Spatially Adaptive Calibrations of AirBox PM$_{2.5}$ Data

TL;DR

This paper introduces a spatially adaptive calibration model for low-cost AirBox PM$_{2.5}$ sensors in Taiwan, improving measurement accuracy by accounting for local environmental factors and enabling reliable air quality estimates across the region.

Contribution

The study develops a novel spatial model with spatially varying coefficients for calibrating IoT sensor data, enhancing accuracy and automatic calibration at new locations.

Findings

Calibration improves prediction accuracy by 37% to 67%.

The model provides reliable PM$_{2.5}$ estimates at any location.

Automatic calibration for new sensors is achieved.

Abstract

Two networks are available to monitor PM$_{2.5}$ in Taiwan, including the Taiwan Air Quality Monitoring Network (TAQMN) and the AirBox network. The TAQMN, managed by Taiwan's Environmental Protection Administration (EPA), provides high-quality PM$_{2.5}$ measurements at $77$ monitoring stations. More recently, the AirBox network was launched, consisting of low-cost, small internet-of-things (IoT) microsensors (i.e., AirBoxes) at thousands of locations. While the AirBox network provides broad spatial coverage, its measurements are not reliable and require calibrations. However, applying a universal calibration procedure to all AirBoxes does not work well because the calibration curves vary with several factors, including the chemical compositions of PM$_{2.5}$, which are not homogeneous in space. Therefore, different calibrations are needed at different locations with different local environments. Unfortunately, most AirBoxes are not close to EPA stations, making the calibration task challenging. In this article, we propose a spatial model with spatially varying coefficients to account for heteroscedasticity in the data. Our method gives adaptive calibrations of AirBoxes according to their local conditions and provides accurate PM$_{2.5}$ concentrations at any location in Taiwan, incorporating two types of measurements. In addition, the proposed method automatically calibrates measurements from a new AirBox once it is added to the network. We illustrate our approach using hourly PM$_{2.5}$ data in the year 2020. After the calibration, the results show that the PM$_{2.5}$ prediction improves about 37% to 67% in root mean-squared prediction error for matching EPA data. In particular, once the calibration curves are established, we can obtain reliable PM$_{2.5}$ values at any location in Taiwan, even if we ignore EPA data.