Supplementary material
Yüklə 262,86 Kb.
|
- Bu səhifədəki naviqasiya:
- Figure S3 and S4.
- Figure S5 and S6.
- Figure S10
- 2.5 Comparison of surface and column model predictions
- Monitoring sites ( n = 68) Mesh blocks ( n
- Predictor 5th 25th 50th
2.3 Model checking Figures S3, S4, S5, and S6 show model residuals and Cook’s distance for the annual column model, annual surface model, monthly column model and monthly surface model, respectively. 
Figure S3 and S4. Model residuals and Cook’s distance for annual column (top) and surface (bottom) models. 
Figure S5 and S6. Model residuals and Cook’s distance for monthly column (top) and surface (bottom) models. We selected the NO2 monitoring station with the highest Cook’s distance from each model and compared it against three randomly chosen comparisons. Tables S7, S8, S9 and S10 show the results for the annual column, annual surface, monthly column and monthly surface models, respectively. We used this information to examine influential monitoring sites and check that their influence had a plausible basis. All of the largest Cook’s distances were well below the commonly used threshold of 1. Table S7. Dependent and standardised independent variables for the monitoring station with highest Cook’s distances in the annual column model (Randwick) and three randomly chosen comparisons. 
Table S8. Dependent and standardised independent variables for the monitoring station with highest Cook’s distances in the annual surface model (Dandenong) and three randomly chosen comparisons. 
Table S9. Dependent and standardised independent variables for the monitoring station with highest Cook’s distances in the monthly column model (Woolloongabba) and three randomly chosen comparisons. 
Table S10. Dependent and standardised independent variables for the monitoring station with highest Cook’s distances in the monthly surface model (Woolloongabba) and three randomly chosen comparisons. 
Figures S7, S8, S9, and S10 show boxplots of df-beta statistics for the annual column model, annual surface model, monthly column model and monthly surface model, respectively. These highlight influential monitoring stations for each variable in the models. These were used in combination with the Cook’s distance data presented above to identify and investigate influential observations. These were then checked for correctness (i.e. no errors in input data) and plausibility (i.e. a real-world basis for their influence [such as proximity to a high emitting industrial NOX source]). Generally, the most influential variable was the industrial site density. These variables were positively skewed with many zero values and occasional high values, which increases the chances of them being influential. Industrial sites are a plausible source of NO2 and none of the changes in the parameter slopes in the df-beta plots were extreme or counter-intuitive.  Figure S7. Boxplot of df-beta statistics for each variable in the annual column model.
Figure S8. Boxplot of df-beta statistics for each variable in the annual surface model. 
Figure S9. Boxplot of df-beta statistics for each variable in the monthly column model. 
Figure S10. Boxplot of df-beta statistics for each variable in the monthly surface model. 2.4 Selected percentiles of predictors For the models to be applicable to areas beyond just those where the NO2 monitors were located, it is important that the predictors span a sufficient range and one that is representative of the areas that they will be applied to (Novotny et al., 2011). We compared the summary statistics for the predictors at the 68 monitoring sites used to build the models (which included a variety of different land use settings) to the summary statistics of predictors in the annual (surface and column) models at the ~350,000 census mesh block centroids spread across the entirety of Australia. Table S11 summarises the results. The statistics were very similar and suggests no evidence that monitoring sites differed markedly from the broader Australian context. We did not compare the statistics for the monthly models due to the large number of results (144). However, the main predictors in the monthly models are similar to those in the annual models (e.g. major and minor roads, impervious surfaces, industrial land use, industrial site density), and all of which are closely matched to the distributions observed around the country. This gives us no reason to suspect issues around representativeness for either of the monthly models. 2.5 Comparison of surface and column model predictions Table S12 presents the summary statistics for NO2 concentrations predicted at the ~350,000 mesh block centroids by the annual surface and column models. The table highlights that the predictions were almost identical. Population-weighted and unweighted mean concentrations are also shown. Table S13 presents the summary statistics for concentrations predicted by the two models at the ~57,000 mesh block centroids that make up the greater Sydney area, which is Australia’s most populous city (4.4 million).
Yüklə 262,86 Kb. Dostları ilə paylaş:
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||