Deep Spatio-Temporal Neural Network for Air Quality Reanalysis

TL;DR

This study tackles the challenge of air quality reanalysis in urban environments by addressing spatial generalization alongside temporal dynamics. It introduces AQ-Net, a hybrid neural network that combines LSTM with multi-head attention, cyclic time encoding, and a neural kNN interpolator to predict $PM_{2.5}$ for both observed and unobserved stations over 168 hours. Across 584 stations in northern China (2013–2017), AQ-Net substantially outperforms baselines on short- and long-term horizons, with consistent gains from cyclic encoding and effective spatial interpolation. The results demonstrate accurate city-scale PM$_{2.5}$ mapping and robust reanalysis capability, offering practical value for health alerts and policy decisions and generalizing to other pollutants and urban contexts.

Abstract

Air quality prediction is key to mitigating health impacts and guiding decisions, yet existing models tend to focus on temporal trends while overlooking spatial generalization. We propose AQ-Net, a spatiotemporal reanalysis model for both observed and unobserved stations in the near future. AQ-Net utilizes the LSTM and multi-head attention for the temporal regression. We also propose a cyclic encoding technique to ensure continuous time representation. To learn fine-grained spatial air quality estimation, we incorporate AQ-Net with the neural kNN to explore feature-based interpolation, such that we can fill the spatial gaps given coarse observation stations. To demonstrate the efficiency of our model for spatiotemporal reanalysis, we use data from 2013-2017 collected in northern China for PM2.5 analysis. Extensive experiments show that AQ-Net excels in air quality reanalysis, highlighting the potential of hybrid spatio-temporal models to better capture environmental dynamics, especially in urban areas where both spatial and temporal variability are critical.

Figures (8)

• Figure 1: Daily mean PM2.5 prediction over northern China using AQ-Net. $\bigcirc$ indicates "visible" stations, which provided historical data for training, whereas $\triangle$ represents "hidden" stations for which only geographic coordinates were available (handled by our neural kNN module). The color scale ranges from blue (low PM$_{2.5}$) to red (high PM$_{2.5}$), highlighting pollution hotspots in specific provinces.
• Figure 2: Overview of the proposed AQ-Net. The input includes historical pollutant concentrations, and visible station coordinates. An LSTM extracts temporal dependencies, enhanced by Multi-Head Attention to highlight critical time steps. After temporal pooling, a neural kNN module performs spatial interpolation for unobserved stations (red markers).
• Figure 3: Comparison of PM2.5 reanalysis for different time slots over seven days. The 4PM-7PM period exhibits greater variability, suggesting increased pollution activity during the late afternoon.
• Figure 4: Visualization of the evolution of attention weights for selected two heads. Head 2 reacts to short-term variations, while Head 1 maintains stable attention, capturing long-term patterns.
• Figure 5: Visualization of the attention heatmap across reanalysis and training days. A diagonal trend suggests the model prioritizes recent observations, while deviations indicate potential long-term dependencies.