Portelli, Beatrice (2024) From machine learning to deep learning: expanding AI applications in environmental sciences. [Tesi di dottorato]

	Documento PDF portelli_beatrice_37_COMPLETO.pdf Visibile a [TBR] Amministratori dell'archivio Download (8MB) | Richiedi una copia
	Documento PDF portelli_beatrice_37_PARZIALE.pdf Visibile a [TBR] Amministratori dell'archivio Download (7MB) | Richiedi una copia

Abstract

In recent years, Artificial Intelligence (AI) has emerged as a transformative tool across various sectors. However, when looking at environmental fields like agriculture, forestry, or soil science, AI techniques are still underutilized due to challenges in data availability, heterogeneity, and model interpretability. This is particularly true for Deep Learning (DL), which is a data-hungry branch of AI leveraging largely "black box" models. This thesis explores how AI and DL models can advance environmental research, emphasizing applications where DL is less common and investigating the unique limitations that impact model effectiveness. The objective is to identify key factors that limit DL adoption, address the obstacles of diverse environmental datasets, and explore the potential of novel AI techniques to improve model performance and interpretability. We define three research questions to guide this work: (1) How can heterogeneous environmental data be effectively integrated to enhance AI predictive capabilities? (2) Can Machine Learning (ML) and DL models yield interpretable insights for environmental applications? (3) What role do novel DL technologies, such as Large Language Models (LLMs), play in advancing environmental research? The thesis comprises four main research chapters. The first chapter focuses on forestry applications, specifically estimating Carbon Stock (CS) and Carbon Sequestration (CSE) using heterogenous data, such as satellite hyper-spectral images, UAV-acquired LiDAR data, geomorphological information, and climatic parameters. By applying ML and DL models, as well as and leveraging explainability techniques, we uncover which features are the most influential for CS and CSE estimation. The second chapter focuses on soil science, integrating data from proximal sensors (electromagnetic induction sensor and gamma-ray detectors) to predict soil composition and Soil Carbon Stock. This chapter highlights again the advantages of data fusion, and also demonstrates that DL models can offer accurate and interpretable insights into soil properties. The third chapter explores wastewater management, employing a Transformer-based LLM (T5) for pollutant detection. Through a "textification" approach, this DL model achieves a higher accuracy compared to other ML and DL models trained on the raw sensor data, indicating that DL can be leveraged in non-standard ways to solve existing problems. Finally, the fourth chapter discusses the use of LLMs in pharmacovigilance, a developed field for human-centered applications which is rapidly spreading to for plant and animal health. We focus on existing pharmacovigilance datasets and perform an in-depth analysis of LLM architectures to derive useful insights for future application in the environmental field. This work contributes to the field of AI applications for environmental sciences by developing methodologies for data integration, interpretability, and proposing novel DL applications. The findings underline the potential of DL technologies to address environmental challenges, despite data limitations and the need for interpretability.

Downloads

Actions (login required)

Modifica documento