51³Ô¹Ï

Topic

Global Sensitivity Analysis for Terrestrial Carbon Cycle Simulations Under Present and Future Climate Conditions

School of Earth and Ocean Sciences

Date & location

• Tuesday, June 3, 2025
• 10:00 A.M.
• Clearihue Building, Room B017

Examining Committee

Supervisory Committee

• Dr. Adam Monahan, School of Earth and Ocean Sciences, 51³Ô¹Ï (Co-Supervisor)
• Dr. Christian Seiler, School of Earth and Ocean Sciences, UVic (Co-Supervisor)
• Dr. Joe Melton, School of Earth and Ocean Sciences, UVic (Member)
• Dr. Julie Zhou, Department of Mathematics and Statistics, UVic (Outside Member)

External Examiner

• Dr. Sara Knox, Department of Geography, McGill University

Chair of Oral Examination

• Dr. Matthew Moffitt, Department of Chemistry, UVic

Abstract

This dissertation assesses sensitivity of simulated land surface carbon and energy fluxes to variations in model input parameters. The terrestrial carbon cycle plays a major role in the atmospheric concentration of CO₂. Anthropogenic CO₂ emissions have substantially perturbed the atmospheric carbon reservoir, and is actively taken up by other reservoirs. Land surface models (LSMs) are used to simulate the exchange of mass, energy and momentum between the terrestrial biosphere and the atmosphere. The terrestrial biosphere has been taking up more carbon than it releases since the late 1960s. LSMs project that the biosphere will continue to take up carbon till early to mid 22^nd century, making it a net carbon sink. But, these models also indicate substantial uncertainties in the strength of the sink. For instance, the latest Global Carbon Budget assessment estimates an inter-model spread of 1 to 3.2 PgC yr⁻¹ during 2014-2023. This spread widens in future projections ranging from 2 to 7 PgC yr⁻¹ by the end of the 21^st century, as noted in Intergovernmental Panel on Climate Change’s Sixth Assessment Report. These large uncertainties limit the reliability of model-based assessments, requiring the need for a deeper understanding of the factors influencing carbon sink projections.

The projections of atmospheric CO₂ is shown to increase drastically due to anthropogenic emissions. Since the terrestrial biosphere uptakes a big fraction of these emissions, reducing uncertainties in the terrestrial carbon sink is important for improving the future carbon predictions and informing mitigation strategies. Some of the mentioned uncertainties in the simulated carbon sink arises from parameter uncertainties. While parameter tuning can help reduce these uncertainties, optimizing all input parameters in a complex, non-linear LSM is computationally prohibitive. Identifying influential parameters and understanding their influence on the model output(s) is an essential step before tuning the parameters. The influence of parameter uncertainties on the terrestrial carbon cycle output variables can be assessed using global sensitivity analysis (GSA). This dissertation applies GSA to the output variables simulated by the Canadian Land Surface Scheme Including Biogeochemical Cycles (CLASSIC).

This research is divided into three parts, each applying GSA to CLASSIC output variables under different conditions. The questions asked are: (1) Is there a common set of parameters that drive the majority of ecosystem output variables simulated at an eddy covariance site?, (2) Which parameters affect the uncertainty of the historical carbon sink for different biomes?, and (3) Which parameters affect the uncertainty of future carbon sink projections for different biomes? To answer these questions, a two-step GSA is used, where the fist step corresponds to a qualitative screening of the parameters, and the second step is a quantitative variance-decomposition based analysis.

This research reveals that only 15–17% of input parameters show any influence on the outputs based on the screening method. The quantitative analysis further narrows this subset, identifying between two and 15 parameters as the most influential for different output variables and statistical measures. The influential parameters vary depending on the meteorological forcing used. The maximum carboxylation rate (vmax) and the canopy extinction coefficient (kn) are the most recurring influential parameters across all forcing scenarios, and statistical measures. Additionally, other photosynthetic parameters, as well as those related to rooting and phenology, play an important role when CLASSIC is forced using reanalysis and Earth system model data. The sensitivity of the terrestrial carbon sink to the uncertainty in vmax reduces by the end of the 21^st century. In many cases the analysis is unable to rank the most influential parameters because of large sampling variations in the sensitivity indices.

Identifying the influential parameters and quantifying their impact on carbon cycle uncertainties is important for improving model predictions, but the computational demands of GSA are substantial. In this study, performing GSA for just seven grid cells required approximately 25 CPU years. Scaling such analyses to a global level using the full model would be computationally prohibitive. However, advancements in machine learning and emulator-based approaches present a promising alternative for GSA and optimization efforts, drastically reducing computational costs by requiring fewer input-output simulations than the full model. These innovations could enable large-scale assessments of parameter uncertainty, ultimately leading to more robust predictions of the terrestrial carbon sink, which will help in the shaping of better mitigation efforts.