Adrian Arfire
With urban populations rapidly increasing and millions of deaths being attributed annually to air pollution, there is a critical need for a deeper understanding of urban air quality. The locality of urban emissions sources, and the specific topography of cities lead to a very heterogeneous pollutant concentration landscape, the details of which cannot be captured by traditional monitoring stations. Although highly accurate, these systems are large, heavy and very expensive, which leads to a very sparse distribution of measurements. Mobile sensor networks hold the potential to allow a paradigm shift in our understanding of urban air pollution, through an augmentation of the spatial resolution of measurements. The road to achieving reliable high quality information from these type of systems is, however, full of challenges. These start from the system design, as the task of developing robust mobile sensing networks for continuous urban monitoring is arduous in itself. The limitations of existing sensor technology is another important source of hard problems. Chemical sensors suffer from many issues that make their use in a mobile scenario non-trivial. These include: instability, cross-sensitivity, low signal-to-noise ratios, and slow dynamic response. The latter problem, in particular, is a tough challenge when considering a mobile scenario, as it leads to significant measurement distortion. The question of maintaining the calibration of chemical sensors is another essential issue that derives from their instability. Finally, the development of appropriate modeling techniques that would enable us to generate high-resolution pollution maps based on mobile sensor network data is a highly difficult problem due to the inherently dynamic and partial coverage of such systems. The aim of this thesis is to show the feasibility of mobile sensor networks for monitoring air quality and their ability to achieve the goal of pushing our understanding of urban air pollution. We have taken a holistic approach, by studying the end-to-end system, and addressing each of the aforementioned challenges. Using public transportation vehicles for mobility, we have developed and deployed a full-scale mobile sensor network for monitoring the air quality in the city of Lausanne, Switzerland. We have carefully considered all steps of the system design process, starting from the choice of targeted pollutants, sensor selection, node design, server architecture, and system operation. For addressing the problem of mobility-caused distortion, we created a rigorous wind tunnel experimental set-up to study this effect and the techniques for mitigating it. We propose using deconvolution for recovering the underlying pollutant signal. Since the performance of this approach is limited by the signal-to-noise ratio of the measurements, we propose using an active sniffer to enhance the quality of the raw signal. On the topic of sensor calibration, we propose two improvements to online rendezvous calibration methodology. The first one is a model-based approach, which considers the use of more sophisticated sensor models, which are more faithful to the complex behavior of chemical sensors. The second one proposes the use of a pre-processing step, in which the mobile data is deconvolved. Finally, we study the problem of generating high-resolution maps based on mobile data. We propose five statistical modeling methods that use a heterogeneous list of explanatory variables.
EPFL2016
