Dataset for probabilistic tail-risk assessment of NORM and heavy metals in estuarine sediments and mining soils

Environmental impact assessments often rely on deterministic point estimates, which can understate tail risks: low-probability events whose ecological and human-health consequences may be severe in heterogeneous environments. This study develops and applies a probabilistic risk framework that examines gamma radiation from naturally occurring radioactive material (NORM) as a potential spatial indicator of co-occurring heavy metals. Ordinary kriging combined with spatial block cross-validation was used to map contamination anomalies and to reduce spatial autocorrelation leakage in validation. Gamma spectra were deconvolved with a softmax-constrained deep autoencoder that enforces non-negative, unit-sum abundances. The resulting measurements were then propagated through Monte Carlo simulation using Latin hypercube sampling (LHS) to estimate upper-tail (95th-percentile) risk. Field data from the Phan Me coal-mining area () and the Ha Long–Ba Che estuarine sediment corridor () in Vietnam were analysed as separate geomorphic domains to avoid confounding from Simpson’s paradox. At Ha Long, the deterministic mean Nemerow pollution index (NPI = 0.51) was below the benchmark of 1.0, whereas the simulated 95th-percentile (P95) NPI was 1.25, indicating that the upper tail of the simulated distribution exceeds this benchmark. At Phan Me, the P95 excess lifetime cancer risk (ELCR) reaches , about 2.3 times the deterministic mean (4.01 × 10⁻⁴); in the Monte Carlo output this indicates a wider upper tail rather than a confirmed hazard. Sobol analysis attributed most of the simulated ELCR variance to a composite source-term intensity variable (total-order index 0.82; first-order index 0.76), whereas exposure and dose-conversion parameters contributed little under the assumed input distributions. Combining radiological spatial indicators, block cross-validation and LHS modelling can reveal upper-tail contamination and risk that mean-based reporting does not show, but the findings rest on a limited number of samples at two sites and require validation.