A neural network-based land use regression model to estimate spatial-temporal variability of nitrogen dioxide

Nitrogen dioxide (NO2) is a kind of highly reactive gas and secondary pollutant mainly from burning fossil fuels, which were predominant species in vehicle exhaust. Since traffic volume density is heavy and large number of temples and restaurants were densely distributed in Taiwan. The high concentration of NO2 may cause adverse effects on respiratory system. To estimate NO2 concentration more accurately, this study aimed to utilize a neural network-based land use regression model to assess the spatial-temporal variability. Daily average NO2 data were collected from 70 fixed air quality monitoring stations in Taiwan main island which were established by Taiwan Environment Protective Administration. Totally, around 0.41 million observations were collected for our analysis. Several datasets were collected for obtaining spatial predictor variables, including EPA environmental resources dataset, meteorological dataset, land-use inventory, landmark dataset, digital road network map, DTM, MODIS NDVI dataset, and thermal power plant distribution dataset. To establish the integrated approach, conventional land-use regression (LUR) was first used to identify the important predictors variables. Then a deep neural network (DNN) algorithm was applied to fit the prediction model. 10-fold cross validation and external data verification methods were used to further confirm the robustness of model performance. The results showed that, the developed conventional LUR model captured 60% of NO2 variation. Of the 11 variables selected by the stepwise variable selection procedure, PM10, SO2, O3 explained 18%, 7% and 5% NO2 variation, respectively. After integrating DNN algorithm with conventional LUR method, the model explanatory power was increased to 85%, with a 25% improved in model performance. Consistent findings were obtained from the 10-fold cross validation, while the cross-validated R² was increased from 61% to 83%, and root-mean-square error (RMSE) was decreased from 6.56 ppb to 4.34 ppb. This study demonstrates the value of incorporating the conventional LUR model and DNN algorithm in estimating spatial-temporal variability of NO2 exposure.