An Ensemble Machine-Learning Model to Predict Historical PM2.5 Concentrations in China from Satellite Data

摘要

Background: The long satellite aerosol data record enables assessments of historical PM2.5 level in developing countries such as China where routine PM2.5 monitoring began only recently. However, most previous models reported decreased prediction accuracy when predicting PM2.5 levels outside the model-training period. This limitation greatly hinders the application of satellite-driven exposure assessments in the research on health effects of long-term PM2.5 exposure. Objectives: We proposed an ensemble machine learning approach that provided reliable PM2.5 hindcast capabilities in China. Methods: Non-random missing satellite data due to cloud cover were first filled by multiple imputation to ensure unbiased long-term exposure estimates. Then the modeling domain, China, was divided into seven regions using a spatial clustering method to control for unobserved spatial heterogeneity. A set of machine learning models including random forest, generalized additive model, and extreme gradient boosting were trained in each region separately. Finally, a generalized additive ensemble model was developed to combine predictions from different algorithms. Results: The ensemble prediction characterized the spatiotemporal distribution of daily PM2.5 well with the cross-validation (CV) R2 (RMSE) of 0.79 (21 μg/m3). The cluster-based sub-region models outperformed national models and improved the CV R2 by ～0.05. Compared with previous studies, our model provided more accurate hindcasts at the daily level (R2 = 0.53, RMSE = 28 μg/m3) and monthly level (R2 = 0.81, RMSE = 13 μg/m3). Conclusions: Our hindcast modeling system allows for the construction of long-term, unbiased historical PM2.5 levels that can support epidemiologic studies on the chronic health effects of PM2.5 in China.