Weekly carbon dioxide exchange trend predictions in deciduous broadleaf forests from site-specific influencing variables

Abstract

Weekly net-ecosystem-exchange (NEE) data recorded and verified over multiple years, together with site-specific influential ecological variables, for distinct deciduous-broadleaf-forest (DBF) sites in North America can reveal useful relationships regarding their functions as long-term carbon dioxide (CO2) sources and sinks. Machine learning (ML) and regression models have greater success at predicting weekly NEE from some DBF sites than others, from the available site variables. In particular, support vector regression (SVR) and extreme gradient boosting (XGG) ML methods perform better than multi-linear regression in the weekly NEE predictions they generate using up to 24 influential variables. The DBF sites studied are distinguished into those that have followed distinctive, albeit fluctuating seasonal NEE trends, and those that are characterized by abrupt fluctuations in NEE across the leaf-on season. ML models predict weekly NEE for the former sites more reliably than for the latter sites. Consideration of the relative influence of the variables on the XGB and regression model NEE predictions identifies which variables are most influential at specific sites. Short wave radiation (in and out) and air temperature are found to be variables exerting substantial influence on the prediction models for the sites studied. From the prediction results and the relative influences of the available environmental variables, it is concluded that complex processes are involved at those sites showing rapid NEE fluctuations in the leaf-on seasons that are not readily detectable from the environmental variables currently being continuously recorded at those sites.