Explainable Machine Learning and Geospatial Assessment of Wildfire Smoke Impacts on Urban Air Quality in Split, Solin, and Kaštela, Croatia

Wildfires increasingly contribute to urban particulate matter (PM) exposure, particularly fine particles (PM2.5), through atmospheric transport processes influenced by meteorological conditions and terrain complexity. This study investigated wildfire impacts on PM10 and PM2.5 concentrations in Split, Solin, and Kaštela (Croatia) using a terrain-aware wildfire transport framework combined with statistical and machine learning (ML) approaches. Daily PM observations (2016–2024) from three air quality monitoring stations were integrated with meteorological data from six stations, wildfire polygons, and a digital elevation model (DEM). A wildfire influence index accounting for fire size, transport distance, wind conditions, and terrain-modified airflow was evaluated using Ordinary Least Squares (OLSs) regression, Random Forest (RF) modelling, and SHAP (SHapley Additive exPlanations) analysis. Results showed stronger wildfire-related effects for PM2.5 than for PM10, while meteorological variables remained the dominant predictors of PM variability. RF models improved predictive performance relative to OLS, achieving R2 = 0.474 for PM2.5 and R2 = 0.416 for PM10. SHAP analysis identified precipitation, temperature, and lagged wildfire transport variables as important predictors. A total of 84 wildfire events were classified as effective wildfires, with most measurable impacts occurring within approximately 30–70 km of monitoring stations, indicating that wildfire impacts on urban air quality in Mediterranean coastal environments are strongly mediated by atmospheric transport and meteorological conditions. The proposed framework demonstrates the potential of explainable and geospatially informed ML for environmental monitoring and wildfire-related urban air quality risk assessment.