Air Quality Prediction via Embedded ML/DL and Quantized Models

In the people’s everyday life, one of the most significant (and unfortunately well-known) environmental problems with substantial impact and effects is <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">air pollution</i>. To this end, extensive research (e.g., conducted by the World Health Organization, WHO) has uncovered a non-negligible correlation between hazardous environments and airborne Particulate Matters (PMs), emphasizing the importance of monitoring pollution to mitigate its adverse effects. In fact, since PMs are microscopic particles of solid or liquid materials suspended in the atmosphere and have negative effects on both weather and precipitations, it has been highlighted that the diffusion of diseases through the inhalation or absorption of contaminants is a concerning outcome of polluted environments. Hence, focusing on the critical need to monitor air quality in common (livable and work) environments (e.g., airports, public transports, or any closed environment) using affordable, compact IoT devices capable of collecting and processing environmental data, as well as forecasting future air quality trends, in this paper we present an experimental evaluation of several Machine Learning (ML), Deep Learning (DL), and quantized DL models designed to accurately predict air pollution (in terms of PM<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</sub> concentration). Furthermore, we conduct a comparative analysis of state-of-the-art DL models, unveiling the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">trade-offs</i> associated with each ML/DL model, and discussing their practical applications and potential deployment on tiny IoT devices for on-site computing. Our results show that Convolutional Neural Network-Bidirectional Gated Recurrent Unit (CNN-BiGRU) is the best performing model. According to the obtained results, using 8-bit quantization on CNN-BiGRU allows to achieve a 66% model size’s reduction with only a 1%drop in average prediction accuracy, making it a balanced and efficient choice for several applications.