Intelligent Approaches to Analysing the Importance of Land Use Management in Soil Carbon Stock in a Semiarid Ecosystem, West of Iran

The effects of different climatic, soil, geometric, and management factors on soil organic carbon (SOC) degradation and sequestration potential was evaluated in the semi-arid zone of Mereg watershed, west of Iran. Two nonparametric methods, viz. Classification and Regression Tree (CART) and feed forward back propagation Artificial Neural Network (ANN) were compared with parametric Multivariate Linear Regression (MLR) in estimation of SOC content. Soil sampling was conducted using randomized systematic method in work unit map by overlying soil, aspect and slope maps. Results indicated that linear models had higher prediction errors. The CART with all variables (physical and management) and the ANN with 31-2-1 topology carried the highest predictive capability, explaining 81% and 76% of SOC variability, respectively. ANN models overestimated SOC content and showed a higher capability to detect the effects of management factors on SOC variations. In all the methods, management factors dominantly controlled SOC stock sequestration or degradation in different land use.