CEM MiddleEast - Abstract

Abstract Information :

Lower-cost air quality sensors (LCS) have significantly transformed the landscape of air pollution monitoring. In India, where the number of regulatory monitoring stations is insufficient, LCS-based monitoring of air pollution has gained traction. Among the various criteria pollutants, fine particulate matter (PM2.5) is the dominant air pollutant in India. Personal Air Quality Systems Ltd. (PAQS) has established an extensive LCS network across several cities in India. These sensors can measure PM2.5, temperature, and relative humidity (RH) with high temporal resolution (every 15 minutes). This presentation focuses on the PM2.5 prediction modeling results derived from data collected by a dense LCS network in Ahmedabad, Gujarat. The network consists of approximately 50 LCS deployed to measure ambient PM2.5 across an urban area of about 300 square kilometers. Alongside the LCS, five regulatory monitoring stations provide high-quality PM2.5 data, resulting in a monitoring density of one station per 5.5 square kilometers on average. Data from these monitors spanning three years (2019-2022) will be used for modeling. The data collection and modeling schema include data from ground (LCS and regulatory data) and satellite measurements, IoT architecture, and an AI engine. The features of the AI engine include spatio-temporal Gaussian processes, reinforced and supervised learning, behavior feedback, and more. The predictor variables include GRASP-retrieved satellite-derived aerosol optical depth (AOD) at a 1 km resolution, meteorological data, geographic features, and other relevant variables. These predictor datasets were interpolated to a 100 m spatial resolution. Hourly AOD values will be reconstructed by combining satellite snapshot AOD data with reanalysis AOD values. The hourly 100 m-resolution predictors, along with point measurements from the LCS, will be used to train the prediction models. A range of machine learning models will be evaluated, with the best-performing model selected to predict high-resolution spatio-temporal PM2.5 concentrations. Model selection and validation will be performed using out-of-sample cross-validation, with a 30%-70% split between training and test datasets. High-resolution PM2.5 maps will be generated, and relevant spatial patterns will be identified. As a next step, we aim to utilize these high-resolution predictions for forecasting future PM2.5 levels. Finally, the AI engine will be capable of delivering precise, personalized solutions, including making accurate binary and complex decisions to mitigate individual exposure levels. The AI engine will be able to generate high-resolution PM2.5 data for regions similar to those used in training, even in the absence of air quality monitoring capabilities. For instance, the AI engine trained using ground data collected over Bahrain will be able to predict PM2.5 for regions in Qatar.