Identification of long-range aerosol transport patterns to Toronto via classification of back trajectories by cluster analysis and neural network techniques

Abstract

In this work, back trajectories of air masses arriving in Toronto were classified into distinct transport patterns by cluster analysis and, for the first time, by a neural network (Adaptive Resonance Theory—ART-2a). Different similarity criteria were used by the two classification techniques, the former relying on the Euclidean distances between trajectories, the latter on the Euclidean angles between trajectories. Nevertheless, both techniques provided similar conclusions as to the location of PM_2.5 emission sources and the level of pollution associated with a given air transport pattern. Both techniques illustrated the cleaner nature of northerly and northwesterly transport patterns in comparison to southerly and southwesterly ones, as well as the effect of near stagnant air masses. In addition, ART-2a resolved a much larger percentage of trajectories than cluster analysis into groups with clearly identifiable transport patterns and compared favourably with cluster analysis with respect to the precision of the classification.

Introduction

Identifying the sources of airborne pollutants is of great importance to the study of fine particulate matter (PM_2.5), which has been linked to adverse health effects [1]. The examination of transport patterns of air masses through the use of back trajectories is commonly performed for source identification. Unlike wind direction, back trajectories provide visualization of not just the local air direction, but transport over a continental scale. While back trajectories are on average accurate to within 20% of the distance traveled, individual back trajectories may be completely incorrect [2]. Thus, a large dataset is required to provide meaningful source identification.

Two approaches have recently emerged for the visualization of air quality data. The first consists of the generation of a probability map of the areas around a receptor site that contribute to its poor air quality days, as characterized by high PM_2.5 and/or trace gas levels (the so-called Potential Source Contribution Function approach). This is the focus of our work in an upcoming publication [3]. The second approach was the focus of the present work and was based on grouping back trajectories with similar distances traveled or similar overall direction. It has been concluded that grouping back trajectories with similar distances traveled or similar overall direction is the best approach for the visualization of air quality data [4]. Cluster analysis, which uses physical distances between trajectories, has typically been used to group similar trajectories during the last two decades via algorithms like average-linkage clustering, Ward’s method and k-means clustering [4]. These algorithms generate different classification results and their interpretation is often subjective. Consequently, there is no one best grouping algorithm [2], [4], [5], [6], [7], [8].

Another grouping technique that has so far been overlooked for the classification of air mass back trajectories is neural network analysis. Neural networks have long been known as useful for analyses in synoptic climatology [9]. Since then, neural networks have found some use in general circulation models as they provide the tightest possible mapping of the complex, non-linear relationships between the atmosphere and the surface environment [10]. However, a study of transport patterns of air masses, a less complex problem, has not emerged.

The neural network approach differs from cluster analysis in that a desired degree of separation must be specified rather than a desired number of clusters. Secondly, the dot product used in neural net classification incorporates an angular component, relatable in this application to wind direction, rather than the geometric distance used in cluster analysis. Finally neural nets are designed to learn: when a novel trajectory is encountered, a neural net will create a new class with this trajectory as its sole member. This feature can be incorporated into some cluster analysis algorithms but it is not fundamental to this method. Hence sufficient differences existed to suggest that the methods might produce different insight.

Air masses arriving in Toronto have diverse histories ranging from clean, fast-moving Arctic air to polluted, nearly stagnant Ohio Valley air. In fact, Southern Ontario PM_2.5 concentrations have been reported to be 2 to 4 times higher under southerly or southwesterly flow conditions than under northerly flow conditions [11]. Thus, the ideal trajectory taxonomy would not group trajectories that pass through both northerly and southerly regions before arriving in Toronto with purely northerly or southerly trajectories. Grouping of air trajectories has been used to study the origins of ozone pollution in Toronto [12]. However, this approach has not previously been applied to particulate matter in the region.

For this reason, the main goal of this work was a comparison of the air pollution information provided by cluster analysis and an artificial neural network, Adaptive Resonance Theory (ART-2a) [13], [14], [15], for back trajectories ending in Toronto during a thirteen month sampling duration. The adaptation of cluster analysis and ART-2a to interpret atmospheric pollutant concentration is described. The inter-cluster variation of atmospheric species concentration was also explored, with special attention devoted to those trajectory groups that displayed abnormally large or small concentrations of atmospheric particulate matter. Some reasons for the dissimilarity of the classifications are also suggested.