Scarrica, Vincenzo Mariano (2024) Development of Machine Learning techniques for the Characterisation and Monitoring of the Coastal Strip. [Tesi di dottorato]

Documento PDF VincenzoMarianoScarricaXXXVIIcycle.pdf Visibile a [TBR] Amministratori dell'archivio Download (22MB) | Richiedi una copia

Abstract

Protecting and preserving our planet is more urgent than ever. Climate change and human-driven pollution are posing severe threats to Earth’s ecosystems, jeopardizing the stability of natural systems upon which all life depends. Recognizing this urgency, the United Nations established the 2030 Agenda for Sustainable Development in 2015—a transforma- tive framework composed of 17 Sustainable Development Goals (SDGs) that tackle global challenges across multiple dimensions, including poverty, hunger, climate change, and environmental sustainability. Among them, Goal 14, Life Below Water, focuses on conserving and sustainably man- aging oceans, seas, and marine resources, highlighting the critical role of marine ecosystems in regulating the Earth’s climate and supporting bio- diversity, while acknowledging the escalating dangers posed by pollution, overfishing, and habitat degradation. Coastal environments, situated at the interface between land and sea, are particularly vulnerable, with pollution and erosion inflicting severe impacts on ecosystems, marine species, and human health. Marine litter, particularly plastics, infiltrates food chains, degrades habitats, and releases toxic substances into the ocean, posing a major threat. Addressing these issues is not only a matter of environmental responsibility but also essential for the sustainable development of coastal communities that rely on marine resources. This thesis focuses on leveraging modern Artificial Intelligence (AI) techniques—primarily machine learning, deep learning, and computer vi- sion—to create a framework for effectively and reliably performing various tasks in coastal marine monitoring. AI enables state-of-the-art solutions through real-time, data-driven interventions for these complex challenges. For example, advanced computer vision empowers unmanned aerial vehi- cles (UAVs) and unmanned surface vehicles (USVs) equipped with high- ii resolution cameras to precisely monitor and assess coastal areas, identifying pollution sources, types of debris, and the extent of environmental damage efficiently. However, achieving these goals presents notable challenges. Key is- sues include the need to adapt to different environments, standardize op- erational protocols, and improve the detection of pollutants such as mi- croplastics, which are challenging to identify using visual methods alone. System performance relies on optimized data acquisition strategies, includ- ing UAV flight altitudes and coverage areas, along with advances in image processing and data analysis. A cloud-based, GPU-powered computing infrastructure is essential to process large amounts of collected data and facilitate real-time analysis and intervention. The thesis is organized into four parts: • Part I introduces the context of coastal marine monitoring, empha- sizing its importance and needs (Chapter 1). It also provides an introduction to Artificial Intelligence and Computer Vision (Chap- ter 2), providing existing solutions in the literature, and identifying what is required to advance the field both technologically and oper- ationally. • Part II describes the methodological contribution to developing the architectures of AI techniques used in different applications. These developments focus on computer vision and cover areas such as Graph Learning for Computer Vision, semantic segmentation, in- stance segmentation, classification, and multi-object tracking (Chap- ter 3). • Part III details the primary applications in which the techniques developed have been implemented. These applications cover both Emerged Environments (Chapter 4) and Submerged Environments (Chapter 5). In emerged environments, the main contributions iii are in beach litter detection, where a complete study protocol, from UAV image acquisition to the definition of flight parameters and lit- ter characterisation requirements, has been developed. In addition, a data set specifically tailored for litter detection was created, as well as a deep learning architecture that could serve as a standard for litter detection modules. For underwater environments, the main contribution is a litter detection and multi-tracking module based on a hybrid approach combining deep learning with classical ma- chine learning and computer vision methods, allowing for consistent monitoring of underwater litter. This case study highlights the need to create custom datasets to properly train and operationalise the proposed approach. • Part IV In the concluding considerations (Chapter 6), it is dis- cussed that while the solutions represent a significant advance, the framework is still evolving. Ongoing research needs to focus on im- proving detection accuracy, refining algorithms, and extending capa- bilities to identify a wider range of pollutants. Integrating these tech- nologies into larger environmental management frameworks could pave the way for more effective, scalable solutions for coastal conser- vation and protection. Moreover, the list of publications produced during the PhD program is reported.

Downloads

Actions (login required)

Modifica documento