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

R. ALVAREZ; H. STEINBACH; A. BONO

SOIL SCIENCE SOCIETY OF AMERICA JOURNAL

SOIL SCI SOC AMER

Año: 2011 p. 965 - 975

0361-5995

Soil quality has been associated to its organic matter content in many agroecosystems. Additionally, as soils can act as sources or sinks for atmospheric carbon, impacting the global warming process, much effort had been performed for understanding the carbon cycle at global, regional and field scales and generate models suitable for carbon flux prediction. The use of simulation models is restricted in developing countries because of the scarcity of available information for model parameterization and validation. As a consequence, in these countries the development of empirical models may be a straightforward approximation for soil carbon budget prediction. We used published data from long term tillage experiments performed in the Pampean Region of Argentina, where C-CO2 emission from organic carbon pools was determined in the field, for developing empirical models suitable for carbon flux emission prediction. We also performed 113 field experiments with corn (Zea mayz L.), wheat (Triticum aestivum L.) and soybean (Glycine max (L.) Merr.), conducted under a vast range of climate, soil and management conditions, along the central portion of the Pampas, to determine crops carbon inputs to the soil. Two empirical modeling techniques were tested, polynomial regression and artificial neural networks, partitioning data sets into 70 % for training and 30 % for validation. Both methodologies generated good models with R2 that ranged from 0.70 to 0.86, and without significant difference between the training and validations sets, indicating excellent ability of the best models fitted for generalization. Nevertheless, neural networks performed a better job than regressions with significantly lower RMSE both for C-CO2 emission and carbon input prediction. Daily C-CO2 emission could be predicted (R2 = 0.86) using soil carbon content, temperature and water level as independent variables by the neural network. Crop carbon input (R2 = 0.85) was estimated using crop type, yield and rainfall during the growing cycle. The models developed were used for the evaluation of the impact of the introduction of soybean in rotations during the 1970-80 decade in some regions of the Pampas. This crop nowadays occupies 60 % of pampean agricultural area. Despite soybean carbon input to the soil is lower than those of wheat and corn, which are replaced in rotations, soil carbon budgets are similar compared to the 1970-80 period, or passed from negative to positive at present, depending of the region considered. These changes were associated to yield increases ascribed to technological improvement which counteract soybean lower biomass production with greater carbon inputs mainly of graminaceus crops.