Air pollution hotspot management using image processing and machine learning techniques.
My research lies in the nexus of computer vision, low-cost sensors and air quality perception, with an overarching theme of localised and efficient urban air quality management. The goal of my research work is to develop an urban air pollution hotspot management plan using machine learning and a perception-based air quality management framework.
Thesis title: Air pollution hotspot management using image processing and machine learning techniques.
Abstract:
Air pollution's sources and characteristics have become more complex in recent years, creating air pollution hotspots across the city. These micro-environments have high temporal and spatial variability, making the identification of such hotspots a challenge. The current monitoring methods are sparsely located and do not provide the high-resolution spatial data required to identify hotspots throughout the city. Additionally, air quality monitoring, being a capital-intensive process, is a challenge in low and middle-income countries. Even though the air quality in metropolitan cities is being monitored to a considerable extent, there is still a lot to be desired. The recent developments in sensor technologies have tried to address this issue, albeit with some trade-offs. With massive potential for growth in new sensing methods, this study attempts to understand the effectiveness of using images for air pollution sensing through machine learning.
Objective 1:
Images captured by cameras, smartphones and CCTV have become ubiquitous and are accessible to people of all demographics. Using images for air quality sensing could make air quality monitoring more accessible and air quality management more effective. This study evaluates the effectiveness of using images for air quality sensing and classification. We developed a dataset of outdoor images, particulate matter, visibility and meteorological data. We then used a convolutional neural network with ResNet architecture to estimate the particulate matter concentration in the ambient air and classify the images. The classification model (accuracy of 0.8) performed better than the regression model (R2 of 0.59). Such image-based models can be used to supplement and augment conventional monitoring systems.
In addition, the study leverages a plethora of existing monitoring methods through data fusion techniques to address the challenge of the lack of fine-grained monitoring data. This objective aims to address the challenges faced by resource-constrained low and middle-income countries in effective air quality management.
Objective 2:
We combine the data from stationary, mobile, satellite monitors and dispersion models, using machine learning-based data-fusion techniques to create a real-time hotspot map of the city. We developed three data fusion models with different combinations of monitoring data, namely, a) mobile + stationary + AERMOD + satellite model, b) mobile + stationary + AERMOD model, and c) Mobile + stationary + satellite model. The performance of the models a) mobile + stationary + AERMOD + satellite model and b) mobile + stationary + AERMOD were similar, with an R2 value of 0.73 and 0.75, respectively. Including satellite data did not improve the model performance much due to frequent missing data caused by cloud interferences and other quality control measures. Mobile + stationary + satellite model performed poorly among the models, and the addition of AERMOD dispersion data improved model performance by 10%. The use of AERMOD dispersion data acts as a supplement to the missing data from other monitoring sources, and the use of mobile monitoring has helped to identify transient hotspots. Fusing data from all the monitoring sources improves data quality and gives a higher resolution of spatial data.
Objective 3:
Once the air quality data is monitored at a fine scale, it must be communicated to the public through the air quality management frameworks. These frameworks include different components like air pollution monitoring, source apportionment, modelling and control to keep air quality levels within the prescribed standards. The public, who are the most important stakeholders, often feel disconnected from these frameworks. The information people receive shapes their perception of air pollution, and the perception, in turn, affects their response to air pollution. However, data dissemination and public participation are often neglected in these frameworks. The study aims to understand the influence of various factors on air pollution perception, segment the population based on perception and develop a participatory air quality management framework with data dissemination as one of the key components. A questionnaire-based online survey was conducted to collect information on perception and five indicators of perception, namely demography, knowledge, precautionary measures, memory and experience of air pollution. After screening, Multivariate Linear Regression and clustering analysis were done on 889 responses. Living in urban and suburban areas, and the use of social media negatively affected air pollution perception. Whereas, frequent exposure to air quality information and past experience of air pollution events improved the perception. People who had good knowledge of air pollution, prevention options and the effectiveness of preventive measures took more preventive actions to protect themselves. K-prototype clustering algorithm was used to segment the population into three groups based on their knowledge, perception, information source and propensity to take preventive actions. The study highlights the need to include data dissemination as a key component in the frameworks, use various communication channels and improve trust in the community by disseminating timely, reliable and actionable information.
Objective 1:
Images captured by cameras, smartphones and CCTV have become ubiquitous and are accessible to people of all demographics. Using images for air quality sensing could make air quality monitoring more accessible and air quality management more effective. This study evaluates the effectiveness of using images for air quality sensing and classification. We developed a dataset of outdoor images, particulate matter, visibility and meteorological data. We then used a convolutional neural network with ResNet architecture to estimate the particulate matter concentration in the ambient air and classify the images. The classification model (accuracy of 0.8) performed better than the regression model (R2 of 0.59). Such image-based models can be used to supplement and augment conventional monitoring systems.
In addition, the study leverages a plethora of existing monitoring methods through data fusion techniques to address the challenge of the lack of fine-grained monitoring data. This objective aims to address the challenges faced by resource-constrained low and middle-income countries in effective air quality management.
Objective 2:
We combine the data from stationary, mobile, satellite monitors and dispersion models, using machine learning-based data-fusion techniques to create a real-time hotspot map of the city. We developed three data fusion models with different combinations of monitoring data, namely, a) mobile + stationary + AERMOD + satellite model, b) mobile + stationary + AERMOD model, and c) Mobile + stationary + satellite model. The performance of the models a) mobile + stationary + AERMOD + satellite model and b) mobile + stationary + AERMOD were similar, with an R2 value of 0.73 and 0.75, respectively. Including satellite data did not improve the model performance much due to frequent missing data caused by cloud interferences and other quality control measures. Mobile + stationary + satellite model performed poorly among the models, and the addition of AERMOD dispersion data improved model performance by 10%. The use of AERMOD dispersion data acts as a supplement to the missing data from other monitoring sources, and the use of mobile monitoring has helped to identify transient hotspots. Fusing data from all the monitoring sources improves data quality and gives a higher resolution of spatial data.
Objective 3:
Once the air quality data is monitored at a fine scale, it must be communicated to the public through the air quality management frameworks. These frameworks include different components like air pollution monitoring, source apportionment, modelling and control to keep air quality levels within the prescribed standards. The public, who are the most important stakeholders, often feel disconnected from these frameworks. The information people receive shapes their perception of air pollution, and the perception, in turn, affects their response to air pollution. However, data dissemination and public participation are often neglected in these frameworks. The study aims to understand the influence of various factors on air pollution perception, segment the population based on perception and develop a participatory air quality management framework with data dissemination as one of the key components. A questionnaire-based online survey was conducted to collect information on perception and five indicators of perception, namely demography, knowledge, precautionary measures, memory and experience of air pollution. After screening, Multivariate Linear Regression and clustering analysis were done on 889 responses. Living in urban and suburban areas, and the use of social media negatively affected air pollution perception. Whereas, frequent exposure to air quality information and past experience of air pollution events improved the perception. People who had good knowledge of air pollution, prevention options and the effectiveness of preventive measures took more preventive actions to protect themselves. K-prototype clustering algorithm was used to segment the population into three groups based on their knowledge, perception, information source and propensity to take preventive actions. The study highlights the need to include data dissemination as a key component in the frameworks, use various communication channels and improve trust in the community by disseminating timely, reliable and actionable information.