A Study of Calibration Model for Light Scattering Method in Monitoring Outdoor PM_2.5 Concentrations

Objective To compare the consistency between light scattering and gravimetric method in the determination of outdoor fine particulate matter (PM_2.5), and to develop calibration models with influencing factors.

Methods Light scattering and gravimetric methods were used to monitor the outdoor PM_2.5 concentrations simultaneously. The consistency of the results from the two methods was analyzed by calculating the relative deviation, correlation coefficient and other indicators. The ratio of the concentrations determined by the gravimetric method and the light scattering method (correction coefficient) was used as the dependent variable, and multiple linear regression (MLR) as well as random forest (RF) were used to develop calibration models for light scattering method. The performance of the models was evaluated by 10-fold cross validation method.

Results A total of 138 sets of valid data were obtained. There was a significant positive correlation between the results of light scattering method and gravimetric method (Spearman r = 0.932, P < 0.001), but there was a systematic deviation that the values from scatting light method were significantly higher than that of the gravimetric method (P < 0.001). After calibration by MLR or random RF model with environmental humidity, temperature and light scattering response value as prediction variables, the accuracy of light scattering method was effectively improved, and RF model performed better than MLR model. After calibration using RF model, the correlation coefficient between the results of the two methods increased from 0.932 to 0.957, and the root mean square error (RMSE) decreased from 13.5 μg/m³ to 7.3 μg/m³, while the mean absolute deviation (MAE) decreased from 10.8 μg/m³ to 5.4 μg/m³.

Conclusion There was a good correlation between the results of light scattering method and gravimetric method, but systematic deviation existed. The results from light scattering method can be effectively calibrated using MLR or RF model, and the later model performed better.