Abstract
Ground-based air quality monitoring networks in Southern Europe are often sparse, leaving large areas, such as rural plains, mountain ranges, and island groups, unmonitored. Standard spatial interpolation methods like Inverse Distance Weighting or Kriging struggle under these conditions, since they cannot easily incorporate physical terrain or provide reliable uncertainty estimates. This paper presents a Hierarchical Bayesian Spatial framework to address the sparse sensor coverage. We model the annual mean concentration of particulate matter (PM2.5) across Greece using data from the European Environment Agency (EEA) e-Reporting network for 2023. We combine ground measurements with terrain elevation from digital elevation models. The continuous spatial surface is discretized using a Finite Element mesh via Stochastic Partial Differential Equations (SPDE). We fit the model deterministically using Integrated Nested Laplace Approximations (INLA). Adding terrain elevation and a sum-to-zero spatial constraint resolved parameter baseline confounding. The model fit improved significantly, increasing the marginal log-likelihood from −41.97 to −39.87. The model achieved an R2 of 0.91 and a Root Mean Square Error (RMSE) of 2.83 μg/m3. The global national baseline concentration was thus estimated at 17.74 μg/m3 (95% Credible Interval: [11.34, 23.88]). Elevation indicated a negative effect (βelev = −7.24 μg/m3 per km), with an 89.2% posterior probability of reducing PM2.5. Spatial autocorrelation persisted across a regional scale of ρ ≈ 297 km. Full posterior variance mapping identified Crete, the Ionian and the Aegean Seas, and Eastern Thrace as zones of maximum predictive uncertainty. These findings provide a quantitative baseline for prioritizing physical sensor deployment in undersampled regions.