Soil pH is a critical property influencing soil health and functions, nutrient availability and microbial activities, and agricultural productivity. Interpreting machine learning models in soil science is a challenge, despite their increasing application. We used a dataset of 2651 soil samples up to 60 cm depth to understand the drivers controlling spatial and depth distribution of soil pH and related them to soil-forming factors in foothills mountains and semi-arid steppes of southern Ural (Russia). Machine learning approach allowed to analyse the effects of key soil-forming factors (climate, topography, vegetation, soil and parent materials) for the predictions and the role of covariates utilizing Shapley values, a game theory-based method to quantify the average marginal contribution of a predictor. The developed models explained 62 %, 56 % and 54 % of the pH variation in 0-20, 20-40 and 40-60 cm, respectively. Climate (precipitation, cloud cover and surface temperature), soil type and elevation were the most important factors of soil pH across all depths. When precipitation in December exceeds 30-35 mm, cloud cover 58-60 % and elevation 400-450 m, the model predicted a lower pH compared with a mean level across all depths. The generated pH maps also revealed a change in soil pH from mountainous forested ecosystems to semi-arid steppe landscapes. These findings are mainly explained by the difference in precipitation-driven leaching and evapotranspiration-induced salt accumulation in soils in the area (9,500 km2). Our study underscores the complexity and non-linearity of the relationships between pH and the environmental variables, providing valuable insights into their variations across both horizontal and vertical spatial dimensions.