfall-runoff model GR6J forced by the SAFRAN reanalysis (named "GR6J" in the text). In the context of climate change the WRL model is run with daily discharges obtained with GR6J forced by one of the 1350 sets of perturbed precipitation, temperature and PET time series. In this later case the regulatory thresholds are calculated on the simulated discharge time series to limit the possible effect of bias in rainfall-runoff modeling. 5 Third, lack of in-depth discussion on the policy implementation. Given that authors use WR in the climate response surface, one can expect that authors should use a lot of space to link their results to drought policy implementation or some information about the adaptation action. However, only a short discussion of WR has been provided at Section 5.5. Given that this is not a methodological paper, these in-depth discussions become the critical point to prove that this paper is worth 10 to be published because readers around the world can learn from this study and apply it to their own drought management policy.
 Discussing the policy implementation is out of the scope of this article. This paper presents a first attempt to simulate water restrictions over a large area in France. This paper aims at promoting the approaches developed in parallel by Brown (named 'Decision Tree Framework") and Prudhomme (named "Scenario neutral approach") and one of the challenges was to 15 define critical thresholds of unacceptable number of days with legally-binding WR for irrigation use. This paper suggests using information provided by insurance (here from a national system of compensation) at the regional scale.
Finally, the structure of the manuscript and English is extremely difficult for readers to follow. The general outline of the paper follows a typical modeling paper while authors introduce their study area and data than their model. However, as I 20 mentioned above, the modeling framework especially for the "Water restriction level modeling framework" is not clear at all. Also, there are general equations list in the results section (Section 5) and irrelevant results (Line 432-474) presented in the result section.
 The scenario-neutral approach is applied at both local and regional scales and results discussed in Sect. 5 before drawing general conclusions in Sect. 6. 25 There are A LOT of grammar errors and typos that make the manuscript hard to read. This is surprising that one of the coauthors is from the UK.
 The text has be screened to correct grammar errors and typos.

Water restrictions under climate change: a Rhone-Mediterranean perspective combining 'bottom up' and 'top-down' approaches"
Sauquet et al.

Anonymous Referee #2
Sauquet and colleagues applied a scenario neutral approach to evaluate the implementation of water use restrictions and their 5 impacts on irrigated agriculture. They applied this approach to 15 catchments in the Rhone-Mediterranean region with minimal human influence. Their methods included calibration of a hydrological model to each catchment, sensitivity analyses, assessment of exposure and clustering to identify basins with common characteristics. Strengths of this work include comparison of results regionally and identification of catchment classes, as well as high quality graphics presenting the results. Areas to for improvement include problem framing, the implementation and communication of the sustainability 10 assessment, and explanation of the clustering process and its value. With a clearer problem framing and improved sustainability assessment I believe the scientific and practical contributions of this work would be clearer.
 Authors agree with this remark and the section explaining the method has been rewritten.
The topic is of interest to HESS readers, and subject to major revision I believe that it would be suitable for publication.

15
Comments 1. The authors make a strong case for why we care about drought risk under climate change. However, the case for why we need to simulate the implementation of water use restrictions should be stronger. The main question I would like to see the authors address here is: how does the simulation of water use restrictions give us a different picture of impacts or ways to mitigate impacts than simulating streamflow alone? 20  Water restrictions simulations complement studies on the impact of climate change on water resources availability and on water use needs. Indeed water needs can only be met first if water resources are available and second if water abstractions are allowed. Regulatory rules are pieces of the puzzle that should be examined. Roughly speaking studying water restrictions is a way to identify additional future constraints on water users. The regulatory aspects have never been deeply examined in France, perhaps due to the recent implementation of DMPs. 25 2. The authors thoroughly review the literature in the scenario neutral and decision scaling methods for assessing climate vulnerability in a bottom-up manner. However, the literature on robust decision making is complementary and should be included in this review. Specifically, there are a few robust decision making studies that assess the performance of existing 2 water management plans [e.g. Lempert and Groves, 2010;Bloom et al., 2013]. The authors should note how their work builds upon or goes beyond these prior works.
 Many thanks. We have added one of the references in the conclusion to make links with RDM.
3. The sustainability assessment is the key link between the occurrence of water use restrictions and impacts. The authors use 5 critical thresholds as a way to measure sustainability. First, I'm not convinced that is a measure of sustainability. Is it serving as a measure of the sustainability of an agricultural economy? Or something else? Please clarify how it meets a reasonable definition of sustainability.
 Sustainability -like vulnerability -has no universal definition. Sustainability assessment is based on the analysis of failures or unacceptable conditions that lead to low crop yield and quality, and consequently to economic losses at such a level that the 10 national system of compensation is initiated. In this application, -we assumed that irrigated farming is not sustainable if restrictions during drought periods are, on average, too severe -i.e. duration with limited or suspended abstraction for irrigation above a critical threshold -to ensure enough water for crops; -since it was not possible to compute the effect of water restrictions on crop yield and quality (no crop modelling was considered here) and on economic losses, we used 'agricultural disaster' notifications as proxies to identify the conditions that 15 would be unacceptable/damaging for farmers activities.
This sustainability is thus indirectly related to agricultural economy (not directly related to losses expressed in euros). We have changed sustainability for failure analysis to clarify the text.
Second, it is not clear how this critical threshold was defined. The authors state that a single critical threshold is applied to all 20 catchments. Is this reasonable given the substantial differences in elevation (and therefore temperatures)? And is the local precipitation factored into this threshold?  Data are collected by the French ministry of agriculture and they are confidential. The year 2011 was the only year when the national system of compensation has been activated with available data between 1958 and 2013 and the duration of water restrictions were derived individually for each catchment and converted in anomalies WR*(2011) with respect to the 25 benchmark value (mean over the period 1958-2013). This dispersion is due to heterogeneity in crops, in irrigation systems, in climate (precipitation, PET, temperature)… at the regional scale leading to locally differentiated sensitivity to water restrictions as well as to biases in WR modelling. Since only the year 2011 it is difficult to conclude on the origin of the dispersion (natural or non-natural). We are convinced that this information is valuable. Finally, simplifying but realistic assumptions are imposed by the lack of detail information; thus only one value was considered despite high dispersion in 30 WR*(2011) values (Table 6): the critical threshold was set to the average WR*(2011) computed on all catchments of the region under agricultural disaster status in 2011 (6.6 10-day periods), and was used for all classes. Note that this value seems realistic: 6.6 10-day periods = 66 days with restrictions = 30% of the time between between the 1 st April and the 31 st October.
3 Lastly, do irrigators or other water users in these catchments have access to other water sources to mitigate impacts (e.g. farm ponds, groundwater)? If so, how does that influence the conclusions?
 More details are given in Section 2.1. In France 80% and 20% of water abstraction are taken from surface water and from groundwater, respectively. In the RM district 10% of water used for irrigation originate from groundwater. Irrigators may 5 have access to small reservoirs (storage capacity usually < 1 Mm 3 ). There is actually a wide discussion about these hydraulic  The discharges simulated by GR6J introduced in the WRL model lead to higher Sensitivity scores than those obtained with observed discharges extracted in the HYDRO database. The reasons for this unexpected result have been investigated.
In particular we have compared the observed and simulated temporal variability in the time series VCN3. A "smoothing" effect in the GR6J simulations compared to observations was initially suspected. Finally no obvious difference in autocorrelation functions was found between observed and simulated time series. One reason could that the period of interest 30 2005-2013with for some basins only three years with stated water restrictionsmay be too short to analyse accurately the relative performance of WRL obtained with OBS and with HYDRO, respectively.

4
The two scores gives a global insight on the performance of the WRL modelling framework and too much weight should not been given to the differences between scores. In this case, we should conclude that the developed WRL modelling framework leads to similar results (moderate performance in detecting stated water restrictions during the period [2005][2006][2007][2008][2009][2010][2011][2012][2013] with both data sources HYDRO and GR6J. The WRL modelling framework provides an overview of the on-going drought and the drought committees are partly free to account for this information to state or to postpone water restrictions. The 5 developed framework is a useful tool to predict water restrictions with no interference of lobbies, i.e. only based on the physical processes.
6. The authors state that the CART analysis can aid sensitivity assessment at unmodelled catchments. Please address in the conclusions if and how this classification can be helpful for water managers or other scientists. 10  The CART algorithm creates the best homogeneous group when splitting the data using through a set of "if-then" logical conditions applied to the most relevant factors, i.e. the decision variables. The result is a decision tree with nodes separating the data into two subgroups. The decision variables known at unmodelled but gauged catchments can be introduced in the chain of rules obtained by CART to finally predictin this applicationthe assignment to one of the four classes.

Anonymous Referee #3
The objective of the study is to develop a risk-based framework to simulate water restrictions (WRs) under climate change in 5 Rhone-Mediterranean district in order to evaluate the vulnerability of current Drought Management Plans (DMPs) to future climate conditions. The proposed framework is based on the assessment of three components: sensitivity of WRs to changes in different climate factors, sustainability of WRs for users and exposure in terms of climate response surfaces. General comments The paper presents an interesting topic. Although the applied methodology seems appropriate to some extent, it is rather unclear in some parts. Overall, I believe that further details should be added to the paper in order to support the 10 interpretations and conclusions drawn from the analyses carried out by the authors.
 Authors agree with this remark and the section explaining the method has been rewritten.

Major comments
Section 3 For the sake of better understanding, I suggest to report the equations of low flow indicators and regulatory 15 thresholds used in the manuscript.
 Changes have been made in Section 3 to better define the variables of interest: -"The low-flow monitoring indicators usually considered are: the daily discharge Qdaily, the d-day maximum discharge QCd, QCd(t) =max(Qdaily(t'),t'∈[t-d+1,t]) and the d-day mean discharge VCd, , with duration d associated with WR decision varying between 2 and 10 days depending on DMPs." 20 -"The threshold associated with WR also varies, generally associated with statistics derived from low-flow frequency analysis, but some being fixed to locally-defined ecological requirements. In the context of DMPs, series of minimum QCd or VCd are calculated by the block minima approach and thereafter fitted to the lognormal distribution. The block is not the year but the month or given by the division of the year into 10-day time-window. The regulatory thresholds are given by quantiles with four different recurrence intervals associated to the four restriction levels. For example, let 25 us consider thresholds based on the annual monthly minima of VCNd. The block minima approach is carried out on the N years of records for each month i, i=1…,12 leading to twelves datasets {min{VCNd(t), month(t)=i, year(t)=j}, 2 j=1,…,N}. The twelve fitted distribution allows the calculation of 48 values of thresholds (=12 months × 4 levels) with four T-year recurrence intervals. To enable comparison of results across all catchments, the same definitions for the monitoring variables and the regulatory thresholds have been adopted for all the catchments. VC3 was selected as the monitoring indicator and the regulatory thresholds are low flow quantiles 10d-VCN3 based on the minimum 3-day mean discharges extracted by the block minima approach considering the fixed 10-day time-windows spanning the year as 5 blocks with return periods, as they are the most common single indicators used in the 28 DMPs of the RM district.
Lastly return periods T of 2, 5, 10 and 20 years will be associated with the "vigilance", "alert", "reinforced alert" and "crisis" restriction levels, respectively, due to their prevalence in the DMPs" Section 4.2 Details on the rainfall-runoff model should be added, with special reference to the way how the influence of reservoirs is taken into account. 10  The GR6J model has six parameters to be fitted (see Figure below): the capacity of soil moisture reservoir (X1) and of the routing reservoir (X3), the time base of a unit hydrograph (X4), two parameters of the groundwater exchange function F (X2 and X5) and a coefficient for emptying exponential store (X6). GR6J is combined with the daily snow module Cemaneige.
The catchment is divided into five altitudinal bands of equal area on which snowmelt and snow accumulation processes are represented. For each band, daily meteorological inputsincluding solid fractions of precipitation -are extrapolated using 15 elevation as covariate and the snow routine is calculated separately. Finally, its outputs are then aggregated at the catchment scale to feed GR6J. The two parameters of Cemaneige are: the parameter controlling snowpack inertia (X1) and the degreeday coefficient controlling snowmelt (X2). No routine to simulate water management (e.g. reservoir) was considered here since discharges of the 106 gauging stations are weakly altered by human actions or naturalized discharges. Section 4.3 The description of the water restriction level modelling is unclear in some parts. For instance, I would expect that the comparison between simulated WRLs driven by GR6J data and historical WRs will provide a lower sensitivity score than the comparison with simulated WRLs driven by HYDRO data (considered as benchmark), but it's not (see Lines 287-290). 5 Could it be a consequence of the fact that the model disregards socio-political aspects of the decision making-process?
 Inputs of the WRL model are daily discharges and precipitation. Outputs are WRL for each of the 21 10-day periods defined between the 1 st April and the 31 st October. VC3(t) is first computed from daily discharge Q(t) every day t, WRL(t) is then deduced by comparing VC3(t) to the four regulatory thresholds and finally a unique representative WR level is assigned to each of the 21 10-day periods, as the median of WRL(t) observed or simulated within that 10-day period. To best match 10 the whole monitoring process stated in most of the DMPs, a simple precipitation correction was applied ("Pcorr", in Fig. 5).
It consists to give a 'no alert' when precipitation during the preceding 10 days exceeds 70% of inter-annual precipitation average, regardless of the WR simulation results. The WRL framework is applied to observed and simulated data of both discharge and precipitation. To assess the performance of the WRL model under current condition against stated WR decisions, the WRL model is run with observed daily discharges extracted from the HYDRO database (named "HYDRO" in 15 the text) and with daily discharges simulated by the rainfall-runoff model GR6J forced by the SAFRAN reanalysis (named "GR6J" in the text). In the context of climate change the WRL model is run with daily discharges obtained with GR6J forced by one of the 1350 sets of perturbed precipitation, temperature and PET time series. In this later case the regulatory thresholds are calculated on the simulated discharge time series to limit the possible effect of bias in rainfall-runoff modeling. 20  The discharges simulated by GR6J introduced in the WRL model lead to higher Sensitivity scores than those obtained with observed discharges extracted in the HYDRO database. The reasons for this unexpected result have been investigated.
In particular we have compared the observed and simulated temporal variability in the time series VCN3. A "smoothing" effect in the GR6J simulations compared to observations was initially suspected. Finally no obvious difference in autocorrelation functions was found between observed and simulated time series. One reason could that the period of interest 5 2005-2013with for some basins only three years with stated water restrictionsmay be too short to analyse accurately the relative performance of WRL obtained with OBS and with HYDRO, respectively.
The two scores gives a global insight on the performance of the WRL modelling framework and too much weight should not been given to the differences between scores. The developed WRL modelling framework leads to similar results (moderate 10 performance in detecting stated legally-binding water restrictions during the period 2005-2013) with both data sources HYDRO and GR6J. The WRL modelling framework provides an overview of the on-going drought and the drought committees are partly free to account for this information, i.e. to state or to postpone water restrictions. The developed framework is a tool to predict water restrictions with no interference of lobbies, i.e. only based on the physical processes.

15
These sentences must be better explained:  lines 295-296-"Furthermore, GR6J performance under low-flow conditions show no statistical link with its WRL modelling performance"  Furthermore, there is no significant link between the GR6J efficiency in simulating low flows (NSE LOG ) and the performance of the WRL (Sensitivity and Specificity scores), since the determination coefficients between NSE LOG and Sensitivity, and between NSE LOG and Sensitivity are lower than 7%. 20  lines 300-301 "possible biases in rainfall-runoff modeling does not affect much the ability of the WR modeling framework to simulate correctly declared or not declared WRs" It seems that despite the difficulties of GR6J model in simulating low-flows accurately, the results of WRL modelling driven by GR6J data are good anyway. How do the authors explain that?  The WRL framework is applied to observed and simulated discharge data available before 31 st December 2013. In this later case the regulatory thresholds are calculated on the simulated discharge 25 time series to limit the possible effect of bias in rainfall-runoff modelling. The possible reasons of comparable performance between GR6J and OBS is that the WRL framework is carried out using regulatory thresholds derived from GR6J outputs and that even if the discharge data are not exactly reproduced by GR6J, their ranking and their relative position to the regulatory thresholds is correctly reproduced.

30
Section 5.3 Vulnerability is computed against a critical threshold. The latter is defined as the difference between the number of WRs simulated by the WR GR6J modelling framework for 2011 and over the baseline period. On the other hand, the Vulnerability Index is computed as the proportion (frequency) of RCM-based simulations that fail above the critical threshold. It sounds like a frequency is compared to a number. I believe that this step must be described in details.
 Indeed there are two measures of vulnerability. Given one specific climate change projection, a catchment could be judged vulnerable if on average the critical threshold is exceeded. The Vulnerability Index is a proportion reflecting the fraction of RCM leading to critical situations on average. This index is introduced here to account for the uncertainty in climate projections in vulnerability assessment. It should be interpreted as conditional probability (risk) with respect to a set of possible future climates and only used as a relative measure to rank the regions, from the less to the most likely impacted 5 regions.
For the same reason, it is not clear how the black dotted lines representing the critical threshold are drawn in Figures 10 and   14.
 The dotted black lines are isopleths connecting points of the response surface with WR*= WR*(2011). Their location 10 in the response surface depends on the shape of the response surface; this is why the dotted lines differ from one catchment to another in Figure 10, and later from one class to another in Figure 14.
Section 5.4 With regard to the hierarchical cluster analysis for catchment classification at regional scale, the authors should specify the catchment characteristics considered to investigate similarity through the Euclidean distance (see line 421-424). 15 Details on the CART model and its implementation should be added.
 CART methods perform successive binary splittings of a given dataset according to decision variables. The algorithm identifies automatically through a set of "if-then" logical conditions the best possible predictors, starting from the most discriminating decision variable to the less important factors, to predict the membership to the one of the four groups. The optimal choices are fixed recursively by increasing the homogeneity within the two resulting clusters. At each step one of the 20 clusters (node) is divided into two nonoverlapping parts.
The list of the potential decision variables by type is: We have included this list in a table.
In lines 273-274 "OBS WRLs are correctly reproduced by both GR6J and HYDRO simulations, but also can be consistent with OBS" (???). This sentence is rather misleading, I wonder if "OBS" at the beginning of the sentence could be a mistake 30 and could be deleted. 7  There is a problem with the phrasing on these lines. "Both GR6J and HYDRO simulations are globally consistent with observed WRLs (OBS). However GR6J and HYDRO results may differ from OBS (e.g. basins 9 to 11 in the Lozère department during the year 2005)." In line 420: ". . . a classification (of what?) was conducted on to define typical response surfaces, . . .". Please specify.
 A classification of the 106 gauging stations based on the 1350 values of WR* was conducted on to define typical 5 response surfaces.
In line 482: "come catchment" to be replaced by "some catchments".
 "are found for come some catchments".
In line 540: replace "precipitations" with "precipitation". 10  "more accurately temperature and/or precipitations" of Zurich, Department of Geography. The students were given the task to select a manuscript in review at one of the EGU journals and to write a review. I discussed this review with the student, and find the comments actually quite valuable.
Therefore, I post the review here in the hope editor and authors will find them useful to improve the manuscript. DMPs in the future will still be effective. The study aims to assess the effectiveness of current DMPs under climate change to be able to revise the DMPs for the most vulnerable basins. They find out that in temperature-sensitive catchments the water restrictions will increase significantly in the short term and that for this reason there is a need to adapt the DMPs. In the catchments where the precipitation determines the water restriction, they see difficulties to adapt the DMPS as the uncertainties in precipitation is high. They state in the conclusion section several points they did not include in their study 20 but could play an additional role besides the analyses of water restriction duration influenced by temperature and precipitation. These are for example socio-economic system stressors like agricultural practices, population growth, water demand, etc. which also should be considered in the DMPs. In my opinion, it is an important topic to discuss the reliability of current decision-making rules regarding water scarcity in the future when climate changes. The method used in this study can give a good overview of where there is a need to rethink the DMPs. But in my opinion, it would be quite important to 25 take the socio-political factors into account in the framework to reproduce water restrictions. A further improvement would be if the economic system stressors would be included to evaluate the DMPs. Therefore the current method has still a lot to improve, and that's why it is not fully clear what the substantial contribution of this paper is.
Further, I think the description of the method of the hydrological modeling and the framework to reproduce the water restrictions could be more detailed.
 Authors agree with this remark and the method needs to be more explained. Water restrictions simulations complement studies on the impact of climate change on water resources availability and on water use needs. Indeed water needs can only be met first if water resources are available and second if water abstractions are allowed. Regulatory rules are pieces of the 5 puzzle that should be examined. Roughly speaking studying water restrictions is a way to identify additional future constraints on water users. The regulatory aspects have never been deeply examined in France, perhaps due to the recent implementation of DMPs. This paper presents a first attempt to simulate water restrictions over a large area in France. This paper aims at promoting the approaches developed in parallel by Brown (named 'Decision Tree Framework") and Prudhomme (named "Scenario neutral approach") and one of the challenges was to define critical thresholds of unacceptable 10 number of days with legally-binding WR for irrigation use. This paper suggests using information provided by insurance (here from a national system of compensation) at the regional scale.  Climate response surface of WR* legally-binding water restrictions level anomalies WR* is a graphic representation summarizing the sensitivity of WR* to climatic drivers. They all suggest an increase in the occurrence of legally-binding water restrictions when precipitation decreases or when temperature increases. Additional temperature increase and its associated PET increase can compensate for precipitation increase and lead to decrease in WR*. The response surfaces 20 differ by their flatness (e.g. the response surface of Class 3 displays the less contrasted shape).
P2-L54: Is the scenario-neutral approach the same as a bottom-up approach? The authors could use the word "bottom up" as well, as they use it also in the title and it is not used in the rest of the paper. Please clarify difference or similarity. The sentence "specifying relevant critical thresholds is the main task involved in bottom-up approaches" was added in 30 section 4.1.
P4-L106 to P5-L120: In section "2.3 Hydrological data" it would be good if the 15 regimes suggested by Sauquet et al.

3
 The classification could be given in Appendix. Sauquet et al. (2008) have defined a classification based on the mean monthly runoff pattern and a map has been published showing the assignment to one class for each basin with drainage area > 50 km².
Groups 1 to 6 are pluvial river flow regimes. The groups mainly differ by the contrast between the maximum and the minimum of monthly streamflow. Nearly uniform flows through most of the year (Group 1) are found where large aquifers 5 moderate flows whereas Group 6 is characterized by very low flow in summer, reflecting the lack of deep groundwater storages in the catchment. Group 7 is representative of Mediterranean river flow regimes where small rivers basins experience hot and dry summers and intense rainy events in autumn. Their runoff pattern therefore exhibits severe low flow in summer and high flow in November. In mountainous areas, uppermost basins display snowmelt-fed regimes (Groups 10, 11 and 12). The lower the outlet is, the lower the contributions of snowmelt to runoff. Groups 8 to 9 are in the transition 10 regime. The seasonal variation of streamflow is affected as much by precipitation timing as by air temperature and topographic influences (on snowpack formation and snowmelt timing). Typically, high flows are observed in spring.

Reference hydrographs representative of the classification of river flow regime for France (after Sauquet et al., 2008)
P5-L121 to P5-L126: In section "2.4 Climate data" Table 2  simulations. This study was used to define the spectrum of changes in temperature and precipitation. Here regional climate projections available in the DRIAS portal are used.  We have modified the paragraph to improve the presentation of the WRL modelling framework: "Water restrictions are 5 decided after consulting drought committees that convene irregularly. The time-step for modelling WRL was chosen to be compatible with the frequency of drought committees estimated from the analysis of the water restriction orders: WRL is thus computed at a regular time step of ten days. VC3(t) is first computed from daily discharge Q(t) every day t, WRL(t) is then deduced by comparing VC3(t) to the four regulatory thresholds and finally a unique representative WR level is assigned to each of the 21 10-day periods defined between the 1 st April and the 31 st October, as the median of WRL(t) observed or 10 simulated within that 10-day period." P7-L173: VC3 was selected, as it is the most common single indicators used in DMPs of the RM district. I might have missed something, but this seems not to be the case for the 15 test catchments chosen for the evaluation of the WR modeling framework. It is not clear for me how you can compare these different low-flow monitoring indicators with each other. This 15 should be described clearer.
 Indeed the decision that lead to selecting VC3 as monitoring variable is was made considering the 28 DMPs and this modality is not prevalent within the 15 test catchments (Figure 3). We have made it clearer in the final version. VC3 was selected as the monitoring indicator and the regulatory thresholds are low flow quantiles 10d-VCN3 based on the minimum 3-day mean discharges extracted by the block minima approach considering the 37 fixed 10-day time-windows as blocks 20 with return periods, as they are the most common single indicators used in the 28 DMPs of the RM district.
P9-L244: Are the 15 catchments used for calibration or only for evaluation? Please clarify.
 They were used both for calibration and for evaluation.

25
P13-L343: It is not clear for me if the perturbation of the climate is based on different climate scenarios as RCP2.6, RCP4.5, RCP8.5 or which exact projection is used. In the reference Terray and Boé, 2013 the authors are using they are also talking of different projections. This needs to be clarified.
 The "delta-change" method was used to provide a set of perturbed climates in scenario-neutral approach. Following ]. (1) ].
(2) 5 P 0 and T 0 + A T are respectively the mean annual changes in equations (1) and (2), with i referring to month 1 to 12, P the phase parameter and A p the semi-amplitude of change (e.g. half the difference between highest and lowest values) in equation (1). The parameters P 0 , P , 0 and T of single-phase harmonic function were fixed with respect to the range of changes suggested by Terray and Boé (2013). Finally 45 precipitation scenarios were created using 9 values of P 0 i.e. [-20; -13.3; -6.6; 0; 6.6; 13.3; 20; 26.6; 33.3] mm.an -1 , by 5 values of Ap i.e. [0; 6.6; 13.3; 20; 26.6] mm.season -1 , while 5 P parameter is fixed to 1 to consider minimum change in January and maximum change in July. Likewise, 30 temperature scenarios were set up with 6 values of 0 i.e. [0; 1; 2; 3; 4; 5]°C.an -1 by 5 values of A T i.e. [-0.5; 0.5; 1.5; 2.5; 3.5]°C.season -1 while T is fixed to 2°C to get maximum change in August. These details are now given in the new version.
P19-L513 to P19-L518: The first two sentences of the conclusion would better fit in the introduction.  The total net water withdrawal is around 6 billion of m 3 in the period 2008-2013 (water abstraction for cooling nuclear plants and hydropower is excluded) with a high proportion of them to support irrigation needs (3.4 billion of m 3 , including 2 billion of m 3 for channel conveyance). Only 10% of water abstracted for irrigation originate from groundwater. Total annual 25 abstracted volumes for drinking water and for water for industrial uses represent 1.6 and 1 billion of m 3 , respectively. P4-L109: I do not understand what the authors mean with "Time series including null values or gaps in the data records above 30% of time were disregarded". Does this mean one null value or 30% null values? Please clarify.
 "Time series with more than 30% of missing values or more than 30% of zero flows were disregarded." 30 P16-L426: In Table 5 in the table description please add where this standard deviation Sd is taken from.
 Table 5 is now referred in Section 5.2.
 The authors would like to thank Jan Seibert and his students for their helpful comments.      Climate change impact studies are usually dedicated to water resources or water needs for the competing users.

86
There are also interests in examining regulatory instruments, such as Drought Management Plans, since these 87 plans state water restrictions imposed to non-priority uses during severe low flows, and climate change is likely 88 to affect water restrictions and modify the access of stakeholders to water resources.

89
The paper develops a framework to simulate legally-binding water restrictions (WR) under climate change in

131
Water management in the RM district is a long-standing issue. Reservoirs have been built to produce energy, 132 to sustain low-flows and to cope with drought effects. As an example, the Serre-Ponçon multi-purpose reservoir   Drought management plans (DMPs) define specific actions to be undertaken to enhance preparedness and 138 increase resilience to drought. In France DMPs include Past and operating regulatory frameworks to be applied 139 in case of drought, named in French "arrêtés cadres sécheresse", were inspected in the 28 departments of the RM 6 district. The past and operating DMPs and the water restriction orders were inspected in the 28 departments of 141 the RM district. They were obtained from:

188
This analysis shows that the implementation of the DMPs has evolved for many departments since 2003, e.g.,

189
with changes in the terminology and a national scale effort to standardize WR levels. Now severity in low-flows 190 is classified into four levels which are related to incentive or legally-binding water restrictions. These measures 191 affect recreational uses, vehicle washing, lawn watering and domestic, irrigation and industrial uses (Table 3).

192
Level 0 (named "vigilance") refers to incentive measures, such as awareness campaign to promote low water 193 consumption from public bodies and general public. Levels 1 to 3 are incrementally legally-binding restriction 8 levels; level 1 (named "alert") and 2 (named "reinforced alert") enforcing reductions in water abstraction for 195 agriculture uses, or several days a week of suspension; level 3 (named "crisis") involves a total suspension of 196 water abstraction for non-priority uses, including abstraction for agricultural uses and home gardening, and 197 authorizes only water abstraction for drinking water and sanitation services. Due to change in the naming of WR 198 levels since their creation one task was dedicated to restate the WR decisions (hereafter "OBS") since 2005 with 199 respect to the current classification into four WR levels.

200
For all catchments, a WR decision chronology was derived, showing a large spatial variability in WR (Fig. 1)

202
WR decisions were mainly adopted between April and October (98% of the WR decisions, Fig. 2), with 62% in

203
July or August, peaking in July.

204
Decisions for adopting, revoking or upgrading a WR measure are taken after consultation of "drought

219
The threshold associated with WR also varies within the district, generally associated with statistics derived 220 from low-flow frequency analysisthe minimum QCd observed or the minimum VCd observed with a T-year 221 recurrence interval (or QCNd(T) and VCNd(T) respectively), but also some being fixed to locally-defined 222 ecological requirements. Generally, return periods T of 2, 5, 10 and 20 years are associated with the 223 "vigilance", "alert", "reinforced alert" and "crisis" restriction levels, respectively.

224
In

231
The block minima approach is carried out on the N years of records for each month i, i=1…,12 leading to

234
The meteorological situation is also examined in terms of precipitation deficit and likelihood of significant

238
Where appropriate other supporting local observations such as groundwater levels, reservoir water levels,

261
The risk-based framework adopted contains three independent components (Fig. 4):

275
(iii) Exposure, as defined by state-of-the-art regional climate trajectories superimposed to the climate 276 response surface The exposure measures the probability of changes occurring for different lead times 277 based on available regional projections..

278
All the components of the framework together contribute to The intersection of all three components defines the 279 vulnerability of the system (including its management) to systematic climatic deviations.

280
The sensitivity analysis was conducted applying a water restriction modelling framework. Climate conditions 281 were generated applying incremental changes to historical data (precipitation and temperature) and introduced as 282 inputs in the developed models to derive occurrence and severity of water restriction under modified climates.

283
The tool chosen here to display the interactions between water restriction and the parameters that reflect the 284 climate changes is a two-dimensional response surface, with axes represented by the main climate drivers. This   threshold T c to be fixed for all the users. Facing this complexity, only the irrigation water use has will been 304 examined here, since it is the sector which consumes most water at the regional scale, with a critical threshold 305 defined for this single water use.

306
The last component of the risk-based framework is the eExposure to changes here. The exposure measures the 307 probability of changes occurring for different lead times based on available regional projections. It is assessed 308 measured using regional projections, visualized graphically by positioning the regional projections in the

391
To best match the whole monitoring process stated in most of the DMPs, a simple precipitation correction was 392 applied ("Pcorr", in Fig. 5). It consists to give a 'no alert' when precipitation during the preceding 10 days

398
In order to evaluate tThe WRL modelling framework was verified in the 15 evaluation catchments (
(2) 487 with P 0 and T 0 + A T mean annual changes in precipitation (1) and temperature (2)

504
A set of 45 precipitation and 30 temperature scenarios was created (Fig. 8), spanning the range of potential

530
Perturbed potential evapotranspiration PET* were derived from temperature data using the formula suggested by

561
"JASO", the latter coinciding with the highest temperatures) have been tested as candidates for the two axes.

562
The response surfaces are exemplified on three of the 15 evaluation catchments (Table 1, Fig. 9):

-
The sensitivity analysis of Water Restriction to climate perturbations is illustrated on three contrasting

-
The Roizonne River basin, in the Alps, typical of summer flow regime controlled by snowmelt, with 574 spring to summer climate conditions dominating changes in low-flows.

575
The response surfaces for three example catchments (Fig. 9)

587
-Sd values may vary significantly from one graph to another (Table 5)  method and Euclidian distance as similarity criteria (Ward 1963) was applied and four classes were identified 599 after inspection of the agglomeration schedule and silhouette plots (Rousseeuw 1987). A manual reclassification 600 was conducted for the few catchments with negative individual silhouette coefficients to ensure higher intra-601 class homogeneity. For each class, a mean response surface and associated Sd was computed, and main climate 602 drivers associated with WR changes identified (Table 5).

603
All suggest an increase in the occurrence of legally-binding water restrictions when precipitation decreases or 604 when temperature increases (Fig. 10). Additional temperature increase and its associated PET increase can 605 compensate for precipitation increase and lead to decrease in WR* with intra-class differences emerging in the

614
To further the regional analysis and help sensitivity assessment at un-modelled catchments, basin descriptors 615 were investigated as possible discriminators of the four classes. A set of potential discriminators -which 616 included measures of the severity, frequency, duration, timing and rate of change in low-flow events (Table 6) Fig. 11 shows the empirical distribution of the three main discriminators, 637 the mean timing  of daily discharge below Q95 and its dispersion r, based on circular statistics, where Q95 is 638 the 95 th quantile derived from the flow duration curve.

639
The classification discriminates catchments primarily on the seasonality of low-flow conditions and the aridity 640 index, with the extreme classes (1 and 4) being particularly well discriminated.

641
Geographically, Class 1 catchments are mainly located along the Mediterranean coast and include the Argens 642 River basin; WR* is mainly driven by changes in precipitation in spring and early summer. Class 1 gathers 643 water-limited basins with small values of AI and a weak sensitivity to climate change in summer. In these dry    The risk-based framework has been applied to the irrigation water use since annual net total water withdrawal 663 for agriculture purposes is ranked first at the regional scale. Note that in the Rhône-Méditerranée district around 664 90% and 10% of water used for irrigation originate from surface water and groundwater, respectively. To 665 complement water needs irrigators may also have access to small reservoirs (storage capacity usually less than 1 666 Mm 3 ). Most of the reservoirs are filled by surface water in winter and release water later in the following 25 summer. Water restrictions are not imposed to these reservoirs but it is assumed here that during severe drought 668 events the majority of them are empty and thus the existence of potential sources auxiliary to surface water on 669 the conclusions has limited influence on the conclusions.

670
We assumed here that irrigated farming is globally under failure if the duration with limited or suspended 671 abstraction is above a critical threshold T c that causes insufficient water for crops. The catchment or area i will 672 be considered more vulnerable than the catchment or area j if the likelihood of failure (i.e., exceeding T c ) for 673 catchment or area i is more than the likelihood of failure for catchment or area j. The critical threshold T c is a 674 value of total number of days with legally-binding water restrictions that needs to be fixed. To move closer to

683
Specifically the 'agricultural disaster' notifications are, issued to each affected department by the agriculture 684 ministry following recommendations from the Prefecture to each department affected by extreme hydro-685 meteorological events, and applied uniformly over the RM district. Whilst 'agricultural disaster' status is a 686 global index that may mask heterogeneity in crop losses within each department, and that reflects losses related 687 to both agricultural and hydrological droughts, it has the advantage of being directly related to economic impact, 688 and uniformly applied across the RM district, hence suitable for a regional-scale analysis. The national system of 689 compensation to farmers is initiated for areas notified under 'agricultural disaster' status.    method and Euclidian distance as similarity criteria (Ward 1963) was applied and four classes were identified 732 after inspection of the agglomeration schedule and silhouette plots (Rousseeuw 1987). A manual reclassification 733 was conducted for the few catchments with negative individual silhouette coefficients to ensure higher intra-734 class homogeneity. For each class, a mean response surface and associated Sd was computed, and main climate 735 drivers associated with WR changes identified (Table 5).

736
The analysis of the Water Restriction response surfaces from 106 catchments identified four classes of 737 catchments organized regionally (Fig. 11). Class 4 regroups snowmelt-fed river flow regimes in the Alps, whilst

757
The performance of the CART model is satisfactory with a misclassification rate of 18%, is parsimonious (five 758 nodes and three variables) and may help as a first guess to assess the sensitivity where discharge levels are used 759 to characterize current hydrological conditions and thereafter to state Water Restriction at the department scale.

760
The empirical distribution of each catchment descriptor is displayed (Fig. 12) for each class, along with that of 761 the mean timing  of daily discharge below Q95 and its dispersion r, based on circular statistics, where Q95 is 762 the 95 th quantile derived from the flow duration curve (see Prudhomme et al. 2015 for calculation details). For 763 the later, a particular representation equivalent to the classical boxplot was adopted.

764
The four classes discriminate well rivers primarily on the basis of the seasonality of low-flow conditions and 765 the aridity index, with the extreme classes (1 and 4) being particularly well discriminated.

766
Class 1 gathers water-limited basins with small values of AI and a weak sensitivity to climate change in 767 summer. In these dry water-limited basins, the mid-year period exhibits the minimal ratio P/PET and changes in      disaster status in 2011 (6.6 10-day periods), and was used as regional critical threshold applied to all 801 classes.Simplifying but realistic assumptions are imposed by the lack of detail information; thus only one value 802 was considered at the regional scale despite high dispersion in WR*(2011) values (Table 7)

807
The response surfaces of each class (Fig. 14) show water restrictions highly (Class 1) to weakly (Class 4) 808 sensitive to precipitation and weakly (Class 1) to highly (Class 4) sensitive to temperature, as suggested by the

-
The RM Water Agency has initiated an unprecedented major initiative that provides guidance for the River

871
The analysis of the past and current DMPs in the RM district shows a decision-making processes highly 872 heterogeneous both in terms of low-flow monitoring variable and regulatory thresholds. In reality, the WR 873 statements follow a set of rules defined in the DMPs (which can be simulated and reproduced automatically) but 874 also expert judgment or lobbying from key stakeholders -which are not accounted for in the WRL modelling 875 framework put in place here. However, the post-processing of GR6J outputs allows detecting more than 68% of 876 severe alerts (more severe than level 1), making the developed framework a useful tool. Our study is a first step 877 towards a comprehensive accounting of physical processes, but does not capture socio-economic factors, also 878 critically important and reaches out to interdisciplinary for completing the modelling framework designed here.

879
The study at the regional scale illustrates an expected difficulty to simulate accurately a regulatory framework.

880
Further improvement is not expected in enhancing hydrological models but in reproducing decision-making 881 processes. The overall performance could be improved by scrutinizing the minutes of the drought committees to 882 better understand the weight of the stakeholders in the final statement.

883
Synthetic scenarios were created from parametric variation of forcing data and integrated in a risk-based 884 framework to derive climate response surfaces showing water restrictions deviations.

885
Our resultsThe sensitivity analysis and the related response surfaces suggest that basins located in the Southern

886
Alps are the most vulnerable responsive basins to climate change and that those experiencing a high ratio P/PET 887 are found the less vulnerableresponsive. The classification method CART has been applied to 106 responses 888 surfaces associated with 106 gauged basins and leads to four classes with different sensitivity. The key-variables 889 known at un-modelled but gauged catchments can be introduced in the decision-tree to finally predict the 890 assignment as a first guess to one of the four classes. Water managers are thus encouraged to monitor in priority 891 and more accurately temperature and/or precipitation when and where the sensitivity of their catchments is found 892 the highest. This may mean efforts to reinforce field instrumentation within these key catchments, but also an 893 opportunity to implement awareness and participatory methods to initiate or to consolidate dialogues between 894 stakeholders from a long term perspective.

895
The impact of climate change on the river flow is expected to be gradual, thus offering opportunities to update,

896
to harmonize and to adapt Drought Management Plans to changes in climate conditions and water needs. Results

897
of our Water Restriction framework show that tAs a consequence, the sustainability need for adaptation of 898 existing drought action plans could differ much from one catchment to another and should take into account 899 intrinsic sensibility to climate change besides 'top-down' projections. Results also show needs to firstly adapt 900 DMPs in temperature sensitive catchments more subject to a significant increase in legally-binding restrictions in 901 the short term. In contrast, the capacity to anticipate new regulations will be challenging where water restrictions 902 are largely driven by precipitation. Regarding long-term relevance of DMPs, robustness of DMPs in these 903 catchments is not warranted given the large uncertainties in precipitation regional projections.

904
Water managers are thus incited to monitor in priority and more accurately temperature and/or precipitations 905 when and where the sensitivity of their catchments is found the highest. This may mean efforts to reinforce field 906 instrumentation within these key catchments, but also an opportunity to implement awareness and participatory 907 methods to initiate or to consolidate dialogues between stakeholders from a long term perspective.

908
The study at the RM scale illustrates the difficulty to simulate accurately a regulatory framework. The overall 909 performance of the WR modelling framework under current conditions is found satisfactory with a probability of 910 detecting events more severe than "alert" (level 1) above 50% but could be improved by scrutinizing the minutes 911 of the drought committees to better understand the weight of the stakeholders in the final statement. A better 912 assessment of the sustainability is required. The risk-based approach was applied to assess the vulnerability of       ( Table 21). The x-abscissa is divided into ten-day periods 1229 for each year spanning the period April-to-October period. Black segments identify updated DMPs.