Application of a widely used denitrification model to Dutch data sets

Abstract

In many modelling studies on N cycling, denitrification is considered by a simplified process model. A widely used model describes denitrification as potential denitrification reduced by the soil conditions nitrate N content ( N), degree of water saturation ( S) and temperature ( T). Hénault and Germon, 2000 [Hénault, C. and Germon, J.C., 2000. NEMIS, a predictive model of denitrification on the field scale. Eur. J. Soil Sci., 51: 257–270.] showed that this model worked satisfactorily for two data sets where parameters had been specifically derived for these data sets. This paper demonstrates that it may not always work well for other data, i.e., Dutch data sets. The model was parameterized for each of eight Dutch data sets, consisting of three sand, two heavy loam and three peat sites. After parameter optimisation the model is not able to predict individually measured actual denitrification rates. However, for sand and loam soils, but not for peat soils, the correspondence between measured and predicted average actual denitrification is good. This means that the calibrated model can predict, for those specific locations, cumulative denitrification rather well, provided that detailed information on soil conditions is available, either from measurements or from simulation models. The parameters and thus the reduction functions differed between the data sets. Parameter values within a class of soils, say sand, loam, and peat, were different. So care should be taken when using parameter values obtained from other studies. Albeit simple in its mathematical formulation, a widely used simplified denitrification process model needs to be parameterized for each location. The parameter optimisation is likely to be influenced by errors in the data. Therefore, an analysis on the effect of errors is presented based on an artificial data set. This data set was constructed consisting of 100 realisations of the soil conditions N, S and T, each drawn from a (log-)normal distribution based on the mineral Dutch data sets. For each realisation the relative denitrification rate ( D r) was computed based on a ‘true’ set of parameters. Next, measurement errors were introduced on D r, N, S and T, either from a uniform or a normal distribution. For each of the 100 data sets the data were perturbed 100 times with newly drawn measurement errors. The perturbed data sets (10,000) were then optimised by the model. The coefficients of variation belonging to the estimates were large. This exercise can only be used to demonstrate possible directions of effects from errors, but it cannot be used to fully judge results obtained from real data sets.