Imputation for spatiotemporal PM2.5 data via Varying-Coefficient Autoregressive Adversarial Network

Abstract

Fine particulate matter (PM2.5) poses risks to environmental health, and missing data due to equipment failures and technical issues hinders pollution analysis. To address this issue, this study proposes Varying-Coefficient Autoregressive Adversarial Network (VCAAN) framework to impute these missing values effectively. First, a Varying-Coefficient Autoregressive (VCA), based on vector autoregression and B-spline approximation of time-varying coefficients, is proposed to capture dynamic spatiotemporal dependencies while reducing model complexity. Next, a Convolutional Discriminative Network (CDN) is designed for spatiotemporal imputation. This network leverages convolutional operations to learn spatiotemporal patterns and assess the quality of the imputed values. In addition, a dynamic adversarial loss weighting mechanism is introduced, enabling VCA and CDN to engage in dynamic adversarial interaction and ultimately converge to a balanced solution. Finally, extensive experiments on Beijing PM2.5 data confirm the proposed method’s superiority, demonstrating its strong adaptability to various missing scenarios and effectiveness even under a high missing rate.