Predict with Confidence: Application of a Bayesian-Based Assimilation Framework to Optimise Wetland CH4 Emissions from a Terrestrial Ecosystem
(2024)- Abstract
- Uncertainty in our calculations of a system reveals the limits of our knowledge. This limitation makes model representations of the Earth system challenging and our predictions less reliable. However, despite appearing chaotic and non-linear, a system always follows a deterministic trajectory, which can often be expressed probabilistically. Building on this concept, this thesis applies data assimilation theories from Bayesian statistics to reduce parameter uncertainties in the LPJ-GUESS model, a widely used Dynamic Global Vegetation Model known for its non-linearity and strong parameter dependency. Given the importance of rising atmospheric methane (CH4 ) concentrations and their climate feedback, the focus is on optimising the model’s... (More)
- Uncertainty in our calculations of a system reveals the limits of our knowledge. This limitation makes model representations of the Earth system challenging and our predictions less reliable. However, despite appearing chaotic and non-linear, a system always follows a deterministic trajectory, which can often be expressed probabilistically. Building on this concept, this thesis applies data assimilation theories from Bayesian statistics to reduce parameter uncertainties in the LPJ-GUESS model, a widely used Dynamic Global Vegetation Model known for its non-linearity and strong parameter dependency. Given the importance of rising atmospheric methane (CH4 ) concentrations and their climate feedback, the focus is on optimising the model’s wetland module to quantify CH4 emissions from high-latitude wetlands. The processes governing CH4 emissions from wetlands are complex, making their implementation in LPJ-GUESS challenging due to numerous parameters and parameter uncertainties. This thesis explores the most uncertain process parameters in the wetland module and develops a Bayesian assimilation framework, informed by Eddy Covariance (EC) flux measurements. Experiencing the failure of the classical Metropolis-Hastings Markov Chain Monte Carlo (MCMC) algorithm due to the complex nature of the problem, I developed a more robust mechanism called the Rao-Blackwellised Adaptive MCMC algorithm (GRaB-AM). A single-site experiment using data from the Siikaneva wetland resulted in approximately a 93.89% reduction in the so-called ’cost function’ which measures the misfit between the observations and the model output. On a broader scale, the algorithm was tested by assimilating EC data from 14 temperate, boreal, and arctic wetlands, and the optimised model was validated against data from five additional wetlands. The experiment achieved around a 95% reduction in the cost function and approximately a 50% reduction in the root mean square error. For wetlands above 45°N, the mean annual CH4 emission estimation of the optimised model was 28.16 Tg C y-1 . These optimised model then used to simulate wetland CH4 emission for the European and Scandinavian domains, which are subsequently propagated through the Lund University Modular Inversion Algorithm (LUMIA) to constrain atmospheric CH4 concentrations. Data assimilation algorithms often struggle due to high variability in observations, observational uncertainty, and temporal correlations. To address this, I have further developed the GRaB-AM to sample error parameters along with model parameters using EC data from Lompoloj¨ankk¨a. The thesis draws an outline of the possibilities of Bayesian-based data assimilation techniques on a highly non-linear model using a single-species example of CH4 for northern wetlands. Optimised quantification of wetland CH4 emissions for the Northern hemisphere above 45°N from the model has been done but on a global scale is yet to be achieved. Issues faced by GRaB-AM, such as equifinality and slow parameter convergence, remain partly unresolved. Future studies should focus on testing and refining the algorithm with different species of data and different models. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/record/b2884935-fa39-4222-8a75-db96d8ee3a9c
- author
- Theanutti, Jalisha LU
- supervisor
-
- Marko Scholze LU
- Johan Lindström LU
- Janne Rinne LU
- opponent
-
- Doctor Kleinen, Tomas, Max Planck Institute for Meteorology, Hamburg, Germany
- organization
- publishing date
- 2024-11-01
- type
- Thesis
- publication status
- published
- subject
- keywords
- Data Assimilation, Methane, Wetlands, Model Uncertainty, Bayesian Statistics, Adaptive MCMC, GRaB- AM, Slice Sampling, Error Propagation, LPJ-GUESS, LUMIA Inversion
- pages
- 229 pages
- publisher
- Lund University
- defense location
- Pangea
- defense date
- 2024-11-01 09:00:00
- ISBN
- 978-91-89187-47-4
- 978-91-89187-48-1
- project
- Predict with Confidence Application of a Bayesian-Based Assimilation Framework to Optimise Wetland CH4 Emissions from a Terrestrial Ecosystem Model
- Model-Data Fusion for improving quantification of methane fluxes from wetlands with LPJ-GUESS
- language
- English
- LU publication?
- yes
- id
- b2884935-fa39-4222-8a75-db96d8ee3a9c
- date added to LUP
- 2024-10-04 14:18:03
- date last changed
- 2025-04-04 14:30:42
@phdthesis{b2884935-fa39-4222-8a75-db96d8ee3a9c,
abstract = {{Uncertainty in our calculations of a system reveals the limits of our knowledge. This limitation makes model representations of the Earth system challenging and our predictions less reliable. However, despite appearing chaotic and non-linear, a system always follows a deterministic trajectory, which can often be expressed probabilistically. Building on this concept, this thesis applies data assimilation theories from Bayesian statistics to reduce parameter uncertainties in the LPJ-GUESS model, a widely used Dynamic Global Vegetation Model known for its non-linearity and strong parameter dependency. Given the importance of rising atmospheric methane (CH4 ) concentrations and their climate feedback, the focus is on optimising the model’s wetland module to quantify CH4 emissions from high-latitude wetlands. The processes governing CH4 emissions from wetlands are complex, making their implementation in LPJ-GUESS challenging due to numerous parameters and parameter uncertainties. This thesis explores the most uncertain process parameters in the wetland module and develops a Bayesian assimilation framework, informed by Eddy Covariance (EC) flux measurements. Experiencing the failure of the classical Metropolis-Hastings Markov Chain Monte Carlo (MCMC) algorithm due to the complex nature of the problem, I developed a more robust mechanism called the Rao-Blackwellised Adaptive MCMC algorithm (GRaB-AM). A single-site experiment using data from the Siikaneva wetland resulted in approximately a 93.89% reduction in the so-called ’cost function’ which measures the misfit between the observations and the model output. On a broader scale, the algorithm was tested by assimilating EC data from 14 temperate, boreal, and arctic wetlands, and the optimised model was validated against data from five additional wetlands. The experiment achieved around a 95% reduction in the cost function and approximately a 50% reduction in the root mean square error. For wetlands above 45°N, the mean annual CH4 emission estimation of the optimised model was 28.16 Tg C y-1 . These optimised model then used to simulate wetland CH4 emission for the European and Scandinavian domains, which are subsequently propagated through the Lund University Modular Inversion Algorithm (LUMIA) to constrain atmospheric CH4 concentrations. Data assimilation algorithms often struggle due to high variability in observations, observational uncertainty, and temporal correlations. To address this, I have further developed the GRaB-AM to sample error parameters along with model parameters using EC data from Lompoloj¨ankk¨a. The thesis draws an outline of the possibilities of Bayesian-based data assimilation techniques on a highly non-linear model using a single-species example of CH4 for northern wetlands. Optimised quantification of wetland CH4 emissions for the Northern hemisphere above 45°N from the model has been done but on a global scale is yet to be achieved. Issues faced by GRaB-AM, such as equifinality and slow parameter convergence, remain partly unresolved. Future studies should focus on testing and refining the algorithm with different species of data and different models.}},
author = {{Theanutti, Jalisha}},
isbn = {{978-91-89187-47-4}},
keywords = {{Data Assimilation; Methane; Wetlands; Model Uncertainty; Bayesian Statistics; Adaptive MCMC; GRaB- AM; Slice Sampling; Error Propagation; LPJ-GUESS; LUMIA Inversion}},
language = {{eng}},
month = {{11}},
publisher = {{Lund University}},
school = {{Lund University}},
title = {{Predict with Confidence: Application of a Bayesian-Based Assimilation Framework to Optimise Wetland CH4 Emissions from a Terrestrial Ecosystem}},
year = {{2024}},
}