Prediction models of polycyclic aromatic hydrocarbons in urban road dust via magnetic properties, particle size, and urban environmental features

This study explores the novel application of artificial neural networks to predict polycyclic aromatic hydrocarbons (PAHs) pollution in urban road dust by integrating magnetic properties, particle size distributions, and urban environmental features. A comprehensive dataset from 284 samples from Warsaw, Poland, included magnetic susceptibility (χ), saturation magnetization (Ms), remanent magnetization (Mrs), traffic intensity, granulometric fractions, and parameters such as building height, building layout, connection to the municipal central heating network, and geospatial coordinates. Principal component analysis (PCA) revealed that ∑PAH16 accumulation patterns are governed by the interplay between magnetic proxies (χ, Ms​, Mrs​), traffic intensity (T), and urban structural configurations, specifically heating grid status (C), building height (H), and building continuity (B; attached vs. detached structures), collectively accounting for 60.21% of the total variance. The predictive performance of the models was evaluated using 5-fold cross-validation. While the Linear Regression (LR) model showed low and unstable accuracy (R 2 ranging from 0.05 to 0.32, mean 0.18), the Random Forest (RF) model provided a significantly more robust framework for capturing the nonlinear relationships between variables. SHAP (SHapley Additive exPlanations) analysis was employed to interpret the RF model, revealing that grain size fraction (F) and geospatial coordinates (LA, LO) were the primary drivers of PAH accumulation. In contrast, factors such as traffic intensity and building layout exhibited a marginal influence. The comparison of modeling approaches revealed a progressive increase in predictive performance as the ability to capture nonlinear and local relationships improved (R 2 ≈0.18 for linear regression, ≈0.26 for ANN, and ≈0.40 for RF), indicating that PAH accumulation is governed by complex, context-dependent interactions rather than simple independent predictors. These findings demonstrate that integrating magnetic properties and urban features using machine learning provides a powerful tool for identifying pollution hotspots and understanding the complex mechanisms underlying the distribution of organic pollutants in urban environments.