Abstract
Accurate and reliable data stream plays an important role in air quality assessment. Air pollution data collected from monitoring networks, however, could be biased due to instrumental error or other interventions, which covers up the real pollution status and misleads policy making. In this study, motivated by the needs for objective bias detection, we propose a hidden dynamic geostatistical calibration (HDGC) model to automatically identify monitoring stations with constantly biased readings. The HDGC model is a two-level hierarchical model whose parameters are estimated through an efficient Expectation-Maximization algorithm. Effectiveness of the proposed model is demonstrated by a simple numerical study. Our method is applied to hourly $PM_{2.5}$ data from 36 stations in Hebei province, China, over the period from March 2014 to February 2017. Significantly abnormal readings are detected from stations in two cities.
Introduction
Accurate air quality data is the backbone of environmental policy, but monitoring networks are prone to "noise." In China, where $PM_{2.5}$ levels are a major public health concern, data from over 1,500 stations can be compromised by hardware failures, poor maintenance, or even intentional interventions—such as using water sprayers near sensors to artificially lower readings. Currently, identifying these "biased" stations relies heavily on manual visual inspection, which is slow, subjective, and inefficient for large-scale networks. This research proposes an objective, automated statistical framework that can flag suspicious sensors by comparing their readings to neighboring stations while accounting for complex weather patterns and local geography.
Methods
The core innovation is the Hidden Dynamic Geostatistical Calibration (HDGC) model, a two-level hierarchical state-space approach:
- The Observation Level: This relates the measured $PM_{2.5}$ to a set of "fixed effects" (such as weather and other pollutants) and a "true" latent pollution state ($z$). It introduces a station-specific calibration parameter, $\alpha_s$. If a station is accurate, its $\alpha_s$ should be similar to its neighbors; if it is significantly lower, the station is likely under-reporting.
- The State Level: This models the "true" pollution as a dynamic random field, accounting for temporal dependence (pollution at a given hour depends on the previous hour) and spatial correlation (nearby stations should observe similar "true" air quality).
To handle the "hidden" nature of the true pollution state and manage missing data, the authors utilize a combination of Kalman filtering/smoothing and the Expectation-Maximization (EM) algorithm to estimate the model's parameters.
Data and Study Area
The model was validated using hourly $PM_{2.5}$ readings from 36 monitoring stations across 8 cities in Hebei Province, China, spanning March 2014 to February 2017. The spatial distribution of these stations is shown in Figure 1, with specific site names detailed in Figure 2.
To ensure the model wasn't just flagging weather anomalies, the researchers included covariates such as $SO_2, NO_2, O_3, CO$, barometric pressure, temperature, dew point, rainfall, wind speed/direction, and the distance of each site to the Taihang Mountains.
Results and Insights
The study confirmed that $NO_2$ (primarily from vehicle exhaust) is the strongest predictor of $PM_{2.5}$ levels in the region. Additionally, high humidity (dew point) promotes secondary particle formation, and proximity to the Taihang Mountains leads to higher pollution accumulation due to blocked air dispersion.
The HDGC model successfully identified two cities—Cangzhou and Langfang—where stations consistently showed lower calibration ($\alpha_s$) values compared to the rest of the province. To validate these findings, the authors examined the ratio of $PM_{2.5}$ to $PM_{10}$. In one flagged Cangzhou site, the ratio was only 38%, while a neighboring city’s site showed 70%. Since these particles usually move together, such a massive discrepancy strongly suggests the $PM_{2.5}$ sensor was biased. The model's low root mean square error (RMSE) of 8.69 $\mu g/m^3$ further validates its ability to capture underlying pollution physics.
Discussion and Conclusion
While effective, the model has limitations. Stations with more than 20% missing data in a single season were excluded, which meant some potentially biased periods (like Langfang in late 2016) could not be fully analyzed. Furthermore, the model assumes linear relationships, whereas atmospheric chemistry can be non-linear. Future work will focus on scaling this approach to the national network of over 1,500 stations using more computationally efficient estimation techniques.
Summary updated