Casolaro, Angelo (2025) Deep Learning for Spatio-Temporal Prediction and Super-Resolution of Primary Air Pollutants. [Tesi di dottorato]

Documento PDF Angelo_Casolaro_38_ciclo_COMPLETO.pdf Visibile a [TBR] Amministratori dell'archivio Download (10MB)

Available Versions of this Item

• L’assessment di personalità in soggetti con Disturbo dell’Identità di Genere attraverso la Swap-200. Uno studio sugli stili di personalità dei pazienti D.I.G. (deposited 23 Nov 2009 13:48)
  • Deep Learning for Spatio-Temporal Prediction and Super-Resolution of Primary Air Pollutants. (deposited 29 Dic 2025 15:57) [Attualmente visualizzato]

Abstract

Air pollution is one of the most pressing environmental and public health issues, driven by industrial processes, transportation, and energy production, and particularly severe in densely populated urban areas. Its spatial and temporal variability, shaped by climatic, geographic, and demographic factors, makes accurate monitoring and prediction highly challenging. Recent advances in deep learning have shown strong potential to address these complexities, consistently outperforming classical machine learning methods in capturing nonlinear relationships and temporal dependencies in air quality data. This PhD thesis contributes to the advancement of the state-of-the-art in air pollution prediction through three main contributions. First, it provides two comprehensive surveys: the former dedicated on temporal prediction and the latter on spatio-temporal approaches, offering a comprehensive assessment of deep learning model capabilities and limitations across different environmental contexts. Second, it addresses a key gap in existing approaches, namely, the inability to quantify predictive uncertainty, by developing probabilistic deep learning frameworks based on Bayesian Neural Networks and Diffusion Models. These models integrate predictive accuracy with uncertainty quantification, thereby enhancing their practical utility for governmental and environmental agencies that depend on both reliable forecasts and confidence intervals for decision-making. Third, it introduces an innovative diffusion-based framework, the Improved Graph-based Implicit Diffusion Model, for environmental super-resolution. This model enables continuous-scale reconstructions of pollutant concentration fields while providing uncertainty estimates, overcoming the limitations of coarse-resolution reanalysis products and sparse ground-based sensors. By incorporating graph attention mechanisms and spatio-temporal loss functions, it achieves fine-grained, uncertainty-aware reconstructions critical for accurate environmental diagnostics and policy support. These contributions advance deep learning for air quality prediction by bridging the gap between high predictive performance and reliable uncertainty quantification, while also extending deep learning applications to super-resolution of environmental datasets. The findings provide both methodological innovations and practical tools, contributing to more informed public health strategies and environmental policies.

Downloads

Actions (login required)

Modifica documento