A Functional Land Use Regression Model for NO2 Concentration in the Italian Alpine Region

Air pollution is a major risk factor for adverse health outcomes. Because monitoring stations are available only at a limited number of locations, predicting the spatial distribution of pollution is essential for exposure assessment. Land Use Regression (LUR) models address this problem by relating pollutant concentrations to geographical and environmental characteristics, such as traffic, land use, and industrial activities. However, traditional LUR models often focus on average concentrations, overlooking crucial intra-day variability that can impact health outcomes. To address this, we introduce a novel functional LUR (FLUR) model designed to estimate hourly NO2 concentrations. Hourly NO2 measurements were derived from a comprehensive dataset of 41 air monitoring stations in the Italian Alpine foothills and mountains collected throughout the year 2023. Our functional penalized regression model considered the hourly daily profile of log-transformed NO2 concentrations as dependent functional variable and a combination of scalar and functional meteorological and spatial variables. Variable selection was carried out through a forward selection approach adapted to functional data. The approach was based on maximising the adjusted explained variance while avoiding concurvity in the selected predictors, taking into consideration the direction of the daily functional effect. The resulting model was then adjusted for spatio-temporal dependence using a Gaussian process smoother and an autoregressive temporal component. Five key predictors influencing hourly NO2 concentrations were identified: all buildings within 1000 m, mean slope within 100 m, primary roads within 2500 m, herbaceous land cover within 2500 m, and average wind speed. Each predictor exhibits distinct temporal patterns of influence throughout the day. The estimated model, validated using a Leave-One-Station-Out Cross-Validation (LOOCV) procedure, reported a good overall fit (adjusted R2 = 64.1%, LOOCV R2= 58.0%) and 95\% LOOCV prediction coverage of 94.5%. Compared with standard LUR models, the FLUR model provides prediction of hourly NO2 exposure by taking to account functional spatial determinants in a complex topographical region.