A machine learning calibration model using random forests to improve sensor performance for lower-cost air quality monitoring

摘要

Low-cost sensing strategies hold the promise of denser air quality monitoringnetworks, which could significantly improve our understanding of personal airpollution exposure. Additionally, low-cost air quality sensors could bedeployed to areas where limited monitoring exists. However, low-cost sensorsare frequently sensitive to environmental conditions and pollutantcross-sensitivities, which have historically been poorly addressed bylaboratory calibrations, limiting their utility for monitoring. In thisstudy, we investigated different calibration models for the Real-timeAffordable Multi-Pollutant (RAMP) sensor package, which measures CO,NO, O, and CO. We explored three methods: (1) laboratoryunivariate linear regression, (2) empirical multiple linear regression, and(3) machine-learning-based calibration models using random forests (RF).Calibration models were developed for 16–19 RAMP monitors (varied bypollutant) using training and testing windows spanning August 2016 throughFebruary 2017 in Pittsburgh, PA, US. The random forest models matched (CO) orsignificantly outperformed (NO, CO, O) the othercalibration models, and their accuracy and precision were robust over timefor testing windows of up to 16 weeks. Following calibration, average meanabsolute error on the testing data set from the random forest models was38 ppb for CO (14 % relative error), 10 ppm for CO (2 %relative error), 3.5 ppb for NO (29 % relative error), and3.4 ppb for O (15 % relative error), and Pearson  versus thereference monitors exceeded 0.8 for most units. Model performance is exploredin detail, including a quantification of model variable importance, accuracyacross different concentration ranges, and performance in a range ofmonitoring contexts including the National Ambient Air Quality Standards(NAAQS) and the US EPA Air Sensors Guidebook recommendations of minimum dataquality for personal exposure measurement. A key strength of the RF approachis that it accounts for pollutant cross-sensitivities. This highlights theimportance of developing multipollutant sensor packages (as opposed tosingle-pollutant monitors); we determined this is especially critical forNO and CO. The evaluation reveals that only the RF-calibratedsensors meet the US EPA Air Sensors Guidebook recommendations of minimum dataquality for personal exposure measurement. We also demonstrate that theRF-model-calibrated sensors could detect differences in NOconcentrations between a near-road site and a suburban site less than 1.5 kmaway. From this study, we conclude that combining RF models with carefullycontrolled state-of-the-art multipollutant sensor packages as in the RAMPmonitors appears to be a very promising approach to address the poorperformance that has plagued low-cost air quality sensors.