| Summary: | Air pollution presents a serious health challenge in urban metropolises. While accurately monitoring and forecasting air pollution are highly crucial, existing data-driven models have yet fully captured the complex interactions between the temporal characteristics of air pollution and the spatial characteristics of urban dynamics. Our proposed Deep-AIR fills this gap to provide fine-grained city-wide air pollution estimation and station-wide forecast, by exploiting domain-specific features (including Air Pollution, Weather, Urban Morphology, Transport, and Time-sensitive features), with a hybrid CNN-LSTM structure to capture the spatio-temporal features, and <inline-formula> <tex-math notation="LaTeX">$1\times 1$ </tex-math></inline-formula> convolution layers to enhance the learning of temporal and spatial interaction. Deep-AIR outperforms compatible baselines by a higher accuracy of 1.5%, 2.7%, and 2.3% for Hong Kong and 1.4%, 1.4% and 3.3% for Beijing in fine-grained 1-hr pollution estimation, and 1-hr and 24-hr forecasts, respectively. Saliency analysis reveals that for Hong Kong, spatial features, including street canyon and road density, are the best predictors for NO<sub>2</sub>, while temporal features, including historical air pollutants and weather, are the best predictors for PM<sub>2.5</sub>. For Beijing, historical air pollutant data, traffic congestion, wind direction and seasonal indicator are the best predictors for all pollutants. PM<sub>10</sub> in Hong Kong is achieving the best estimation and forecast accuracy, whilst CO in Beijing is achieving the best results.
|
|---|