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

• DOI: 10.1021/acs.est.8b02917
• Publication Date: 2018-10-24T23:04:26Z
• Authors: Qingyang Xiao; Howard H. Chang; Guannan Geng; Yang Liu
• Keywords: Hindcast; Gradient boosting; Ensemble forecasting; Ensemble Learning; Data set

The long satellite aerosol data record enables assessments of historical PM2.5 level in regions where routine monitoring began only recently. However, most previous models reported decreased prediction accuracy when predicting levels outside the model-training period. In this study, we proposed an ensemble machine learning approach that provided reliable hindcast capabilities. missing were first filled by multiple imputation. Then modeling domain, China, was divided into seven using a spatial clustering method to control for unobserved heterogeneity. A set including random forest, generalized additive model, and extreme gradient boosting trained each region separately. Finally, model developed combine predictions from different algorithms. characterized spatiotemporal distribution daily well with cross-validation (CV) R2 (RMSE) 0.79 (21 μg/m3). cluster-based subregion outperformed national improved CV ∼0.05. Compared studies, our more accurate out-of-range at ( = 0.58, RMSE 29 μg/m3) monthly 0.76, 16 Our system allows construction unbiased levels.