Abstract
Street-level imagery is an appealing, low-cost complement to air-quality stations, but a camera is only useful when its pixels respond to the pollutant of interest. For particulate matter they do: aerosols scatter light along the line of sight, so distant buildings, sky boundaries, and horizon regions lose contrast as aerosol loading rises. Swin-AQI is a depth-aware Swin Transformer for monocular PM2.5 and PM10 estimation from street-level urban images: a MiDaS-derived depth-coherence bias steers post-backbone attention, learnable Fourier features encode the capture hour, and Beer–Lambert and Koschmieder terms keep predictions consistent with visibility physics. We supervise with SILAM atmospheric fields. On 1,977 daylight images from Ho Chi Minh City, Swin-AQI reaches R²=0.80 for PM2.5 and R²=0.76 for PM10 on a random held-out test split — a within-distribution result, since only 35 distinct label combinations span the whole dataset — outperforming every CNN and transformer baseline under a standard single-head MSE objective on the same images and labels. Our ablations are candid about why: hour-of-day and uncertainty weighting contribute the most, while depth-aware attention and the physics terms add smaller but separable gains. The scope is deliberately constrained too: Swin-AQI is a single-city, moderate-pollution estimator whose gaseous outputs stay contextual.