Improving the simulation of permanent grasslands across Germany by using multi-objective uncertainty-based calibration of plant-water dynamics

https://doi.org/10.1016/j.eja.2022.126464Get rights and content

Highlights

  • One single objective and two multi-objective calibration schemes were applied.

  • Grassland model parameters were estimated using multi-objective calibration scheme

  • The uncertainty ranges of the parameters were quantified.

  • The productivity of grasslands under different numbers of cutting regimes were estimated

Abstract

The dynamics of grassland ecosystems are highly complex due to multifaceted interactions among their soil, water, and vegetation components. Precise simulations of grassland productivity therefore rely on accurately estimating a variety of parameters that characterize different processes of these systems. This study applied three calibration schemes – a Single-Objective (SO-SUFI2), a Multi-Objective Pareto (MO-Pareto), and, a novel Uncertainty-Based Multi-Objective (MO-SUFI2) – to estimate the parameters of MONICA (Model for Nitrogen and Carbon Simulation) agro-ecosystem model in grassland ecosystems across Germany. The MO-Pareto model is based on a traditional Pareto optimality concept, while the MO-SUFI2 optimizes multiple target variables considering their level of prediction uncertainty. We used measurements of leaf area index, aboveground biomass, and soil moisture from experimental data at five sites with different intensities of cutting regimes (from two to five cutting events per season) to evaluate model performance. Both MO-Pareto and MO-SUFI2 outperformed SO-SUFI2 during calibration and validation. The comparison of the two MO approaches shows that they do not necessarily conflict with each other, but MO-SUFI2 provides complementary information for better estimations of model parameter uncertainty. We used the obtained parameter ranges to simulate grassland productivity across Germany under different cutting regimes and quantified the uncertainty associated with estimated productivity across regions. The results showed higher uncertainty in intensively managed grasslands compared to extensively managed grasslands, partially due to a lack of high-resolution input information concerning cutting dates. Furthermore, the additional information on the quantified uncertainty provided by our proposed MO-SUFI2 method adds deeper insights on confidence levels of estimated productivity. Benefiting from additional management data collected at high resolution and ground measurements on the composition of grassland species mixtures appear to be promising solutions to reduce uncertainty and increase model reliability.

Introduction

Grasslands occupy 40.5% of the world’s total terrestrial area (White et al., 2000) and contribute greatly to terrestrial biodiversity and carbon storage, which accounts for approximately 28–37% of the global soil organic carbon pool (Lal, 2004). A recent study in tree-sparse grasslands such as California has demonstrated that grasslands are more reliable carbon sinks than forest and wetland ecosystems with respect to climate change (Dass et al., 2018). Economically less profitable than arable crops (Wimberly et al., 2017), permanent grasslands are usually cultivated where arable cropping is limited by, e.g., steep slopes, the risk of frequent flooding, adverse climatic or hydraulic conditions (Lei et al., 2016), and poor soils. In Germany, managed grassland areas cover more than 28% of agricultural land (Griffiths et al., 2020, Statistisches Bundesamt, 2019) and represent remarkable values for fodder production, biodiversity conservation, recreation, and ecosystem services. In fact, about 50% of plant species in Germany depend on grassland habitats (BFN, 2014). Grasslands also contribute to the replenishment of groundwater resources, as they allow more water to percolate into deeper soil zones compared to forests or cropland (von Wilpert et al., 2016). Permanent grasslands (in the form of meadows or pastures), which comprise more than 90% of Germany’s grassland areas, have been under pressure from land use changes since at least 1990. Since then, the total area has been reduced by 12% (Statistisches Bundesamt, 2019) as a result of land use changes in favor of intensive agriculture. At the same time, low-productivity grasslands on marginal sites have been abandoned for economic reasons, threatening some rare plant communities (BFN, 2014). The various types of grassland in Germany nevertheless still provide a large range of ecosystem services and biodiversity, if managed appropriately. Therefore, preservation of grassland as an important part of agricultural land use in Germany remains of crucial importance, and a better understanding of their environmental functioning is required to manage these areas more sustainably, especially in the context of global warming.

The dynamics of grassland ecosystems are highly complex, due to the multifaceted interactions among their soil, water, and vegetation components. Equipped with understanding of these interrelations, it is easier to design grassland management strategies that do not interfere with the provision of ecosystem services. Over the last few decades, various process-based agro-ecosystem models have been developed to simulate plantsingle bondsoilsingle bondwater relations and to explore the productivity and functioning of grassland systems. Such model application studies have focused on different aspects, such as the spatial distribution of grassland productivity (Chang et al., 2015b, Zheng et al., 2020), the impacts of environmental factors and climate change on grassland productivity (Gomara et al., 2020, Graux et al., 2011), and their potential as sinks for carbon storage and the mitigation of greenhouse gas emissions (Chang et al., 2015a, Jones and Donnelly, 2004, Sandor et al., 2018). In these studies, the most commonly simulated variables are aboveground biomass (AGB) or net primary productivity, providing a synthetic indicator to evaluate the productivity of grassland and the applied models were calibrated and validated based on vegetative variables, such as AGB. However, vegetative variables are tightly linked to water dynamics in soils constraining ecosystem productivity (Archontoulis et al., 2020, Tang et al., 2018), and thus a model calibrated using solely vegetation variables may not be sufficient to predict soil-related services or disservices and may result in unrealistic representations of soil processes. Including soil moisture as a constraint to plant growth in the calibration process is critical to add explanatory power to the plant-related target variables. The main difficulty in this context is the interdependency of soil water and vegetation dynamics and the simultaneous consideration of parameters and variables related to both vegetation and soil processes in the calibration. Multi-objective approaches help the model to simulate various observation data concurrently and close to reality (Houska et al., 2017), since, in contrast to single criterion calibration procedures, they also take into account the existence of compensating effects when calibrating the model. Considering multiple variables combined with global optimization routines/algorithms in calibration decreases the risk for the model getting trapped in local minima, which may lead to a better fit for one variable, but to an unsuitable parameter set representing other processes (i.e., right fit for wrong reason). To account for this feedback, it is crucial to implement methods for model calibration capable of taking into account information that describes the interconnected processes. In most studies with integrative aspects, different observation variables are combined as a single objective (SO) which subsequently forms a basis to estimate the model’s parameters. The multi-objective (MO) optimization perspective (Groh et al., 2018, Wöhling et al., 2015), which examines the trade-offs between different conflicting objectives, has received less attention in the application of grassland models, and to the best of our knowledge, we offer the first study of an integrative perspective in grassland simulations.

Process-based agro-ecosystem modeling is based on the simultaneous simulation of different interconnected biophysical processes, which depends on various parameters describing behaviors of various conceptual processes and the way they relate to each other (Fenicia et al., 2007). Some model parameters are extremely difficult to measure or even cannot be directly measured and often have to be estimated inversely, through an optimization procedure. Such procedures are intended to minimize deviations between simulated and observed target variables, such as AGB, soil moisture (SM), or leaf nitrogen concentration. Since the processes in agro-ecosystem models are closely interlinked, these minima are often only local, and optimization procedures may reveal only one out of many existing minima, but not necessarily the global minimum (or maximum). This means that the parameter values identified are associated with high levels of uncertainty, which in turn may lead to poor model predictions. Reducing the uncertainty of these parameters requires methods which consider (A) multiple components of the model in an MO calibration framework and (B) quantification of the confidence level of output variables.

MO optimization approaches use tradeoffs to determine a set of non-dominated parameters that cannot be improved for one objective without compromising the other objective. During optimization, these approaches incorporate multiple objectives on the basis of different information such as: (1) multi-variable data representing different interrelated processes, (2) multi-site data, or (3) multi-response models considering independent criteria of one model aspect (Efstratiadis and Koutsoyiannis, 2010, Kamali et al., 2013). Therefore, instead of converging around a single optimum, MO approaches spread search in parameter space in a way to detect a number of feasible parameter sets (solutions) with acceptable trade-offs along the Pareto front (Proximedia, 2018). Given the fact that different parameters activate different processes in the model (Efstratiadis and Koutsoyiannis, 2010), exploring the trade-off between different objectives (considering various variables) assists models in finding a more realistic and robust parameter estimates. Furthermore, evaluating the trade-off among solutions located on the Pareto fronts provides additional indications on probable limitations of a model (Efstratiadis and Koutsoyiannis, 2010). For example, an irregular shape of a Pareto-front can be a sign for an ill-posed model, or a significant trade-off points to a probable ill model parameterization. A proper evaluation of these additional information obtained from MO approaches help modelers to better assessment of model performance and consistency and find robust solutions to reduce the uncertainty in the model.

The most common algorithms for MO calibration include MO particle swarm optimization (Kennedy and Eberhart, 1995), MO genetic algorithm (Fonseca and Fleming, 1993), and MO complex evolution (Yapo et al., 1998). While most MO methods offer superior performance compared to SO calibration (Kamali et al., 2013), they also suffer from their inability to provide information on the uncertainty of model predictions. Bayesian approaches can account for parameter uncertainty in optimization, and their superior performance vis-à-vis SO approaches has recently been demonstrated for eco-hydrological models (Tang et al., 2018, Wöhling et al., 2013) where mostly two variables (LAI and SM) have been considered. However, their application for calibration of grassland models have been limited mostly to SO approaches where the errors from different variables were aggregated to overall one single function (Höglind et al., 2016, Korhonen et al., 2018). The complex dynamics of grasslands, strongly vary depending on the types of their management (intensively and extensively managed grasslands), species composition and prevailing soil. Therefore, we assume that model calibration will benefit a lot from considering multiple key variables e.g., LAI as a model state variable, AGB and SM. Despite the importance, the current studies lack reflection on uncertainty-based calibration and parameter estimation approaches which capitalize on: (1) information that describes the multiple variables and find the trade-off among different variables; (2) the uncertainty ranges of each variable; and (3) the comparison of these approaches on intensively and extensively managed grasslands.

In this paper, we aim to (1) implement an uncertainty-based MO calibration procedure to a process-based models simulation model for grassland (Nendel et al., 2011) to understand how it helps to improve the simulation of different soil and vegetation processes; (2) analyze the implications (scaling/uncertainty) derived from the simulation of managed grasslands in Germany, and use that information to quantify the level of uncertainty associated with biomass estimates across different regions of Germany. The method proposed and applied accounts for all sources of uncertainty, such as model input, model structure, model parameters, and measured data, and does not disentangle different types of uncertainty. The suitability of the method is tested against data from grassland sites with two/three cutting events per growing season (extensively managed) and sites with four/five cutting events per growing season (intensively managed). Overall, five experimental sites located in different parts of Germany were selected. In this paper, we maintain our assumption that grassland management (i.e. intensively and extensively managed grasslands) relates directly to only the number of cutting events.

Section snippets

Data sources

Two data sets were used in this study. For the sensitivity analysis, model calibration and model validation, we used data from five experimental grassland sites that have been originally designed for different projects (see Table 1). The second data set includes national-scale soil and climate data for German-wide simulations of grasslands (see Section 2.4 for details).

The five experimental sites are: (1) Braunschweig (to understand the increasing importance of grassland areas for sustainable

Sensitivity analysis

The Morris sensitivity method for experimental sites did not show differences in the identified most sensitive parameters among case studies and the result for all were found to be similar to each other. The σ and μ indices were calculated according to daily information within this simulation time. Fig. 4a shows the maximum σ values during the growing season for different parameters in relation to the LAI, AGB, and SM state variables. We screened parameters based on the calculated σ and μ

Effects of the sensitivity analysis

In this paper, we applied a two-step sensitivity analysis procedure, i.e., screening and time-varying ranking, which the literature has recommended as a systematic approach to sensitivity analysis (Pianosi et al., 2016). The time-varying ranking of parameters provided useful information for the prioritization of parameters to be estimated. Sensitivity analysis is a function of boundary conditions as well as time (Paleari and Confalonieri, 2016). Ignoring this fact can result in the

Conclusion and future prospect

We have presented the first simulations of grasslands across with number of cutting events per growing season the whole area of Germany. To the best of our knowledge, there is no study available that compares grassland productivity in light of different numbers of cutting regimes over such a large territory. We have concluded that, to obtain reliable estimates on grassland productivity, model calibration based on both vegetation and soil-related components is essential, and should be more

Funding acquisition

This work was supported by the innovation funding program of the German Federal Ministry of Food and Agriculture (BMEL; Project No. 2818300716).

CRediT authorship contribution statement

Bahareh Kamali: Conceptualization, Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization, Project administration. Tommaso Stella: Validation, Formal analysis, Investigation, Resources, Writing – review & editing. Michael Berg-Mohnicke: Validation, Formal analysis, Investigation, Resources, Writing – review & editing. Jürgen Pickert: Validation, Formal analysis, Resources, Data curation, Supervision.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This work was supported and funded by the innovation funding program of the German Federal Ministry of Food and Agriculture (BMEL; Project No. 2818300716).

References (69)

  • C. Nendel et al.

    Testing farm management options as climate change adaptation strategies using the MONICA model

    Eur. J. Agron.

    (2014)
  • L. Paleari et al.

    Sensitivity analysis of a sensitivity analysis: we are likely overlooking the impact of distributional assumptions

    Ecol. Model.

    (2016)
  • F. Pianosi et al.

    Sensitivity analysis of environmental models: a systematic review with practical workflow

    Environ. Model. Softw.

    (2016)
  • R. Sandor et al.

    The use of biogeochemical models to evaluate mitigation of greenhouse gas emissions from managed grasslands

    Sci. Total Environ.

    (2018)
  • H.J. Smit et al.

    Spatial distribution of grassland productivity and land use in Europe

    Agr. Syst.

    (2008)
  • I.M. Sobol

    Distribution of points in a cube and approximate evaluation of integrals

    USSR Comput. Maths. Math. Phys.

    (1967)
  • X. Specka et al.

    Analysing the parameter sensitivity of the agro-ecosystem model MONICA for different crops

    Eur. J. Agron.

    (2015)
  • Y.T. Tang et al.

    A Bayesian alternative for multi-objective ecohydrological model specification

    J. Hydrol.

    (2018)
  • J. Wang et al.

    Parameter sensitivity analysis of crop growth models based on the extended fourier amplitude sensitivity test method

    Environ. Model. Softw.

    (2013)
  • M.C. Wimberly et al.

    Cropland expansion and grassland loss in the eastern Dakotas: new insights from a farm-level survey

    Land Use Policy

    (2017)
  • P.O. Yapo et al.

    Multi‐objective global optimization for hydrologic model

    J. Hydrol.

    (1998)
  • C.S. Zhan et al.

    An efficient integrated approach for global sensitivity analysis of hydrological model parameters

    Environ. Model. Softw.

    (2013)
  • G. Zhao et al.

    Demand for multi-scale weather data for regional crop modeling

    Agr. For. Meteor.

    (2015)
  • Abbaspour, K.C., 2015. SWAT-CUP; SWAT calibration and uncertainty programs - A user...
  • Agrometeorological Research Center (ZAMF), Germany,...
  • S.V. Archontoulis et al.

    Predicting crop yields and soil-plant nitrogen dynamics in the US Corn Belt

    Crop Sci.

    (2020)
  • BFN, 2014. GRÜNLAND-Report. Alles im Grünen Bereich? Federal Agency for Nature Conservation: Grassland conservation in...
  • BGR, 1995. Institute for Geosciences and Natural Resources,...
  • J.F. Chang et al.

    The greenhouse gas balance of European grasslands

    Glob. Change Biol.

    (2015)
  • J.F. Chang et al.

    Modeled changes in potential grassland productivity and in grass-fed ruminant livestock density in Europe over 1961-2010

    PLOS One

    (2015)
  • P. Dass et al.

    Grasslands may be more reliable carbon sinks than forests in California

    Environ. Res. Lett.

    (2018)
  • DWD, 1994. German Weather Service:...
  • A. Efstratiadis et al.

    One decade of multi-objective calibration approaches in hydrological modelling: a review

    Hydrol. Sci. J.

    (2010)
  • F. Fenicia et al.

    A comparison of alternative multiobjective calibration strategies for hydrological modeling

    Water Resour. Res

    (2007)
  • Cited by (2)

    View full text