Learning Neighborhood-Scale Cross-Dependencies Among Air Pollutants, Meteorology and Land Cover Using Mobile Sensing and Transformers

This study integrates high-resolution mobile air-quality sensing with a transformer-based masked autoencoder to explore fine-scale spatial relationships among pollutants, meteorology, and land cover across the Weizmann Institute campus. Using a custom-built miniaturized broadband cavity-enhanced spectrometer (mBBCEAS) for NO₂, along with PM₁, PM₂.₅, O₃, and meteorological sensors, the authors conducted 66 mobile surveys across the study area. A transformer-based masked autoencoder, pretrained on synthetic data, accurately reconstructed full pollutant and meteorological fields from only 25% of the data (R²=0.89) and correctly classified concentration and meteorological intensity levels into ten categories (F1 = 92.9% with one-bin tolerance), revealing that local land cover and wind conditions strongly modulate pollutant gradients over tens of meters. Attention analysis uncovered previously hidden spatial patterns and key variables, offering a data-driven basis for optimizing urban air-quality sampling strategies.