Refining Air Pollution Exposure Estimates: A Comparison of Citywide and Neighborhood Land Use Regression Models in Toronto

Land use regression (LUR) models assess air pollution exposure but often struggle with transferability (predicting concentrations in areas without measurements) and generalizability (capturing spatial patterns across neighborhoods). This study evaluated transferability and generalizability of Toronto City LUR models for black carbon (BC) and ultrafine particles (UFP) using mobile monitoring data. Models were developed using multiple linear regression (MLR) and XGBoost under three spatial configurations: Toronto City (TC), Toronto City minus a neighborhood (TCM-NB), and neighborhood-specific (NB). Transferability of TCM-NB models and generalizability of TC models were tested using neighborhood-specific data and compared to NB models. XGBoost outperformed MLR, achieving <i>R</i>² of 0.77 for UFP and 0.54 for BC in TC models compared to 0.32 and 0.27 for MLR. TC models exhibited poor generalizability, with <i>R</i>² dropping to 0.1 in certain neighborhoods. Similarly, TCM-NB models exhibited limited transferability, with MLR slightly outperforming XGBoost (<i>R</i>² of 0.3 vs 0.2). Hyperparameter tuning with spatial cross-validation improved XGBoost transferability and generalizability, with <i>R</i>² increases of up to 0.2 for both UFP and BC depending on the neighborhood. These findings highlight the importance of monitoring campaigns covering diverse urban environments and adopting tailored modeling approaches to capture neighborhood-specific pollution sources to advance air pollution exposure assessment.