STCM: A spatio-temporal calibration model for low-cost air monitoring sensors

Abstract

Air pollution is one of the most common health-threatening factors, potentially causing millions of deaths every year. Static stations typically collect air pollution data by utilizing a small number of costly and high-quality sensors and numerous low-cost micro stations. However, data collected from micro stations usually suffers from large noise and thus calibration would be necessary for operating good air pollution governance. Point-to-point models and sequence-to-point models have the potential limitations of either failing to mine latent patterns embedded in historical time series or ignoring spatial dependency within a certain region. To address these issues, we propose a novel method called Spatio-Temporal Calibration Model (STCM) based on dual encoders. STCM consists of long-term encoder, short-term encoder, and decoder modules. The long-term encoder encodes historical reference data via GRU and extracts the trend, periodicity, and adjacency of the target pollutant through a temporal attention mechanism. The short-term encoder then reflects real-time conditions through a spatial attention mechanism, quantifying dynamic station-wise correlations between micro stations and the static station. The decoder ultimately integrates outputs of dual encoders and generates calibration results of all micro stations. STCM has been experimentally justified by comparing against nine baseline methods based on two real-world datasets.