Predictive Soil Mapping of Key Soil Properties in Western Ghats, India

Abstract

A study was conducted to map the key soil properties such as pH, organic carbon (OC), cation exchange
capacity (CEC), sand, silt, clay and bulk density (BD) in part of Western Ghats, South India using three
machine learning algorithms (random forest, cubist and support vector machine). Primary and secondary
terrain attributes, vegetation indices and bioclimatic variables were used as environmental variables for
prediction of soil properties. Equal-area quadratic splines were fitted to 173 soil profile datasets collected
over the study area to estimate the soil properties at six soil depths (0-5, 5-15, 15-30, 30-60, 60-100 and
100-200 cm) as per GlobalSoilMap specifications. The models were calibrated using 80% of the samples
(138) and validated using remaining 20% of the samples (35). The accuracy of the performance was
assessed based on coefficient of determination (R2), concordance correlation coefficient (CCC), root mean
square error (RMSE) and mean error (bias). The random forest model outperformed other two models with
high R2 and minimal RMSE for most of the soil properties. The model explained 41, 42, 31 and 36% of
variation for surface pH, CEC, OC and BD, respectively. The high resolution (250 m) predicted soil
properties aid the policy makers to revert the land degradation process and to preserve soil quality by
executing suitable land use policies.