3.1. Sub-Basins Selected for Calibration
As stated in the modelling protocol, the criterion for selecting outlets for calibration was that they had an unsatisfactory rating in at least one of the objective functions presented in Table 2
over the simulation period (1984–2015). Figure 4
illustrates the sub-basins used for calibration (top figure), and examples of the temporal evolution of observed and simulated monthly discharge, as well as the values of performance indicators. Considering the 78 outlets selected, 23 performed satisfactory or better in all objective functions following the classification suggested by Moriasi et al. [46
] and Thiemig et al. [47
]. Hence, these sub-basins outlets were not used for the calibration process. It is clear that most of the outlets that performed well are located in the southern parts of the basin, especially in the Iguaçu sub-basin (VI) and adjacent areas of the Paraná (IV) and Paranapanema (V) sub-basins. This goodness-of-fit between measured and simulated discharge is mainly due to a large number of precipitation stations that are located over the sub-basins, with a low percentage of missing data (see Figure 2
). For instance, in the streamflow of the Upper Iguaçu River (Figure 4
b), the model has a good representation of the average, minimum and maximum discharge values. Regarding the statistical indices, SWAT has provided more than satisfactory results with 7.8 (very good), 0.77 (very good), 0.74 (good), 0.86 (good), and 0.51 (good) for
, respectively. The remaining figures of the temporal evolution of the outlets that yielded good performance are available in Figure S1 in the Supplementary Materials
3.2. Calibration and Validation Performance
a–e show the spatial distribution of the values of the objective functions used to evaluate the goodness-of-fit of measured discharge data estimated by SWAT. The performance of the monthly simulations for the calibration (1984–2004) and validation (2005–2015) period ranged from very good to unsatisfactory. It is clear that after the calibration process, the model has a good representation of monthly discharge values for most of the outlets of the UPRB that are located mainly in the Grande (II), Tietê (III), Paranapanema (V), and Iguaçu (VI) sub-basins. On the other hand, Paraná (IV) and Paranaiba (I) were the sub-watersheds that had the highest number of outlets with unsatisfactory simulations. This can be attributed to the low density of rain gauges mainly on the Ivinheima and Sucuriú river basins located on the Paraná sub-basin (IV), on the western side of the basin.
The indices R2
a) and PBIAS (Figure 5
b) present the best hydrological performance for all sub-basins, with 92% and 86% of the outlets showing satisfactory or better performances. For R2
, 61 (78%) of the outlets performed better than satisfactory with values of up to 0.91 over the Paraná and Sapucaí rivers. Similarly, the PBIAS index gave more than half of the outlets (63%) a better than satisfactory rating.
The rating of the KGE index (Figure 5
d), which is based on the equal weighting of three different components (correlation, bias, and variability), shows that 88% of the outlets performed better or equally satisfactory. Only 9 outlets produced unsatisfactory simulations. The maximum value obtained for KGE was for the Grande river with 0.87.
Finally, the NSE (Figure 5
c) and RSR (Figure 5
e) indices were those with the highest number of outlets with unsatisfactory simulation. However, the percentage of satisfactory stations was still high. Considering NSE > 0.5 or RSR < 0.7 for a satisfactory simulation, the model reached this criterion in 76% and 74% of the outlets, respectively. One of the reasons that explain why these indices performed slightly below the others is the low quality of the simulations of the base flow. This limitation is underlined by previous studies that evaluated the hydrological routines of the SWAT model [51
]. SWAT simulates two types of aquifers: shallow (unconfined) aquifers, which contribute to return flow to streams within the catchment, and deep (confined) aquifers, which are responsible for the flow outside the basin (amount of water used, for example, for irrigation and water supply) and are considered water sinks in the system [26
]. Once the model calculates the groundwater, studies that present difficulties in representing transfers associated with these types of water may present an unsatisfactory performance for the base flow prediction with the SWAT model. For instance, Srivastava et al. [52
] found a NSE value of −0.16 in the predictions of monthly base flows. Similarly to the current study, Wu and Johnston [53
] simulating long-term periods found it difficult to simulate dry seasons with the model. In this case study, the SWAT model performed better in simulating wet seasons than dry seasons.
shows the comparison between the observed data and simulated values for the temporal evolution of the monthly discharge in the calibration and validation (1984–2015) period. The plots show the final outlets of the main rivers of the UPRB. Even though the model did not have a good estimate of the discharge at some outlets in the basin, these did not have a significant effect on the final outlet of the main rivers of the basin, due to their contribution area. This could be explained by the difference among the magnitudes of discharges. For instance, a closer examination of the long-term monthly mean discharge at the final outlets of the rivers shows that the Paranaíba river has a discharge of 2465 m3
, while the Da Prata river, one of its tributaries has an average discharge about 71 m3
, which represents 3% of Paranaiba river. The fact that the simulation for the Da Prata river performed an unsatisfactory simulation in R2
, NSE, and RSR indices did not impact the quality of the performance of the Paranaíba. Similar cases occur in other major rivers of the UPRB. Figure S2, in the Supplementary Materials
, shows the remaining graphs of the temporal evolution of the discharge on outlets after the calibration process.
shows the objective function values from the final outlets of the main rivers for the calibration (1984–2004) and validation (2005–2015) periods. PBIAS ranged from satisfactory to very good simulation both for calibration (mean = −7.86) and validation (mean = −15.5) for the five rivers. High values of R2
, greater than 0.80, were found in both calibration and validation results, indicating a very good correlation between the monthly observed and simulated discharges. In the calibration period, the NSE and KGE ranged from 0.56 (satisfactory) to 0.73 (good), and from 0.55 (satisfactory), to 0.77 (good), respectively. In the validation period, the NSE and KGE ranged from to 0.51 (satisfactory) to 0.73 (good), and from 0.55 (satisfactory), to 0.67 (satisfactory), respectively. Regarding the RSR index, values between 0.52 (good) and 0.66 (satisfactory) were found during the calibration process. For the validation process, only the Paraná river represented an unsatisfactory simulation with RSR value of 0.71. The remaining objective functions values of the outlets over the UPRB can be found in the Supplementary Materials (Table S2)
As a whole, the calibration and validation of the outlets of UPRB provided promising results as indicated by acceptable values of statistical indices. The performance is better or comparable to other SWAT applications over Brazilian watersheds. For instance, Creech et al. [54
] reported NSE ranging from 0.42 to 0.75 and from 0.42 to 0.77 for monthly discharge calibration and validation periods of the São Francisco River, the largest basin in the northeast of Brazil. On the other hand, considering small basins, Rocha et al. [55
], modelling São Bartolomeu Stream Watershed, showed values of NSE and R2
indices between −1.19 and 0.91, and 0.22, and 0.96, respectively. In addition, the results presented here agree with the range found in previous works where SWAT was calibrated for large basins worldwide. For example, Pagliero et al. [56
] estimated the monthly flow for representative regions of the Danube basin found NSE ranging from 0.22 to 0.75 and R2
ranging from 0.68 to 0.88. Another study performed by Easton et al. [57
] for the Upper Blue Nile Basin showed values of R2
ranging from 0.73 to 0.92 and NSE from 0.53 to 0.92.
One of the strengths of the current work is that the simulation was performed at a high spatial resolution, with the basin being divided into 5187 sub-basins and further into 44,635 HRUs. In addition, the project was built for a long-term simulation over 37 years (1979–2015). These spatial and temporal resolutions were not found in previous studies of large-scale SWAT applications. For instance, Jha et al. [58
] simulated the streamflow of the Upper Mississippi River, which has an area around 447,500 km2
, and discretized the basin into 119 sub-basins. These represent around 3760 km2
of the average sub-watershed area, compared 179 km2
for the current study basin. Pagliero et al. [56
] defined 4663 HRUs (10% of our HRUs basin) over the Danube Basin, which has a drainage area of about 803,000 km2