Using Bayesian model selection and calibration to improve the DayCent ecosystem model
Date
2020
Authors
Gurung, Ram B., author
Ogle, Stephen M., advisor
Paustian, Keith, committee member
Parton, William J., committee member
Breidt, F. Jay, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
Process-based biogeochemical models have been developed and used for decades to predict the outcomes of real-world ecological processes. These models are based on a theoretical understanding of relevant ecological processes and approximated using highly complex mathematical equations and hundreds of unknown parameters—requiring calibration using physical observations of the system. These models are then used to test scientific understanding, estimate pools and fluxes, make predictions for future scenarios, and to evaluate management and policy outcomes. To provide a better understanding of the ecological processes, these models need to be simple, make accurate predictions, and account for all sources of uncertainty. The focus of this dissertation is to develop a Bayesian model analysis framework to meet the goal of developing simple and accurate models that fully address uncertainty. This framework includes variance-based global sensitivity analysis (GSA) to identify influential model parameters, a Bayesian calibration method using sampling importance resampling (SIR) to estimate the posterior distribution of unknown model parameters and hyperparameters, and a Monte Carlo analysis to estimate the posterior predictive distribution of model outputs. The framework accounts for all sources of uncertainty, including the remaining uncertainty over the fitted parameters. Additionally, Bayesian model selection is also implemented in the framework to determine the most appropriate level of complexity during model development. The framework is applied to improve the DayCent ecosystem model in agricultural applications. The DayCent model was improved with several model developments, including NH3 volatilization, the release of nitrogen (N) from controlled-release N fertilizers (CRNFs) and the inhibition of the biological process of nitrification and delay the transformation of NH+4 to NO-3 with nitrification inhibitor (NIs). The model development incorporates key 4R management practices that mitigate NH3 and N2O emissions in fertilized upland agricultural soils. In addition, I recalibrated the soil organic matter submodel to improve estimation of soil organic carbon (C) sequestration potentials to a 30 cm depth for several management practices, including organic matter amendment, adoption of no-till management, and addition of synthetic N fertilizers. The results showed that the DayCent model predictions of C sequestration and reduction in N2O flux as well as NH3 volatilization from several management practices were consistent with the field observations. The model result suggested that addition of organic amendments and adoption of no-till are viable management option for C sequestration, however, the addition of synthetic N fertilizer did not produce a significant level of C sequestration. For NH3 volatilization, the model also adequately captures the reduction potential of urease inhibitor along with the incorporation of urea by mechanical means or with immediate irrigation/rainfall. The model also shows promising results in mitigating N2O emissions with both CRNFs and NIs in comparison to field observations. The model prediction focuses on estimating greenhouse gas (GHG) mitigation potential and estimation of uncertainty arising during model prediction—enhancing DayCent as a tool for scientific understanding, regional to global assessments, policy implementation, and carbon emission trading. Overall, the model improvements enhanced the ability of the DayCent model in providing a stronger basis to support policy and management decisions associated with GHG mitigation in agricultural soils.
Description
Rights Access
Subject
Bayesian model calibration
enhanced efficiency N fertilizers
soil organic carbon
ecosystem modeling
ammonia volatilization
nitrous oxide