Spatio-Temporal Prediction of Personal Exposure to \text Using a Network of Stationary and Wearable Air Quality Monitors

• Published in: 2024 IEEE Conference on Pervasive and Intelligent Computing (PICom) pp. 90 - 97
• Main Authors: Arvind, D K, Rao, R, Bates, C A
• Format: Conference Proceeding
• Language: English
• Published: IEEE 05.11.2024
• Subjects: Accuracy; Air quality; Atmospheric modeling; Biomedical monitoring; distributed sensors; Estimation; Global Positioning System; machine learning; Monitoring; Random forests; Spatial resolution; spatio-temporal estimation of airborne particulates; Urban areas; wearable sensors
• DOI: 10.1109/PICom64201.2024.00019

Machine learning methods were exercised to predict personal exposure to airborne particulates of aerodynamic sizes less than \mathbf{2.5}\boldsymbol{\mu} \mathbf{m}(\mathbf{PM}_{\mathbf{2.5}}) . A labelled dataset of \mathbf{PM}_{\mathbf{2.5}} values, tagged with time-stamp and GPS location at 1-minute resolution, was collected by subjects going about their normal lives in Delhi, India and in London, England, using the wearable Airspeck-P(ersonal) sensor to monitor personal exposure to airborne particulates. Contemporaneous datasets of \mathbf{PM}_{\mathbf{2.5}} values were also gathered from a network of municipal stationary Air Quality Monitoring Stations (AQMS) at 1-hour resolution, and a network of solar-powered, fixed Airspeck-S(tationary) monitors at 5-minute resolution in Delhi and up to 30-minute resolution in London. Five machine learning models were trained on the labelled stationary network datasets to predict the personal exposure to PM2.5 at a given time (at minute-level resolution) and a GPS location. The \mathbf{PM}_{\mathbf{2.5}} dataset derived from a network of stationary Airspeck-S monitors were trained to predict personal \mathbf{PM}_{\mathbf{2.5}} exposure of subjects in Delhi, and the \mathbf{PM}_{\mathbf{2.5}} dataset from a network of AQMS in London were trained to predict personal \mathbf{PM}_{\mathbf{2.5}} exposure in London. Extra Trees model followed by Random Forest demonstrated the best accuracy in both cases with a Mean Absolute Percentage Error (MAPE) of less than 20%. These methods can be used to predict personal exposure to airborne particulates in cities endowed with a rich network of public AQMS, and a network of low-cost stationary Airspeck air quality monitors would suffice in places without such an infrastructure.