Development of high-resolution particle concentration prediction model - an application of remote sensing and machine learning

According to WHO, air pollution is associated with 7 million premature deaths annually. Lack of air pollution monitoring systems is one of the major challenges in developing regulations and policies for air pollution control as well as health risk assessments in majority of the developing nations. Ground based monitoring stations are sparsely distributed in major cities only. The major reason for lack of monitoring stations is the heavy cost associated with their establishment and operations, whereas low-cost sensors come with different variety of challenges related to the calibration, accuracy and reliability. This problem could be solved by combining remote sensing data with machine learning to estimate particulate matter concentrations. Modern satellite sensors hold the potential to provide high-quality aerosol optical depth (AOD) data at a resolution of 10m × 10m. Current approaches for estimating particle (PM_2.5) concentrations rely on AOD with meteorological data such as relative humidity, ambient temperature, wind speed, etc. However, the prediction accuracy decreases dramatically if these results are extrapolated over a wider timeline and a broader region. This is owing to the fact that traditional methods take many assumptions based on the training data set. We implemented and validated the conventional method of predicting particle concentration to evaluate prediction accuracy using five distinct machine learning models. The light gradient boosting regression model generated the highest prediction accuracy, i.e., 92.46% for a training data set of one year (January 2019 – December 2019) for the city of Mumbai, with a resolution of 0.5° × 0.625° (latitude × longitude). This work presents a hybrid approach combining the physics-based relations and statistical methods, to predict surface level concentration. It uses the vertical distribution of aerosols along with the optical properties like single scattering albedo and angstrom exponent for determining particle characteristics and meteorological parameters for a greater prediction accuracy over a wider timeline and a broader region. Since AOD provides a measure of total particles above a location, we employed data from multi-angular satellite sensor (CALIOPS) to generate vertical distribution profiles and ultimately surface-level concentration. Also, the physics based empirical relations are considered while determining the input parameters for model training, which significantly increases the prediction accuracy of model. When the particle size distribution curve was combined with the surface level concentration from vertical distribution profile, a more accurate surface level PM_2.5 concentration was obtained. Unlike previous approaches that make several assumptions based on the location of training data, this method, by removing those assumptions, is valid over a broader area and a wider timescale.