An integrated model combining random forests and WRF/CMAQ model for high accuracy spatiotemporal PM2.5 predictions in the Kansai region of Japan. (1st October 2021)

• Record Type: Journal Article
• Authors: Thongthammachart, Tin; Araki, Shin; Shimadera, Hikari; Eto, Shinnosuke; Matsuo, Tomohito; Kondo, Akira
• Is Part Of: Atmospheric environment. Volume 262(2021)
• Publication Date: 2021-10-01
• Subjects: Air pollution -- Chemical transport model -- Random forests -- Land use regression; RF random forests -- CTM chemical transport model -- CMAQ community multiscale air quality -- CV cross-validation -- RMSE root mean square error -- with-CMAQ LURF with-CMAQ land use random forest -- without-CMAQ LURF without-CMAQ land use random forest -- LUR land use regression -- MLR multiple linear regression -- GEOS-Chem Goddard Earth Observing System–Chem model -- ML machine learning -- LURF land use random forest -- LULR land use linear regression -- WRF weather research and forecasting model -- PBL planetary boundary layer -- VIF variance inflation factors -- ME mean errors; Air -- Pollution -- Periodicals; Air -- Pollution -- Meteorological aspects -- Periodicals; 551.51
• DOI: 10.1016/j.atmosenv.2021.118620
• Languages: English
• ISSNs: 1352-2310

Abstract

Abstract: Accurate spatial and temporal prediction of PM2.5 ambient concentration is crucial to appropriate exposure assessment. We develop a spatiotemporal land use regression model by integrating a random forests (RF) technique and the Community Multiscale Air Quality (CMAQ) modeling system to accurately estimate daily PM2.5 levels in the Kansai region of Japan, which is affected by long-range transport in the Asian continent and by local pollution. The most important advantage of RF is that it captures nonlinearity among the target air pollutants and the predictor variables including land-use variables, meteorological variables, and CMAQ-estimated PM2.5 concentration. We compare the predicting performances of the land use random forests (LURF) models with and without CMAQ variables to determine their effectiveness. A cross-validation (CV) technique that calculates the coefficient of determination (R 2 ) and root mean square error (RMSE) is performed to evaluate their prediction performances through spatial and temporal CVs. The performance of the with-CMAQ LURF model was superior to that of the without-CMAQ LURF model. Moreover, we evaluated the PM2.5 prediction performances of the with-CMAQ LURF and the with-CMAQ land use linear regression (LULR) models via CV to determine the efficiency of the non-linear model. Accordingly, the with-CMAQ LURF model is preferable for PM2.5 estimation compared to that of the with-CMAQ LULR model. In addition, the with-CMAQ LURF model exhibits higher PM2.5 predictability than the CMAQ model, as indicated by the higher model-R 2 and lower model-RMSE values. Our findings demonstrate that the CMAQ-simulated PM2.5 level integrated into the LURF is advantageous in accurately estimating PM2.5 concentration, which is influenced by long-range transport and local pollution. Graphical abstract: Image 1 Highlights: Land use random forest (LURF) model is developed from a random forests algorithm. Daily average PM2.5 concentrations are estimated by combining LURF and CMAQ models. Performance of the model with-CMAQ LURF is more accurate than that without-CMAQ. With-CMAQ LURF model accurately predicts the spatiotemporal variability of PM2.5 . With-CMAQ LURF model estimates PM2.5 from long-range transport and local sources.