Soil Depth Prediction through Soil-Landscape Modelling Using Machine Learning in the Vidarbha Region of Central India Moharana P.C.*, Naitam R.K., Kumar Nirmal, Nagaraju M.S.S., Biswas H., Patil N.G. ICAR-National Bureau of Soil Survey and Land Use Planning, Nagpur, 440033, India *Corresponding author (Email: pravashiari@gmail.com)
Online Published on 02 August, 2024. Abstract Spatial variation of soil depth is poorly understood and rarely mapped. It is still difficult to predict soil depth with a small sample size in a large area with complex landscapes. In the present study, the spatial distribution of soil depth was investigated using machine learning (ML) techniques for an area of ~4.9 lakh ha in the Washim district of the Vidarbha region, Maharashtra. Thirty-one environmental variables were selected, and 150 sampling points were used for soil depth estimation. Three ML algorithms namely, random forest (RF), cubist, and extreme gradient boosting (XGBoost) were evaluated to map the soil depth distribution. Landsat data and terrain attributes were used as environmental variables. Among the covariates, relative slope positions (RSP), channel network base level (CNBL), channel network distance (CND), and valley depth (VD) were identified as the most influential variables in soil depth modeling. The RF model prediction was the best with R2 (Calibration) of 0.93 and R2 (Validation) of 0.26. Predicted soil depth varied from 21 to 127 cm in RF, 2.47 to 225 cm in XGBoost and 14 to 161 cm in Cubist models. This demonstrates that the model and variables selection are effective ways to predict soil depth. The study presents new ideas to predict soil depth by digital mapping techniques, which will be useful in comprehensive land use planning for the Vidarbha region by policymakers and planners. Top Keywords Digital soil mapping, Machine learning algorithms, Soil depth, Vidarbha regions of India. Top |