1. Introduction
Karst springs play a fundamental role in large-scale human water supply, both from a strategic and socio-economic point of view [
1]. Quantifying actual and future project spring discharges is extremely important to manage water resources in karstic areas, especially in view of the effects related to climate change. According to the Intergovernamental Panel on Climate Change (IPCC), the mean temperature is expected to increase globally throughout the 21st century, with a consequent increase in the frequency and magnitude of heat waves, drought periods and intense precipitation events in many regions of the world [
2]. The prediction of the response to climate change in hydrogeological systems becomes vital, especially if exploited for drinking water supply [
3] or the sustainability of groundwater dependent ecosystems [
4].
The application of traditional groundwater flow models for the prediction of fractured media is complicated due to the duality of the flow systems [
5]; this is particularly true for karst systems [
6,
7]. A typical karst system is formed by a fractured rock matrix, which includes both diagenetic rock micro pores and small fractures of tectonic origin, as well as a network of widely articulated karst conduits [
1,
8]. This dual nature leads to a hybrid flow behavior, Darcian within the rock matrix, and turbulence in the conduits. In addition, flow in the conduits can be assimilated to full pipe (pressurized) flow or to free surface flow, based on the channel saturation. Attempts to define physically based, distributed models incorporating some or all of these characteristics have been made by many authors [
9,
10,
11,
12,
13,
14,
15]. However, the definition of a conceptual model that could depict the system heterogeneity reliably, in terms of both 3D geometry and parameterization, can be very data demanding, time consuming, and costly [
16]. Speleological exploration (e.g., [
17]) and artificial tracers (e.g., [
18]) can furnish information on spring catchment areas and the main geometric properties of the karst system. Additional insights regarding different water sources and water residence time in the aquifer can be derived from natural tracer analyses, such as water isotopes, major ions, trace elements, and dissolved organic carbon (e.g., [
19,
20,
21,
22,
23]). In-situ hydraulic methods, such as pumping tests, can provide knowledge on the degree of confinement of the aquifer and its quantitative parameters (e.g., [
24,
25]). Non-invasive geophysical methods can also be used to gather information on aquifer geometry and boundaries (e.g., electric and seismic methods, [
26,
27]), to identify major voids and conduits (e.g., gravimetric and geomagnetical methods, [
28]), and to recognize preferential infiltration pathways (e.g., self-potential methods, [
29]). In addition, to allow model calibration, a continuous monitoring of at least precipitation and spring discharge is essential [
16].
When the main goal of the study concerns the response of a karst spring to meteorological events, the continuous monitoring of precipitation and discharge could suffice for a modeling setup. Input-output models can be used to relate the precipitation contribution to the outflow through empirical equations based on lumped parameters (e.g., [
30,
31]), neural networks (e.g., [
32]), or input-response functions (e.g., [
33,
34]). An increasing number of studies involving such modelling approaches have been documented for karst environments in recent years (e.g., [
35,
36,
37,
38,
39]). Due to the lack of physical processes representation, a limitation of these models concerns their loss of reliability if applied in conditions that largely differ from calibration [
40,
41]. Parameter transferability has been largely tested (e.g., [
42,
43,
44]), suggesting an increase in simulation errors and uncertainties with increasing differences in mean rainfall. Also, performance losses appear larger when transferring parameters from wet to dry periods than vice versa. Vaze et al. [
42] concluded that calibrated models could be used for climate change impact studies if mean rainfall changes between future and calibration periods fall within a −15% to +20% range.
Climate projections are derived from Global Circulation Models (GCMs) forced according to the representative concentration pathways (RCP, [
45]) proposed by the IPCC [
2]. Usually, these models are run at large horizontal resolutions (≥ 80 km) and are therefore not able to reproduce climate states at regional or local scales reliably [
46]. To obtain climate projections at a finer resolution by reproducing the physics of the processes, dynamical downscaling through Regional Climate Models (RCMs) is performed (e.g., [
47,
48]). With the ongoing climate modeling experiments—for example, the Coordinated Regional Climate Downscaling Experiment (CORDEX,
http://www.cordex.org/), WorldClim dataset project (
http://www.worldclim.org/), and the Climate Change, Agriculture, and Food Security research program (CCAFS,
http://ccafs-climate.org/downscaling/)—future climate projections are made available over several domains worldwide with a horizontal resolution of up to 12.5 km. Some studies (e.g., [
49,
50]) demonstrated that dynamical downscaling can be applied to achieve an even larger detail (≤4 km). However, these methods are computationally very intensive and often show performance losses over complex terrains [
46,
51]. The computationally inexpensive alternative is statistical downscaling, which provides a predictive framework of local conditions (with almost no bias with historical records) through statistical relationships with regional climatic aspects. The hypothesis is that the relationship between large-scale atmospheric circulation and local scale dynamics is constant [
52,
53,
54]. Therefore, statistical downscaling allows for producing large ensembles of climate change projections for extended time periods [
46,
55].
The overarching goal of this study is to propose a methodological approach to evaluate possible variations in the discharge regime of a karst spring in basins affected by climate change. The selected study area is the hydrogeological basin of the Nossana Spring (northern Italy), which is heavily exploited for drinking water. Using a control data period covering 1998–2017, the specific objectives of this study are: (i) quantification of the expected changes in precipitation and temperature in the study area, for twenty-year periods till 2100, as resulting from the statistical downscaling of RCMs run under three RCPs (2.6, 4.5, 8.5); (ii) calibration and validation of a lumped-parameter model (GR4J with CemaNeige [
56,
57,
58]) based on observed data; (iii) recognition of possible limits in the future utilization of the spring as a drinking supply based on simulated discharge and actual warning thresholds.
2. Study Area
The Nossana Spring is located in the central Prealps of Seriana Valley, within the Lombardy Region of the Province of Bergamo (Northern Italy) (
Figure 1). Its hydrogeological basin covers an area of about 80 km
2 and is characterized by large differences in altitude, ranging from 474 m a.s.l. (Nossana Spring) to 2512 m a.s.l. (Pizzo Arera Mountain).
The spring is managed by the public company UniAcque S.p.A. and feeds the main local water distribution system, serving over 300,000 people. The observed mean annual discharge (1998–2017) is 3.77 m3 s−1. For the same period, the observed daily variability is extremely high, with values ranging from 0.55 m3 s−1 to 18.00 m3 s−1. In a single day, discharge varied up to 12 m3 s−1, suggesting high flow velocities characterizing the system.
To manage the water resource, UniAcque S.p.A. set two warning thresholds based on average water demand according to the period of the year (one threshold for the cold-wet season and one threshold for the hot-dry months). Below these discharge thresholds, there is a need to integrate the spring resources pumping water from deep wells in the Serio River Valley. This entails different issues, from costs increase to quality decrease. Moreover, nowadays the additional discharges are limited (0.50 m3 s−1). During wet, cold periods, the warning threshold is set at 1.32 m3 s−1; during dry, hot periods, the threshold is set at 1.52 m3 s−1. These two discharge rates include a share for UniAcque S.p.A. (0.80 and 1.00 m3 s−1, respectively) and a share for environmental flow and use of the downstream municipalities (0.52 m3 s−1). Water that is not used by UniAcque S.p.A. flows into the Serio River, the main watercourse of the Seriana Valley.
Many studies have defined a hypothetical hydrogeological catchment for the spring [
59,
60], considering the geological, geomorphological, and tectonic evidences of the area [
61,
62,
63,
64]. The karst massif that characterizes the area consists of calcareous and calcareous-silicoclastic series (medium-late Triassic). The main calcareous formation is represented by “Calcare di Esino” (middle-late Trias), constituted by deeply karstified massive or stratified limestone, with a maximum thickness of about 800 m. Karst evidences (e.g., dolines and tunnels) are mainly recognized in the western and north-eastern area of the basin. The late-Triassic marl formations with low permeability (Breno Formation, Metamorphic Bergamasco Limestone and Gorno Formation) link the massif to its southern, western, and eastern borders along the tectonic contacts with the Calcare di Esino formation. Three main fragile structures mark the limits of the massif along the northern side (“Valtorta-Valcanale thrust”), west side (“Grem fault”), and southeast side (“Clusone fault”). More specific geological information about the study area can be found in literature [
63,
65,
66].
According to Vigna and Banzato [
67], based on the geological and hydrographic characteristics described above, Nossana can be classified as a karst spring fed by a dominant drainage system. These systems are usually characterized by a high permeability, which is linked to a considerable karstification process. The outflow network is well-organized with a series of main natural tunnels and secondary conduits that rapidly discharge incoming infiltration water. The most important features are the very low or complete absence of a classic phreatic zone (indeed, the presence of a phreatic zone has not been verified for Nossana yet) and the very high flow velocity.
The climate of the study area is wet and temperate. According to data recorded at the meteorological station of Clusone (591 m a.s.l.), located at around 4 km east of Nossana Spring, rainfall is distributed throughout the year. Mean annual rainfall calculated for the period 1998–2017 is 1384 mm, with November as the wettest month (average of 160 mm) and January as the driest (average of 65 mm). May, June, October, and November yield similar amounts of rainfall. Conversely, the winter lowest values indicate a change from liquid (rainfall) to solid precipitation (snow/ice). Average monthly temperatures vary between 1.4 °C (January) and 21.7 °C (July), with an overall mean annual temperature of 11.5 °C. According to Köppen–Geiger climate classification [
68], the climate of the area can be classified as Cfb; namely, temperate with a wet summer, the mean temperature of the warmest month below 22 °C, and at least four months above an average of 10 °C.
6. Discussion
Groppelli et al. [
94,
95] produced five future climate scenarios by statistical downscaling of three GCMs (2010–2060), forced under the A2 IPCC emission scenario [
96], over the Oglio basin, the main valley east of the Seriana Valley. The A2 emission scenario, derived in the framework of 4th IPCC report [
96], is somewhat comparable to RCP8.5, derived in the framework of the 5th IPCC report [
2]. For the ten years around 2050 (in comparison to 1990–1999), they found an increase in mean annual temperature at 2000 m a.s.l. in the range 1.6–4.8 °C. At a lower elevation (around 600 m a.s.l.), for a comparable period (2041–2060) and scenario (RCP8.5), in this study a temperature increase in the range 1.8–2.8 °C was found. Larger increases could be expected only toward the end of the century (up to 5.8 °C). In terms of mean annual precipitation, Groppelli et al. [
94,
95] found variations between −15% and 40%, while in this study results suggest more limited changes (from −7% to 15%). The 18.5% decrease was obtained for the period 2041–2060, but under RCP4.5.
It is recognized that the selection of a rainfall-runoff modeling approach implies the assumption that the karst system will not undergo profound changes in the next 80 years, as pointed out by Hartmann et al. [
16]. Specifically on the utilized rainfall-runoff model (GR4J with CemaNeige as snow accounting routing), it was found that the CemaNeige snow routine including hysteresis did not increase the performance of the model in terms of simulated discharge. This is in agreement with Riboust et al. [
80], who found that accounting for hysteresis improved model simulations limited to snow cover area.
Gattinoni and Francani [
60] calibrated a distributed equivalent porous media model (MODFLOW) for the simulation of Nossana Spring discharges and applied it to build spring depletion curves under variable recharge. The calibrated model and the depletion curves were then used to assess changes in spring behavior according to expected regional changes in temperature and precipitation, for the period 2080–2099 in comparison to 1980–1999, according to GCM forced under the A1B emission scenario [
96]. All the models considered in their study suggested a decrease in mean annual precipitation (from 4% to 27%) and an increase in temperature (from 2.2 to 5.1 °C), which is in contrast with the results of this study obtained for the comparable RCP4.5 scenario (precipitation changes between −8% and 10%; temperature changes between 2.0 and 2.6 °C). However, due to the maximum expected precipitation decrease and evapotranspiration increase, they suggested a possible decrease in the spring discharge up to 40%, which is similar to the 39% decrease calculated in this study for the period 2041–2060 (Mod_2, see
Figure 8b) for a precipitation decrease of 18.5% (
Table 5) and a temperature increase of 2.1 °C. Gattinoni and Francani [
60] also suggested that changes of the precipitation regime during the recharge season (March to November) are expected to have the largest influence on spring discharge. This is proven by the results of this study, which showed how the changes in precipitation partitioning between January and May will change the 30-day average discharge, anticipating and decreasing the observed late spring high flows. In addition, the projected rainfall decreases affecting the summer period (June to September) will reduce Nossana Spring discharges during this season. Higher recharge, in comparison to the recent past, can be expected in October, November, and December.
7. Conclusions
Climate change directly impacts the water cycle and consequently groundwater resources. Therefore, planning their management is essential for their protection and responsible use. In this study, a comprehensive methodological approach for the evaluation of future climate and discharge variations (up to 2100, in four twenty-year intervals) at Nossana karst spring was presented. The method is based on the statistical downscaling of three bias-corrected EURO-CORDEX RCMs forced by three different emission scenarios (RCP2.6, RCP4.5, RCP8.5) and the implementation of a rainfall-runoff model (GR4J) extended with a snow accounting routine (CemaNeige). Today, Nossana Spring is exploited by UniAcque S.p.A. for the supply of drinking water to 300,000 people (including the town of Bergamo). Simulated discharges were evaluated in terms of mean flow and in comparison to actual water demands (expressed as discharge warning thresholds), to provide projections that can be operationally useful to the service company.
Based on the presented results, the following can be concluded:
The considered bias-corrected EURO-CORDEX RCMs have very good skills in reproducing observed temperature climatology (NSE > 0.95 and relative MAE < 10%) over the study area, while larger errors persist regarding precipitation (NSE between 0.25 and 0.65, relative MAE between 10% and 20%);
According to the downscaled RCMs data, in comparison to 1998–2017, mean temperature will likely increase throughout the rest of the XXI century, from 0.7 °C in 2021–2040 (RCP4.5, Mod_2) to 5.8 °C in 2081–2100 (RCP8.5, Mod_1);
Downscaled RCMs data do not show a clear trend in precipitation. For all twenty-year periods and RCP scenarios, there are single RCMs projecting increasing and decreasing rainfall (except 2021–2040, RCP2.6, all increasing). Variations in mean annual rainfall varies between −18.5% (2041–2060, RCP4.5, Mod_2) and 15.1% (2041–2060, RCP8.5, Mod_2);
A pronounced decrease of precipitation is expected in the summer period after 2060, as most RCM-RCP combinations show;
Mean discharges are generally projected to decrease in comparison to observed flow (3.77 m3 s−1) since changes in mean annual precipitation usually do not balance increases in evapotranspiration rates due to higher temperatures;
Variability in the projected mean discharges is mainly linked to the meteorological input rather than the rainfall-runoff model parameterization;
The maximum number of consecutive days below the warning thresholds was recognized as the best index to evaluate the spring low flow conditions;
After 2060, the length of the periods with discharge lower than the warning thresholds is expected to increase. These periods could last up to 64 days (86%) longer than in 1998–2017.
The present study indicates that additional water resources might be needed to satisfy the population water demand in the Nossana Spring area, especially after 2060. Therefore, the presented approach can be considered useful to provide indications regarding the management of the resource in the mid- long-term period. In fact, UniAcque S.p.A. will soon start to investigate possible alternative sources for water supply. However, it is evident that the results include large uncertainties derived from the climate input variability, especially precipitation, and as new emission scenarios are produced, the study should be updated.
Furthermore, the study focused on the quantitative aspects of the spring behavior; however, it is possible that a change in the discharge regime could also influence the chemical-physical characteristics of the water and its quality. A chemical-isotopic monitoring of the spring could improve the conceptual model of the aquifer system and the knowledge of the system response. This would allow a refinement of the method, raising it to a qualitative and quantitative prevision of the resource behavior.