PhD

I hold a Master’s in Computer Science and proudly serve as a tenured computer science teacher at a Rome Secondary School. In November 2022, I embarked on an exciting journey as a full-time PhD candidate in Data Science at the prestigious Sapienza University of Rome.

PhD proposal.

My research proposal, submitted with enthusiasm and dedication, emphasized the groundbreaking use of Graph Neural Networks (GNNs) for weed recognition, introducing a revolutionary perspective in the field of computational botany. Central to the proposal were the unique challenges posed by recognition in complex environments such as agricultural fields, characterized by the presence of diverse plant species and varying soil substrates. The importance of enhancing the network’s ability to capture long-range dependencies in these images, considering the potential for multiple species and diverse environmental contexts, emerged as a key element of the research. The realm of plants, with its intrinsic complexity and often accompanying epistemic uncertainty, was portrayed as a fascinating and particularly challenging domain. My proposal underscored the urgency of encoding a robust knowledge prior into a neural network, a fundamental prerequisite for successfully addressing the unique challenges of this environment. At the core of my research lay the perspective of pushing the boundaries in understanding and applying GNNs, exploring the revolutionary potential of this technology in the context of botany and weed classification. The challenge of effectively integrating the intricacies of the plant world into a predictive model represented a captivating research opportunity, promising to contribute significantly to our ability to manage and preserve agricultural ecosystems sustainably.

Published paper: “ArcheoWeedNet: Weed Classification in the Parco archeologico del Colosseo”.

This paper summarizes the development of a weed monitoring system in the Parco archeologico del Colosseo (hereinafter, Parco) using Deep Learning (DL) techniques to recognize forty-one species of plants now present in the area. The project is part of SyPEAH (System for the Protection and Education of Archaeological Heritage), a platform designed to safeguard the Parco by its Authority. This study emanates from an extended phase of the photographic collection spanning ten months. This endeavour facilitated the compilation of a dataset comprising nearly 5,000 photographs depicting the flora of pertinent significance. In the paper, we detail the first version of the system, consisting of a neural network trained to predict the species of plants and the materials on which they grow. We also describe transfer learning techniques aimed at improving performance. The present system attains recognition accuracy exceeding 90% for common species, enabling near real-time monitoring of the entire Park’s flora through image analysis using supplied fixed and mobile devices. It will support proactive interventions for maintenance. The paper details data analysis and neural network design and envisions future developments. URL Cite

Published paper: “Convolutional Neural Networks for the Detection of Esca Disease Complex in Asymptomatic Grapevine Leaves”.

The Esca complex is a grapevine trunk disease that significantly threatens modern viticulture. The lack of effective control strategies and the intricacy of Esca disease manifestation render essential the identification of affected plants before symptoms become evident to the naked eye. This study applies Convolutional Neural Networks (CNNs) to distinguish, at the pixel level, between healthy, asymptomatic and symptomatic grapevine leaves of a Tempranillo red-berried cultivar using Hyperspectral imaging (HSI) in the 900–1700 nm spectral range. We show that a 1D CNN performs semantic image segmentation (SiS) with higher accuracy than PLS-DA, one of HSI data’s most widely used classification algorithms. URL Cite

Next projects.

In the upcoming months, my research endeavours will reach new heights as we embark on an exhilarating exploration of diverse neural network architectures. While Convolutional Neural Networks (CNNs) have been instrumental thus far, the horizon beckons us to delve into the realms of Transformers and Graph Neural Networks (GNNs). An exciting prospect involves revisiting my initial research proposal, focusing on implementing a Visual Graph Neural Network architecture. Simultaneously, my passion for advancing our understanding of weed and pest dynamics is leading to a parallel investigation of GNNs in various contexts, given their remarkable applicability and generalization prowess. The allure of novel research avenues extends to the exploration of generative architectures, not only within the domain of plant images but also venturing into entirely different realms, such as the creative landscapes of music. For the past few months, and ongoing into the present, my primary focuses include the delving into the intricate realm of Parkinson’s disease recognition, the extension of the scope of two seminal papers within species and disease recognition, the exploration of the transformative potential of generative AI in the realm of plant sciences and the robustness of datasets that underpin our research efforts. This evolving portfolio underscores my dedication to forefront innovations and the dynamic momentum propelling my investigation into unexplored domains within the vast realm of artificial intelligence and plant science. As the ongoing orchestration of exploration persists, the amalgamation of these heterogeneous research trajectories holds the potential to enhance our scientific comprehension significantly and substantively contribute to the overarching domains of medicine, environmental science, and creative technology.