Detection of the effects of management and physical factors on forest soil carbon stock variability in semiarid conditions using parametric and nonparametric methods

Forest soils in western parts of Iran are being degraded by inappropriate management. The soil organic carbon (SOC) stock was dominantly affected by this type of degradation. On the other hand, SOC is an important sink for atmospheric carbon dioxide and can play a key role in global warming. This study was conducted to evaluate the effects of 15 different physical and 8 different management factors on the SOC content and to determine relative importance of these exploratory variables for SOC estimation in a semiarid forest using multiple least-squares regression, tree-based model, and neural network model. Results showed that the CART model with all physical and management variables and 24-2-1 neural networks had the highest predictive ability that explained 81 and 76% of SOC variability, respectively. Neural network models slightly overestimate SOC content. ANNs have a higher ability to detect the effects of management variables on SOC variability and the advantage of CART was to distinguish the effects of physical variables. In both methods the management system dominantly controlled SOC variability in these semiarid forest conditions.