Advanced Air Quality Management with Machine Learning

Air pollution has been a significant health risk factor at a regional and global scale. Although the present method can provide assessment indices like exposure risks or air pollutant concentrations for air quality management, the modeling estimations still remain non-negligible bias which could deviate from reality and limit the effectiveness of emission control strategies to reduce air pollution and derive health benefits. The current development in air quality management is still impeded by two major obstacles: (1) biased air quality concentrations from air quality models and (2) inaccurate exposure risk estimations

Inspired by more available and overwhelming data, machine learning techniques provide promising opportunities to solve the above-mentioned obstacles and bridge the gap between model results and reality. This dissertation illustrates three machine learning applications to strengthen air quality management: (1) identifying heterogeneous exposure risk to air pollutants among diverse urbanization levels, (2) correcting modeled air pollutant concentrations and quantifying the bias of sources from model inputs, and (3) examine nonlinear air pollutant responses to local emissions. This dissertation uses Taiwan as a case study, due to its well-established hospital data, emission inventory, and air quality monitoring network.

In conclusion, although ML models have become common in atmospheric and environmental health science in recent years, the modeling processes and output interpretation should rely on interdisciplinary professions and judgment. Except for meeting the basic modeling performance, future ML applications in atmospheric and environmental health science should provide interpretability and explainability in terms of human-environment interactions and interpretable physical/chemical mechanisms. Such applications are expected to feedback to traditional methods and deepen our understanding of environmental science.