Oasis farmlands are the core grain production carrier in arid regions, where soil heavy metal contamination directly threatens regional agricultural product safety and resident health. This study developed an integrated framework of stratified spatial sampling, machine learning clustering and multi-scale statistical modeling to investigate heavy metal spatial characteristics, environmental driving mechanisms and health risks in the Hotan Oasis, based on 35 paired soil and maize tissue samples collected across three depth layers (0-20 cm, 20-60 cm, 60-80 cm). Self-organizing maps (SOM) effectively identified deep-soil anomalous samples, providing empirical evidence for pronounced small-scale vertical heterogeneity of heavy metal distribution. The combination of conservative risk assumption and two-dimensional Monte Carlo simulation (2D-MCS) enabled dual quantification of parameter uncertainty and inherent variability in health risk assessment: adult risks stayed within safety thresholds, while the 95th percentile of non-carcinogenic risk (1.16) for sensitive child subpopulation exceeded the threshold, and total carcinogenic risk (TCR, 4.4 & times;10-5) falls within the risk-concern range. Geographically Weighted Regression (GWR) outperformed the global Ordinary Least Squares model by addressing distinct spatial non-stationarity of driving factors, especially in deep soil layers. Linear Mixed-Effects Models confirmed Cd as the only element with significant depth dependence. Redundancy Analysis revealed TP/ TN stoichiometry and pH (saline-alkaline stress proxy) as key regulators of heavy metal bioavailability, despite a 15.3% explanatory power. This study provides methodological support and data reference for stratified contamination control and targeted health risk management in arid oasis farmlands.