Accurate soil depth information is critical for sustainable crop planning in semiarid regions, where water availability and rooting depth directly influence agricultural productivity. Despite the growing application of machine learning (ML) techniques in digital soil mapping, a clear quantitative assessment of their performance relative to conventional soil mapping approaches is still lacking, particularly in the Central India where soil variability is high and data scarcity is common. This study addresses these gaps by evaluating four ML algorithms, Random Forest (RF), Extreme Gradient Boosting (XGB), Support Vector Regression (SVR), and Cubist, for soil depth prediction in Yavatmal district, Maharashtra (∼13,601 km 2 ). A total of 655 geo-referenced soil profile was collected and used alongside 32 environmental covariates. The XGB produced the highest depth (10–183 cm), followed by Cubist (1.5–178 cm) and RF (11–146 cm). NDVI consistently emerged as the most influential predictor. Model performance was evaluated using independent validation, where RF yielded the highest R 2 (0.43) followed by Cubist (R 2 = 0.39), XGB (R 2 = 0.38), and SVR (R 2 = 0.34). Prediction interval maps (90% confidence) were also generated to quantify model uncertainty. The comparison between conventional and digital mapping revealed that, in digital mapping, the areas of shallow (25–50 cm) and moderately shallow (50–75 cm) soil classes increased by 79.85% and 48.33%, respectively, compared to conventional mapping. The resulting soil depth maps were linked with agronomic planning. These data-driven outputs offer a valuable tool for spatially targeted soil management and precision agriculture in semiarid regions. • ML models improved soil depth mapping accuracy over conventional methods • NDVI and terrain factors were key drivers of soil depth variability • Soil depth maps support site-specific crop planning in semiarid regions
Naitam et al. (Fri,) studied this question.