Regional mapping of soil heavy metals via a novel deep learning approach: A case study in the Yangtze River Delta, China.

Accurate regional mapping of soil heavy metals is increasingly supported by multi-source environmental covariates, yet predictive performance is often limited by sparse field sampling that cannot capture regional heterogeneity. We developed DLTL, a novel deep learning framework that integrates kriging-based virtual sample augmentation with transfer learning to improve mapping under limited observations. For six metals (Zn, Cu, Cr, Cd, Pb, and As) in the Yangtze River Delta, China, DLTL outperformed benchmark models, improving R2 by 17.36-42.99% over an augmented deep learning baseline, 24.00-95.34% over ordinary kriging (OK), and 34.50-133.93% over random forest (RF). Optimal performance was achieved with 6-km virtual sampling, full-network fine-tuning, and terminal-layer replacement. The shapley additive explanations (SHAP) analysis revealed that soil properties (37-42%) and climate (23-32%) dominated model explanations, with element-specific contributions from anthropogenic indicators and PM2.5. The resulting 1-km resolution maps better delineated localized hotspots, identifying a larger fraction of high ecological-risk areas (3.20% vs 0.07% for RF) while reducing false-positive exceedance of childhood carcinogenic risk (0.09% vs 1.50% for OK). DLTL provides a scalable solution for regional contamination mapping and risk screening in covariate-rich but sample-poor settings.