Sampling-bias-aware machine-learning mapping of geogenic indoor radon hazard in Jeollanam-do, Korea using measured-low negatives

초록

Indoor radon is a naturally occurring hazardous exposure controlled by the coupled effects of source-rock radioelement content, regolith and fracture pathways, terrain setting, soil properties, and the spatial distribution of buildings. We develop a sampling-bias-aware machine-learning framework for mapping geogenic indoorradon hazard susceptibility across Jeollanam-do, South Korea, by integrating lithology-resolved frequency-ratio encoding of geological, geomorphometric, pedological, and geochemical factors with four supervised classifiers: XGBoost, CNN, CNN-LSTM, and a miniature U-Net. The study addresses two recurring limitations that can distort hazardous-exposure mapping from residential environmental monitoring data. First, rather than relying on unmeasured random pseudo-absences, the negative class is constructed from measured-low indoor-radon survey points (those with measured indoor radon below the 148 Bq m-3 reference level used to define positive sites) separated from positive sites by 1000 or 1500 m buffers, thereby grounding both classes in observed environmental measurements. Second, Land Use is used diagnostically to identify residential survey clustering; because its single-factor AUC mainly reflects monitoring design rather than radon generation or transport, it is excluded from the final geogenic hazard predictor set. Across ten random seeds, XGBoost attained the highest mean test AUC (0.81 +/- 0.01; the single model used to generate the susceptibility map reached 0.8140) for the selected 1500 m buffer and nine-direction augmentation setting; the CNN, CNN-LSTM, U-Net and unweighted FR baseline reached mean AUCs of 0.79, 0.77, 0.77 and 0.70 respectively, so non-linear learning improves hazardous-area screening beyond an unweighted FR overlay. The XGBoost-CNN difference was not statistically significant (DeLong p = 0.44; the paired-bootstrap 95% CI of the AUC difference includes zero), so the tabular and spatial architectures are treated as statistically comparable. The resulting susceptibility-score map assigns 412.8 km2, or 3.3% of the province, to the Very High class. Spatial overlay with the KIGAM lithology raster highlights Bulguksa porphyritic granite, the Permian Choongnam Group, Cretaceous dacitic units, and sheltered Quaternary alluvial lowlands as high-susceptibility domains. The resulting map should therefore be used as a lithology-aware hazardous-exposure screening layer for radon-priority investigation, radon-resistant construction planning, mitigation prioritisation, and future stratified monitoring, rather than as a calibrated dwelling-level exceedance probability map.

키워드