De Lucia, Marica (2025) Application of Machine Learning Techniques and Geospatial Data for High-Resolution Intra-Urban Air Quality Monitoring and Assessment. [Tesi di dottorato]

Documento PDF tesi_de_lucia.pdf Download (51MB)

Abstract

Air pollution represents one of the greatest environmental and health threats of our time. Accurate monitoring and prediction of particulate matter concentrations and other pollutants are fundamental for developing effective mitigation strategies. Artificial intelligence offers an innovative approach to improve existing monitoring data and develop advanced predictive models for atmospheric particulate matter (PM), including its chemical speciation, potentially contributing to filling these important knowledge gaps. Methods: This study developed the HReBLOCK (High Resolution Ensemble Block Level Of City Knowledge) framework to address three different tasks. The first two tasks enabled the prediction of PM10 and PM2.5 concentrations in the metropolitan area of Bari at an intra-urban spatial resolution of 300m, using machine learning techniques to optimize satellite AOD data, meteorological and geographical parameters. The third task developed the foundations for PM2.5 speciation (EC, OC, SO4, NO3 , DUST) in the context of the PTA-Rome of the NASA MAIA mission. The models were validated with data from ground monitoring stations and enriched with explainability techniques. The AOD improvement model increased Sentinel-3 Synergy’s correlation with AERONET ground data by 47%. For particulate prediction, the framework achieved robust performance in spatial cross-validation, with determination coefficients (R2) of 0.81 ± 0.01for PM10 and 0.80 ± 0.04 for PM2.5, and RMSE values of approximately 4.90 ± 0.31μg/m3 and 2.90 ± 0.04μg/m3 respectively. Regarding PM2.5 speciation, a comparison was conducted with the GRM algorithm of the NASA MAIA mission, demonstrating that the addition of spatio-temporal variables allowed for achieving comparable results for each chemical component, with R2 values ranging from 0.53 to 0.70. The approach developed through artificial intelligence techniques significantly expands the capabilities of atmospheric pollution monitoring and prediction, extending them also to areas without ground monitoring stations. The high spatial resolution maps generated represent an essential tool for environmental and health authorities, enabling targeted interventions to protect the most vulnerable populations. The ability to estimate not only the overall concentrations of particulate matter but also its chemical composition constitutes a fundamental advancement towards more precise epidemiological studies, capable of assessing the specific hazards of different pollutant components and deepening the understanding of atmospheric pollution effects on public health.

Downloads

Actions (login required)

Modifica documento