Artificial neural networks and adaptive neuro-fuzzy inference systems for prediction of soil respiration in forested areas southern Brazil.

Abstract

The purpose of this study was to estimate the temporal variability of CO2 emission (FCO2) from O2 influx into the soil (FO2) in a reforested area with native vegetation in the Brazilian Cerrado, as well as to understand the dynamics of soil respiration in this ecosystem. The database is composed of soil respiration data, agroclimatic variables, improved vegetation index (EVI), and soil attributes used to train machine learning algorithms: artificial neural network (ANN) and an adaptive neuro-fuzzy inference system (ANFIS). The predictive performance was evaluated based on the mean absolute error (MEA), root mean square error (RMSE), mean absolute percentage error (MAPE), agreement index (d), confidence coefficient (c), and coefficient of determination (R2). The best estimation results for validation were FCO2 with multilayer perceptron neural network (MLP) (R2 = 0.53, RMSE = 0.967µmolm-2s-1) and radial basis function neural network (RBF) (R2 = 0.54, RMSE = 0.884µmolm-2s-1) and FO2 with MLP (R2 = 0.45, RMSE = 0.093mgm-2s-1) and RBF (R2 = 0.74, 0.079mgm-2s-1). Soil temperature and macroporosity are important predictors of FCO2 and FO2. The best combination of variables for training the ANFIS was selected based on trial and error. The results were as follows: FCO2 (R2 = 16) and FO2 (R2 = 29). In all models, FCO2 outperformed FO2. A primary factor analysis was performed, and FCO2 and FO2 correlated best with the weather and soil attributes, respectively.