Modeling city-wide intra-urban variations of ultrafine particles via low-to-middle cost sensors: Comparisons between land use regression and machine learning approaches

Accurate exposure assessment for pollutants such as black carbon (BC) and ultrafine particles (UFPs) remains challenging in cities lacking dense monitoring networks. This study develops city-wide prediction models for BC, UFPs, PM 10 , and PM 2.5 in Zurich city, Switzerland based on the results of a small-scale monitoring campaign using low-to-middle cost sensors. Two conventional land use regression (LUR) methods including LM-LUR & generalized additive models (GAM), and two non-linear machine learning (ML) methods including random forest (RF) & XGBoost were used in models training. In total, 16 high-resolution (50 m × 50 m grid) models were developed and externally validated. The study further examined the spatial extrapolation of predictions from a small-scale modeling domain covering a 500 m 2 sampling area to a city-wide prediction domain of 87.88 km 2 . Among the four modeling approaches, RF demonstrated the most robust and consistent performance, with a repeated leave-one-out cross-validation (LOOCV) R² ranging from 0.74 to 0.87 and externally validated prediction errors within 2.46 % – 39 % across pollutants. While non-linear models outperformed conventional LUR approaches, LM-LUR and GAM exhibited inconsistent performances when applied to the larger prediction domain. Despite the limitations of a small-scale monitoring campaign in representing temporal variation, the intra-urban spatial variability captured by the integrated campaign and ML modeling approach provide valuable insights into future air pollution exposure and health studies in under-monitored regions. • Modeling based on small-scale monitoring campaign & low-to-middle cost sensors. • Intra-urban spatial variability of BC, UFPs PNC, PM 10 , PM 2.5 was mapped. • Two conventional LUR & two machine learning models were developed. • 16 city-wide extrapolations of pollutant maps (50 m × 50 m) were investigated. • RF outperformed others with R² of 0.74–0.87 & 2.4 %–39 % external validation error.