Evaluating the prediction success of soil organic carbon stock in pasture land using different modeling performance metrics

Many national and international initiatives depend on detailed spatial data on changes in soil organic carbon stock (SOC stock) at various scales to support policies aimed at land degradation neutrality and climate change mitigation Developing tools to accurately model the spatial distribution of SOCstock at national scales is a priority for both monitoring soil organic carbon (SOC) changes and contributing to global carbon cycle studies. The primary goal of this study was to evaluate and compare various spatial performance metrics used to assess the accuracy of predicting soil SOC and SOCstock content in a semi-arid pasture. Soil samples were taken from 0-20 cm soil depth at 150 random sampling points. Spatial structure of SOCstock and SOC were modelled by ordinary kriging The soil pH varied from slightly acidic (6.34) to neutral (7.19), and salinity was not an issue in the study area. Lime content, with an average of 2.04%, stands out as the most variable soil property, with a coefficient of variation (CV) of 61.76%. The carbon stock ranged from 23.46 to 65.36 tons ha-1, with an average carbon stock of 43.28 tons ha-1 calculated. In the study area, SOC (%) and stoniness (%) had the shortest autocorrelation distance (21.00 m), while bulk density had the longest (27.00 m). The prediction errors indicated that parameters in the random sampling did not result in better predictions using the OK technique.The results indicated that SOC content can exhibit significant spatial variability even within a small area, highlighting the need for site-specific management in semi-arid pastures. In order to achieve high accuracy and success in modeling, metrics of the performance such as RRMSE, RMSE and MAPE should be used that minimize the effect of the relevant soil property measurement unit.