Joint features random forest (JFRF) model for mapping hourly surface PM2.5 over China

Ambient PM2.5 exerts strong regional pattern for its ability long-range transport, implying that including the features of surrounding stations may improve accuracy machine-learning based model to estimate surface from satellite-retrieved aerosol optical depth (AOD). However, most current models either just use single point features, or simply average observed adjacent on a fixed spatial proportional relationship. The question how properly take advantage retrieving is still not well addressed. Here we propose an integrated algorithm called joint random forest (JFRF) which includes complex feature differences with and observation learn dynamic relations target pixel, rather than weighted (WAF) only by as traditional (with WAF) used. Results cross validation suggest better performance JFRF (R2 = 0.61–0.8; RMSE 15.97–20.91 μg/m3) (ΔR2 0.09–0.3). also exhibits 0.05–0.11), particularly in regions large AOD gradient (accounts 33% total test set), great significance accurately representing heterogeneity (e.g., pollution edging hot spots areas). And exclusion significantly reduced −0.07 ∼ −0.1). Therefore, our study demonstrates important further helping estimating PM2.5.