Exploration of intra-city and inter-city PM2.5 regional calibration models to improve low-cost sensor performance

Low-cost PM_2.5 sensors often suffer from environmental cross-sensitivities, requiring regular calibration across a wide range of concentrations. This is typically achieved by co-locating LCS with regulatory stations and using statistical models. However, this approach becomes challenging in regions with limited regulatory monitoring stations or access. To address this challenge, we explored building separate calibration models for the pseudo-regional component of the total PM_2.5 concentration, which represents background concentration, and the hyper-local component of the total concentration. This is based on the premise that the regional concentration is consistent across a given region and therefore direct co-location is less necessary, and the idea that the local concentration is not influenced by geographic properties and therefore can be calibrated based on co-location elsewhere. In this work, we used publicly-available PurpleAir data for 2022 from five different cities in South Asia and North America, and built city-specific calibration models for the regional concentrations using multiple linear regression. We tested the model performance in the city the model was built in (intra-city models; trained and cross-validated in the same city) and in other cities (inter-city models; trained and cross-validated in different cities). The regional calibration model reduced the normalized root mean square error (nRMSE) of both intra-city models, from 51% to 26%, and inter-city models, from 55% to 25% compared to PurpleAir reported concentrations. Overall, the results of this work demonstrate the potential for improved transferability of calibration models and provides evidence that calibration models built for regional concentration and local concentration separately may be a viable solution when deploying in places with limited regulatory monitoring or access to monitoring stations.