Pascarella, Antonio Elia (2025) Harnessing artificial intelligence for sustainable development: applications in biofuels and carbon sequestration. [Tesi di dottorato]

Documento PDF Pascarella_Antonio_Elia_Ciclo_37.pdf Download (11MB)

Abstract

This thesis explores the application of Artificial Intelligence (AI) in support of the United Nations Sustainable Development Goals (SDGs), with a specific focus on renewable energy (SDG 7), environmental sustainability through forest carbon monitoring (SDG 15), and improving AI robustness in data-scarce conditions typical of the energy and environmental context. AI technologies are increasingly vital in addressing global challenges, but their practical application presents several technical, ethical, and methodological obstacles. This research investigates these challenges while proposing novel AI-driven solutions across biofuels, forest monitoring, and model robustness in scarce and missing data scenarios. After introducing in the first chapter the use of Artificial Intelligence to tackle sustainable development goals, the second chapter of the thesis examines how AI can support SDG 7 by optimizing biofuel production through biomass pyrolysis. The study creates the PYRIS dataset, which includes 1,137 records of pyrolysis experiments from the literature, and leverages machine learning models such as XGBoost to predict bio-liquid yields. With the use of missing data imputation techniques like Generative adversarial networks, the XGBoost model achieved a mean absolute error (MAE) of 3.45, an R-squared value of 0.66 on the entire dataset, and an MAE of 2.28 and an R-squared value of 0.80 on a reduced dataset without missing values. These results highlight the potential of AI to optimize pyrolysis processes, improve biofuel yields, and support the broader transition to clean energy. The third chapter of the thesis focuses on SDG 15, addressing the challenge of monitoring forest carbon sequestration on a global scale using AI and remote sensing techniques. Two approaches are presented: the first employs a pixel-wise regressive UNet architecture (ReUse) that integrates Sentinel-2 satellite data with ESA's Climate Change Initiative (CCI) Biomass Project data to estimate above-ground biomass (AGB). This method was validated in Italy's Astroni WWF nature reserve, which accurately predicted a pre-fire carbon stock of 18,748 tonnes and a post-fire reduction of 10,104 tonnes. The second approach uses active learning to minimize the need for field samples during model recalibration. This method achieved an RMSE of 28.8 t/ha when using all samples, with a slight increase to 30.0 t/ha and 30.7 t/ha when the number of samples was reduced, demonstrating that AI can reduce the reliance on costly field measurements while maintaining accuracy. The fourth chapter of the thesis addresses the challenge of missing data imputation, a common issue in fields such as biofuels and environmental applications, where data collection is often costly and incomplete. The thesis introduces the Distribution-Aware Masking Autoencoder (DAMA), a novel framework that uses self-supervised learning and distributional constraints to impute missing data. DAMA outperformed traditional imputation methods in "missing completely at random" (MCAR) scenarios and remained competitive in "missing not at random" (MNAR) scenarios. By incorporating distributional priors into the imputation process, DAMA ensures that predictions remain physically plausible, a critical requirement for real-world applications. This thesis demonstrates the potential of AI to advance sustainable development in clean energy, environmental monitoring, and model robustness. By assessing practical applications and proposing AI-driven solutions, this work contributes to achieving sustainable and equitable progress toward the SDGs.

Downloads

Actions (login required)

Modifica documento