Uncertainty propagation analysis of an N 2O emission model at the plot and landscape scale

Abstract

Nitrous oxide (N 2O) emission from agricultural land is an important component of the total annual greenhouse gas (GHG) budget. In addition, uncertainties associated with agricultural N 2O emissions are large. The goals of this work were (i) to quantify the uncertainties of modelled N 2O emissions caused by model input uncertainty at point and landscape scale (i.e. resolution), and (ii) to identify the main sources of input uncertainty at both scales. For the Dutch western fen meadow landscape, we performed a Monte Carlo uncertainty propagation analysis using the INITIATOR model. The Monte Carlo analysis used novel and state-of-the-art methods for estimating and simulating continuous-numerical and categorical input variables, handling spatial and cross-correlations and analyzing spatial aggregation effects. Spatial auto- and cross-correlation of uncertain numerical inputs that are spatially variable were represented by the linear model of coregionalization. Bayesian Maximum Entropy was used to quantify the uncertainty of spatially variable categorical model inputs. Stochastic sensitivity analysis was used to analyze the contribution of groups of uncertain inputs to the uncertainty of the N 2O emission at point and landscape scale. The average N 2O emission at landscape scale had a mean of 20.5 kg N 2O-N ha − 1 yr − 1 and a standard deviation of 10.7 kg N 2O-N ha − 1 yr − 1, producing a relative uncertainty of 52%. At point scale, the relative error was on average 78%, indicating that upscaling decreases uncertainty. Soil inputs and denitrification and nitrification inputs were the main sources of uncertainty in N 2O emission at point scale. At landscape scale, uncertainty in soil inputs averaged out and uncertainty in denitrification and nitrification inputs was the dominant source of uncertainty. This was partly because these inputs were assumed constant across areas with the same soil type and land use, which is probably not very realistic. Experiments at landscape scale are needed to assess the spatial variability of these fractions and analyze how a more realistic representation influences the uncertainty budget at landscape scale. This research confirms that results from uncertainty analyses are often scale dependent and that results for one scale cannot directly be extrapolated to other scales.