Machine learning for prediction of soil CO emission in tropical forests in the Brazilian Cerrado.

Soil CO emission (FCO) is a critical component of the global carbon cycle, but it is a source of great uncertainty due to the great spatial and temporal variability. Modeling of soil respiration can strongly contribute to reducing the uncertainties associated with the sources and sinks of carbon in the soil. In this study, we compared five machine learning (ML) models to predict the spatiotemporal variability of FCO in three reforested areas: eucalyptus (RE), pine (RP) and native species (RNS). The study also included a generalized scenario (GS) where all the data from RE, RP and RNS were included in one dataset. The ML models include generalized regression neural network (GRNN), radial basis function neural network (RBFNN), multilayer perceptron neural network (MLPNN), adaptive neuro-fuzzy inference system (ANFIS) and random forest (RF). Initially, we had 32 attributes and after pre-processing, including Pearson's correlation, canonical correlation analysis (CCA), and biophysical justification, only 21 variables remained. We used as input variables 19 soil properties and climate variables in reforested areas of eucalyptus, pine and native species. RF was the best model to predict soil respiration to RE [adjusted coefficient of determination (R adj): 0.70 and root mean square error (RMSE): 1.02 µmol m s], RP (R adj: 0.48 and RMSE: 1.07 µmol m s) and GS (R adj: 0.70 and RMSE: 1.05 µmol m s). Our findings support that RF and GRNN are promising for predicting soil respiration of reforested areas which could help to identify and monitor potential sources and sinks of the main additional greenhouse gas over ecosystems.