The physically-based distributed WRF-Hydro modelling system, including the Noah land surface model with multiple parameterization options (Noah-MP) and the hydrological extension of the WRF atmospheric model (Weather Research and Forecasting model), has recently been widely used for water balance investigations, streamflow and coupled land-atmosphere simulations. Despite the multiple available physical parameterizations in the model, equations for simulating particular losses from the water balance are missing, and a grid-based calibration of distributed parameters across multiple watersheds has not been studied. To fill these gaps, this study aims: (i) to analyze the impact of soil, runoff, groundwater and vegetation parameters on water balance components; (ii) to improve baseflow and transpiration equations; and (iii) to test a grid-based calibration approach for distributed model parameters, using streamflow observations. The WRF-Hydro groundwater model was improved through the introduction of a groundwater loss factor and the Jarvis stomatal conductance model was modified to account for nocturnal transpiration. The grid-based calibration was performed for three pa-rameters (infiltration, hydraulic conductivity and percolation) for 19 spatially-distributed classes, with the Parameter Estimation (PEST) software. The study area includes 31 small mountainous watersheds (5-115 km2) in Cyprus, in the Eastern Mediterranean. A two-year period (2011-2013) was used for calibration and a five-year period (2013-2018) for the evaluation. The baseline model set-up overestimated streamflow, on average, by 50 % in 2011-2012 and more than 100 % in 2012-2013. Overall, streamflow and evapotranspiration (ET) could vary by about +/- 30 % from the baseline simulation, using different model parameters and model options. The simulation of groundwater losses as a function of groundwater level reduced total streamflow, on average, by 30 %. The use of the proposed Jarvis equation for nocturnal transpiration increased the total ET, on average, by 25 %. The grid-based approach facilitated the calibration of the distributed parameters over the area of the 31 watersheds. The median Nash-Sutcliff Efficiency (NSE) was 0.49 during calibration, but 0.02 in the drier eval-uation period. The calibrated WRF-Hydro model reproduced the annual variability of ET and the improved groundwater and transpiration equations reduced the substantial streamflow overestimation of WRF-Hydro. The model performance during dry years demonstrated the need for representation of more processes that occur in semi-arid environments with ephemeral streams and are not included in WRF-Hydro and Noah-MP. The grid-based WRF-Hydro parameterization can be applied to the full study area for fully-coupled atmospheric-hydro-logic simulations.