High-precision, multi-tier categorical soil classification faces critical bottlenecks, including the neglect of spatial context by conventional pixel-based models, the error cascade propagation phenomenon in multi-level classification networks, and the disruption of geophysical directional anisotropy by traditional geometric data augmentation. To address these challenges, in this study, we propose a multi-level soil classification model based on MTSC-ResNet-Trans, which organically couples residual convolutional blocks with a 3-layer Transformer encoder to model long-range spatial dependencies. The framework integrates a geospatial-safe data augmentation pipeline to preserve the topological fidelity of absolute geographic coordinates alongside four decoupled parallel multi-task classification heads to substantially suppress inter-level error propagation. Evaluated in the Southern Foothills of Qilian Mountains using 18 environmental covariates, the framework achieves an Overall Accuracy of 0.8931 at the Great Group level under conventional random splitting. Under a distance-stratified spatial evaluation—which isolates the contribution of spatial autocorrelation to accuracy estimates—the framework maintains robust performance, with MTSC-ResNet-Trans consistently outperforming pixel-based baselines (Random Forest) by approximately 3.7 percentage points even at spatial separation distances exceeding 400 m. This protocol transparently decomposes predictive accuracy into a component attributable to spatial proximity and a component reflecting reduced spatial proximity performance. Across the four taxonomic levels, accuracy decay is suppressed to 5.11%. Although spatial-block cross-validation indicates a lower regional extrapolation accuracy (OA = 0.7389 ± 0.0671), the decoupled parallel framework provides an effective and robust baseline for high-resolution regional digital soil mapping.