Many low-cost sensors (LCSs) are distributed for air monitoring without any rigorous calibrations. This work applies machine learning with PM2.5 from Taiwan monitoring stations to conduct in-field corrections on a network of 39 PM2.5 LCSs from July 2017 to December 2018. Three candidate models were evaluated: Multiple linear regression (MLR), support vector regression (SVR), and random forest regression (RFR). The model-corrected PM2.5 levels were compared with those of GRIMM-calibrated PM2.5. RFR was superior to MLR and SVR in its correction accuracy and computing efficiency. Compared to SVR, the root mean square errors (RMSEs) of RFR were 35% and 85% lower for the training and validation sets, respectively, and the computational speed was 35 times faster. An RFR with 300 decision trees was chosen as the optimal setting considering both the correction performance and the modeling time. An RFR with a nighttime pattern was established as the optimal correction model, and the RMSEs were 5.9 ± 2.0 μg/m3, reduced from 18.4 ± 6.5 μg/m3 before correction. This is the first work to correct LCSs at locations without monitoring stations, validated using laboratory-calibrated data. Similar models could be established in other countries to greatly enhance the usefulness of their PM2.5 sensor networks.
Keywords: PM sensing device; efficient in-field PM2.5 correction; in-field calibration; particle sensing correction; random forest model.