Modeling terrestrial net ecosystem exchange using machine learning techniques based on flux tower measurements

Abstract

Identifying the complex relationships of Net Ecosystem Exchange (NEE) of CO2, as an underlying factor of land surface, and atmosphere interactions is extremely important to the dynamic of carbon fluxes. Assessment of the model-based estimation of land-atmosphere carbon flux across various plant functional types (PFTs) can support the accurate identification of the carbon cycle and the adaptation and mitigation of climate change programs. Five different machine learning methods named Multiple Linear Regression (MLR), Support Vector Machine (SVM), Decision Tree (DT), Gradient Boosting Machine (GBM) and Random Forest (RF) were used to predict daily NEE magnitude. In this study, 24 sites classified into four PFTs of Deciduous Broadleaf Forest (DBF), Evergreen Needle-leaf Forest (ENF), Mixed Forest (MF) and Grassland (GRA) were examined through ground-based flux tower data. The numbers of sites were six, four, six and eight for DBF, ENF, MF and GRA respectively, while measurement periods varied from two to thirteen years. The model calibration and validation were carried out using 70%and 30% of the data-set, respectively. The models’ performances were assessed using statistical indices including the coefficient of determination (R2), the Nash-Sutcliffe efficiency (NSE), bias error (Bias) and root mean square error (RMSE) through Python software. Based on statistical indices, the models showed different levels of capability when analyzing data from the DBF, ENF, MF and GRA sites. Among the models, RF showed the best performance, MLR showed the poorest performance, while SVM, GBM and DT models all had moderate responses. The effect of both air and soil temperatures, as the state variables, were examined to assess model performance. Whether soil temperature is included in the model plays a more important role in the performance of the models in grassland than in forest. Soil temperature inclusion, as an input variable, improved the models’ performance about 14% in grassland, while it improved performance 2.4%, 2.4% and 3.5% in ENF, MF and DBF, respectively. Finally, to assess the models' performances, the NEE behavior in terms of over- or under- estimation was investigated across each PFT and over various phenological periods. The results indicate that high uncertainty occurs between the 140th and 220th days of the Julian calendar for forested areas and between the 120th and 210thdays for grassland.