Sensitivity analysis of a catchment-scale nitrogen model

There are now considerable expectations that semi-distributed models are useful tools for supporting catchment water quality management. However, insufficient attention has been given to evaluating the uncertainties inherent to this type of model, especially those associated with the spatial disaggregation of the catchment. The Integrated Nitrogen in Catchments model (INCA) is subjected to an extensive regionalised sensitivity analysis in application to the River Kennet, part of the groundwater-dominated upper Thames catchment, UK The main results are: (1) model output was generally insensitive to land-phase parameters, very sensitive to groundwater parameters, including initial conditions, and significantly sensitive to in-river parameters; (2) INCA was able to produce good fits simultaneously to the available flow, nitrate and ammonium in-river data sets; (3) representing parameters as heterogeneous over the catchment (206 calibrated parameters) rather than homogeneous (24 calibrated parameters) produced a significant improvement in fit to nitrate but no significant improvement to flow and caused a deterioration in ammonium performance; (4) the analysis indicated that calibrating the flow-related parameters first, then calibrating the remaining parameters (as opposed to calibrating all parameters together) was not a sensible strategy in this case; (5) even the parameters to which the model output was most sensitive suffered from high uncertainty due to spatial inconsistencies in the estimated optimum values, parameter equifinality and the sampling error associated with the calibration method; (6) soil and groundwater nutrient and flow data are needed to reduce uncertainty in initial conditions, residence times and nitrogen transformation parameters, and long-term historic data are needed so that key responses to changes in land-use management can be assimilated. The results indicate the general difficulty of reconciling the questions which catchment nutrient models are expected to answer with typically limited data sets and limited knowledge about suitable model structures. The results demonstrate the importance of analysing semi-distributed model uncertainties prior to model application, and illustrate the value and limitations of using Monte Carlo-based methods for doing so.