Overcoming data scarcity in soil depth prediction: A machine learning approach for India’s Black Soil Region (BSR)

Soil depth (SoD) is a fundamental property controlling ecosystem services, agricultural productivity, and hydrological processes, yet its spatial prediction remains a challenge in data-scarce regions. This study demonstrates the effectiveness of a Quantile Regression Forest (QRF) model to predict the spatial distribution of SoD and, crucially, quantify its prediction uncertainty using a limited dataset in the black soil region (BSR) of Amravati, Maharashtra, India. Ninety-two soil profiles were integrated with a suite of environmental covariates derived from terrain, climate, and remote sensing data. Key predictors were identified through recursive feature elimination. The QRF model explained 86% of the variance (R² = 0.86) with a root mean square error of 19.99 cm (sqrt-transformed) and 12.4 cm (back-transformed). A Lin's concordance correlation coefficient (CCC) of 0.93 indicated excellent agreement between predicted and observed values. The resulting map revealed distinct patterns: deeper soils in depositional valleys and plains (>150 cm) and shallower soils on erosional plateau tops and hillslopes (<50 cm). Predictive uncertainty was lowest in well-sampled alluvial plains and highest in sparsely sampled steep landscapes. The QRF model successfully handled non-linear relationships and provided robust, interpretable predictions from sparse data. The high-resolution SoD map with quantified uncertainty is a vital tool for optimizing agricultural water use, preventing land degradation, and implementing targeted soil conservation practices in this rainfed agricultural system.