An artificial neural network approach for predicting soil carbon budget in agroecosystems

Abstract

Soil quality has been associated with its organic matter content. Additionally, much effort has gone into understanding the C cycle and generating models suitable for C flux prediction. We used published data from long-term tillage experiments performed in the Pampas of Argentina, where CO2–C emissions from organic C pools were determined in the field, for developing empirical models suitable for C flux emission prediction. We also performed 113 field experiments with corn (Zea mays L.), wheat (Triticum aestivum L.), and soybean [Glycine max (L.) Merr.] to determine crop C inputs to the soil. Two empirical modeling techniques were tested: polynomial regression and artificial neural networks. Both methodologies generated good models with R2 ranging from 0.70 to 0.86. Nevertheless, neural networks performed better than regressions, with significantly lower RMSE values for both CO2–C emissions and C input prediction. Daily CO2–C emissions could be predicted by the neural network (R2 = 0.86) using soil C content, temperature, and moisture level as independent variables. Crop C inputs (R2 = 0.85) were estimated using crop type, yield, and rainfall during the growing cycle. The models were used for evaluating of the impact of soybean introduction in rotations during the 1970 to 1980 decade. Despite soybean C inputs to the soil being lower than those of wheat and corn, which were replaced in rotations, soil C budgets are similar compared with the 1970 to 1980 period, or changed from negative to positive at the present. These changes were associated with yield increases ascribed to technological improvement that resulted in greater C inputs from graminaceous crops.