In this study, bacterial α-diversity, represented by Chao1 richness and Shannon diversity, was quantified and mapped across the acidic red and yellow soils of southern China by integrating a comprehensive meta-database of 920 georeferenced observations with multi-source environmental data and machine-learning models. Chao1 and Shannon were modeled using Random Forest, eXtreme Gradient Boosting, and Support Vector Machine algorithms. Among these, Random Forest consistently achieved the highest predictive accuracy and robustness (R² > 0.83). SHapley Additive exPlanations (SHAP) analysis revealed that soil pH acted as the fundamental filter for both diversity dimensions, while they were further shaped by distinct environmental controls. The Normalized Difference Vegetation Index (NDVI) negatively constrained species richness, reflecting the prevalence of intensively managed monoculture systems in this region, whereas temperature exerted a strong negative control on community evenness. Spatial predictions showed pronounced regional contrasts among the Southwest, Middle–Lower Yangtze, and Southern China agricultural zones, with uncertainty patterns closely linked to terrain complexity and data density. The results demonstrate that bacterial richness and evenness respond to distinct yet interacting environmental controls, highlighting the necessity of multidimensional diversity assessment. This study provides a scalable and interpretable framework for digital soil mapping of microbial diversity, offering spatially explicit insights for managing acid-sensitive agroecosystems under environmental change.