France presents a large panel of hydrological regimes resulting from various physiographical contexts (e.g. geology, topography, see Fig. a) but also from the large diversity of regional climates oceanic influences in the West, Mediterranean in the South-East,. Catchment regimes can be considered from a hydroclimatic perspective, as provided by the framework. The Budyko model, which provides a simplified representation of the mean annual water balance, allows for the analysis of interactions between energy and water limitations and their impact on catchment hydrology. More specifically, the ratio between total precipitation (P) and potential evapotranspiration (PET) provides a humidity index that can be used as a proxy for different types of interactions and various impacts on runoff. P/PET ratio values lower than one characterize water-limited regions, where dry conditions prevail and sensible heat fluxes are typically high. In contrast, P/PET ratios greater than one denote energy-limited regions, which are generally wetter and exhibit lower sensible heat fluxes . In France, most regions are energy-limited, especially mountainous areas (Alps, Jura, Vosges in the East, Pyrénées in the South, Massif Central in the center, see Fig. b). Water-limited regions are found in the South (Mediterranean region), in a large area northwest of the Massif Central and in some small areas. As reported by or , the projected evolution of hydrological changes can be interpreted according to this distinction.

Hydrological regimes can also be classified based on the seasonal cycle of inter-annual mean monthly discharges, following the work of . propose seven regimes ranging from snow-dominated regimes in mountainous areas to mixed regimes downstream and/or rainfall-dominated regimes in lowlands (Fig. c, d). Snow-dominated regimes are mainly characterized by a monthly peak flow in spring due to snowmelt which can be particularly large for high-elevation mountainous catchments. Three different snow-dominated regimes are considered: (1) a regime with a dominant nival contribution (N) found in the highest regions of the Alps, (2) a Nivo-Pluvial regime (NP) observed at downstream locations, in mid-elevation areas (e.g. Pyrénées, PreAlps regions) and in southeastern France where Mediterranean events can be important, and (3) a Pluvial-Nival (PN) regime mainly found in downstream locations and low-elevation mountainous areas (Massif Central, Vosges, Jura, Corsica).

Rainfall-dominated regimes present a seasonality mostly derived from the seasonality of precipitation and evapotranspiration losses, with high flows in winter and low flows in summer. The winter/summer contrast depends on the catchment. It is very large in the Western and northeastern parts of France, where catchments have a low storage capacity, large in the North and the South-West, moderate in the Center and in the East, and very small in regions where river streamflows are sustained all year long by important aquifers (e.g. in the Paris Basin region). This diversity is represented by four regimes: highly contrasted pluvial (HCP), contrasted pluvial (CP), pluvial (P), and non-contrasted pluvial (NCP). Different hydrological responses to climate change are expected according to these regimes e.g.,.

The Explore2 dataset, spanning the period 1976–2099, was generated using various scenarios and models shown in Fig. . A detailed description is provided in . The 36 climate projections correspond to a subset of the large CMIP5-EUROCORDEX ensemble and have been obtained with three emission scenarios 10 with the RCP2.6, 9 with the RCP4.5, and 17 with the RCP8.5, see. It is based on six CMIP5 GCMs downscaled by nine RCMs. As summarized in Table , it considers 17 GCM/RCM combinations and is therefore incomplete (e.g. not all GCMs have been used to force a given RCM); however, the selected projections ensure a diverse range of climate models and include at least two simulations for each GCM and RCM. This selection enhances the robustness of the uncertainty component estimates.

The BAM ADAMONT was applied to adjust climate projections for some systematic errors using the meteorological reanalysis SAFRAN as the reference dataset. Adjusted climate projections are produced on a regular horizontal 8 km grid covering Metropolitan France and part of Switzerland. In this study, we focus on the 8981 grid points that cover Metropolitan France. Reference evapotranspiration (ET0) is obtained from the Penman-Monteith formula with a parameterization for short grass. The net radiation is derived from the Hargreaves equation .

Explore2 hydrological projections offer a “natural reference hydrology” as a foundation for future studies on water allocation. Streamflow simulations are provided in the absence of any water management interventions – such as abstractions, releases, or reservoir operations – thus representing natural flow conditions. In this study, we assess an ensemble of hydrological projections obtained for 1735 simulation points with four HMs of different types (Table ) covering almost all the country, namely CTRIP, GRSD, ORCHIDEE and SMASH (Fig. ). The Explore2 MME provides hydrological simulations with nine HMs, but the five other HMs cover only specific regions of France and for fewer catchments . SMASH and GRSD HMs are conceptual “bucket” type models. CTRIP and ORCHIDEE are land surface models and physically based. No single model provided discharge simulations for all simulation points across the study area. Some simulation points could not be reliably paired with their corresponding locations in the hydrographic network (or were excluded due to large differences in simulated versus physical catchment areas). All HMs used the same meteorological forcings: the reanalysis SAFRAN as the reference dataset for the period 1976–2005 and adjusted scenarios for historical and future climate experiments. Calibration of each HM, if required, was model-specific and carried out automatically or manually based on the expertise of the model developers. To assess model performance, all HMs have been evaluated within a standardized framework, focusing on their ability to replicate various observed hydrological signatures for the current climate seefor evaluations.

We carried out our analyses for a selection of indicators representative of projected changes in climate and hydrology. Indicators considered in the present work are surface air temperature (TAS), total precipitation (PR), reference evapotranspiration (ET0) at a seasonal scale (summer and winter), annual maximum daily precipitation (RX1D), annual minimum monthly flow (QMNA), mean annual daily discharge (QA), and annual maximum daily discharge (QJXA).

To characterize how a given climate or hydrological indicator is projected to change, we use the QUALYPSO method . QUALYPSO focuses on the climate change response (CCR) projected for the considered indicator. The CCR projected by a given simulation chain defines, for any future time, the absolute or relative difference of the long-term climate response for this future time compared to the reference period. The QUALYPSO method is implemented in an R package and was specifically designed to tackle the main challenges of uncertainty partitioning in unbalanced and incomplete MMEs. More information is provided in Appendix , including an illustration of the different steps in Fig. .

QUALYPSO aims to estimate, for any future time, the CCR of each modeling chain, the mean CCR of the MME, the dispersion between the CCRs of individual chains, the sources of uncertainty explaining this dispersion, and the additional potential dispersion in future realizations of climate indicators that can result from IV.

Figure  shows the 5 %, 50 %, and 95 % quantiles of the CCR for the climate indicators, for the RCP8.5. These results confirm those reported in previous publications based on EUROCORDEX ensembles e.g.. Projected climate change depends on the indicator, season, and region. It is contrasted and uncertain.

Temperature is projected to increase significantly, especially in summer. Depending on the region, the median projected increase varies from 2.5 to 4.5 °C in winter and from 3.5 to 6 °C in summer. Summer warming is projected to be much greater in mountainous areas (Alps, Pyrénées, Massif Central, Vosges) and in the south, particularly in the South-East. The dispersion between modeling chains is important. The 5 %–95 % interquantile range is around 1.5 °C in winter, up to 4 °C in summer. The largest warming projected in summer is greater than 7.5 °C.

In summer, precipitation is projected to significantly decrease, from -10 % to -40 % for the median CCR depending on the region. Although there is considerable dispersion among the projections, they largely agree on the direction of change. However, a few exceptions exist, notably in the Rhône River valley and some eastern regions, where some projections indicate a slight increase. The most extreme projections lead to very large decreases, -60 % on average over France and up to -85 % in the South.

In contrast to summer, winter precipitation is mostly projected to increase (but less than the projected decrease in summer). The median change is around 0 % in the South, South-East, and it exceeds 30 % over the northern half. The inter-model dispersion remains significant, although it is less pronounced than in summer. Models almost always agree on the sign of the changes and, except in the South, even the 5 % quantile projects slight increases.

The median reference evapotranspiration (ET0) is projected to increase, from 20 % to 30 % over most of France (and slightly more in summer). If all models agree on the sign of changes, the inter-model dispersion is large. In summer, the 5 % quantile ranges from 10 % to 20 % and the 95 % quantile is larger than 40 %. In general, results roughly follow those obtained for temperature, but spatial patterns of changes do not match the N–W, S–E gradients obtained for temperature. The South-East and the mountainous areas do not appear as hot spots of changes.

The median annual maximum daily precipitation is projected to increase. Projected increases vary a lot from one location to another, from 0 % to 20 % in a large southern part to (20 %, 30 %) in northern areas. Regardless of the region, the inter-model dispersion is considerable. In the south, for instance, models do not agree on the sign of the changes, and the most extreme chains project decreases down to (-40 %, -20 %) and large increases up to (40 %, 60 %).

The 5 %, 50 %, and 95 % quantiles of the CCR for the hydrological indicators, for the RCP8.5 (Fig. ) are contrasted, uncertain, and depend on the region.

Low flows are mostly projected to decrease, from -10 % in the North to (-60 %, -45 %) in the South for the median CCR. While the inter-model dispersion is very large, projections agree on a decrease for most southern regions (with the exception of the Alps). The driest chains lead to very large decreases, greater than -60 % (except for a small area in the North where they “only” reach -45 % to -60 %). Conversely, the most humid ones project large increases (40 % or more) in the Center-North.

For median annual discharges, while slight increases are projected (often less than 15 %) in the North, decreases (up to -30 %) are projected in the South. With the exception of the Pyrénées and the southern Alps, where most models project a decrease, modeling chains do not agree on the sign of change. The inter-model dispersion is very important. Large decreases are projected for the driest chains (up to -75 % in the South-East) while large to very large increases are projected for the most humid ones (up to 75 % in the North).

Results for high flows are roughly similar. The median high flow is projected to slightly increase (up to 15 %–30 %) in the North and to decrease in the South. Regardless of the region, the inter-model dispersion is large, and models do not agree on the sign changes. Large decreases are projected for the driest chains (up to -75 % in the South) while large to very large increases are projected for the most humid ones (up to 75 % in the North). We refer to for further analyses of projected high flows from the Explore2 dataset.

Figures  and compare the CCR uncertainty (CCRU) and the internal variability IV for the climate and hydrological indicators, respectively. Except for summer temperatures, IV is large to very large and is often of the same order as, or even greater than CCR uncertainty.

For seasonal temperature, year-to-year fluctuations around the CCR are of the order of 1 °C (Fig. ). IV is much lower than CCR uncertainty in summer but similar in winter. Results for ET0 are roughly similar (Fig. S5 in the Supplement). IV is rather large, slightly smaller than CCR uncertainty in summer, and slightly larger in winter.

For precipitation, IV is very large (>30 %), much larger than CCR uncertainty (between 10 % and 30 %), especially in winter and along the Mediterranean sea (>50 %, Fig. ). For maximum annual precipitation, IV is large (often greater than 40 %, see Fig. S5 in the Supplement). It is also much larger than CCR uncertainty (10 %–20 %).

For the three hydrological indicators, IV is logically also very large (Fig. ). For annual flows, IV is larger than 30 %–40 % in many regions. For low and high flows, it is higher than 40 % almost everywhere (higher than 100 % for low flows locally). It is often more than two times larger than CCR uncertainty.

Figures S1 and S2 in the Supplement show the 5 %, 50 %, and 95 % quantiles of the CCR projected for the climate indicators, for the lower emission scenarios RCP2.6 and RCP4.5 respectively. For temperature and related indicators, changes projected with lower emission scenarios are generally smaller than those under RCP8.5, while exhibiting similar spatial patterns. In contrast, projections for precipitation and hydrological indicators do not necessarily lead to the same conclusions. For these indicators, inter-model spread remains substantial, with frequent disagreement among models regarding the sign of change. Moreover, the direction of projected changes does not always align with those under RCP8.5. For instance, under RCP2.6, winter precipitation is projected to increase in the South-East, and some model chains also indicate increases in summer.

This pattern is also observed in hydrological projections. Changes projected for lower emission scenarios are not necessarily smaller than changes projected for RCP8.5 (Figs. S3 and S4 in the Supplement). The mean annual flows are projected to increase everywhere in RCP2.6 and almost everywhere in RCP4.5. Low flows are still projected to decrease (much less) for RCP4.5 but they are projected to increase for RCP2.6, regardless of the region. Projected changes for high flows are roughly the same for all scenarios.

Figure  shows the main uncertainty components for seasonal temperature and precipitation changes and confirms the results obtained for Europe e.g.. The relative contributions of the different uncertainty sources to CCR uncertainty depend on the indicator and region.

For temperature projections, scenario uncertainty is by far the main contribution for all seasons. A non-negligible contribution of climate model uncertainty is found in summer. It comes from GCMs for coastal areas (Atlantic and Channel) and from RCMs for mountainous areas and in the East.

For summer precipitation, the main contributions come from RCMs and then from scenarios (between 20 % and 40 %). In winter, they mostly come from GCMs or from RCMs in mountainous areas and in the Mediterranean region. Scenario uncertainty has a moderate contribution in the North-East. For maximum precipitation, the main contributions come from GCMs and RCMs (Fig. S5 in the Supplement).

For summer ET0, results closely follow those for summer temperature (Fig. S5). However, the contribution of scenario uncertainty is smaller, whereas the contribution of RCMs becomes much larger (>20 %–40 %). For winter ET0, the contribution of scenario uncertainty is smaller to the benefit of RCM uncertainty, which reaches up to 40 %–60 % in the West.

With the exception of temperature, the contribution of residual variability to CCR uncertainty is moderate to important, which means that a moderate to large part of the CCR uncertainty cannot be explained by the additive effects of scenario, GCM or RCM uncertainty. This is the case for precipitation in both seasons, for instance. In summer, the unexplained fraction of CCR uncertainty varies from 10 %–20 % in the West to 20 %–40 % in the East. In winter, it varies from 20 %–40 % in the East to 40 %–60 % in the South-East. show that interactions between scenarios and climate models can be important in France for winter precipitation (see their Fig. S7). In particular, scenario/GCM and GCM/RCM interactions can partly explain this important residual variability in the current study.

Figure  presents the main uncertainty components for hydrological changes. These results are generally consistent with previous studies e.g.. The extent to which each source of uncertainty contributes to the spread in projected changes also depends on the region and indicator.

For low flows, CCR uncertainty comes from emission scenarios, especially in southern France (>40 %–60 %), from RCMs, especially in the East (>20 %–40 %), and to a lesser extent from HMs in the North-West and in the Alps (>20 %–40 %). The uncertainty from GCMs is significantly lower than that from RCMs. However, it remains notable in the northern regions, at the expense of RCP uncertainty, especially in the energy-limited areas. In France, except in mountainous areas, low flows occur in summer when snow and soil moisture are depleted. They are mainly driven by soil processes, evapotranspiration, and sometimes groundwater support. Low flows are thus controlled by the climate conditions that determine the summer water budget and by the choice of the HM.

For mean annual flows, CCR uncertainty mostly comes from RCMs, especially in the East (>40 %–60 %), from GCMs (everywhere except in the South-East), and from emission scenarios in the southern half of France (>20 %–40 %). A non-negligible contribution of HMs is obtained in the Paris Basin region (>10 %–20 %).

For high flows, CCR uncertainty is predominantly attributable to climate models, with GCMs and RCMs contributing to a comparable extent. This substantial contribution can partly be explained by the considerable uncertainty in projections of extreme precipitation (Fig. ), but may also reflect uncertainties in other climatic variables. High flows are driven by a combination of factors, including the magnitude and spatiotemporal distribution of precipitation, as well as antecedent hydrological conditions resulting from preceding precipitation events . Using the GRSD hydrological model, demonstrate that projected changes in soil saturation exert an influence on high flows comparable to that of projected changes in intense rainfall. In the current study, the pronounced sensitivity to hydrological model selection in eastern France further supports the role of antecedent conditions in shaping high flow responses.

The unexplained fraction of CCR uncertainty can also be large for hydrological indicators. It varies from 10 %–20 % in the South to 20 %–40 % in the North for low flows, mostly between 20 %–40 % for mean annual flows, from 20 %–40 % to more than 60 % in a number of sites for high flows (Fig. ).

The hydrology of French rivers is expected to undergo marked changes in the coming decades, but the high uncertainty surrounding future precipitation makes the nature of these changes mostly unpredictable. However, the projected increase in temperature is robust and, as a consequence, some changes are very likely, especially for some specific hydroclimatic regimes. As already shown by previous studies in the area e.g. and by additional Explore2 analyses , warming will profoundly modify the hydrological regimes where discharges are significantly influenced by snow accumulation and melt. Snow-dominated regimes will shift to mixed regimes, while mixed regimes will become predominantly rainfall-driven. As a result, snowmelt floods will occur earlier in the spring and be less intense.

Figure  aggregates the different results shown in this study by hydrological regime (obtained for the reference period 1976–2005). As indicated above, median CCRs (Fig. A) highlight more pronounced changes for low flows than for mean and high flows. For low flows, strong decreases are obtained across most hydrological regimes (around -40 %), and scenario uncertainty emerges as the dominant source of variability (Fig. B, first row), primarily due to the strong dependence of evapotranspiration losses on temperature, as previously discussed. There are two exceptions. The first concerns catchments with a non-contrasted pluvial (NCP) regime, where low flows are likely to remain supported by deep aquifer contributions regardless of the emission scenario. In this case, the main source of uncertainty arises from climate models, but hydrological models (HMs) also contribute due to their varying representations of deep groundwater storage. The second exception involves catchments with a nival (N) regime, where low flows typically occur in winter due to a high proportion of solid precipitation during this season (Fig. C). Under warmer conditions, this proportion is expected to decrease, leading to increased winter discharges. As HMs vary in their representation of snow processes, HM uncertainty plays a particularly significant role in this case.

For mean annual and high flows, the combined uncertainty from GCMs and RCMs is relatively consistent across all regimes. In contrast, the contributions of scenario and HM uncertainties vary by regime. For mean annual discharges, scenario uncertainty is minor in rainfall-dominated regimes but more significant in snow-dominated ones. HM uncertainty is generally low, except in catchments where low flows make up a substantial portion of the mean annual flow (i.e., P and PNC regimes). For high flows, scenario uncertainty is negligible across all regimes. HM uncertainty, while also similar across regimes, is more pronounced due to differing representations of runoff generation in the models – particularly in snow-dominated regimes, where high flows are primarily driven by snowmelt.

A qualitative comparison between Figs. B, and shows that projected changes are only partially influenced by whether catchments are water- or energy-limited. In the South, the decrease of mean annual flows, which highly depends on the interplay between precipitation and evapotranspiration, tends to be larger in water-limited areas. In the North, mean annual flows tend to decline in water-limited catchments, while they increase in energy-limited ones. However, projected changes in high and low flows appear largely independent of the hydroclimatic regime. For example, the significant reduction in low flows projected for the South is widespread across different regimes. A notable exception is the case of high flows in the South, which tend to increase more in water-limited areas. Across most locations and hydrological variables, the sources of uncertainty contributing to CCR uncertainty are quite similar between water- and energy-limited regimes, especially regarding HM uncertainty. An exception occurs in the energy-limited catchments of the North, where HM uncertainty is greater for projections of both low and mean annual flows compared to other regions.

Our results show that IV can be a substantial source of uncertainty in hydrological projections and are in agreement with most other analyses where IV has been considered e.g.. The influence of IV is strong when precipitation is a key driver and is particularly pronounced for low and high flow projections (Fig. ). For the three variables considered here, IV is larger than CCR uncertainty at the end of the century, and much larger for projected periods that are not far (not shown).

Our results show that IV should not be overlooked or disregarded in hydrological impact studies. IV expresses the range of possible future realizations around the climate response. IV is likely to cause substantial deviations from the climate response, potentially leading to unusual or critical years or sequences of years. Disregarding IV could be detrimental for interpretations and use of projections. It could prevent a fair evaluation of the resilience and robustness of natural and anthropogenic systems to climate change, variability, and extremes .

As shown in for Europe, a large contribution of GCM (resp. RCM) uncertainty to CCR uncertainty is often due to a single or a small number of individual GCM models (resp. RCM models). This can be easily identified from the maps of the main effects obtained for each individual model. For summer temperature, for instance (Fig. ), the large RCM contribution comes mainly from the higher warming signature of HadREM3-GA7-05 and the lower warming signature of WRF381P. For winter precipitation (Fig. ), the important GCM uncertainty comes mainly from the “wet signature” of IPSL-CM5A-MR; the large RCM contribution in mountainous areas comes mainly from the large differences between four RCMs (RACMO22E and RCA4 are much drier than HIRHAM5 and WRF381P).

A large contribution of a specific climate model to the CCR uncertainty in climate projections frequently leads to similar contributions for other climate or hydrological indicators. For example, the deviating signals produced by HadREM3-GA7-05 and WRF381P for projected changes of summer temperature are reflected in similar divergent patterns for changes in summer ET0 (Fig. S8 in the Supplement) and low flows (Fig. S12 in the Supplement).

A single or a small number of HMs can also lead to a large contribution of HM uncertainty to CCR uncertainty. As shown by the maps of the main effects obtained for each individual HM (Fig. ), the important HM uncertainty obtained for mean annual flows in the Paris Basin region (lowland area around Paris) is mainly due to the “more humid” (or “less dry”) signature of ORCHIDEE. For high flows, the important HM uncertainty in the North-East is mainly due to CTRIP. HM main effects vary more regionally (or by regime) for low flows than for mean and high flows. Higher uncertainty for low flows might be expected where more complex processes are involved (e.g. snowmelt for the nival regime, groundwater-river exchange for the non-contrasted pluvial regime) and are represented differently by the different HMs. High flows are mostly driven by precipitation and are more easily represented by HMs when similar antecedent conditions are simulated.

Note that the main effects of the different model categories often have different spatial scales. Spatial patterns of the main effects of GCMs are mostly large scale, presenting smooth variations, whereas those of RCMs and HMs, often conditioned by local topography for RCMs or by specific physiographic features for HMs, are more local, often depicting rough variations. The spatial variation of these main effects can be related to the effective resolution of the corresponding models, which is often larger than the resolution provided by the simulations .

Models with very large main effects compared to the other ones can thus be identified using ANOVA approaches. These deviating models inflate the corresponding uncertainty component. As large uncertainties may dampen important information about future changes, it is important to examine these deviating models carefully and understand the reasons behind their atypical behavior. If the models deviate for the wrong reasons (numerical artifacts, bugs, model inadequacy), a reasonable choice could be to discard them from the MME. Large deviations by a single model can also indicate that other models are overlooking or oversimplifying a key process that could become critical under future climate conditions. This raises concerns about model transferability, particularly for models based on empirical or highly parameterized approaches.

In Explore2, recommendations were provided to help end users select or filter available projections for a catchment. Diagnostic assessments were carried out for all catchments to evaluate the performance of surface hydrology and hydrogeological models when driven by the SAFRAN reanalysis . These diagnostics complement the uncertainty analysis and aim to support the selection of a subset of models for prospective studies, rather than to rank them. Users are encouraged to consult summary sheets at different spatial scales, from local to regional, and to make choices based on their specific needs. Model selection can then depend on intended data use (e.g., streamflow only or additional variables) and on each model’s ability to deliver the required variables at the desired locations.

CCRs and related uncertainties depend on the future period. Results obtained for other future periods, specifically 2020–2050 (near future) and 2040–2070 (mid-century), are presented in . As the analysis focuses on the CCR of projections, the dispersion between the main effects of the different models in each category is almost zero for the near future and increases with projection lead time, for all emission scenarios. The uncertainty of each model category and, in turn, total CCR uncertainty thus also increases with projection lead time. As a consequence, CCR uncertainty tends to be much lower than IV in the near future.

To highlight how results vary over time throughout the 21st century, a summary sheet was created for all stations included in the project, detailing projected changes and associated uncertainties. This summary sheet facilitates the identification of key sources of uncertainty, notably including substantial contributions from individual models to the overall model-related uncertainty. Furthermore, it provides a comprehensive depiction of the range of potential future projections, capturing both systematic changes in climate response and stochastic variations arising from internal climate variability. For illustrative purposes, the summary sheet developed for the Seine catchment at Bazoches-Lès-Bray is presented in Fig. . It summarizes the different results shown above, notably:

Panels A and B: A decrease in low flows and an increase in high flows under the scenario RCP8.5 (panel A), with an overall agreement of the MME (80 % of the CCR exhibiting the same sign for the projected changes).

Explore2 projections are obviously not free of limitations, partly due to model imperfections. Some known limitations are reported here.