https://hdl.handle.net/10012/22458

Peatlands represent a type of wetland, that has accumulated a layer of organic material or peat, resulting in high organic carbon (C) accumulation in this ecosystem. Peatlands hold up to one-third of the global organic C stock, and specifically peatlands located at high latitudes in the Northern Hemisphere, or northern peatlands, store a large proportion of the of the total peatland organic C stock. However, this organic C may be in jeopardy, as warmer temperatures may lead to increases in both the C uptake and output. There is some uncertainty as to how northern peatlands C sink function will be impacted by climate warming, and conventional models of peatland C cycling have been constrained to site-specific applications rather than national-scale analyses. Moreover, in-situ measurements are limited in northern peatlands, due to the remoteness of these sites, but also due to equipment limitations under low temperature and light conditions. In this thesis, I specifically looked at one component of C flux, the net ecosystem exchange (NEE) of CO2. To enable forecasting of NEE of CO2 fluxes in peatlands under future scenarios, or to generate real-time estimates where no in-situ measurements exist, machine learning algorithms trained on in-situ CO2 fluxes from the eddy covariance (EC) technique must be applied. Remotely sensed or gridded climate data products represent potentially important inputs to these modeling applications, as they are widespread both geographically and temporally enabling flux estimation for broader geographic domains. In Chapter 2, I explored the possibility of using the remotely sensed and modeled Soil Moisture Active Passive Level 4 Global Daily EASE-Grid Carbon NEE (SMAP-NEE) data product to determine NEE in Canadian peatlands. I acquired nine years (2015–2023) of SMAP-NEE data for five peatlands. I also acquired a subset of year-round eddy covariance NEE (EC-NEE) measurements within this time frame at each of the five peatland sites. The analyses showed that the SMAP-NEE data product reports a stronger growing season (GS) sink and a weaker non-growing season (NGS) source than the EC-NEE measurements. As a result of this finding, I used the relationship between SMAP-NEE and EC-NEE to produce a Corrected-SMAP-NEE dataset, which provides an estimate of seasonal and annual CO2 budgets. The data analyses of the Corrected-SMAP-NEE dataset showed that NGS CO2 emissions represent a variable proportion (33%–256%) of the GS CO2 uptake, and when these NGS emissions were accounted for, the annual CO2 sink strength was reduced proportionally. Furthermore, this study showed that longer growing seasons were consistent with greater annual net CO2 uptake at these five peatland sites from 2015-2023. The findings highlight the importance of considering the NGS when evaluating annual northern peatland C budgets. This chapter also provides evidence that existing algorithms leveraging remotely sensed and gridded climate data products to model NEE need improvement for peatlands. In Chapter 3, I compiled year-round measurements of EC-NEE from 15 Canadian peatland sites and coupled these target data with 34 hydroclimatic predictor variables (features) from remote sensing and gridded climate data products. The models were trained, validated, and tested using four algorithms: ElasticNet Regression (EN), Light Gradient-Boosting Machine (LGBM), Random Forest Regression (RFR), and Support Vector Regression (SVR). A comprehensive feature importance and selection workflow including hierarchical clustering, Gini importance, and minimum redundancy maximum relevance (mRMR) analysis was followed. Model performance stabilized at eight features, which were (relative importance shown in parentheses): evapotranspiration (40%), shortwave radiation (19%), burn area index (11%), normalized difference snow index (10%), snow water equivalent (7%), climate water deficit (4%), wind speed (4%), and soil moisture (4%). I found the best performing model to be the RFR model with these eight features (R2 = 0.76; RMSE = 0.31 g C m−2 day−1). I also assessed the generalizability and transferability of the top-performing model via a leave-one-ecoregion-out sensitivity analysis as well as on six external validation sites. The RFR model had the highest generalizability within the Taiga Plains ecoregion and the lowest generalizability within the Boreal Plain ecoregion. When testing the model on the external validation sites, performance metrics were comparable to the internal testing data for sites outside of the ecoregions represented in the training data. The findings demonstrate that the eight-feature models can be confidently upscaled to national extents, offering a clear pathway to improve Canada’s spatially explicit CO2 emission inventories.

https://hdl.handle.net/10012/22458