Modelling and prediction of soil organic carbon using digital soil mapping in the Thar desert region of India

In the present study, the distribution of soil organic carbon (SOC) was investigated using digital soil mapping for an area of ~29 lakhs ha in Bikaner district, Rajasthan, India. To achieve this goal, 187 soil profiles were used for SOC estimation by Quantile regression forest (QRF) model technique. Landsat data, terrain attributes and bioclimatic variables were used as environmental variables. 10-fold cross-validation was used to evaluate model. Equal-area quadratic splines were fitted to soil profile datasets to estimate SOC at six standard soil depths (0–5, 5–15, 15–30, 30–60, 60–100 and 100–200 cm). Results showed that the mean SOC concentration was very low with values varied from 1.18 to 1.53 g kg-1 in different depths. While predicting SOC at different depths, the model was able to capture low variability (R2 = 1–7%). Overall, the Lin's concordance correlation coefficient (CCC) values ranged from 0.01 to 0.18, indicating poor agreement between the predicted and observed values. Root mean square error (RMSE) and mean error (ME) were 0.97 and 0.16, respectively. The values of prediction interval coverage probability (PICP) recorded 87.2–89.7% for SOC contents at different depths. The most important variables for predicting SOC concentration variations were the annual range of temperature, latitude, Landsat 8 bands 2, 5 and 6. Temperature-related variables and remote sensed data products are important for predicting SOC concentrations in arid regions. We anticipate that this digital information of SOC will be useful for frequent monitoring and assessment of carbon cycle in arid regions.