End-to-end deep learning for pollutant prediction using street view images

Abstract

Urban outdoor air pollution poses significant threats to public health, underscoring the need for accurate pollutant concentration estimation. This study presents an innovative end-to-end pollutant prediction model (E2EPPM) that directly predicts pollutant levels from street-level imagery. To the best of our knowledge, the model is the first to achieve pollutant concentration estimation at the street scale solely based on street-view images collected through mobile monitoring. The framework comprises three stages: feature extraction from image sequences, temporal feature extraction across sequences, and pollutant concentration fitting based on the Kolmogorov-Arnold theorem. E2EPPM was evaluated using datasets and street-view images from Augsburg, Beijing, and Hotan, focusing on particulate matter (PM_1.0, PM_2.5, PM₁₀, particle number concentration (PNC), lung-deposited surface area (LDSA), ultraviolet PM (UVPM), black carbon (BC), and brown carbon (BrC)) and four gaseous pollutants (CO, NH₃, SO₂, O₃). The model achieved over 90 % explanatory power for all pollutants. Ablation experiments demonstrated its effectiveness, while SHAP analysis identified vegetation, buildings, sky, and vehicles as the most significant contributors to pollutant levels. E2EPPM offers a robust approach for urban air quality assessment, providing actionable insights for urban planning and pollution exposure evaluation.