Learning neighborhood-scale cross-dependencies among air pollutants, meteorology and land cover using mobile sensing and transformers

Abstract In this work, we integrate high-resolution measurements of air pollutants with a transformer-based masked autoencoder to explore fine-scale spatial relationships among pollutants, local meteorology, and land cover across a heterogeneous urban area. Using a custom-built miniaturized broadband cavity-enhanced spectrometer (mBBCEAS) for NO 2 , along with PM 1 , PM 2.5 , O 3 , and meteorological sensors, we conducted 66 mobile surveys across the 1.1 km 2 study area over three seasons. A transformer-based masked autoencoder, pretrained on synthetic data, accurately reconstructed full pollutant and meteorological fields from heavily masked, multi-variable observations ( R ² = 0.89) and precisely classified concentration and meteorological intensity levels into ten quantile-based categories (F1 = 92.9% with one-bin tolerance). Attention-derived feature relevancy revealed fine-scale transport and identified key drivers, including winds and land cover, that strongly modulate ground-level pollutant gradients over tens of meters. The results further demonstrate the feasibility of optimized, data-driven urban air-quality sampling strategies.