Impact of calibration metric selection and spatial heterogeneity in soil parameters on the realism of distributed hydrological models

Distributed hydrological models are useful tools to explicitly simulate the spatial heterogeneity of water and energy fluxes and storages. Nevertheless, their parameters are typically calibrated using streamflow-based objective functions that integrate information on the spatial variability of physical processes into a single metric. Additionally, these models contain several soil parameters that can be distributed in space, affecting the spatial representation of hydrological variables. Here, we examine the implications of streamflow-based calibration metric selection and spatial heterogeneity in soil parameters on the realism of model simulations, with emphasis on spatial patterns. To this end, we conduct several calibration experiments in six pilot basins with different hydrological regimes (two snowmelt-driven, two mixed-regime, and two rainfall-driven basins), in central-southern Chile, using the Variable Infiltration Capacity (VIC) model coupled with the mizuRoute routing model. In each experiment we assess, for a given calibration objective function, the effects of distributing individual soil parameters using a spatial regularization strategy based on principal component analysis of physiographic and soil characteristics (elevation, slope, clay content, sand content and bulk density), defining the case of spatially constant soil parameters as the benchmark (i.e, only meteorological forcing data and vegetation attributes are spatially distributed). To evaluate simulated spatial patterns, we use satellite remote sensing data of soil moisture from the ESA-CCI product, and fractional snow-covered area, actual evapotranspiration (ET), and land surface temperature from MODIS products. The results show that similar streamflow performance metrics can be achieved with different combinations of regularized soil parameter and calibration metric; however, the simulated spatial patterns can be considerably different, without clear connections with the hydrological regime. Further, a streamflow-based calibration is insufficient to represent the seasonality of other variables, especially in water-limited catchments, where important shifts (e.g., up to five months) in peak ET can be obtained compared to the reference product.