Predictive mapping of indoor radon in Spain using graph neural networks

Abstract

Indoor radon exposure represents a major environmental health risk and is a leading cause of lung cancer among non-smokers. Because direct radon measurements are time-intensive and costly, radon risk assessment in Spain largely relies on national radon potential maps that characterise geogenic hazard but do not account for building-specific factors that strongly influence actual indoor radon concentrations. In this dissertation, graph neural networks (GNNs) are investigated as a data-driven framework for predicting indoor radon levels in Spain by jointly modelling geogenic context and building-level characteristics. The proposed approach integrates heterogeneous information sources, including geogenic information such as radon potential and lithology, together with building attributes such as floor level, construction material, building age, and season of measurement. Spatial relationships between observations are represented at the postal-code level, enabling explicit modelling of spatial dependencies that are ignored by traditional tabular methods. Several graph-based representations are proposed, ranging from aggregated postal-level graphs to building-level homogeneous and heterogeneous architectures, and are evaluated on binary and multiclass radon classification tasks. In this setting, the study is structured around two complementary evaluation objectives. First, different graph model variants are assessed under spatial cross-validation and region-wise hold-out experiments in order to identify architectures that generalise well across regions. The results indicate that, given the limited dataset size, simpler homogeneous graph models consistently outperform more complex heterogeneous variants, underscoring the importance of aligning model complexity with data availability. Second, the selected homogeneous graph model is compared with strong non-graph tabular baselines, including Random Forests (RF), Support Vector Machines (SVM), and the national radon potential (RP) map. In the binary classification setting, the GNN consistently outperforms RP, although the improvement remains modest, while the trainable tabular baselines do not substantially improve upon the RP map. In contrast, in the multiclass setting, the performance of RP deteriorates sharply, whereas all trainable models (both graph-based and tabular) achieve markedly superior results. This highlights the limitations of geogenic-only indicators for fine-grained indoor radon categorisation. Overall, this work demonstrates that graph neural networks offer a principled and effective framework for indoor radon prediction when spatial dependencies and building-level factors are central to the task, improving upon geogenic-only indicators such as radon potential in settings where additional contextual information is available. At the same time, it underscores that graph-based models are not universally superior and must be evaluated carefully against simpler alternatives.