Spatio-temporal modelling of $\text{PM}_{10}$ daily concentrations in Italy using the SPDE approach

TL;DR

This study develops a Bayesian spatio-temporal model using the SPDE approach to accurately interpolate daily PM10 concentrations across Italy, incorporating meteorological and aerosol data, validated by cross-validation and practical applications.

Contribution

It introduces a novel spatio-temporal interpolation method for PM10 using SPDE with AR1, including 12 monthly models and providing detailed uncertainty measures.

Findings

High correlation between predicted and observed PM10 values (0.79-0.91)

Accurate daily PM10 maps with uncertainty estimates for Italy

Effective reproduction of large-scale data features without artifacts

Abstract

This paper illustrates the main results of a spatio-temporal interpolation process of $\text{PM}_{10}$ concentrations at daily resolution using a set of 410 monitoring sites, distributed throughout the Italian territory, for the year 2015. The interpolation process is based on a Bayesian hierarchical model where the spatial-component is represented through the Stochastic Partial Differential Equation (SPDE) approach with a lag-1 temporal autoregressive component (AR1). Inference is performed through the Integrated Nested Laplace Approximation (INLA). Our model includes 11 spatial and spatio-temporal predictors, including meteorological variables and Aerosol Optical Depth. As the predictors' impact varies across months, the regression is based on 12 monthly models with the same set of covariates. The predictive model performance has been analyzed using a cross-validation study. Our results show that the predicted and the observed values are well in accordance (correlation range: 0.79 - 0.91; bias: 0.22 - 1.07 $\mu \text{g}/\text{m}^3$; RMSE: 4.9 - 13.9 $\mu \text{g}/\text{m}^3$). The model final output is a set of 365 gridded (1km $\times$ 1km) daily $\text{PM}_{10}$ maps over Italy equipped with an uncertainty measure. The spatial prediction performance shows that the interpolation procedure is able to reproduce the large scale data features without unrealistic artifacts in the generated $\text{PM}_{10}$ surfaces. The paper presents also two illustrative examples of practical applications of our model, exceedance probability and population exposure maps.