Air pollutant prediction by spatial-temporal information reconstruction and fusion

Accurate prediction of air pollutant concentrations is crucial for urban public health and environmental management. While previous studies have explored the influence of various factors and spatial-temporal relationships on air pollution, they often fall short in precisely capturing the complexity of influencing variables and in thoroughly understanding the spatial-temporal evolution patterns of pollutant concentrations. To address this issue, we develop a novel Multi-spatial-temporal Dynamic Information Fusion (MSTDIF) model based on a parallel learning mechanism. Specifically, we select information on pollutant-related influencing factors within and among cities and reconstruct them into Intra-city Spatial-Temporal (Intra-ST) and Inter-cities Spatial-Temporal (Inter-ST) information patterns, respectively. Furthermore, we design the Intra-city Spatial-temporal Multi-factors Interaction (Intra-SMI) module to extract multivariate features from the Intra-ST pattern and establish the Inter-cities Spatial-temporal Pollutant Synergy diffusion (Inter-SPS) module to extract spatial-temporal synergistic diffusion features from the Inter-ST pattern. By fusion operation integrating cutting-edge spatial-temporal mining techniques and decision networks, these two types of features are effectively fused to achieve accurate prediction. Extensive experiments conducted on real-world datasets covering six major air pollutants demonstrate that the MSTDIF model significantly outperforms nine state-of-the-art benchmark models in terms of both prediction accuracy and interpretability of pollutant spatial-temporal dynamics.