Digital mapping of soil organic carbon at multiple depths using different data mining techniques in Baneh region, Iran

This study aimed to map SOC lateral, and vertical variations down to 1m depth in a semi-arid region in Kurdistan Province, Iran. Six data mining techniques namely; artificial neural networks, support vector regression, k-nearest neighbor, random forests, regression tree models, and genetic programming were combined with equal-area smoothing splines to develop, evaluate and compare their effectiveness in achieving this aim. Using the conditioned Latin hypercube sampling method, 188 soil profiles in the study area were sampled and soil organic carbon content (SOC) measured. Eighteen ancillary data variables derived from a digital elevation model and Landsat 8 images were used to represent predictive soil forming factors in this study area. Findings showed that normalized difference vegetation index and wetness index were the most useful ancillary data for SOC mapping in the upper (0–15cm) and bottom (60–100cm) of soil profiles, respectively. According to 5-fold cross-validation, artificial neural networks (ANN) showed the highest performance for prediction of SOC in the four standard depths compared to all other data mining techniques. ANNs resulted in the lowest root mean square error and highest Lin's concordance coefficient which ranged from 0.07 to 0.20 log (kg/m³) and 0.68 to 0.41, respectively, with the first value in each range being for the top of the profile and second for the bottom. Furthermore, ANNs increased performance of spatial prediction compared to the other data mining algorithms by up to 36, 23, 21 and 13% for each soil depth, respectively, starting from the top of the profile. Overall, results showed that prediction of subsurface SOC variation needs improvement and the challenge remains to find appropriate covariates that can explain it.