The Carapelle basin (Figure 1), situated in northern Apulia (SE Italy), encompasses a drainage area of 506 km² with a main channel extending 52.16 km. The topography is characterized by a mean elevation of 466 m above sea level (a.s.l.), ranging from 120 to 1,089 m a.s.l., with an average basin slope of 8.2% and a main channel slope of 1.8%. Originating in the Campanian Apennine region, the river traverses the Daunia Hills before transitioning to a braided morphology in the alluvial plain (Abdelwahab et al., 2016). The hydrological regime exhibits significant short-term variability, with minimal flows during summer months (June–September) and peak discharges in winter and early spring.

The areas of the Carapelle catchment belong to the irrigated area of the Capitanata Reclamation Consortium (literally Consorzio di Bonifica della Capitanata). The territory has a high agricultural vocation (arable land >75%; olive groves about 7%), especially in the areas of Ortanova, Ordona, and Carapelle municipalities, forests (5%) is a minor land use (Figure 1). Therefore, the Carapelle River basin supports significant agricultural water demands and, in the last years has undergone intensive agricultural transformation primarily dependent on groundwater irrigation (Autorità di Bacino, 2015; Consorzio di Bonifica Della Capitanata, 2022). However, unrestricted groundwater extraction over the past two decades has led to declining water tables (Consorzio di Bonifica Della Capitanata, 2022). To address these water management challenges, a major water infrastructure project has been proposed: the Palazzo d’Ascoli reservoir, with a planned capacity of 67 million m³, of which 40 million m³ would be allocated for agricultural irrigation across 30,000 hectares, with the remainder serving industrial needs. The reservoir will have several implications on the river ecosystem and an accurate evaluation of the ecological flow will be required. Currently, there is no explicit information on environmental flow requirements (Consorzio di Bonifica Della Capitanata, 2022).

Dominant erosion processes include sheet wash and concentrated water erosion, while gully erosion is negligible. The area is prone to landslides, particularly in clay-flysch geological units susceptible to rainfall-induced slope movements (Wasowski et al., 2010). Bank erosion is prevalent, especially in upstream reaches. The basin experiences a Mediterranean climate, characterized by wet autumns and winters contrasting with dry springs and summers (Milella et al., 2012). Annual precipitation ranges from 450 to 800 mm, with March and November being the wettest months and August the driest. A monitoring station near Ordona (41°17′50.347”N, 15°36′2.583″E) is equipped with dual gauging systems: an electromechanical and ultrasound stage meter for streamflow measurements, and an infrared optical probe (Hach-Lange Solitax) for suspended sediment concentration (SSC) measurements, both recording at 30-min intervals. Daily sediment loads were calculated for the 2007–2011 study period, with brief interruptions for maintenance. Detailed station specifications are available in Gentile et al. (2010).

The methodology applied in this study is composed of three main stages: hydrological modelling; downscaling and bias correction (BC) of data from CMIP6 climate models; and analysis of the CC impacts on flow regime and sediment (Figure 2). In the first stage the SWAT model was set up for the study area, calibrated and validated. Subsequently, in the second stage, meteorological data (precipitation, minimum and maximum temperatures) were extracted from the three selected climate projections downscaled and, bias-corrected. The calibrated SWAT model was then run using daily observed climatic data from 2000 to 2020, and daily climatic modelled data from 2030 to 2050, which constituted the baseline and the near future (NF) periods, respectively. Finally, in the third stage, the modelled series of streamflow and sediment were analyzed to assess the CC impact on hydrology and sediment.

The CMIP6 (O’Neill et al., 2016) data archive is distributed through the Earth System Grid Federation (ESGF) though many national centers have either a full or partial copy of the data. A quality-controlled subset of CMIP6 data are made available through the Climate Data Store (CDS) for the users of the Copernicus Climate Change Service (C3S). This subset consists of 51 core variables from 9 of the most popular CMIP6 experiments, which can be used to assess plausible future changes in these variables under various socio-economic pathways. The CDS subset has undergone a rigorous quality control procedure to ensure a high standard of dependability, making it more reliable than the main CMIP6 ESGF archive, which comes with limited quality assurance and may contain metadata errors or omissions. For this reason, CDS¹ offers access to over large number of CMIP6 General Circulation Models (GCMs) with various variables. From this extensive collection, three models as detailed in Table 1, were selected. Then, historical and future climate simulations for the period 1985 to 2065, focusing on 3 parameters: minimum temperature (Tas_min), maximum temperature (Tas_max), and precipitation (pr) for the specified basin, were extracted. The analysis carried out considered one prospective scenario: SSP2-4.5, a business-as-usual pathway. In this scenario, the number 2 represents a medium pathway for human development without climate policy, while “4.5” indicates the projected radiative forcing (W/m²) by the year 2,100 (El-Rawy et al., 2023).

Bias Correction (BC) procedures are used to minimize the discrepancies between modelled and measured climatic data, making simulated climate data more accurate for hydrological models. The bias in data can be influenced by various factors such as mean, standard deviation, temporal distribution, frequencies, and extreme occurrences. Several BC methods are available, differing in applicability and limitations, and are widely discussed in climate impact studies (Ehret et al., 2012; Teutschbein and Seibert, 2012; Chen et al., 2013; Rajczak et al., 2016; Rohith and Jayakumar, 2017; Turco et al., 2017; Yin et al., 2020). These methods are classified into two main categories: 1-simple parametric transformations (e.g., linear scaling, power transformation, and linear regression) (Fang et al., 2015; Tschöke et al., 2017; Gudmundsson et al., 2012), which correct bias using various transformation functions; and 2-distribution-derived transformations (quantile scaling QS, Chen et al., 2013, quantile mapping QM, Maraun et al., 2010, and distribution mapping DM, Teutschbein and Seibert, 2012), which correct bias based on probability distributions fitted to observed data and simulations. Distribution-derived methods generally perform better than simple transformations (Chen et al., 2013; Fang et al., 2015). QS and QM use empirical distributions, while DM uses parametric distributions, making it more suitable for extrapolation and outperforming in general (Nanjegowda and Parambath, 2022). CMhyd tool (Rathjens et al., 2016) is used in this work to perform the bias correction. It is a tool for extracting and bias-correcting data from global and regional climate models, written in Python and uses various packages. It provides climate data representative of specific gauge locations in a basin. CMhyd identify biases in historical climate data to adjust future simulations, assuming the correction remains valid under future conditions. However, the effectiveness of these corrections for future conditions remains uncertain. Due to its higher performance as previously mentioned, the DM technique was adopted in this work for BC process.

DM (Teutschbein and Seibert, 2012) aligns climate values simulated by Regional Climate Models (RCMs) with observed values through a transfer function that adjusts key climate variables’ distributions. Known as ‘probability mapping’, ‘quantile-quantile mapping’, or ‘statistical downscaling’, this method is widely recognized in climate science (Sennikovs and Bethers, 2009; Block et al., 2009; Ines and Hansen, 2006; Johnson and Sharma, 2011; Boé et al., 2007; Sun et al., 2011). For modelling precipitation distributions, the Gamma distribution, characterized by shape (α) and scale (β) parameters, is effective. The shape parameter (α) influences the distribution’s profile, and the scale parameter (β) affects the spread, impacting the probability of extreme events. Temperature data is typically modeled using the Gaussian distribution, defined by location (μ) and scale (σ) parameters, to capture symmetric and continuous data patterns. A detailed description of the method can be found in Teutschbein and Seibert (2012).

In this study, the SWAT model, a semi-distributed continuous hydrological model developed by the United States Department of Agriculture (USDA) (Arnold et al., 1998; Arnold et al., 2012), was used to assess streamflow and sediment transport. The revised (Abbaspour et al., 2015) version of SWAT 2012 (Winchell et al., 2013) was implemented and run at daily time-step from 2004 to 2011. The first 3 years (2004–2006) were used as warm-up period. SWAT divide the basin first in sub-basins and then into hydrological response units (HRUs), which are the minimum territorial unit of the model, characterized by unique combination of soil characteristics, land use and slope. The study area was divided into 115 sub-basins, using a threshold of 200 ha and, into 451 HRUs, using percentage thresholds of 25% for slope, 25% for land use and, 5% for soil characteristics. The maintaining of the proportions between the main soil types, slope classes, and land uses was verified (Ricci et al., 2020; Ricci et al., 2022). For each HRU, the SWAT model simulates the surface runoff using the modified Soil Conservation Service-Curve Number method (SCS-CN) (USDA-SCS, 1972) and the sediment yield using the Modified Universal Soil Loss Equation (MUSLE) (Williams, 1975). The Manning’s equation and a simplified version of Bagnold’s stream power relationship were used to computes the rate and the velocity of flow and, sediment degradation or deposition, respectively (Bagnold, 1977; Neitsch et al., 2011). The Hargreaves method (Hargreaves, 1975), which requires temperature and solar radiation as input, was used to estimate potential evapotranspiration. SWAT requires several measurements and input data for its implementation, calibration and validation, which can limit its applicability (Abdelwahab et al., 2018). The input data used in this study to configure the model is detailed in Table 2. Farming operations data were retrieved for winter wheat and olive groves and the management files (.mgt) was modified accordingly. Winter wheat was planted in November and harvested in July. Ploughing (25–40 cm depth), which in the study area is generally perpendicular to contour lines even in mountain areas (>15%) was performed in August. Harrowing was carried out in October. Two fertilizing operation were conducted in December (fertilizer grade: 25–15-00; amount 180–220 kg ha⁻¹) and February (urea; amount 100–140 kg ha-1). A 4-year crop rotation (wheat, wheat, wheat, clover) was considered. For olive groves, three shallow tillage operations (ploughing and harrowing) were performed every 2 months starting in April and two fertilizer applications were adopted, in December (manure) and in spring (fertilizer grade: 26–00-00; amount 90–110 kg ha⁻¹). The crop was yield in November (Abdelwahab et al., 2016; Ricci et al., 2022).

Following Arnold et al. (2012, 2015), the SWAT model was first calibrated for streamflow and then for sediment load. Two sub-datasets were prepared for observed daily streamflow and sediment load (2007 to 2011). The first subset (January 2007 to April 2009) was used for the calibration, and the second (from January 2010 to December 2011) was used for the validation. Data from May to December 2009 were not available due to a malfunctioning of the station (Ricci et al., 2020). The Sequential Uncertainty Fitting version 2 (SUFI2) method, available in the SWAT-CUP tool (Abbaspour et al., 2015), was utilized to determine the sensitive parameters for both the calibrated variable. A further description of the sensitivity analysis, calibration, and validation are available in Ricci et al. (2018). Subsequently, SWAT-CUP was used manually adjust the model parameters and refine the match between simulated and observed data (Jeong et al., 2010; Ricci et al., 2020) and to identify the optimal set of parameters through an automated process, targeting a Nash-Sutcliffe Efficiency (NSE) greater than 0.5 as the objective function. Table 3 lists the calibrated parameters for both streamflow and sediment load. The USLE P-factor, which represents conservative management practices, was set to 1.0 because no conservation practices were implemented in the basin. To evaluate the model’s performances at daily scale, coefficient of determination (R2) and NSE > 0.5 and Percent Bias (PBIAS) ≤ 25% for runoff, and ≤ 55% for sediment were considered (Moriasi et al., 2007; Molina-Navarro et al., 2017).

There were substantial differences in the ability of GCMs to reproduce temperature and precipitation data under current climate conditions. To evaluate the efficacy of BC procedure, a range of statistical parameters were employed, for both daily Min and Max temperature and daily precipitation time series, including monthly means (μ), and 90th percentiles. It is important to note that this assessment was limited to the historical period (2000–2014), as direct validation of BC methods for future conditions is inherently unfeasible. For precipitation specifically, the analysis incorporated additional metrics: the probability of wet days (Prwet), and the average precipitation intensity on wet days (iwet). These characteristics of bias- corrected Climate Model simulated variables were compared to observational data on a month-by-month basis throughout the year. The un-corrected CMCC model simulations (Figure 3) showed an underestimation of the mean monthly PCP values through most of the months while it overestimated it during February and from August to October. The CMCC model had constant tendency to predict more wet days while it often simulated many low-intensity rain events. MPI and EC-EARTH models showed similar results, with higher mean monthly values shown by EC-EARTH, compared to MPI model. In general, the EC-EARTH showed higher mean monthly precipitation values in the 1st half of the year compared the 2 other models while in the 2nd half the CMCC showed the highest values. The precipitation-BC was able to improve the raw GCM simulations to some extent. With all models, there was a very slight a bias present in the mean of daily precipitation as in the case of MPI in the summer months (under-estimation) and EC-EARTH (slight overestimate during summer months) (Figure 3).

There were slight differences in the daily precipitation series’ probability of wet days and the precipitation intensity between the 3 models. However, the major dissimilarities were identified in the 90th percentiles which showed larger variability ranges and still had biases of a similar magnitude to the uncorrected precipitation. Regarding the temperature there were substantial differences in the ability of the 3 GCMs un-corrected simulation to reproduce temperature under current climate conditions. While the EC-EARTH and MPI models showed a tendency to overestimate the mean monthly maximum temperature (Tmax) and minimum temperature (Tmin) all over the year (except December), the CMCC model overestimated both minimum and maximum temperatures in the 1st half of the year from January to July while it showed an underestimation of the mean monthly temperature from August to November (Figure 3). In general, the highest values of Tmax were provided by CMCC model during the summer months (January-August). After the BC (Figure 4) the GCMs showed a very close match in the mean monthly temperatures with all biases perfectly corrected, while the 90th percentile showed a very slight variation between the 3 models and the observed data. Overall, the DM BC method was able to correct raw values of both temperature and precipitation for the 3 GCM models adopted in this study.

The analyzed CMIP6 GCMs project a consistent increase in mean monthly Tmax and Tmin for all months, with the strongest warming occurring in winter and the weakest in summer. The average annual rise of Tmax and Tmin projected for the NF period was 1.5°C and 1.3°C, respectively, compared to the baseline scenario, indicating a pronounced warming projected by CMCC model especially during Autumn and winter months. The trendline Analysis Mann-Kendall test at a significance level 5%, showed a significant positive trend for both Tmax (p < 0.0001 Kendall’s Tau = 0.622) and Tmin (p < 0.0001 Kendall’s Tau = 0.669) projected by the CMCC model on annual scale. The MPI model projected a consistent increase in mean monthly Tmax and Tmin for all months. Generally, the average annual Tmax projected for the NF were 1.3°C and 1.1°C by MPI and EC-EARTH respectively, while for Tmin both Models showed a rise by almost 1°C for the forementioned period. The Mann-Kendall trendline Analysis test showed a significant positive trend for both Tmax (p < 0.0001 Kendall’s Tau = 0.435) and Tmin (p < 0.0001 Kendall’s Tau = 0.525) projected by the MPI model on annual scale. However, The EC-EARTH showed a very weak significant positive trend line only for the Tmin (p = 0.000 Kendall’s Tau = 0.346) on annual scale while for Tmax there was not a significant trend. Changes in mean monthly precipitation show consistent reduction in the future.

Under CMCC model projections, the mean monthly precipitation decreases by 9%(August) to 41.2% (April) with an average reduction of 24% for all months except July that was the only month with a precipitation increase by 34.1%. The minimum mean monthly value was projected in June (11 mm), while July (76.2 mm) unexpectedly showed the highest projected value for the period 2030–2050 even higher than the winter months. The accumulated annual precipitation varied significantly during the NF with the highest value expected (741.8 mm) in 2050 and the lowest (183.4 mm) in 2043, with an overall average for the entire NF period of 371.8 mm, which is 38% less than the baseline.

The MPI model showed a decline in monthly precipitation from 4% (October) to 23.8% (April) in all months except February and May, which witnessed an increase by 6.9 and 6.3%, respectively. The minimum mean monthly value was projected in August (5.7 mm), while December (57 mm) showed the highest projected value for the period 2030–2050. Generally, the MPI projection showed higher values of precipitation than baseline during winter and Autumn (average of 48 mm). The accumulated annual precipitation varied significantly in NF with the highest value expected (1,008 mm) in 2048 and the lowest (183.4 mm) in 2039, with an overall average for the entire NF period of 466.2 mm which is 23% less than the baseline.

The EC-EARTH model showed a general decline by 5% in October to 16.7% in March. The minimum mean monthly value was projected in August (11.4 mm), while November (75 mm) showed the highest projected. The accumulated annual precipitation varied significantly during the NF period with the highest value expected (799 mm) in 2048 and the lowest (248.8 mm) in 2038, with an overall average for the entire NF period of 539 mm which is 10% less than the base scenario average Annual value. Among the analyzed projections, EC-EARTH showed the highest precipitation values either as average monthly or accumulated annual for the NF period predictions. Though a trend analysis Mann-Kendall test, it was found that neither CMCC nor EC-EARTH models showed any significant trend for precipitation on annual basis. While the MPI model showed a significant negative trend on annual scale.

The SWAT model exhibited satisfactory performance for the two simulated variables, streamflow and sediment load at the outlet, in both the calibration and validation periods (Figures 5, 6). The streamflow was underestimated in the calibration period (PBIAS +5.3%) while overestimated in the validation period (PBIAS -17.2%). Despite the overall performance, SWAT tended to overestimate slightly the normal flow and the low-flow (Figure 5) whereas several peaks were underestimated, particularly during the calibration period (Figure 6). In the dry year (2011), SWAT generally overestimated low-flow, while peak flows were relatively well simulated (Figure 6). For instance, the modeled highest peak flow, which was recorded in the wettest month of the simulation period (November 2010), was 6% lower than observed (simulated 79 m³s⁻¹; observed 94 m³s₋₁). The average yearly rainfall was 662 mm, ranging from 542 mm (2007) to 888 mm (2010). The average yearly actual evapotranspiration was 470 mm (71% of the rainfall). This value is comparable to those reported by Romanazzi et al. (2015). The average yearly surface runoff and total water yield (Surface runoff + baseflow + lateral flow) were 86 mm (17% of the rainfall) and 180 mm (27% of the rainfall), respectively.

The sediment load was overestimated in the calibration period (PBIAS -2.8%) and underestimated in the validation period (PBIAS +5.1%). The simulated sediment loads generally matched the observed data, except some peaks that were underestimated (Figure 6). The highest sediment peak, which was recorded in November 2010, was underestimated by the model (49.5%; simulated 1.52 t ha⁻¹; observed 3.07 t ha⁻¹). Soil erosion in the area mainly occurs in winter (December to April) due to frequent rainfall events. The average annual specific sediment load was 5.95 t ha⁻¹ yr.⁻¹, ranging from 1.46 t ha⁻¹ yr.⁻¹ (2007) to 8.71 t ha⁻¹ yr.⁻¹ (2010).

The streamflow simulated by the SWAT model by using daily observed climatic data (2000–2020, baseline) and model projections (CMCC, EC-EARTH, and MPI, respectively) (2030–2050, NF) were compared. The climatic models adopted in this study predicted a reduction in the average yearly rainfall compared to the baseline. Specifically, at the basin scale, rainfall was 678.5 mm for the baseline, 386.5 mm for CMCC (−43%), 602.0 mm for EC-EARTH (−11.3%) and 529.0 mm for MPI (−22%). A reduction of the actual evapotranspiration (477.8 mm yr.⁻¹, baseline) was predicted in the future (305.2 mm yr.⁻¹ for CMCC, 420.3 mm yr.⁻¹ for EC-EARTH and 396.2 mm yr.⁻¹ for MPI). Similarly, a reduction of the total water yield (Surface runoff + baseflow + lateral flow) was simulated (from 168.44 mm yr.⁻¹, in the baseline, to 54.77 mm yr.⁻¹, 156.23 mm yr.⁻¹ and 109.33 mmyr⁻¹ for CMCC, EC-EARTH and MPI, respectively).

To analyze the variability of the flow regime components under climate change, some Indicators of Hydrological Alterations (IHAs) were assessed (Richter et al., 1996; Leone et al., 2024). The mean flow at the outlet, which was equal to 0.89 m³s⁻¹ for CMCC, 2.48 m³s⁻¹ for EC-EARTH and 1.74 m³s⁻¹ for MPI, compared with the baseline (2.66 m³s⁻¹) showed an important contraction. Similarly, a contraction of the magnitude of monthly flow for the wet season (from December to April) was predicted for the three climate model projections (Figure 7). CMCC model predicted the largest reduction of flow that was quantified for the wet months at 81%. The highest reduction (86.3%) was simulated in February (median flow was 2.67 m³s⁻¹ for the baseline and 0.365 m³s-1 for CMCC). The average reduction of monthly flow was 21.5% for EC-EARTH and 47.0% for MPI. The highest reduction (51.2% EC-EARTH and 57.4% MPI) was predicted in March when the median value was 2.89 m³s⁻¹for the baseline and 1.41 m³s⁻¹ and 1.23 m³s⁻¹ for EC-EARTH and MPI, respectively. Similar behavior can be evidenced for the 10th and the 25th percentile. Considering the 90th percentile, both EC-EARTH and MPI showed higher values than the baseline. Specifically, EC-EARTH overtake the baseline in January (baseline =3.91 m³s-1, EC-EARTH = 4.02 m³s⁻¹), February (baseline = 4.33 m³s⁻¹, EC-EARTH = 4.47 m³s⁻¹) and December (baseline = 4.16 m3s-1, EC-EARTH = 5.44 m3s⁻¹), while MPI showed high values in April (baseline = 4.60 m3s⁻¹, MPI = 4.76 m³s⁻¹). These increases in monthly streamflow can be attributed to flood events.

Changes in the annual extreme streamflow of 1-, 30-, and 90-day duration were analyzed (Figures 8A,B). CMCC and MPI showed similar values of these indicators, with a general reduction compared to the baseline. The depletion of the median value of the daily streamflow was 43.7 and 38.8% for CMCC and MPI, respectively. Reductions in the 90-day maximum flow were predicted, especially by CMCC and MPI (4.46 m³s⁻¹ for the baseline, 1.85 m³s⁻¹ for CMCC, and 2.5 m³s⁻¹ for MPI; Figure 8B). CMCC showed values lower than MPI and EC-EARTH (10th, 25th, and 75th percentile in Figure 8). Reduction in 1-day maximum flow is predicted by the CMCC model (i.e., 10th and 25th percentiles were 33.7 m³s⁻¹ and 39.8 m³s⁻¹ for the baseline, and 0.64 m³s⁻¹ and 7.57 m³s⁻¹ for future, respectively) (Figure 8A). Contrarily, EC-EARTH showed the median values of the three IHAs higher than the baseline and a larger interannual variability (i.e., larger interquartile range) (Figures 8A,B). An increment of 35.7% was predicted for the median value of 1-day maximum flow (median value and the 75th percentile were 50.3 m³s⁻¹ and 75.3 m³s⁻¹ for the baseline, and 68.3 m³s⁻¹ and 104.9 m³s⁻¹ for EC-EARTH, respectively) (Figure 8B). EC-EARTH showed a higher occurrence of high values of 30- and 90- days maximum flow, with respect to the baseline (Figures 8A,B). These results could be due to the increase of occurrence of large floods predicted by EC-EARTH.

Changes in the specific sediment load (t ha⁻¹) at the basin scale and in the sediment yield (t ha⁻¹) at the sub-basins scale were assessed by comparing the daily sediment simulated data by the SWAT model for the baseline and for the future by using the three climate model projections (CMCC, EC-EARTH, MPI). A reduction of the specific sediment load (Table 4) associated with the reduction of the streamflow was predicted. At the annual scale, the specific sediment load simulated for the baseline was 6.84 t ha⁻¹ yr.⁻¹, whereas CMCC predicted a reduction of 52.8% (3.23 t ha⁻¹ yr.⁻¹), EC-EARTH of 18.1% (5.60 t ha⁻¹ yr.⁻¹) and MPI of 41.7% (3.99 t ha⁻¹ yr.⁻¹). At the monthly scale, the specific sediment load for the baseline was generally higher than the future projection. However, in July, CMCC (0.35 t ha⁻¹ yr.⁻¹) and MPI (0.24 t ha⁻¹ yr.⁻¹) showed simulated values higher than baseline (0.07 t ha⁻¹ yr.⁻¹), EC-EARTH showed sediment yield higher than baseline in April (baseline = 0.19 t ha⁻¹ yr.⁻¹ and EC-EARTH = 0.30 t ha⁻¹ yr.⁻¹), in May (baseline = 0.08 t ha⁻¹ yr.⁻¹ and EC-EARTH = 0.29 t ha⁻¹ yr.⁻¹), in June (baseline = 0.10 t ha⁻¹ yr.⁻¹ and EC-EARTH = 0.12 t ha⁻¹ yr.⁻¹), November and December.

A visual comparison was carried out to analyze the spatial distribution of soil erosion risk areas at the sub-basin scale between the baseline (2000–2020) and NF (2030–2050) (Figure 9). To better draw the maps, the annual sediment yield values were divided into six classes, considering the ranges obtained for the baseline (0.01 to 38.6 t ha⁻¹). In the baseline 33 of the total 115 sub-basins, whose HRUs are characterized by the presence of wheat cultivation and slopes >15% (Figure 1), were under high erosion rate conditions (average > 10 t ha⁻¹) (Figure 9). Specifically, sub-basins N. 59 and N. 60, whose dominant HRUs are wheat and the slope is close to 20%, showed the highest sediment yield values, respectively 30.6 t ha⁻¹ and 38.64 t ha⁻¹. Flat areas instead had an average rate of sediment yield <3 t ha⁻¹.

The model projections analyzed here showed a general reduction of the annual sediment yield: CMCC ranged from 0.01 to 22.8 t ha⁻¹ (average reduction 53%), EC-EARTH from 0.01 to 30.4 t ha⁻¹ (average reduction of 12%), and MPI from 0.01 to 25.3 t ha⁻¹ (average reduction 42.5%). For CMCC the number of sub-basins under high erosion rate conditions (average > 10 t ha⁻¹) is decreased to 5, while for MPI to 10. Several sub-basins in steep slope areas (e.g., central and eastern parts of the basin) changed class from 10 to 30 t ha⁻¹ to 5–10 t ha⁻¹ (Figure 9). Sub-basins N. 59 and N. 60 in Figure 9 showed a decrease of about 41% (18.3 t ha⁻¹ and 22.4 t ha⁻¹, respectively) for CMCC, and 34% (20.1 t ha⁻¹; 5.3 t ha⁻¹) for MPI. EC-EARTH showed a distribution of the high-risk erosion areas similar to the baseline: only 10 sub-basins changed class from 10–30 t ha⁻¹ to 5–10 t ha⁻¹ (Figure 9). Indeed, the number of sub-basins under high erosion rate conditions (average > 10 t ha⁻¹) remained high (23). Sub-basins N. 59 and N. 60 showed a decrease of about 21% (N. 59 24.27 t ha⁻¹; N.60 30.43 t ha⁻¹).

The present work confirmed that the SWAT model can simulate the complex hydro-sedimentary response in the Mediterranean environment (Borrelli et al., 2021; Ricci et al., 2022) and the climate change impacts on hydro-sedimentary response. The statistical performances for the streamflow and the sediment load simulations were satisfactory for the calibration period (2007–2009). For the validation period (2010–2011), the statistical indices showed lower values, especially for the streamflow, which was overestimated. This can be attributed to errors in predicting low flow after the biggest flood event, recorded in November 2010 (Figure 2), which caused an increase in the CN values due to the high antecedent soil water conditions (Ricci et al., 2020; Dash et al., 2023). Another source of uncertainty, which can lead to discrepancies between observed and simulated data, can be represented by the not-optimal location of the weather station, associated with the convective character of rainfall events (Mtibaa and Asano, 2022; Ricci et al., 2022). In the validation period, SWAT underestimated the sediment load, especially the main observed peak (November 2010). This can be related to the MUSLE equation, which computes average soil loss values and does not consider some processes, such as streambank erosion, which are very common during peak flows in the Mediterranean environment (Panda et al., 2021). Ricci et al. (2020) pointed out that particular attention must be reserved when defining the sub-basins and HRUs in the Mediterranean areas to improve the model performances. This work confirmed the importance of the sub-basin delineation (Chen et al., 2021).

The identified limitations in SWAT’s performance during the baseline period may have important implications for interpreting future scenario predictions. Several key uncertainties should be considered. The model’s tendency to underestimate peak sediment loads during extreme events suggests that future projections may similarly fail to capture the full extent of sediment transport during intense rainfall. Additionally, the simulated overestimation of streamflow following a major flood, due to the high antecedent soil water content, suggests that post-event flow conditions could be overestimated in future scenarios. Another limitation is SWAT’s inability to effectively model streambank erosion processes, which are common in the study area peak flows. Hence, as a result, total sediment yield may be underestimated.

The model projections selected in the present study (MPI, EC-EARTH, CMCC) showed differences in rainfall prediction (coefficient of variation = 0.22), whereas they showed a good agreement in temperature increase. These results confirmed previous studies that recognized Global Circulation Models as the largest sources of uncertainty when analyzing the impact of CC on water resources (Vetter et al., 2017; Hattermann et al., 2018). Hence, the results of the SWAT model simulations showed differences in the magnitude of annual (Coefficient of variation = 0.47) and monthly streamflow (Coefficient of variation up to 0.67 in June and December) among the model projections. While uncertainty affects the results, the climate projections indicated that CC will have important implications on water resources with alterations in flow and sediment regimes. Specifically, reductions in streamflow were predicted by the model projections (up to 67% on a yearly basis) with a shift toward drier conditions. Among the model projections, CMCC forecasted the worst impact on water resources (−67%) whereas the EC-EARTH (−7%) showed a minor deviation compared to the baseline. The highest deviation from the baseline of the maximum streamflow for different durations (90-day max; 30-day max) was detected for the CMCC model and the lowest was found for the EC-EARTH. Concerning the maximum streamflow for one-day duration (1-day max), EC- EARTH showed an increase of the high values (+16% on average) compared to the baseline, contrarily CMCC showed a reduction of the highest values (−35%). MPI model projection also showed a negative rate of change in the mean annual 1-day max (−28%) and it generally showed an intermediate behavior compared to the CMCC and EC- EARTH for all the IHAs.

Concerning the sediment regime, overall the three model projections showed a reduction in annual sediment load (52.8% for CMCC, 18.1% for EC-EARTH, and 41.7% for MPI). A monthly redistribution of sediment transport can be highlighted. An average reduction of 60% was predicted by the models in January, March, and October. In July a remarkable increase in sediment load under CMCC (0.35 t ha⁻¹ yr.⁻¹) and MPI (0.24 t ha⁻¹ yr.⁻¹) compared to baseline (0.07 t ha⁻¹ yr.⁻¹) was simulated. The EC-EARTH predicted an increase in sediment load in April, May, and November, with respect to the baseline. Specific sediment loads ranged from 0.3 t ha⁻¹ yr.⁻¹ to 1.09 t ha⁻¹ yr.⁻¹ for EC-EARTH and from 0.08 t ha⁻¹ yr.⁻¹ to 0.68 t ha⁻¹ yr.⁻¹ for the baseline. Slopes and wheat cultivations were the main factors influencing the sediment yield production at the sub-basin level. Model projections showed a general reduction of the annual sediment yield and different spatial distributions of high-risk erosion areas. The sub-basins under a high erosion rate decreased. This change involved mainly the sub-basins located in the central and eastern parts of the basin, where the average slope ranges from 15 to 18%. EC-EARTH showed a minor reduction (12%) and a distribution of the high-risk erosion areas similar to the baseline. Some persistent hotspots can be highlighted. Sub-basins with slopes >18% (e.g., 49, 59, 60, 66, 108) maintained high erosion rates (>10 t ha-1) under all climate models.

These results confirmed previous studies analyzing CC impacts in basins under the Mediterranean climate. Indeed, a reduction in mean annual values of maximum streamflow for different duration was predicted for the Celone River basin (Italy; De Girolamo et al., 2022), for Ben Guerir River basin (Brouziyne et al., 2021) and the Merguellil basin (Tunisia; Abouabdillah et al., 2010) for the NF. A SWAT+ model-based study on the Sulcis watershed (Sardinia, Italy) (Pulighe et al., 2021) predicted notable reductions in water yield, surface runoff, groundwater recharge, and baseflow by the late 21st century under both RCP4.5 and RCP8.5 scenarios, primarily driven by increased temperatures and decreased precipitation, leading to enhanced evapotranspiration losses. Blöschl et al. (2019), who analyzed changing climate in Europe, found a negative rate of change in the mean annual flood discharge (highest peak discharge recorded per year) per decade in Candelaro River (Apulia, Italy), from 12 to 24%. On the other hand, the increment of the maximum streamflow for 1 day, evidenced especially by the EC-EARTH model, can be related to an increase in PCP intensity, as already stated by other authors (Noto et al., 2023; Ghanghas et al., 2024). This can lead to heightened soil erosion (Borrelli et al., 2020).

The findings of this study underscore that climate-driven changes are not merely prospective but are actively manifesting, necessitating immediate integration of these considerations into water resources management and territorial-environmental planning frameworks. The reduced water availability is only one of the implications of the CC on water resources. This study showed that hydrological and sediment regime alterations are expected in the future. As a result of the reduced magnitude of the annual flow, 1-day maximum flow, 30-, and 90-days maximum flow, river morphology and physical habitat conditions could change with several implications on the integrity of river ecology (Arthington, 2012; Todaro et al., 2022). For instance, species diversity, abundance and richness could be reduced, and non-native species could become dominant (Poff and Ward, 1989). Changes in nutrient cycles and transport could occur as pointed out by Ramião et al. (2024). The authors detected a decline in nutrient exports in a humid Portuguese Mediterranean catchment due to reduced runoff and soil erosion from decreased rainfall, reinforcing the link between climate change, sediment transport, and nutrient cycling. However, the potential for extreme hydrological variability remains. Indeed, studies on Mediterranean flood risk (Tramblay et al., 2023), found increasing trends in extreme precipitation-driven flash floods despite declining annual rainfall. The Carapelle River basin, as reported in the General Reclamation plan of the Capitanata areas (Consorzio di Bonifica Della Capitanata, 2022), is already experiencing significant water scarcity with declining groundwater levels from intensive agricultural extraction and insufficient surface water supplies (https://www.bonificacapitanata.it/schemi-idrici/ accessed on 7/2/2025). While plans exist for building the Palazzo d’Ascoli reservoir (67 million m³ capacity) to supplement water resources, climate change is expected to exacerbate the competition for water resources in the future.

Hence, water resources management options and mitigation measures (i.e., wastewater reuse, precision agriculture) are needed urgently to address the challenge of maintaining the river ecosystem and meeting the growing demand for freshwater resources. This study can provide to water manager’s specific details about the areas within the basing at risk of erosion also in the CC context. Management strategies, such as Best Management Practices (BMPs) or Nature Based Solutions (NBSs), should be promoted to achieve the goals of the EU Soil Strategy of the European Green Deal and of the Nature Restauration Law, which aim to address land degradation and have all EU ecosystems in healthy conditions by 2050 (Panagos et al., 2022; EU, 2024). No-tillage, contour farming, and cover crops are agricultural practices, already implemented in European countries, which contribute to maintaining a minimum soil cover and reduce the number and the depth of the tillage thus reducing sheet and rill erosion (Panagos et al., 2016; Ricci et al., 2020; Fendrich et al., 2023). In the proximity of the stream network and steep slope areas, riparian buffers, buffer strips, or land use changes in permanent crops, such as pasture or forest, are recognized as effective measures (Englund et al., 2021; Cole et al., 2020; Ricci et al., 2020). Related to this, it is important to highlight that BMPs and NBSs are multifunctional by definition, hence their implementation can lead to several other positive outcomes for the Agro-environmental ecosystem such as improved water quality, increased nutrients in the soil (Ricci et al., 2022; Netti et al., 2024) or increase the net greenhouse gas removal (Ran et al., 2018; Montanarella and Panagos, 2021). However, social, attitudinal, and other contextual factors, including the small size of farms, and the significant investments from the private and public sectors required for their implementation led to the failure of adopting BMPs or NBSs (Panagos et al., 2016; Brown et al., 2021; Ricci et al., 2020). To enhance the adoption of BMPs and NBSs, the most recent EU Common Agricultural Policy (CAP) 2023–2027 adopted the concept of eco-schemes, or a program of funding addressed to each farmer who voluntarily implements climate-sensitive and nature-sensitive practices (Panagos et al., 2022).