Digital mapping of topsoil organic carbon content in an alluvial plain area of the Terai region of Nepal

Topsoil (0–20 cm) soil organic carbon content (SOC, g kg⁻¹) was predicted and mapped at 30 m resolution on an intensively cultivated alluvial plain in the Sarlahi district of Nepal. It has been reported that SOC content in this region has decreased to alarmingly low levels; however, no documented work on the mapping of SOC at high spatial resolutions has been found. We compared the performance of the Stepwise-Multiple-Linear-Regression-Kriging (SMLRK) and Random Forest (RF) techniques for mapping SOC content. Environmental covariates of SOC were selected following the SCORPAN framework of digital soil mapping. Prediction uncertainty was quantified using Quantile Regression Forest technique. Predicted SOC was further calculated for different soil and land cover unit combinations. It was found that RF performed better than SMLRK as evidenced by more favourable error statistics. Silt content, distance to major river systems, precipitation during coldest quarter of the year, and NDVI for January-February were found to be the most important variables affecting SOC content in this region. Mean SOC of the study area was predicted as 10.98 ± 0.01 g kg⁻¹. Calcaric Phaeozems under Shrubs (14.85 ± 0.09 g kg⁻¹) and forests (14.54 ± 0.02 g kg⁻¹) had the highest mean SOC contents. Cultivated lands within irrigation command areas had less SOC content (9.41 ± 0.01 g kg⁻¹) than those outside the areas (11.84 ± 0.01 g kg⁻¹). The SOC spatial distribution map could guide efforts to prioritize SOC content enhancement activities in this area. It was concluded that in such low-relief alluvial regions, the processes of silt deposition due to the proximity to river systems are key predictors than other variables, and that the RF technique could predict SOC content better than SMLRK. The success of activities to raise SOC contents might thus be subject to the control of silt deposition from the upstream hills.