We used Hydrological Predictions for the Environment (HYPE) version 5.9, a semi-distributed hydrological model, for the assessment of water resources and water quality at small and large scales (Lindström et al., 2010) in order to simulate the hydrological response of Costa Rican catchments. The HYPE model can be considered as the evolution of the distributed Hydrologiska Byråns Vattenbalansavdelning (HBV) model (Lindström et al., 1997). HYPE was developed by the Swedish Meteorological and Hydrological Institute (SMHI) as the operative model for drought and flood forecasting across Sweden (Pechlivanidis et al., 2014). Moreover, HYPE was recently applied to other climatic regions (Andersson et al., 2017; Arheimer et al., 2018; Berg et al., 2018; Lindström, 2016; Pugliese et al., 2018; Tanouchi et al., 2019), including a global-scale application (Arheimer et al., 2020).

The HYPE model allows simulating the water balance and nutrient fluxes at a daily or sub-daily scale using precipitation and temperature as forcings (SMHI, 2018). The model structure (Fig. 2a) describes the major water pathways and fluxes, ensuring mass conservation at the catchment and sub-catchment scale. Furthermore, each sub-catchment is divided into the most fundamental spatial soil and land use classes (SLCs) depending on the classification of soil types, land cover, climate, and elevation, as shown in Fig. 2b. The SLCs in HYPE provide the capability to predict streamflows in ungauged basins since the parameters that regulate the fluxes and storages are linked to each SLC, with a maximum of three layers of different soil thickness, as shown in Fig. 2b. Water bodies such as lakes and rivers may be considered as an SLC, where lakes can be defined as natural lakes or regulated dams with multiple water outputs. For full details of the HYPE model, see the description by Lindström et al. (2010) and the open-access code references located at https://hypeweb.smhi.se/model-water/ (last access: 10 June 2021).

For this study, Costa Rica was divided into 605 catchments (Fig. 1a) with 12 SLCs obtained from the spatial combination of soil types and land cover maps shown in Fig. 1b and c, respectively. Outlet lakes (which discharge to downstream catchments) and internal lakes (which discharge into the main river or tributaries) were set up as different SLCs to consider the water bodies that regulate the streamflow. The largest water body in Costa Rica is the Arenal reservoir, located in the San Carlos River catchment (Fig. 1b). The Arenal reservoir is an artificial lake used for hydropower purposes with an average surface area of 87.8 km2 and a depth that ranges from 30 to 60 m. The Arenal reservoir was implemented as a natural lake since its operational rules are confidential and therefore unknown.

Soil thickness varied for different SLCs, with a maximum soil thickness of 3 m under forests and a minimum of 2 m for bare soil cover, following Arheimer et al. (2020). Furthermore, delimited catchments were classified according to their elevation and location (Pacific Basin and Caribbean Basin), applying regional factors to correct the hydrological behavior of lowland and mountainous catchments with similar SLCs, resulting in six defined regions.

Daily time series of precipitation from CHIRPS and temperature from NOAA for the period 2000–2014 were extracted for each catchment using GRASS GIS (Neteler et al., 2012), where datasets were resampled to 1 km using the nearest-neighbor criteria and spatially averaged for each catchment. The climatological forcings were resampled due to the small size of some catchments (area of ∼1 km2). Arheimer et al. (2020) recommended computing the average of the nearest grids to obtain the forcings instead of deriving the data from the nearest pixel.

Rainfall estimations from satellites are subject to systematic errors that may produce uncertainty in hydrological simulations (Goshime et al., 2019; Grillakis et al., 2018; Infante-Corona et al., 2014; Wörner et al., 2019). The CHIRPS product already incorporates a bias correction procedure but uses only a few concentrated ground stations in Costa Rica. The performance of CHIRPS estimating annual water balances is shown in Fig. S1a and c in the Supplement, with frequent underestimation in the monitored catchments. Therefore, we applied a linear scaling to further correct for the bias between the product and ground precipitation from 75 available stations across Costa Rica (Fig. 1a). The corrected precipitation was estimated as 1CHIRPSc(t)=CHIRPS(t)⋅BF, where CHIRPSc is the bias-corrected precipitation at time t, CHIRPS is the original precipitation at time t, and BF is the bias factor. The bias factor was estimated as 2BF=μ(P)μ(CHIRPS), where μ(P) is the mean of the historical precipitation from the ground stations, and μ(CHIRPS) is the mean of the historical precipitation from CHIRPS. Note that μ(P) and μ(CHIRPS) were computed using the common study period. The simple linear bias correction was preferred over more complex methods due to the lack of a long common period for all stations. Therefore, we used the individual records of more than 60 available stations covering a period from 1980 to 2010 to better capture the complex topography and resulting rainfall patterns.

Some monitored catchments exhibiting higher annual streamflow than annual precipitation could not be corrected due to groundwater contributions from neighboring catchments (Genereux and Jordan, 2006; Genereux et al., 2002), under-catch at rainfall gauges (Frumau et al., 2011), and the insufficient number of precipitation stations to correct the CHIRPS database at a national scale. Nevertheless, errors in climatological data have been found to be the most common issue for water balance modeling in Central America (Westerberg et al., 2014; Birkel et al., 2012). In that sense, an additional approach was implemented to reduce the unrealistic relationship between streamflow and precipitation, which consisted of the creation of virtual stations at the catchment centroid, where the new bias factor was computed as 3BF2=μQty+μAETyμCHIRPSy, where μ refers to the mean value, Qty is the annual streamflow from 1990 to 2003, AETy is the MODIS annual actual evapotranspiration from 2001 to 2014, and CHIRPSy is the CHIRPS annual precipitation from 1990 to 2014. BF2 adjusts the long-term precipitation volume to ensure that the water balance is preserved, avoiding underestimation of streamflow and evapotranspiration. Four streamflow gauges in addition to those shown in Table 1 were used to cover additional spatial area at high elevations for the correction of satellite-based precipitation. These streamflow gauges were omitted from the model calibration–validation procedure due to their shorter records (less than 7 years). The location of the four streamflow gauges and their catchments are shown in Fig. 1b.

Finally, BF points from precipitation stations and BF2 from virtual points were interpolated using the inverse distance weighted (IDW) method (Shepard, 1968), with an exponent value of 2 at the original CHIRPS resolution. The interpolated map of the bias factor was used to spatially correct the time series of CHIRPS across Costa Rica.

HYPE incorporates four methods for PET estimation (SMHI, 2018). After initial tests, we found that the monthly PET signal from MODIS in Costa Rica can be reproduced by only using temperature as forcing, where PET is computed as 4PET=cevp⋅cseason⋅temp-ttmp⋅1+cevpcorr, where PET is the daily potential evapotranspiration (in millimeters), temp is the daily mean air temperature (∘C), cevp is an evapotranspiration parameter that depends on the land use (mm ∘C-1 d-1), ttmp is a threshold temperature for evapotranspiration (∘C), cevpcorr is a correction factor for evapotranspiration, and cseason is a factor computed as 5cseason=1+cevpam⋅sin⁡2⋅π⋅dayno-cevpph365, where cevpam is a correction factor, dayno is the day of the year, and cevpph is a factor used to correct the phase of the sine function in order to correct the potential evapotranspiration (set as zero in this study). To deal with the coarse spatial resolution of the temperature database (0.5∘), a correction factor that depends on catchment elevation was computed (SMHI, 2018): 6tempc=temp-tcelevadd⋅elev100, where tempc is the corrected air temperature (in ∘C), temp is the original air temperature (∘C), tcelevadd is a calibrated parameter that corrects temperature (∘C 100-1 m-1), and elev is the mean catchment elevation (m). Since only few (<10) temperature station records were available, a bias correction procedure was not possible, but measured temperature closely followed the environmental lapse rate (Esquivel-Hernandez et al., 2017). The parameters cevp, cevcorr, cevpam, and tcelevadd are part of the Monte Carlo simulation and their ranges are shown in Table 3.

Figure 3 shows the workflow adopted for model calibration, which involves a qualitative parameter sensitivity analysis to find the most suitable range of values for the automatic calibration. The initial parameter ranges were obtained from manual iterations of one parameter at a time to facilitate automatic calibration (Infante-Corona et al., 2014).

We considered four model configurations to analyze the effect of including PET and AET in model calibration:

model configuration 1 (M1), calibrated using only daily streamflow (Qt);

The streamflow records were divided into the period from 1991 to 1999 for calibration and from 2000 to 2003 for validation. The PET and AET calibration period was established as 2002 to 2010 and the validation period as 2011 to 2014. In both cases, for model warm-up we ran it for 2 years prior to calibration since our modeling tests showed that using 2 years was enough to stabilize the effects of initial conditions of water content in soil layers, rivers, and reservoirs. The 13 monitored catchments were used for streamflow calibration. For PET and AET calibration steps, only the 130 catchments within the 13 monitored catchments were used since our tests showed that using the 605 catchments did not significantly increase the model performance but it did increase the calibration time by a factor of 5. The simulations of the 605 catchments were used to compute the metrics for the calibration and validations periods.

A total of 86 parameters were used to build the HYPE model structure consisting of 36 parameters linked to four soil types, 24 parameters linked to four land cover classes, 6 parameters for the general structure, 12 parameters for the regional correction of PET and temperature, and 8 parameters for lake discharge. The Monte Carlo (MC) routine for parameter sampling and sensitivity analysis included in HYPE was used for calibration, and the model configurations were run 10 000 times for each step, except for M1, which used 20 000 runs to cover more parameter combinations since this configuration only used daily streamflow. Despite the lower computational efficiency of the MC with respect to other optimization schemes (such as gradient-based methods), the MC routines are more flexible in accounting for multiple parameters sets in complex models (Beven, 2006). The 10th and 90th percentiles of the resulting parameters from the best 100 runs were used to constrain the parameters for the next calibration step.

The CHIRPS product was evaluated with ground records using the false alarm rate (FAR, computed with Eq. 7), probability of detection (PD, computed with Eq. 8), and threat score (TS, computed with Eq. 9): 7FAR=falsealarmshits+falsealarms,8PD=hitshits+misses,9TS=hitshits+falsealarms+misses, where hits are days with precipitation detected by CHIRPS and ground rain gauges, false alarms are days where precipitation was detected only by CHIRPS, and misses are days where precipitation was detected only by rain gauges.

The model performance was evaluated using the Kling–Gupta efficiency (KGE; Kling and Gupta, 2009), computed as 10KGE=1-(r-1)2+(∝-1)2+(β-1)2,11r=CC=covxo,xsσoσs,12∝=σsσo,13β=μsμo, where subscripts “o” and “s” correspond to observations and simulations, respectively. μ is the mean, x is the time series (streamflow, actual evapotranspiration, or potential evapotranspiration), σ is the standard deviation, r and CC are the correlation coefficient, α is the agreement between amplitudes, and β is the bias. KGE was chosen as the objective function for calibration since it equally captures maximum and minimum flows (e.g., Arheimer et al., 2020; Pechlivanidis and Arheimer, 2015; Rajib et al., 2018a; Rakovec et al., 2016; Xiong and Zeng, 2019), and has been described as a relatively balanced metric with slightly more focus on high flows (Garcia et al., 2017). Non-transformed data were used given the advice of Santos et al. (2018) against using log-transformed discharge with the KGE for low flow evaluation. Furthermore, other statistical criteria were computed to facilitate assessing the performance of the model configurations, such as the Pearson correlation coefficient (computed with Eq. 11), mean absolute error (MAE, computed with Eq. 14), Nash–Sutcliffe efficiency (NSE, computed with Eq. 15), root mean square logarithmic error (RMSLE, computed with Eq. 16), and relative bias (Eq. 17): 14MAE=1n∑i=1nxs(i)-xo(i),15NSE=1-∑i=1nxs(i)-xo(i)2∑i=1nxo(i)-μo2,16RMSLE=∑i=1nlog⁡xs(i)-log⁡xo(i)2n,17Bias=μs-μoμo. Furthermore, hydrological signatures were calculated to independently assess how well the calibrated model configurations reproduced different hydrological criteria. The hydrological signatures used in this study are shown in Table 4 and the Budyko curve was constructed from the aridity index (AI) and evaporative index (EI).

Finally, the non-parametric Kruskal–Wallis (Kruskal and Wallis, 1952) and Mann–Whitney (Mann and Whitney, 1947) tests were used to detect statistically different performance.

Comparing precipitation from CHIRPS with annual streamflow and streamflow plus evapotranspiration (assuming long-term balance P-Qt-AET=0) showed an underestimation of annual precipitation (as shown in Fig. S1a and c), leading to unrealistic water balance values. The interpolated bias correction factor (BF, Fig. 4a) showed an overestimation of CHIRPS rainfall in blue and underestimations in red. The BF ranged from 0.65 to 1.57 with an average of 1.06±0.14, where the greater disagreements between the ground precipitation and satellite-merged precipitation were observed along the Pacific Basin. The underestimation reached 30 % to 35 % in the north of the Gulf of Nicoya and in the southwest of the Providencia catchment. Underestimation of CHIRPS across the Caribbean slope was mainly observed in the Terron Colorado and Cariblanco catchments, with a BF between 1.2 and 1.4. Moreover, the largest overestimation of CHIRPS was observed for the Guanacaste region (BF = 0.65–0.8), downstream of the Tacares catchment (BF = ∼0.8), and to the southeast of Costa Rica (BF = 0.8–0.85).

For modeling purposes, we evaluated the temporal synchronicity of rainfall versus streamflow (Fig. 4b) using cross-correlation between daily streamflow and catchment-scale daily precipitation from CHIRPS, where the x axis corresponds to the lag time in days. Most of the monitored catchments exhibited the highest correlation within lag time zero, indicating that the hydrological response of catchments tends to occur within the same day. Nevertheless, the Cariblanco and Rancho Rey catchments exhibited poor correlation (ρ<0.3), which means a lack of synchrony between daily satellite-merged precipitation and streamflow.

The bias correction improved the annual precipitation where CHIRPSc was consistent with annual streamflow, and the long-term water balance was mostly preserved, as observed in Fig. S1b and d. Figure S2 shows the MAE normalized by mean precipitation for CHIRPS and bias-corrected CHIRPS (CHIRPSc), both with respect to the 75 precipitation stations, where boxplots correspond to the variability of normalized MAE estimated by each point. The average normalized MAE at a daily scale was estimated at 1.2±0.12 mm mm-1 for CHIRPS and 1.17±0.08 mm mm-1 for CHIRPSc, 0.30±0.09 and 0.27±0.07 mm mm-1 at a monthly scale, and 0.16±0.09 and 0.11±0.04 mm mm-1 at an annual scale, respectively.

Figure S2d shows the probability of success and failure of CHIRPSc at detecting rainy or dry days with respect to ground stations, where the probability was computed from a single time series merged from the 75 station records. Furthermore, Fig. S2e to g show the false alarm rate (FAR), probability of detection (PD), and threat score (TS), respectively. Results indicated that CHIRPSc detected true rainy and dry days with a similar probability (0.31 to 0.34) as in situ observed rainfall, whereas the FAR ranged from 0.15 to 0.38 with the greater values (i.e., incorrect detection of dry days as rainy days by CHIRPS) to the southeast, and the PD showed greater values (i.e., better performance of CHIRPS at detecting rainy days) for the Pacific Basin (median of ∼0.69) in comparison with the rain gauges in the Caribbean Basin (median PD of ∼0.54). The TS showed similar results to PD (Fig. S2g), with better performance of CHIRPS for the Pacific Basin (median of ∼0.53) than for the Caribbean Basin (median of ∼0.46). Such low capacity of CHIRPS to detect rainy days in the Caribbean Basin could affect the performance of the hydrological model during peak flows.

Figure 5 shows the comparison of the model configurations' performance for the calibration (dark blue) and validation (light blue) periods. Simulated daily streamflow for the 13 gauged catchments was similar for baseline configuration (M1) and M2 during the calibration period (1991–1999; Fig. 5a) with a mean KGE of 0.54±0.09 and 0.53±0.08, respectively. Nevertheless, M1 showed a median NSE of 0.33 in comparison to M2, which showed a median NSE of 0.25. The comparison of the correlation coefficient, median absolute error, and NSE metrics is shown in Fig. S3. Moreover, metrics for M3 and M4 were slightly poorer due to a larger dispersion across the sample, with a mean KGE of 0.45±0.2 and 0.47±0.17, respectively, and a mean NSE of 0.23±0.2 and 0.21±0.21. For the validation period (2000–2003), the mean KGE decreased by ∼0.08, but with similar performance for NSE.

The configuration M2 best reproduced monthly streamflow for the calibration period (Fig. 5b), with a mean KGE of 0.67±0.11, whereas the configurations M1, M3, and M4 showed a mean KGE of ∼0.60, also driven by larger dispersion along the KGE scale. The NSE also supports M2 as the best monthly streamflow predictor for the calibration period (Fig. S3), with a median value of 0.54, in contrast to M1, which showed a median NSE of 0.43. The four model configurations preserved performance for the validation period, and in some cases, the KGE even increased, as was the case for the Palmar, Caracucho, and El Rey catchments (not shown). Nevertheless, the Rancho Rey catchment exhibited poor performance during the validation period (KGE < 0 and NSE < -2) for daily and monthly scales since the four configurations overestimated streamflow. In the following sections, we present more details for Rancho Rey that can explain the catchment behavior and its performance.

Figure 5c and d show the effect of including AET and PET in the calibration steps, and the KGE was computed by aggregating the complete domain (605 nested catchments). The calibration consisted of 130 nested catchments within the monitored catchments. Furthermore, M1 and M2 were only plotted for comparison purposes since these configurations were calibrated with streamflow. From Fig. 5c, we observed that simulated monthly AET for the calibration period (2002–2010) improved for M3 and M4 with a mean KGE of ∼0.49±0.17 with respect to M1 with a mean KGE of 0.29±0.29 and M2 (0.04±0.33). The higher performance of AET was also observed for M3 and M4 according to the correlation coefficient and MAE (Fig. S3). Surprisingly, the baseline configuration (M1) showed a slightly better performance of simulated monthly PET, with a mean KGE of 0.64±0.09, whereas M3 and M4 showed a mean KGE of ∼0.61±0.10, and M2 a mean KGE of 0.43±0.28 (Fig. 5d). The performance of monthly AET and monthly PET was similar for the validation period (2011–2014; Fig. S3).

The results from Fig. 6 suggested that the best performance of daily and monthly streamflow for the calibration period (2001–2009) was obtained for catchments in the southeast of Costa Rice, such as the Palmar, Caracucho, El Rey, and Guapinol catchments, with KGEs higher than 0.55 (NSE > 0.4, as shown in Fig. S4) for daily streamflow and higher than 0.8 (NSE > 0.63) for monthly streamflow. Nevertheless, the mid-Pacific Basin also resulted in the Tacares and Providencia catchments exhibiting the worst performance for monthly streamflow and the configurations M3 and M4 with KGE < 0.3 and NSE < 0 (Fig. 6g and h).

The spatially distributed KGE on the last two panels of Fig. 6 shows the improvement by including AET and PET in the calibration steps (Fig. 6k, l, o and p), unlike the case of daily and monthly Qt, where no significant improvements were observed using the four calibration procedures. The calibrated monthly AET simulated with M1 showed low efficiency (KGE < 0.2 and NSE < 0) for ∼182 catchments of the Pacific Basin but an acceptable performance (KGE > 0.6 and NSE > 0.2) for monthly PET. M2 exhibited poor performance across the simulation domain for AET and low efficiency of PET in the southeast Caribbean. Additionally, M3 and M4 showed similar results with acceptable performance (KGE > 0.6 and NSE > 0.2) for ∼179 catchments, most of them located in the northeast. The median MAE and the median correlation coefficient for M3 and M4 were ∼11 % lower (MAE = ∼15 mm) and ∼39 % higher (CC > 0.63) than for M1. Surprisingly, the simulated PET with M3 and M4 was similar to the PET from M1. The performance of the calibrated water level on the Arenal reservoir was relatively low for all configurations (KGE ∼ 0.35, NSE < -0.1, and CC ∼ 0.36), affected by the unknown quantity of withdrawals from the reservoir during the driest months (April–July).

Figure 7 shows the model parameter ranges from the 100 best-fit simulations resulting from the last calibration step for each configuration. The red dots from Fig. 7 correspond to the optimal parameters used for modeling, where multiple red dots and boxplots for each model are shown by soil type and land use.

A large dispersion with a coefficient of variation (CV = SD/mean) greater than 0.35 was observed for runoff response parameters (srrate, fraction for surface runoff; srrcs, recession coefficient for surface runoff) and baseflow parameters (rrcs1, recession coefficient for uppermost soil layer; rrcs2, recession coefficient for lowest soil layer). The impacts of monthly streamflow on calibration were observed for the general model parameters of rivvel (river velocity) and rrcs3 (deep layer recession coefficient), with constrained posterior parameter distributions for configurations M2 and M4 and higher velocities and greater baseflow discharge for M2 with respect to M4.

The soil type and land use coverage influence the calibrations' parametrization. M2 and M4 showed constrained distributions of parameters srrate and rrcs1 for clay-loam soil (third class), the most frequent soil type in the monitored catchments (Fig. 1b). The bottom panel in Fig. 7 shows the spatial distribution of the srrate parameter, with similar values for M2, M3, and M4 and the most frequent soil classes (clay and clay-loam).

The soil parameters that regulate the soil water content (wcwp, wcep) showed similar distributions with the median value of the fraction of soil water available for evapotranspiration (wcfc). The effective porosity (wcep) was slightly higher for configurations M1 and M2, but the final parameters (red dots) differed between the models. Furthermore, for M3 and M4, the parameters lp and cevpam exhibited constrained distributions with a CV of 0.12 and 0.11, respectively. In comparison, M1 and M2 showed CV values of ∼0.25 and ∼0.28 for lp and cevpam.

The step-wise calibration improved model performance in different aspects. Figure 8 shows the comparison of the hydrological simulations for two monitored catchments contrasting the best simulation with the highest KGE performance (Palmar catchment) and the worst simulation with the lowest KGE performance (Rancho Rey catchment).

The Palmar catchment exhibited acceptable performance (KGE > 0.5 and NSE > 0.45) for daily streamflow, but all configurations underestimated the highest peak flows during the calibration and validation periods. For the Rancho Rey catchment, the observed highest peak flows during 1997 were 3 times larger than simulated peak flows. Underestimation of simulated peak flows was related to the poor capabilities of CHIRPSc to detect heavy storms since observed peak flows were not associated with large precipitation amounts (Fig. 8a). In the Palmar and Rancho Rey catchments, the baseline (M1) underestimates low flows by 1 and 2 orders of magnitude during the dry season, respectively (Fig. 8d and j). Moreover, daily streamflow for M2 exhibited a mean relative bias (Eq. 17) of -0.046 with respect to M1 due to the underestimation of peak flows in all monitored catchments, and M3–M4 showed a mean relative bias of 0.156. The mean relative bias with respect to baseline (M1) using the logarithm of daily streamflow ranged from 0.24 to 0.29; hence, configuration M1 tended to generate lower flows during the dry period.

At a monthly scale, streamflow was preserved by the model configurations in several catchments, except for Rancho Rey, where simulated streamflow was on average 2 times larger than the observed streamflow during the rainy season (Fig. 8l). Such overestimation indicated that the bias factor was insufficient to correct the global precipitation product or for large discharge measurement errors. Furthermore, all configurations reproduced the seasonality of AET and PET from MODIS (not shown), but M3 and M4 underestimated the AET and PET in Palmar while showing good performance for AET in Rancho Rey. Moreover, simulated monthly soil moisture (SM) content was independently compared with the catchment average soil moisture content from the Land Parameter Retrieval Model (LPRM; Owe et al., 2008) product for the period 2012 to 2016. The simulated SM for M1 followed the seasonal behavior of the LPRM product in the Palmar catchment, matching the absolute LPRM % SM content. The LPRM product uses SM from the upper 5 cm against the 50 cm of the upper layer defined for all model configurations. However, all model configurations show a high correlation (CC > 0.7) in both catchments (Palmar and Rancho Rey), matching the seasonality.

The observed and simulated flow duration curves for all monitored catchments are shown for the period 1991–2003 in Fig. 9. The baseline (M1, dashed red line) underestimated the median and low flows in several catchments (Guardia, Rancho Rey, Guatuso, Terron Colorado, Caracucho, El Rey, Guapinol), with a median RMSE of 1.15 considering all monitored catchments, 2 times larger than other model configurations. M2 (dashed blue line) exhibited the best performance for median and low flows (with a median RMSE of 0.42), whereas M3 (orange line) and M4 (blue line) showed similar results to M2. Higher efficiencies for median and low flows were obtained for catchments that exhibited higher cross-correlation with precipitation (Fig. 4b), as was the case for Palmar, Caracucho, and El Rey, among others.

The simulated and observed hydrological signatures are shown in Fig. 10, where simulations covered the period 1991–2014, and observations covered different periods depending on available records. The comparison of hydrological signatures by monitored watershed is shown in Tables S1 and S2 in the Supplement. The simulated long-term mean annual water balance (Prec-Qt-AET) was mostly closed in all catchments (∼0 mm), with average values of -1.80 mm (M1), -0.62 mm (M2), -0.79 mm (M3), and -1.17 mm (M4); hence the water balance was significantly improved when compared to the observed water balance. Indeed, the observed water balance (Prec-Qt-AET) yielded values from -800 to 600 mm. Such variability in observed water balances may be related to the short common period of data but also to the discrepancies between the data sources (in situ, interpolated and merged, remotely sensed), as can be observed in Fig. S4. Figure S4 shows how the long-term water balance using the observed data differs from the Budyko curve in all monitored catchments (Westerberg et al., 2014), while the simulations fit the theoretical curve.

The spatial distribution of baseflow indices (BFI) derived from M2, M3, and M4 exhibited similarities with respect to the observations. Simulated BFI showed an overwhelming groundwater contribution to streamflow with relatively similar average values of 0.70, 0.69, 0.68, and 0.74 for model configurations M1, M2, M3, and M4, respectively.

Larger differences were observed in the northwest and southwest when comparing the BFI of M1 with respect to other configurations, whereas M4 resulted in larger contributions of baseflow to streamflow in coastal areas of the Caribbean. Similar spatial patterns were obtained for the streamflow coefficient (SC=Qt/Prec), with low values in the drier northwest and higher values for catchments that receive more rainfall (Fig. 1e). In contrast, M2 indicated that a lower amount of precipitation became streamflow, with an average value of 0.49, in comparison with M1, M3, and M4, which showed medians of 0.52, 0.57, and 0.57, respectively. Moreover, M1 and M2 followed the spatial patterns of the observed SC due to their higher streamflow performance.

The EI and AI were similar for M3 and M4 due to their similar model parameters. The spatial distribution of observed EI from MODIS was reproduced by M3 and M4, whereas AI spatial patterns were preserved by M1. In addition, M1 and M3–M4 showed similar spatial patterns for EI and AI across the north, but differences were observed in the south, where M1 indicated lower water availability attributed to higher evaporative ratios (higher EI). M2 simulated the driest catchments, with an average value of EI and AI of 0.50 and 0.63, respectively, whereas M1 showed median values of 0.47 and 0.59, and M3–M4 values of 0.42 and 0.51.

Daily precipitation from CHIRPS was preferred over other global precipitation products because of a relatively higher spatial resolution and good performance across different climates and biomes (e.g., Bayissa et al., 2017; Ullah et al., 2019; Zambrano-Bigiarini et al., 2017). Nevertheless, CHIRPS showed a large bias and more rainy days with respect to ground precipitation across Costa Rica (Fig. 4a). The results suggested that at large scales, the precipitation bias was compensated for since the mean bias factor (BF) was ∼1, but underestimation of precipitation was observed in mountainous regions as wells as large overestimations in the drier northwest (Fig. 4). Similarly, results from Chile by Zambrano-Bigiarini et al. (2017) indicated that an evaluation and even correction of global products is necessary and may be related to the poor density of ground gauges used to generate CHIRPS in those regions, especially across regions with steep topography (Funk et al., 2015). The latter authors also reported underestimation during extreme rainfall events and overestimation during rainy days.

Our simple, linear bias correction of CHIRPS showed better performance at monthly and annual scales and solved the water balance inconsistencies of most catchments (Fig. 10). However, the cross-correlation between daily precipitation and streamflow remains unchanged by the bias correction. Not surprisingly, our results showed that catchments with highly correlated streamflow and daily precipitation exhibited better performance than catchments with low correlations. Several studies highlighted those meteorological forcings are the largest source of uncertainty in hydrological modeling (e.g., Arheimer et al., 2020; Dal Molin et al., 2020; Lin et al., 2018; Wörner et al., 2019), whereas more complex bias correction techniques (e.g., quantile mapping) may improve the results (e.g., Goshime et al., 2019; Wörner et al., 2019). However, the lack of matching daily streamflows with precipitation inputs and intense rain events might persist. Nonetheless, Infante-Corona et al. (2014) suggested that global products can achieve better streamflow simulation results than sparse ground precipitation data, whereas Westerberg and Birkel (2015) found that in situ precipitation in Costa Rica may require corrections to achieve better model results.

The global CPC temperature dataset used was not bias-corrected due to the lack of sufficient in situ measurements. Temperature can introduce large errors in hydrological simulations if used for the estimation of potential evapotranspiration and actual evapotranspiration (Andersson et al., 2015). We corrected the temperature data set using elevation and a lapse rate parameter (Eq. 6) during the calibration steps. The corrected temperature closely followed the environmental lapse rate of 6 ∘C temperature decrease per 1000 m elevation gain and improved model performance (KGE) by on average ∼10 %. Global evapotranspiration estimates showed differences compared to ground estimations in different regions due to the influence of, e.g., irrigation, vegetation dynamics, and uncertainty in climatological forcings (Pan et al., 2020; Velpuri et al., 2013). We used the PET and AET from MODIS 16 for calibration and evaluation due to the good performance reported in different applications around the world (e.g., Lin et al., 2018; Mu et al., 2013; Pan et al., 2020; Rajib et al., 2018b; Tang et al., 2011; Velpuri et al., 2013). Nevertheless, MODIS AET has shown poor performance at the point scale in different regions (e.g., Liu et al., 2015; Weerasinghe et al., 2020), but better performance when aggregated at the catchment scale (Velpuri et al., 2013). Additionally, Wohl et al. (2012) recognized that dense vegetation and frequent cloudiness in the humid tropics are challenges for satellite monitoring of AET. Unfortunately, the low density of eddy covariance towers and lack of comparative studies for tropical climates are limiting factors in the validation of MODIS 16 in Central America. Nonetheless, among the few existing studies available, Esquivel-Hernandez et al. (2017) compared the MODIS 16 PET product against 10 Priestley–Taylor station data-derived PET estimates in Costa Rica and found relatively small errors from -0.33 to +0.36 mm yr-1.

Simulated daily streamflow showed reasonable performance using the four model configurations, where M2 (calibrated Qt monthly + Qt daily) improved low flows simulations in comparison with M1 (Fig. 9). Moreover, our configuration using a step-wise calibration for Costa Rica resulted in improved streamflow performance compared to the global model by Arheimer et al. (2020). The main shortcomings of the four configurations were the underestimated larger peak flows (Fig. 8), but such errors were associated with the precipitation input rather than model capabilities. The spatial comparison of streamflow simulations indicated that catchments in the southwest performed best (KGE > 0.5 and NSE > 0.45, Fig. 6) compared to other areas. The southwestern Pacific is characterized by moderate precipitation seasonality (Fig. 1d) with a low bias in the precipitation product (Fig. 4a) and better performance in detecting rainy and dry days (Fig. S2), compared with the tropical climate gradient of the dry to humid tropics in Costa Rica. Furthermore, we found overestimated monthly streamflows in the drier northwestern region of Costa Rica (Rancho Ray and Guardia catchments with higher mean daily and monthly simulated streamflows, as is shown in Table S2). Previous studies have noted that HYPE overestimates streamflow in dry environments (Arheimer et al., 2020). Finally, the streamflow overestimation could also be related to the precipitation bias of CHIRPSc and possibly to the nonuniform spatial distribution of our streamflow observational sample, with additional “wetter” catchments used for calibration. Indeed, despite bias correction, precipitation overestimation persisted in drier environments (Fig. 4), associated with the lack of ground precipitation records to correct the CHIRPS product in headwater catchments, such as Rancho Rey. Furthermore, the low correlation of daily precipitation of CHIRPS with rain gauges resulted in low NSE values (<0.2) for daily streamflow due to the unsynchronized peak flows, mainly in the Caribbean slope.

Results from the Kruskal–Wallis test suggested median KGE in AET and PET from M3–M4 were statistically different compared to the baseline configuration M1 (the statistic H∼183 and p value < 0.05 for AET, and H∼72 and p value < 0.05 for PET). Nevertheless, the Kruskal–Wallis and Mann–Whitney tests indicated that median and the distribution of KGE for daily and monthly streamflows are similar among all model configurations (p values > 0.05). Hence, statistical significance using AET and PET for calibration is related to improvements in water partitioning in soil layers (Figs. 6 and 7) rather than runoff generation.

Such multi-objective calibration trade-offs were previously observed by, e.g., Zhang et al. (2018). Larger improvements were obtained for AET simulation of M3 and M4, whereas M1 (only daily Qt calibration) showed similar PET performance in both periods (calibration and validation). The worst PET and AET simulations were observed for M2, since the monthly aggregation ignores an accurate representation of spatial water partitioning to match the monthly hydrograph (Rajib et al., 2018b). Our results also suggested that low flows were improved using PET and AET for calibration (Fig. 9), where FDC exhibited an average RMSLE (Eq. 16) value of ∼0.5±0.22 compared to 1.1±0.53 from M1, constraining vertical fluxes and regulating discharge from soil layers (Massari et al., 2015; Rakovec et al., 2016). The constrained posterior parameter distributions using MODIS PET, and AET in calibration steps decreased the variability of parameters related to evapotranspiration processes (lp, cevpam). However, the soil water content parameters (wcfc, wcwp, wcep) ranges were similar among model configurations (Fig. 7). Additionally, including monthly Qt into the calibration routine also constrained parameters related to soil layer discharge (srrate, rrcs1, rrcs3) for the most frequent soil types on the monitored catchments. Such results highlight that remotely sensed PET and AET are useful at constraining some parameters and that the combination of data sources representing different modeled hydrological processes helps constrain model uncertainties, particularly for large-scale domains (Rajib et al., 2018b). Despite the partly different calibration periods used due to the limited data availability, similar record lengths (8 years of calibration and 3 years of validation) resulted in consistent results from M1 to M4.

Despite the generally reasonable performance of our model configurations, we found some issues when comparing our results with previous efforts, such as those of Birkel et al. (2012), who modeled the streamflow in the Sarapiqui River basin (data not used in this study for calibration) with the HBVlight model (Seibert, 2005) for the period from 1983 to 1991 and obtained an NSE of 0.74 after rainfall correction of the underestimated observed precipitation. Our configuration M4 resulted in an NSE = -5.6 due to rainfall underestimation (Frumau et al., 2011). In contrast, our model configuration M4 showed lower improvements with respect to the global product of Arheimer et al. (2020). For example, we obtained a KGE = 0.73 and NSE = 0.46 for the streamflow simulation of the Palmar catchment, in comparison to the global product showing a KGE of less than 0.3. Such results reflect that more data are required to improve the streamflow response at local scales.

A hydrological model useful for water management should be able to mimic streamflow seasonality and to realistically represent the large-scale physical processes of the water partitioned by vegetation interception and the soil matrix into evapotranspiration and discharge (Arheimer et al., 2020; Kwon et al., 2020; Pechlivanidis and Arheimer, 2015; Rajib et al., 2018b; Rakovec et al., 2016; Xiong and Zeng, 2019). We, therefore, independently evaluated the four configurations using a range of hydrological signatures (Table 4) following Westerberg and McMillan (2015) in an attempt to single out the sought-after well-balanced model for use in decision-making. However, using multiple signatures also complicated the interpretation of simulations since daily streamflow (Qtd) and monthly streamflow (Qtm) indicated improvements in different configurations (see Tables S1 and S2).

Significant spatial variations in hydrological signatures were observed between M1–M2 and M3–M4 since implementing a spatial calibration of AET improved the representativeness of the more complex large-scale climate gradient. Similar results were found in catchments in the United States (Lin et al., 2018; Rajib et al., 2018a) and worldwide (Arheimer et al., 2020). The model configurations M3 and M4 better reproduced the spatial variability between the Pacific and Caribbean basins and the north–south gradient of the AI and EI (Esquivel-Hernandez et al., 2017). Furthermore, the resulting hydrological signatures of M3 and M4 were consistent with previous small catchment-scale studies that showed that runoff coefficients tend to be larger than the evaporative index (Dehaspe et al., 2018; Gómez-Delgado et al., 2011). Results also suggested that the event streamflow response is dominated by quick near-surface soil water discharge (Dehaspe et al., 2018), with streamflow being fed by groundwater during dry periods resulting in BFI values exceeding 0.7 (Birkel et al., 2012). In contrast to Westerberg et al. (2014), who calibrated Central American catchments using FDC information, we used the observed FDCs as an independent hydrological signature (Fig. 9). The configurations M2 to M4 outperformed the baseline (M1), supporting the notion that only streamflow used for calibration is not enough to produce a well-balanced model.