A neural network-based land use regression model to estimate spatial-temporal variability of SO₂

One of the air pollutants from industrial waste and transportation combustion is Sulfur Dioxide (SO₂). Previous studies have shown that SO₂ has a serious impact on human health, in particularly on respiratory problems. By taking Taiwan as the study area, this study aimed to assess the spatial-temporal variability of SO₂ using a neural network-based land use regression model. Daily SO₂ observations during 2000 to 2018 were obtained from 73 monitoring stations established by Taiwan Environmental Protection Agency (EPA). Totally, around 0.48 million observations were collected for our analysis. Several databases were used to collect the spatial predictor variables, including EPA environmental resources database, meteorological database, land-use inventory, landmark database, digital road network map, DTM, MODIS NDVI dataset, and thermal power plant distribution database. To establish the integrated approach, conventional land-use regression (LUR) was first used to identify the important predictors variables. After that, a deep neural network (DNN) algorithm was applied to fit the prediction model. The results showed that, the adj-R² obtained from the conventional LUR approach was 0.37. Of the 15 variables selected by the stepwise variable selection procedure, PM10, nearest thermal power plants, and NO₂ are important variables that increased the SO₂ exposures with the explanatory ability up to 18%, 6%n and 4%, respectively. Compared to the conventional LUR approach, by combining DNN algorithm can improve the model explanatory ability up to 21% (adj-R²=0.59). The results of 10-fold cross validation and external data verification confirmed that the value of the adj-R² after combining both approaches increased from 0.37 to 0.59, and RMSE decreased from 2.48 ppm to 2.01 Findings of this study confirm that the combination of LUR, and DNN algorithm can improve the prediction performance level and the explanatory abilities in assessing spatial-temporal variability of SO₂ exposure.