Developments and future prospects of land-use regression modelling

Abstract

In recent years, land-use regression (LUR) modelling has become a popular method to assess the spatial variation of outdoor air pollution for large geographical areas. The technique is often used to estimate exposure estimates for individual study subjects in large epidemiological studies. Land-use regression studies combine measurements from an air pollution monitoring network in the study area with predictor variables, obtained through geographic information systems (GIS). Stochastic modelling is used to determine which of these variables best predict the pollution concentrations measured at each location of the network, after which the resulting model can be applied to estimate individual exposure levels for non-network locations.

Current approaches for modelling intra-urban contrasts in pollution commonly use population density, land use, physical geography and various traffic-related variables to predict pollution concentrations. The objective of this thesis is to present an overview of several promising expansions to regular LUR models and assess the requirements and possibilities for their application in exposure assessment modelling for epidemiological studies. We also discuss some current problems in the application of LUR models to predict accurate concentrations, and suggest ways to overcome these problems by integrating different new elements into LUR modelling.

Recent developments have been able to increase explained spatial variability of LUR models further by incorporating predictor variables like winter woodsmoke in residential areas, which make up a significant part of particulate matter pollution in several colder parts of the developed world. Street configuration (e.g. street canyons) has been dealt with by incorporating freely available remote sensing data. Furthermore, several papers present an increased temporal resolution, while “hybrid” models attempt to use emission data in combination with meteorological changes, as has been done earlier in dispersion modelling.

Besides discussing this recent progress, we will also explore the practicalities and possible impacts of including extra predictor variables in future LUR modelling which were not included before. Related to pollution trapping by street canyons, we briefly look at the similar effect of street trees in urban areas. Vertical dispersion in high-rise urban areas has been studied, but was so far never included in LUR modelling. Stagnation points where traffic jams often occur have been known to be bad locations for air pollution, while the possibilities of incorporating this into LUR models have not been widely examined. Finally, a recent focus on inland shipping emissions has led to an increasing demand for LUR studies to include shipping as an additional pollution source.

This review concludes that the most promising new factors to include in LUR modelling are those that significantly improve model accuracy, with minor investments to define and derive the additional predictor variables. The approaches for increasing the temporal resolution on the basis of a continuous sampling point are relatively simple, but provide effective improvements for short-term exposure assessment. More promising additions are those relating to improving the model representations of vertical gradients in concentration. Currently problems exist especially in high-rise urban areas, where a 3 dimensional representation could allow for the inclusion of street configuration as well as a drop in concentration with increasing vertical distance to the source. Other factors like residential woodsmoke and shipping emissions might be relevant to include for areas where this is known to be an issue. Similarly, the integration of emissions and meteorological data could be appropriate only where data on emissions and wind direction are available on a detailed scale, and a keen interest exists for short term exposure estimates.