Soil color is an integrative indicator of pedogenesis and key physicochemical properties. Deep-layer soil color prediction supports practical uses such as depth-targeted fertilization in precision agriculture, depth-resolved carbon stock estimation, and soil health assessment. However, high-confidence 3D (horizontal-vertical) soil color mapping remains challenging and has often been overlooked, leading to limited accuracy. In this study, we explored prediction of 30 m-resolution soil color maps across eight soil depths (1 m profile) in Shaanxi province, China. We constructed and compared four random forest (RF) models: global depth-function RF (GDF) and independent depth-wise RF (IndDW), each trained with a base feature set (Base) or with profile-derived intrinsic properties (pH, SOM, and CEC; PSC). The results demonstrated that GDF models both outperformed IndDWs across depths in both class-wise balanced accuracy and global metrics (OA, Kappa, and Macro-F1). The GDF-PSC model was identified as the optimal model for predicting soil color across different depths, achieving an overall accuracy (OA) of 0.54-0.59, a Kappa coefficient of 0.42-0.58, and a Macro-F1 score of 0.54-0.59. Moreover, SHAP analysis revealed that climate and biological factors, including organism abundance index (OAI), annual precipitation (PRE), total solar radiation (TSR), and land surface temperature (LST), dominated throughout the soil profile, while geological factors such as bedrock depth (BD), parent material (PM), lithology (LIT), and elevation (Ele) had limited explanatory power. We also identified a 15-25 cm soil color transition zone in which OAI, PRE, LST, and evaporation (EVP) were most influential. Regionally, dull yellowish-orange was the dominant soil color across all layers among seven soil colors in the Loess Plateau subregion, and orange soil predominated in the Loess Plateau and Daba Mountains at 0-25 cm but was concentrated in southern Shaanxi below 25 cm. In contrast, grayish-brown was the least distributed, mainly found in the Guanzhong Plain. For most soil colors, high-confidence predictions accounted for over 52.14% of their respective initially mapped areas, with paleyellow as the exception. Our predictions provide a reliable basis for precise soil health diagnostics and agricultural management.