Considerable Sample-Imbalance-induced Biases in Ground-Level PM2.5 Estimation from Satellite and Machine Learning in U.S. (Student Poster #26)

Abstract Description: Accurate estimation of ground-level PM2.5 concentrations using satellite-derived data is critical for understanding the impacts of air quality on human health, ecosystems, and climate change. PM2.5, which includes particulate matter smaller than 2.5 μm, is a key pollutant associated with serious health risks, including respiratory and cardiovascular disease and premature mortality. However, current monitoring networks are limited by sparse spatial coverage, particularly in rural and underserved regions, which hinders comprehensive exposure assessments and introduces significant biases in air quality assessments. Recent advances in machine learning and numerical simulation offer promising opportunities to address these gaps by leveraging diverse datasets and innovative modeling approaches. This study develops a novel testbed that integrates numerical simulation and machine learning to improve the accuracy of PM2.5 prediction over different spatial regions. Models are trained on grid cells corresponding to ground-based monitoring sites and evaluated at unmonitored sites, optimizing both model structures and training datasets. By incorporating spatio-temporal features, particularly through CNN-based models such as ResNet, this approach significantly reduces PM2.5 overestimation and improves exposure assessment in downwind areas. In addition, the integration of data sets from networks such as CASTNET and IMPROVE helps to mitigate sample imbalances and biases, improving accuracy in rural and low-pollution areas. The use of high-resolution spatio-temporal data allows detailed analyses of pollutant transport pathways and improves model robustness. The results show significant improvements in PM2.5 retrieval, highlighting the critical role of advanced machine learning techniques in addressing complex atmospheric dynamics. This study not only fills critical gaps in surface monitoring networks but also supports environmental justice by identifying disparities in pollution exposure. It provides a robust framework for informed policy decisions that can improve public health protection and reduce environmental inequities. The methodology has potential for broader applications, providing a scalable approach to addressing air quality challenges globally.

Keywords: Satellite retrievals, PM2.5, testbed, numerical model, machine learning