Figure 2: Physical characteristics of the analyzed watershed: a) morphology, b) land use classification (CLC, 2018), and c) soil types (DGADR, 2013).

Climate dataset

The time-series for calibration and validation purposes, including daily precipitation and max/min temperatures, were obtained from the dataset “Iberia01” (Herrera et al., 2012, 2016), which provides a dense network of stations over the Iberian Peninsula. All the data were extracted in seven grid points located inside or in the proximity of the analyzed basin. The high reliability of the spatial pattern of the reported variables is discussed by Herrera et al. (2019). These climate data were utilized for calibration and validation procedures as they overlap the simulation’s period (1985-1989) of available data for streamflow and sediment transport.
As a reliable representation of precipitation’s intensity and distribution is one of the predominant factors affecting the simulation results of hydrological processes, the choice of a representative climate dataset for the future scenarios is crucial to obtain accurate and reliable SE estimates. The datasets contemplated are based on currently available Regional Climate Models (RCMs), forced by different Global Climate Models (GCMs), which were used in the Fifth Assessment Report (AR5) of the Intergovernmental Panel on Climate Change (IPCC). The datasets were made available by the World Climate Research Program’s CORDEX initiative (www.euro-cordex.net). In this study, the selected climate model, here called KNMI (i.e., RACMO22E driven by ICHEC-EC-EARTH), is in agreement with Soares et al. (2017), who assessed the performance of EURO-CORDEX historical (HIST) simulations to represent temporal and spatial patterns of precipitation over Portugal. Data were extracted from the RCM within 7 grid points, covering the period from 2020 to 2040.

2.3 Modelling Framework

SWAT model has the main purpose of simulating and predicting the effects of land management practices and CC on the hydrological cycle using different timescale input/output (daily, monthly, and yearly). The main step of the model consists in the creation of the hydrologic response units (HRUs), referring to all those portions of a territory characterized by unique land use, morphological, and soil attributes combination (Neitsch et al., 2000). All the model’s outputs such as runoff, evapotranspiration, aquifer recharge, sediment, and nutrient loadings from each HRU are obtained using the input of climate, soil properties, topography, vegetation, and land management practices and further summarized to obtain the sub-basins loading. The outputs of runoff and sediment yield are calculated using a modified version of the curve number method (Brakensiek et al., 1984; USDA, 2004) or the Green–Ampt infiltration method, and the Modified Universal Soil Loss Equation (MUSLE) (Williams, 1995), respectively.
The SWAT model for the Guadiana river sub-basin was realized using ArcSWAT interface on ArcGIS 10.2 environment. The regime of SE for the whole watershed was evaluated with sequential steps.
First, all the physical characteristics of morphology, land cover, and soil properties were evaluated and used as main inputs for the setup of the SWAT model and to define the HRUs spatial distribution in the whole basin. The watershed automatic delineation tool divided the area of 37,233 km2 in 99 sub-basins further discretized in 3000 HRUs. The HRUs were obtained intersecting five slope classes (<5, 5-10, 10-15, 15-20, >20), seven land cover categories (Fig. 2b), and six soil groups (Fig. 2c). The CLC classification was reclassified to match with the vegetation cover types present in the SWAT default database. Specifically, URMD was used to describe artificial settlements, OATS and OLIV for rainfed crops and olive plantations respectively, FRSE for rainfed forest, PINE for coniferous and mixed forest, while the “montado” system was represented using a combination of OAK and PAST (30% and 70%, respectively) using the SWAT code WPAS. Some parameters characterizing the typical Mediterranean vegetation were updated according to Nùnes et al. (2008). Concerning the soils’ properties, all the information about Ks, AWC, texture, soil organic carbon (SOC), BD, and soil albedo utilized for the simulation are resumed in Table S1.
The meteorological data are the other necessary input for a proper SWAT evaluation. For this model, the meteorological data from “Iberia01” were used to run the model using daily data of precipitation and temperature to estimate runoff and actual evapotranspiration (AET) via the Hargreaves formula (Aschonitis et al., 2017). After the set-up procedure, the SWAT simulation, performed monthly, was divided into three blocks: i) a warm-up period of four years (1980-1984), ii) a calibration procedure from 01/1985 to 06/1987, and iii) a validation phase for the period 07/1987-12/1989 on four hydrometric stations for streamflow. For SE only two hydrometric stations were available (Monte da Ponte and Oeiras) with scattered daily data for the period 1984-1989. In particular, 38 daily data of SE were available, which were evenly split to perform calibration and validation analysis. Finally, the SWAT model was forced to simulate the HIST (1980-2000) and future period (2020-2040) using the chosen RCM (KNMI) under two different emission pathways (RCP 4.5 and 8.5).
The robustness of the methodology was investigated through an extensive calibration/validation procedure. A preliminary “trial and error calibration” was applied along with an “auto-calibration and validation”. In the trial and error calibration, the fitted values of some parameters, such as groundwater ”revap” coefficient (GW_REVAP), deep aquifer percolation fraction (RCHGDP), soil evaporation factor (ESCO), plant uptake compensation (EPCO), and deep aquifer percolation fraction (RCHGR_DP) were manually adjusted considering the results obtained from Nùnes et al. (2017) which performed a SWAT application in a nearby watershed. The other parameters responsible for streamflow and sediment load processes utilized for the auto-calibration were identified through an extensive literature review (Chen et al., 2019; Khelifa et al., 2017; Serpa et al., 2015), and are reported in Table 2. The standalone software SWAT-CUP via the Sequential Uncertainty Fitting version 2 (SUFI-2) algorithm (Abbaspour, 2015) was used for the auto calibration/validation. During the calibration phase, the SUFI-2 algorithm tries a different combination of the chosen parameters within their fixed range of variation (generally ± 25% of the initial value) and calculates the effect on the various fitting between the observed and simulated variables.
Three well known statistical indices were utilized to evaluate the simulation’s results according to the performance values suggested by Moriasi et al. (2007) (Table S3): i) the coefficient of determination (R2), ii) the Nash-Sutcliffe efficiency (NSE), and iii) the percent of bias (PBIAS); while the P-factor and R-factor values were investigated to account for model fit and uncertainties (Abbaspour et al., 2004).
A total of three thousand calibration runs, divided in six interactions of five hundred runs each, were performed until a satisfactory calibration was obtained. The values of SE in t/ha/year were obtained for the whole watershed. Furthermore, the values of SE for each HRU were spatialized and classified to obtain three SE susceptibility maps as to highlight the main changes over time and to identify those areas where SE will increase/decrease in the future. Finally, the results were analyzed to identify the mean yearly SE rate for each land cover category identified in the watershed.