Spatial Assessment of Soil Organic Carbon Density Through Random Forests Based Imputation

Regional estimates of soil carbon pool have been made using various approaches that combine soil maps with sample databases. The point soil organic carbon (SOC) densities are spatialized employing approaches like regression, spatial interpolation, polygon based summation, etc. The present work investigates a data mining based spatial imputation for spatial assessment of soil organic carbon density. The study area covers Andhra Pradesh and Karnataka states of India. Field sampling was done using stratified random sampling method with land cover/use, soil type, agro-ecological regions for defining strata. The spatial data at 1 km resolution on climate, NDVI, land cover, soil type, topography was used as input for modeling the top 30 cm Soil Organic Carbon (SOC) density. To model the SOC density, a Random Forest (RF) based model with optimal parameters and input variables has been adopted. Experiment results indicate that 500 number of trees with 5 variables at each split could explain the maximum variability of soil organic carbon density of the study area. Out of various input variables used to model SOC density, land use / cover was found to be the most significant factor that influences SOC density with a distinct importance score of 34.7 followed by NDVI with a score of 12.9. The predicted mean SOC densities range between 2.22 and 13.2 Kg m^?2 and the estimated pool size of SOC in top 30 cm depth is 923 Tg for Andhra Pradesh and 1,029 Tg for Karnataka. The predicted SOC densities using this model were in good agreement with the measured observations (R?=?0.86).