Low cost Internet of things based sensor networks for air quality in cities

Air pollution is a major public health concern, with over 7 million deaths globally attributed to it annually, as stated by the World Health Organization (WHO) in 2018. Existing real-time Air Quality (AQ) monitoring stations are expensive to install and maintain; therefore, such air quality monitoring networks are sparsely deployed and lack the measurement density to develop high-resolution spatiotemporal air pollutant monitoring. The data generated also lacks accuracy, but still, they have great potential to complement the existing air quality assessment framework. Therefore, this thesis aims to propose a comprehensive architecture for utilizing low-cost sensors in air pollution monitoring. The thesis presents a novel approach to deploy a low-cost sensor network in a city and use a hybrid convolutional-long short-term memory (Conv-LSTM) model for spatiotemporal prediction of air pollution. This approach utilizes both convolutional layers to capture spatial patterns in the sensor data and LSTM layers to capture temporal dependencies. The use of a hybrid model allows for the simultaneous capture of both spatial and temporal patterns in the data, resulting in more accurate predictions compared to models that only utilize one or the other. The research also explores the use of statistical models such as Seasonal Autoregressive Integrated Moving Average (SARIMA) and Nonlinear Autoregressive with exogenous inputs (NARX) models for air quality forecasting, presenting a comparison of the proposed hybrid model with other such state-of-the-art statistical and machine learning models. The results show that the proposed Conv-LSTM model outperforms these approaches in terms of prediction accuracy and robustness and, therefore, is a promising approach for spatiotemporal prediction of air pollution using low-cost sensor data. Additionally, the thesis proposes a general solution to analyze how the noise level of measurements and hyperparameters of a Gaussian process model affect the prediction accuracy and uncertainty of low-cost sensor data. The thesis further presents an extensive evaluation of the proposed hybrid model using real-world data from the low-cost sensor network deployed in Sheffield, and the results demonstrate the effectiveness of the proposed approach. Finally, the real-world studies present the integration of low-cost sensor data into a decision-making system, social and behavioural changes driven by such sensors and the impact of these results on driving policy changes to achieve the World Health Organization’s (WHO) 2021 target for air quality.