Crops as sensors: Using crop yield data to increase the robustness of hydrologic and biogeochemical models

摘要

Hydrologic models are most commonly calibrated only for streamflow at one or multiple stations within the watershed. However, it has been argued that in distributed hydrologic models with large numbers of parameters, it is possible to get good streamflow statistics with a completely incorrect representation of the internal watershed processes. Consequently, these models fail when they are used for other purposes, for example, biogeochemical prediction. Specifically, in watersheds dominated by agricultural land use, there is information on crop yield that is rarely used for constraining hydrologic fluxes. However, crops can potentially act as distributed sensors in the landscape and help in constraining evapotranspiration fluxes and improving model consistency. Our main goal in this paper is to address this important process understanding and show how using the information on crop yield helps to identify the proper model structure, thereby improving the robustness of hydrologic models. Using a 32,660 km2 agricultural watershed in Iowa as a case study, we used stepwise model refinement to show how the consideration of additional data sources can increase model consistency. We first developed a hydrologic model using the Soil and Water Assessment Tool that provided excellent monthly streamflow statistics at eight stations within the watershed. However, comparing crop yield measurements with modeled results revealed a strong underestimation in model estimates, and it was apparent that the model was getting the "right" streamflow by lowering the evapotranspiration flux and allowing the crops to die. Adding crop yield as an additional calibration target allowed us to identify the models' structural deficiency and alter the potential evapotranspiration estimation method, thereby considerably improved model predictions of crop yield, while not altering streamflow predictions considerably. The modified model was able to (i) better capture variations in nitrate loads at the catchment outlet with no calibration and (ii) reduce parameter uncertainty, model prediction uncertainty, and equifinality. These findings highlight that using additional data sources to improve the hydrological consistency of distributed models increases their robustness and predictive ability.