|Ph.D Student||Chen Shimon|
|Subject||Development of a Physically-Sound and Portable Air Quality|
Regression Model with a High Spatio-Temporal
|Department||Department of Civil and Environmental Engineering||Supervisor||Professor David Broday|
|Full Thesis text|
A novel air pollution modelling approach has recently been introduced, called Optimized Dispersion Modelling (ODM). In ODM, a dispersion model is formulated and its parameters are regressed to produce the best fit between modelled and observed ambient concentrations of primary air pollutants. Only a single ODM formulation has been published prior to this study: the Traffic-Only ODM (TO-ODM). It was applied to nitrogen dioxide and displayed similar validation fit measures to interpolation and Land Use Regression, while producing results at high temporal resolution and manifesting important dispersion characteristics. However, the TO-ODM was only applied to specific times of the day and over a limited geographical region. The main reasons for these limitations were: (1) lack of high-quality and high-coverage traffic data; (2) a dispersion scheme which assumed a homogeneous wind field; (3) accounting only for emissions from traffic; and (4) high computational complexity. All these limitations were tackled in my research.
Most air quality models use some traffic-related variables as an input. Previously, vehicular activity was estimated through sporadic counts, traffic assignment models, or by summing nearby road lengths. These methods produced poor or no estimates for nights, weekends and holidays. Emerging technologies, such as vehicle-mounted GPS transceivers, allow for more accurate and extensive traffic estimation. In this work, I studied traffic volumes that were derived from such a technology, using them as a proxy for vehicular emissions within the ODM framework.
In order to incorporate industrial emissions a new term was added to the original TO-ODM. It expresses the non-monotonous change in ground-level concentrations as the distance from an elevated emission source increases. Using the new Traffic-Industry ODM (TI-ODM), the relative impact of each sector on ambient nitrogen oxides concentrations was assessed over the coastal plain of Israel. The TI-ODM out-performed both Inverse-Distance-Weighing interpolation and the TO-ODM in predicting ambient nitrogen oxides concentrations.
Several existing state-of-the-art models are based on Gaussian-like dispersion. This formulation has several advantages which are useful for the generalization of the ODM framework. Hence, for the first time, the Gaussian dispersion model was re-formulated as a nonlinear regression model and applied for estimating ambient nitrogen oxides concentrations at a high spatiotemporal resolution. Model results over the same area had a better spatial correlation with the ambient measurements than those of the TI-ODM. Overall, the two models performed similarly. This successful implementation paved the way for a new type of model: a Dynamic ODM (DynODM). This Gaussian puff-like model is the first air quality regression model with memory terms that relate subsequent time-points through the spatiotemporal patterns of the puff. The DynODM was able to display important characteristics resulted by heterogeneous wind fields, and had a better cross-validated mean spatial correlation with the ambient measurements than all the other tested models. However, it produced few extreme over-estimations which caused the sum of squared errors to be extremely high.
The advancements made in this study enable the application of regression modelling to a plethora of new settings and scales, enabling unprecedented flexibility in air pollution exposure assessment.