Using land-use machine learning models to estimate daily NO2 concentration variations in Taiwan

It is likely that exposure surrogates from monitoring stations with various limitations are not sufficient for epidemiological studies covering large areas. Moreover, the spatiotemporal resolution of air pollution modelling approaches must be improved in order to achieve more accurate estimates. If not, assessments will applicable future health risk assessments. To deal this challenge, study featured Land-Use Regression (LUR) models use machine learning assess spatial-temporal variability Nitrogen Dioxide (NO 2 ). Daily average NO data was collected 70 fixed quality stations, belonging Taiwanese EPA, on main island Taiwan. Around 0.41 million observations 2000 2016 were used analysis. Several datasets employed determine spatial predictor variables , including EPA environmental resources dataset, meteorological land-use inventory, landmark digital road network map, terrain model MODIS Normalized Difference Vegetation Index database, and power plant distribution dataset. Regarding analyses, conventional LUR Hybrid Kriging-LUR performed first identify important variables. A Deep Neural Network Random Forest, XGBoost algorithms then fit prediction based selected by models. Lastly, splitting, 10-fold cross validation, external verification, seasonal-based county-based validation methods applied verify robustness developed The results demonstrated proposed captured 65% 78%, respectively, variation. When algorithm further incorporated hybrid-LUR, explanatory increased 84% 91%, respectively. outperformed all other integrated methods. This demonstrates value combining an estimate exposure. For practical application, associations specific land-use/land cover types final can management planning emission reduction strategies. • Estimating long-term daily concentration Land-use patterns included using regression. most contributed predictors identified stepwise variable selection. Explanatory 0.65 0.91. XGboost RF DNN algorithms. Capsule: coupled variations reached 91%