Parameter-induced uncertainty quantification of soil N<sub>2</sub>O, NO and CO<sub>2</sub> emission from Höglwald spruce forest (Germany) using the LandscapeDNDC model

Abstract

Abstract. Assessing the uncertainties of simulation results of ecological models is becoming of increasing importance, specifically if these models are used to estimate greenhouse gas emissions at site to regional/national levels. Four general sources of uncertainty effect the outcome of process-based models: (i) uncertainty of information used to initialise and drive the model, (ii) uncertainty of model parameters describing specific ecosystem processes, (iii) uncertainty of the model structure and (iv) accurateness of measurements (e.g. soil-atmosphere greenhouse gas exchange) which are used for model testing and development. The aim of our study was to assess the simulation uncertainty of the process-based biogeochemical model LandscapeDNDC. For this we set up a Bayesian framework using a Markov Chain Monte Carlo (MCMC) method, to estimate the joint model parameter distribution. Data for model testing, parameter estimation and uncertainty assessment were taken from observations of soil fluxes of nitrous oxide (N2O), nitric oxide (NO), and carbon dioxide (CO2) as observed over a 10 yr period at the spruce site of the Höglwald Forest, Germany. By running four independent Markov Chains in parallel with identical properties (except for the parameter start values), an objective criteria for chain convergence developed by Gelman et al. (2003) could be used. Our approach showed that by means of the joined parameter distribution, we were able not only to limit the parameter space and specify the probability of parameter values, but also to assess the complex dependencies among model parameters used for simulating soil C and N trace gas emissions. This helped to improve the understanding of the behaviour of the complex LandscapeDNDC model while simulating soil C and N turnover processes and associated C and N soil-atmosphere exchange. In a final step the parameter distribution of the most sensitive parameters determining soil-atmosphere C and N exchange were used to obtain the parameter-induced uncertainty of simulated N2O, NO and CO2 emissions. These were compared to observational data of the calibration set (6 yr) and an independent validation set of 4 yr. The comparison showed that most of the annual observed trace gas emissions were in the range of simulated values and were predicted with a high certainty (Residual mean squared error (RMSE) NO: 2.5 to 21.3 g N ha−1 d−1, N2O: 0.2 to 21.4 g N ha−1 d−1, CO2: 5.8 to 12.6 kg C ha−1 d−1). However, LandscapeDNDC simulations were sometimes limited to accurately predict observed seasonal variations in fluxes.

Simulation of CO2 Fluxes in European Forest Ecosystems with the Coupled Soil-Vegetation Process Model “LandscapeDNDC”

“…Therefore, the possibility of adjusting respiration fluxes at the ecosystem scale, which originate from plant (autotrophic) as well as soil (heterotrophic) respiratory processes, was limited. With respect to those processes, we thus relied on a previous parameter calibration study of the soil biogeochemistry sub-module [43]. In the current investigation, the parameter derivation of the plant physiology module reveals that: (a) parameters are specific to particular ecosystem properties (see Section 4.1); and (b) generally defined (species-specific) model parameters can still describe forest gas exchange across a multitude of sites (see Sections 4.2 and 4.3).…”

“…We have used the Metropolis algorithm to make a random walk through the whole parameter space (defined by literature and expert knowledge as given in Table 2) and have derived a database with thousands of parameter sets (between 7500 and 19,000 depending on site). All parameter sets in this database were ranked using the normal distribution function in order to determine the probabilities of discrepancies between simulation and observations [43]. The ranking was done separately for GPP and NEE and the highest score for the average value was taken to select the "best" parameter set.…”

Understanding the DayCent model: Calibration, sensitivity, and identifiability through inverse modeling

“…Greenhouse gases (GHG) released from the soils of terrestrial ecosystems are highly variable in space and time due to the interaction of climatic drivers and ecosystem processes involved in carbon (C) and nitrogen (N) transformation associated with production and consumption of GHGs (Müllera et al, 2002;Rahn et al, 2012;Wrage et al, 2001). Field measurements that capture the high temporal and spatial variability of N 2 O fluxes (Bouwman et al, 2002;Parkin, 2008;Snyder et al, 2009) or the high spatial variability of soil organic carbon (SOC; Conant and Paustian, 2002;Kravchenko and Robertson, 2011) are expensive, time intensive, and unable to capture the full range of ecological and environmental conditions.…”

First 20 years of DNDC (DeNitrification DeComposition): Model evolution

“…A DNDC-based physiology module for agricultural crop growth (including grassland) and a PnET-based forest growth module allow land use change in a transient way to be described (Haas et al, 2013). Modules, derived from physical and chemical principles that describe soil environmental conditions, soil-chemistry integrating microbial C and N turnover processes and vegetation dynamics are integrated within the model (Rahn et al, 2012). The model can be applied at site scale and three-dimensional region simulations.…”