Application of artificial neural networks to estimate soil organic carbon in a high-organic-matter Mollisol

Soil organic carbon (SOC) has a key role in the global carbon (C) cycle. The complex relationships among the components of C cycle makes difficult the modelling of SOC variation. Artificial neural networks (ANN) are models capable to determine interrelationships based on information. The objective was to develop and evaluate models based on the ANN technique to estimate the SOC in Mollisols of the Southeastern of Buenos Aires Province, Argentina (SEBA). Data from three long term experiments was used. Management and meteorological variables were selected as input. Management information included numerical variables (initial SOC (SOCI); number of years from the beginning of the experiment (Year), proportion of soybean in the crop sequence; (Prop soybean); crop yields (Yield), proportion of cropping in the crop rotation (Prop agri), and categorical variables (Crop, Tillage). In addition, two meteorological inputs (minimum (Tmin) and mean air temperature (Tmed)), were selected. The ANNs were adequate to estimate SOC in the upper 0.20 m of Mollisols of the SEBA. The model with the best performance included six management variables (SOCI, Year, Prop soybean, Tillage, Yield, Prop agri) and one meteorological variable (Tmin), all of them easily available and with low level of uncertainty. Soil organic C changes related to soil use in the SEBA could be satisfactorily estimated using an ANN developed with simple and easily available input variables. Artificial neural network technique appears as a valuable tool to develop robust models to help predicting SOC changes.