Chapter 1: Introduction

Chapter 2: Vegetation Influence and Environmental Controls on Greenhouse Gas Fluxes from a Drained Thermokarst Lake in the Western Canadian Arctic

Chapter 3: Controls on Carbon Dioxide and Methane Fluxes from a Low-Center Polygonal Peatland in the Mackenzie River Delta

Chapter 4: Modeling Interannual Variability of Carbon Fluxes at a Low-Center Polygon Ecosystem in the Mackenzie River Delta

Chapter 5: Conclusions

Algorithms that build predictive models from experience.

• Neural Network (NN) are universal approximators
• Often treated as "black box models"
• Single layer with H "hidden" nodes and M independent inputs:

$f(X,w)= \sum_{h=1}^Hβ_hg(\sum_{m=0}^M\gamma_{hm}x_m)$

• Gap filling is essential for calculating C budgets from EC observations
• Flux partitioning is the standard approach for NEE, no standard for NME
• Not practical during polar summer
• NN methods offer a flexible alternative

Advance our understanding of the growing season C exchange in the Mackenzie Delta Region.

1. Measure CO₂ and CH₄ fluxes and use gap-filling to estimate NEE and NME at Illisarvik and Fish Island
 
2. Use NN models to study controls of ecosystem scale NEE and NME
1. Identify flux drivers and map functional relationships
2. Investigate the influence of spatial heterogeneity
 
3. Conduct a temporal upscaling experiment to study the possible influence of climate variability on growing season NEE and NME at Fish Island

Neural Network (NN) models were used to:

1. Identify dominant drivers of CO₂ and CH₄ fluxes:
• Iteratively trained models, adding one new input at a time
• Climate & soil observation
• Source area fractions
2. Gap-fill CO₂ and CH₄ fluxes to estimate NEE and NME

Long-term observations are needed to better contextualize results.

1. Compared to young DTLB (< 50 yrs) studied in Alaska:
• Greater shrub cover and higher CO₂ uptake
• Drier soil conditions and lower CH₄ emission
 
2. Illisarvik will continue to evolve, following one of two trajectories:
• Sedge dominated > NME will increase significantly
• Shrub dominated > NME will remain low

Similar approach to Chapter 2, but a more robust NN was used to identify flux drivers. Weights Method (Gevery et al. 2003):

1. Start by training over-parametrized models
• 21 climate, soil, and source area inputs
2. Calculate Relative Importance (RI) from the sums the squared partial derivatives
• Pruned inputs with RI < 2.5%
• Limited to one pruning iteration for expedience
3. Retrain model and gap-fill CO₂ and CH₄ fluxes to estimate NEE and NME
• Derivatives were inspected to ensure response functions were plausible

Spearman Correlation Analysis

Temporal upscaling indicated significant interannual variability in growing season C exchange.

1. Climate warming could reduce the CO₂ sink strength
• Highlights need for long-term monitoring
 
2. Compared to two similar LCP ecosystems with long-term records:
• Weaker CO₂ sink and stronger CH₄ source than LCP ecosystem near Utqiagvik, AK.
• Stronger CO₂ sink and CH₄ source than LCP ecosystem at Samoylov Island, RU.

Neural Networks are not "black-box" models when applied carefully.

1. The weights method helps us understand what the models are doing
• Ability to map response functions is a key advantage over other common methods (i.e.random forests)
• To date, had not been applied to EC data
 
2. Can be paired with footprint modelling to account for spatial heterogeneity
 
3. Meticulous application is the key to success
• Bootstrapping, cross validation and early stopping

1. Long-term flux data needed:
• Interannual variability can have a significant impact on net fluxes
 
2. Cold season dynamics:
• Extended senescent periods could offset growing season uptake
 
3. Methane fluxes:
• Pressure pumping
 
4. Network architecture
• Multi-layer models to simulate GPP and ER separately
• Time-lagged models to account for CH₄ dynamics

If our priority is just extracting information, the research only serves to perpetuate colonialism (Liboiron 2021).

• Future work in the Arctic should focus on long-term flux monitoring
• Short, resource intensive campaigns are not sustainable
• Prioritize involvement of local stakeholders
• Technical training, employment, & education
• Community support = longer-term viability

I would like to thank:

• My supervisors and committee members
• Friends and family
• Funding agencies
• NSERC
• CFI