W15: ISEM workshop on Dealing with uncertainty and sensitivity issues in process-based models of carbon and nitrogen cycles in northern forest ecosystems

organised by Guy R. Larocque, Jagtar Bhatti, Andrew M. Gordon, Nancy Luckai


Title: Dealing with uncertainty and sensitivity issues in process-based models of carbon and nitrogen cycles in northern forest ecosystems

Authors: Guy R. Larocque et al.

Abstract: Many process-based models on carbon (C) and nitrogen (N) cycles have been developed for northern forest ecosystems. They are widely used to evaluate the long term decisions in forest management dealing with effects like particulate pollution, productivity and climate change. Regarding climate change, one of the key questions that has sensitive political implications is whether northern forests will sequester atmospheric carbon or not. Whilst many process-based models have been tested for accuracy by evaluating or validating against observed data, very few have dealt with complexity or incorporated procedures to estimate the uncertainty associated with their predictions or their sensitivity to input factors in a systematic inter-model comparison fashion. In general, models differ in their underlying attempts to match natural complexities with assumed or imposed model structure and process formulations to estimate model parameters, to gather data and to address issues on scope, scale and natural variations. Uncertainty may originate from model structure, estimation of model parameters, data gathering, representation of natural variation and scaling exercises. Model structure relates to the mathematical representation of the processes modelled and the type of state variables that a model contains. The modelling of partitioning among above- and below-ground C and N pools and the interdependence of these pools remain a major source of uncertainty in model structure and error propagation. Most soil carbon models use three state variables to represent the different types of soil organic matter (SOM). This approach results in creating three artificial SOM pools, assuming that each one contains carbon compounds that have the same turnover rate. In reality, SOM consists of many different types of C compounds with turnover rates that can widely differ. Uncertainty in data and parameter estimates are closely linked. In particular, data uncertainty may be associated with the high variation in the biomass, productivity and soil organic matter in forest ecosystems and may be incomplete for model initialization, calibration, validation and sensitivity analysis of generalized predictor models. The scale level also affects the level of uncertainty, as the errors in the prediction of the carbon and nitrogen dynamics differ from the site to the landscape levels and across climatic regions. If the spatial or temporal scale of a model application is changed, additional uncertainty arises from neglecting natural variability in time and space. Uncertainty issues are also intimately related to model validation and sensitivity analysis. The estimation of uncertainties is needed to inform decision process, in order to detect the possible corridor of development. Uncertainty in this context is an essential measure of quality for stakeholder and decision makers.

Title: Modelling carbon exchange of Canadian boreal forests: Model development, validation, and sensitivity analysis

Authors: Changhui Peng, Xiaolu Zhou, Jianfeng Sun

Abstract: This paper presents the process-based model of TRIPLEX-Flux for estimating net ecosystem productivity (NEP) and the sensitive analysis of model response by simulating CO2 flux in old black spruce site of BOREAS in central Canada. The objective of the research was to examine the effects of parameters and input variables on model responses. The validation using 30 minutes interval data of NEP derived from tower and chamber measurements showed that the modelled NEP had a good agreement with the measured NEP (R2>0.65). The sensitivity analysis demonstrated different sensitivities between morning and noonday, and from the current to doubled atmospheric CO2 concentration. Additionally, the comparison of different algorithms for calculating stomatal conductance shows that the modeled NEP using the iteration algorithm is consistent with the results using a constant Ci/Ca of 0.74 and 0.81, respectively for the current and doubled CO2 concentration. Varying parameter and input variable values by ?10% resulted in the model response to less and equal than 27.6% and 27.4%, respectively. Most parameters are more sensitive at noonday than in the morning except those that are correlated with air temperature suggesting that air temperature has considerable effects on the model sensitivity to these parameters/variables. The air temperature effect was greater under doubled than current atmospheric CO2 concentration. In contrast, the model sensitivity to CO2 decreased under doubled CO2 concentration.

Title: Performance evaluation of organic matter decomposition models based on carbon, nitrogen and matter dynamics in litterbags

Authors: Paul Arp, Jagtar Bhatti, Chengfu Zhang

Abstract: Several models (CENTURY, Yasso, ROMUL, DOCDOM) were compared with a readily calibrated 3-compartment model for projecting matter, C and nitrogen remaining in litter bags over 7 years, across the wide range of climate conditions and litter type of the Canadian Intersite Decomposition Experiment (CIDET). The 3 compartments refer to fast, slowly and very slowly decomposing portions of the litter, using initial water-extractable and acid-hydrolyzable portions to estimate initial mass and N of the fast fraction, and the initial ash content to estimate these quantities in the slow and very slow fractions. Mean July and January air temperatures, and annual precipitation were used to capture the influence of climate on litter decomposition. Following model initialization based on initial mass and N contents, the model required altogether - 6 predictor variables and 12 calibration parameters. The model was constrained to conform to general expectations concerning the endpoint of litter decomposition: approaching 0 mass and N at a C/N ratio similar to that of humified matter. ModelMaker software was used for model formulation and parameter estimation. Best-fitted results generated: r2 = 0.92 for mass remaining; = 0.81 for carbon/nitrogen ratio; = 0.80 for nitrogen concentrations, with values for N and C/N lower because of error propagation. The other models use more compartments, more predictor variables and more parameters, and also differ in terms of model structure and assumptions. Differences in that regard are due to various ways to incorporate expectations dealing with decomposition dynamics based on (a) litter composition, and (b) how to formulate actions and quantities of microbial biomass. Expectations resulting from climate expectations also differ, but are more similar than dissimilar.

Title: Upscaling Terrestrial Carbon Dynamics From Sites to Regions With Uncertainty Measures: The GEMS Experience

Authors: Shuguang Liu, Jinxun Liu, Zhengpeng Li, Thomas R. Loveland, Mingshi Chen, Larry L. Tieszen

Abstract: Upscaling the spatial and temporal changes of carbon stocks and fluxes from sites to regions is challenging owing to the spatial and temporal variances and covariance of driving variables and the uncertainties in both the model and the input data. Although various modeling approaches have been developed to facilitate the upscaling process, few deal with error transfer from model input to output, and error propagation in time and space. We develop the General Ensemble Biogeochemical Modeling System (GEMS) for upscaling carbon stocks and fluxes from sites to regions with measures of uncertainty. GEMS relies on site-scale biogeochemical models (e.g., the Erosion-Deposition-Carbon Model (EDCM) and CENTURY) to simulate the carbon dynamics at the site scale. The spatial deployment of the site-scale model in GEMS is based on the spatial and temporal joint frequency distribution of major driving variables (e.g., land cover and land use change, climate, soils, disturbances, and management). At the site scale, GEMS uses stochastic ensemble simulations to incorporate input uncertainty and to quantify uncertainty transfer from input to output. Using data assimilation techniques, GEMS simulations can be constrained by field and satellite observations or census data including estimates of net primary productivity from the Moderate Resolution Imaging Spectroradiometer (MODIS), grain yield and cropping practices, and forest inventories. The modeling philosophy embedded in GEMS makes it ideal for incorporating and assimilating information with various uncertainties at a range of spatio-temporal resolutions. The application of GEMS to quantify the contemporary terrestrial carbon dynamics in the United States is presented as an example of GEMS applications.

Title: A framework for assessing uncertainty in ecosystem models

Authors: Martin Wattenbach, Pia Gottschalk, Pete Smith

Abstract: In addition to their use as research tools, ecosystem models have been used more frequently in the last two decades to support policy decisions and inform stakeholder consultations. Models have been central to the work of the Intergovernmental Panel of Climate Change (IPCC) and the International Geosphere-Biosphere Programme (IGBP). The usefulness of results from model simulations for any purpose is determined by their quality and the uncertainty accompanying model outputs. In model evaluation, however, a broad variety of different approaches to define uncertainty still exists and these have not been standardized so far. In contrast, field research has already defined standard uncertainties. Here, we define uncertainty based on statistical methods like standard deviation of a number of independent measurements as type A uncertainty, and define uncertainty based on scientific judgement as type B uncertianty. We are proposing three other categories of model uncertainty. Baseline uncertainties that originate from type A and B uncertainties in measurements used to determine inputs to the model are termed type C uncertainties. Further uncertainty arises from the scenarios constructed to run the model, which cannot be defined precisely. This category of uncertainty incorporates type C uncertainty but includes that element of future scenarios that cannot be predicted. Uncertainty also arises from not knowing precisely the true value of internal parameters of the model equations; this is refered to as type E uncertainty. Here we propose a framework for expressing the quality of model outputs in terms of a quantification of type C uncertainty for descriptive model uses and type D uncertainty for predictive model uses, each with associated type A and B uncertainty ranges. Internal parameter uncertainty (type E) should be treated separately as it refers to model structure itself, and which we assume as valid when assessing uncertainty due to other factors.

Title: Improving the Use of Data for Quantifying Uncertainty in Parameters and Predictions of Forest Dynamic Models by Bayesian Approach

Authors: Alexander Komarov, Pavel Grabarnik, Alexey Mikhaylov, Oleg Chertov

Abstract: Ecological models may include components which are stochastic: climatic parameters, spatial structure and so on. Therefore, the model outputs possess inevitable uncertainties. Moreover, the output uncertainty depends on variability and/or uncertainty of parameters and initial values. Uncertainty analysis which studies how input variability influences a variability of output is an important part of a model building allowing separating different sources of uncertainties. Another task which is associated with highly structured multiparametrized model is sensitivity analysis, goal of which is to characterize how the model outputs respond to changes in the inputs. Both uncertainty and sensitivity analysis can be conducted by means of the Monte Carlo procedure. To use a model in specific context it may be necessary to calibrate the model by using some observed data. Calibration is a reduction of uncertainty of input parameters and it is a key stage of a model building. An effective approach based on Bayesian estimation was proposed recently [1,2] allowing to incorporate a prior knowledge of parameter variability. There is a certain difficulty when applying the Bayesian calibration for parameters of highly complicate models, which is a case of spatially explicit individual-based models [3]. For such models the likelihood, connecting data (output) and parameters (input) in probabilistic form, is either impossible or computationally prohibitive to obtain. Recently, there was proposed a Markov chain Monte Carlo method for generating samples from a posterior distribution without the use of the likelihood [4]. We discuss uncertainty and sensitive analysis and Bayesian calibration issues by example of a model of growth and cycling of elements in boreal forest ecosystems EFIMOD [3]. 1. Gertner, G.Z., Fang, S., Skovsgaard, J.P. (1999). Ecological Modelling, 119: 249-265. 2. Van Oijen, M., Rougier, J. & Smith, R. (2005). Tree Physiology 25: 915-927. 3. Komarov, A., Chertov, O., et al., (2003). Ecological Modelling, 170: 373-392 4. Marjoram, P. et al., (2003). Proc. Natl. Acad. Sci. USA, 100: 15324-15328.

Title: Using the Monte Carlo method to quantify uncertainty in predictions of a soil carbon cycle model in balsam fir (Abies balsamea (L.) Mill.) and black spruce (Picea mariana (Mill.) B.S.P.) forest ecosystems in the boreal forest

Authors: Guy R. Larocque, David Pare, Robert Boutin, Valerie Lacerte

Abstract: The majority of process-based models of the carbon cycle in forest ecosystems are deterministic. Very few components have been implemented in these models to represent the uncertainty that may result from model structure, parameter estimates and natural variation in forest ecosystems. There are many sources of stochastic variation in the carbon cycle of forest ecosystems, including soil organic matter quality, climatic conditions, vegetation type and carbon fluxes, such as litterfall. Therefore, uncertainty in model predictions may be significant, which may affect the degree to which a model is sensitive to relatively small variations in the inputs. As the development of forest management policies will rely more and more on the use of models, it is essential that policy makers have good estimates of the uncertainty level in the predictions. Several approaches based on Monte Carlo simulations can be used to quantify uncertainty. However, for a complex dynamic model that contains several state variables and fluxes, the application of Monte Carlo methods can be cumbersome. We discuss uncertainty and sensitivity issues by applying the Monte Carlo method to a soil carbon cycle model in balsam fir (Abies balsamea (L.) Mill.) and black spruce (Picea mariana (Mill.) B.S.P.) forest ecosystems in the boreal forest. Gaussian random distributions are computed on key parameters of the model.

Title: Quantifying uncertainties associated with estimates of greenhouse gas emissions and removals from Canada's managed forests

Authors: Thomas White, Werner Kurz, Graham Stinson

Abstract: Canada is developing the National Forest Carbon Monitoring Accounting and Reporting System (NFCMARS) to meet international reporting obligations under the United Nations Framework Convention on Climate Change (UNFCCC) and the Kyoto Protocol. The system integrates information about forest area, age and species composition, growth rates, and natural and anthropogenic disturbances in a modeling framework that simulates the forest carbon cycle at stand and landscape scales. The system reports, at regional and national scales, the emissions and removals of greenhouse gases in Canadas managed forests. Approaches to quantify the uncertainties associated with these estimates are being developed. The difficulties associated with meaningfully quantifying uncertainty for national-scale estimates of greenhouse gas emissions and removals from Canadas forests arise from the wide range of input data sources that are combined in the analysis and the degree to which statistical uncertainties can be defined for these data. The paramerization of ecological processes is typically based on studies that are localized and with known error structure. Uncertainties are introduced when scaling up or outside of the domain within which these parameters were developed. Other input data  such as forest inventories compiled for operational and planning purposes  cover larger spatial scales. Information characterizing the uncertainties associated with these data, however, is often anecdotal or based on expert judgment. We present and discuss a framework that we use to assess the appropriate spatial and temporal scales at which to evaluate the effect of specific sources of uncertainties on the overall estimates of greenhouse gas emissions and removals.

Title: Impacts of Forest Disturbances on the Carbon Cycle in the Laurentian Plains and Hills of the United States.

Authors: Jinxun Liu, Shuguang Liu, Linda S. Heath, Thomas R. Loveland, Larry L. Tieszen

Abstract: Land cover changes over large areas can be monitored using remote sensing (RS) technology. These changes are one of the key driving forces for ecosystem carbon (C) dynamics. We applied the General Ensemble Biogeochemical Modeling System (GEMS) using a top-down, remote sensing (RS) driven mechanism to estimate forest C fluxes in the Laurentian Plains and Hills ecoregion in the northeastern United States for the period of 1972-2000. Disturbances such as forest stand replacement were detected on 30 randomly-located 10-km by 10-km sampling blocks using Landsat imagery at 60-m resolution. Spatially explicit modeling of carbon dynamics in GEMS was organized using the joint frequency distribution of major controlling variables. Each unique combination of these variables forms a simulation unit. For each forest simulation unit, a Monte Carlo process was used to initialize forest type, forest age, forest biomass, and soil C based on county level Forest Inventory and Analysis (FIA) data and the State Soil Geographic (STATSGO) data. Ensemble simulations were performed for each simulation unit to incorporate input uncertainty into the simulations, providing uncertainty estimates for model outputs. The results include forest ecosystem C budgets, land cover change impacts, and uncertainties of model outputs. Keywords: carbon budget, land cover change, biogeochemical modeling, uncertainty