Evaluating modeling approaches for estimating net ecosystem exchange in Canadian peatlands

Abstract

Peatlands are globally significant carbon reservoirs, yet models for peatland carbon cycling are often limited to the site level. To enable the prediction of carbon dioxide (CO 2 ) fluxes in peatlands where no in situ measurements exist, machine learning algorithms must be trained on in situ CO 2 flux measurements from multiple sites. In this study, year-round eddy covariance (EC)-derived net ecosystem exchange (NEE) measurements and 32 hydroclimatic predictor variables were compiled for 21 Canadian peatland sites spanning seven ecoregions. A comprehensive feature selection workflow carried out on the predictor variables (features) identified that model performance stabilized at four features, which were: evapotranspiration, burn area index, normalized difference water index, and modeled soil moisture. Four machine learning algorithms: ElasticNet Regression (EN), Light Gradient-Boosting Machine (LGBM), Random Forest Regression (RF), and Support Vector Regression (SVR) were trained and evaluated using held-out test data. The best performing model was the LGBM model, which was then assessed for generalizability via a leave-one-ecoregion-out sensitivity analysis, which highlighted the necessity of ensuring predictor variables fall within the range of the training data before applying this model framework to additional sites. These findings offer a framework for regional scaling to improve Canada’s spatially explicit CO 2 emission estimates.