Semi-parametric methods for long-term trend detection in air quality: The North-Eastern Italian Alps case

Atmospheric pollution has received increasing attention in recent years, particularly in regions such as Northern Italy's Po Valley, where the presence of extended industrial settings, high population density and orographic factors complicate efforts to maintain pollutant concentrations within the mandated standards. This work focuses on assessing air pollution trends in the North-East of Italy, using daily measurements of NO, NO$_2$, O$_3$, PM$_{10}$, and PM$_{2.5}$ collected between 2009 and 2023 from 40 monitoring stations located in foothill and mountainous areas. By means of an integrated regression modelling framework based on Generalised Additive Models that combines Tobit likelihood with Box-Cox transformation, we extract trends on long-term variations accounting for seasonal patterns and weather predictors. Model performance varied by pollutant, with higher accuracy for gases ($R^2$ up to 0.86 for O$_3$) than for particulate matter ($R^2$ of approximately 0.66). Variability, evaluated by means of Shapley values, was driven mainly by autoregressive persistence (0.53--0.57) and seasonality (0.15--0.38), with wind effects being relevant for PM (contributing approximately 0.15). Trends showed strong declines in NO (up to $-50\%$) and PM (around $-30\%$), while O$_3$ concentrations increased by roughly $20\%$. Meteorological anomalies contributed only marginally to the overall variance. The high variability observed in the detected long-term changes underlined the high heterogeneity of air pollution evolution across the study area. This indicated that short-term factors played a dominant role, whereas long-term changes were smaller in magnitude and exert a weaker influence on local air pollution.