Resolving stratified Pb and Cr hazards in 3D at an abandoned steel complex: A heterogeneity-learning approach to...

Lead (Pb) and chromium (Cr) are priority soil hazards: Pb is a potent neurotoxicant and Cr (VI) is a recognized carcinogen, both persisting for decades and threatening groundwater and human health through ingestion and leaching. At large industrial sites, the delineated three-dimensional extent of these metals directly governs remediation volume, cost, and residual exposure risk. To resolve hazard distributions at decision-relevant resolution, we developed D-MSN, a deep-learning-enhanced extension of the Mean Surface with Stratified Nonhomogeneity (MSN) model. D-MSN integrates masked spatial attention, gated residual networks, and meta-polynomial trend learning to jointly capture in-stratum correlation, between-stratum heterogeneity, and cross-stratum dependency. Applied to Pb and Cr at an abandoned integrated iron-and-steel complex in China (2.81 km², 12,900 samples spanning coking, sintering, ironmaking, steelmaking, rolling, and captive power generation) under stratified spatial-block cross-validation, D-MSN outperformed inverse distance weighting, 3D ordinary kriging, 3D-MSN, and DKNN, reducing MAE and RMSE by ∼35% and ∼38% relative to ordinary kriging. The model-derived hotspots spatially coincide with known process units, revealing diffuse Cr signatures near slag-handling and material-storage areas and compact Pb plumes near sintering, blast-furnace, and captive-power-plant operations. Unlike conventional smoothing-based interpolation methods, D-MSN preserves localized hotspot structures and supports source-oriented interpretation. Coupling D-MSN with Monte Carlo dropout further yields probabilistic exceedance maps under GB 36600-2018, identifying decision-uncertain volumes that deterministic interpolators systematically conceal. The practical implications of stratified-heterogeneity-aware modeling for risk-based contaminated site management are highlighted by the fact that the interpolation method alone changes the delineated remediation volume by about 36% in comparison to standard kriging, with cost implications on the order of 10⁸ CNY.