Abiotic stresses such as drought and salinity pose serious challenges to plant health, leading to impaired morphological development and crop productivity. These stresses affect processes like photosynthesis and water balance, often resulting in visible symptoms such as reduced growth and leaf chlorosis. While traditional laboratory methods for detecting plant stress are accurate, they are typically time-consuming, destructive, and limited in throughput. In contrast, image-based phenotyping offers a faster, non-invasive approach to monitor plant health under changing environmental conditions. This thesis, titled "Morphophysiological analysis of plant responses to different environmental conditions and definition of machine learning algorithms for predictive and automated analytics activities", presents an integrated framework combining laboratory-based biochemical analyses with advanced image processing and machine learning techniques. The study focuses on plants subjected to drought and salt stress, including model and crop species like Arabidopsis thaliana, different landraces of common bean (Fagiolo d’acqua and Fagiolo della levatrice), and two dwarf tomato cultivars (Tiny Tim and Micro-Tom). Visible light (RGB) and chlorophyll fluorescence (ChlFl) images —collected using both custom and automated platforms—were used to extract digital traits linked to stress responses, including color- and geometry-based features. In parallel, physiological and biochemical markers such as fresh and dry weight, electrolyte leakage percentage, as well as pigment, proline, malondialdehyde, and relative water contents, were measured as ground truth indicators of stress. Areas, NPQ (non-photochemical quenching) and several color indices (i.e. Chroma indices) were among the most informative digital traits measured. By analyzing the full pixel distribution on false-colored images, subtle changes in plant appearance and physiology were detected, often preceding the onset of visible symptoms. These parameters showed high effectiveness not only in identifying the presence of stress but also in classifying its intensity, and Random Forest models achieved 76% precision in distinguishing between drought and salt stress. Notably, as the complexity of the classification task increased, image-derived traits proved more informative than laboratory-derived attributes. By bridging plant biology, imaging technologies, and artificial intelligence, this thesis advances the understanding of plant stress response and contributes to the development of a more sustainable and “smart” approach to crop growth and productivity. The findings validate image-based phenotyping as a robust, scalable tool for non-invasive monitoring of plant stress, and the demonstrated potential for early stress detection opens the door to more timely crop management and targeted breeding strategies.
Gli stress abiotici, come la siccità e la salinità, rappresentano un’importante minaccia per la salute delle piante, compromettendone il loro sviluppo e la loro produttività. Questi stress influenzano processi cruciali come la fotosintesi e l’equilibrio idrico, causando spesso sintomi visibili, come crescita ridotta e clorosi fogliare. Nonostante i metodi di laboratorio tradizionalmente impiegati per il rilevamento dello stress siano accurati, sono generalmente dispendiosi in termini di tempo, distruttivi e poco adatti ad analisi su larga scala. Di contro, la fenotipizzazione basata su immagini rappresenta un approccio più rapido e non invasivo per monitorare lo stato di salute delle piante. Questa tesi, dal titolo "Analisi morfofisiologica delle risposte delle piante a diverse condizioni ambientali e definizione di algoritmi di apprendimento automatico per attività predittive e analitiche automatizzate", propone un approccio integrato che combina analisi biochimiche, d’immagine e machine learning per monitorare e prevedere la risposta delle piante agli stress abiotici, in particolare siccità e salinità. Lo studio si concentra su specie modello e specie di interesse agronomico, tra cui Arabidopsis thaliana, due varietà locali di fagiolo (Fagiolo d’Acqua e Fagiolo della Levatrice) e due cultivar nane di pomodoro (Tiny Tim e Micro-Tom). Le immagini sono state acquisite tramite sistemi di imaging nello spettro del visibile (RGB) e con sensori adatti a rilevare la fluorescenza della clorofilla (ChlFl), utilizzando piattaforme sia manuali che automatizzate. Sono stati estratti quindi tratti digitali, come parametri colorimetrici e geometrici, correlabili alla risposta agli stress. Parallelamente, sono stati misurati indicatori fisiologici e biochimici comunemente utilizzati come marcatori di stress: peso fresco e secco, electrolyte leakage, concentrazione dei pigmenti, contenuto di prolina e malondialdeide, e contenuto idrico delle foglie. Tra i tratti digitali più informativi sono emersi i parametri dimensionali (Area), i valori di quenching non-fotochimico (NPQ) e gli indici cromatici (i.e. Chroma indices). L’analisi della distribuzione dei pixel su immagini a falsi colori ha permesso di rilevare precocemente variazioni morfologiche e fisiologiche impercettibili all’occhio umano. Questi parametri si sono rivelati altamente efficaci sia nel rilevamento della presenza di stress, sia nella classificazione della sua intensità. I modelli, basati su algoritmi come Random Forest, hanno raggiunto una precisione del 76% nella distinzione tra siccità e salinità. Inoltre, con l’aumento della complessità dei modelli, i tratti derivati dalle immagini hanno mostrato un potere discriminante superiore rispetto ai parametri biochimici tradizionali. Questo studio, combinando dati di biologia/fisiologia vegetale, tecnologie di imaging e intelligenza artificiale, contribuisce ad una comprensione più profonda delle risposte delle piante agli stress abiotici e supporta un’agricoltura più sostenibile e “smart”. I risultati ottenuti confermano la validità del phenotyping basato sull’analisi d’immagine come strumento solido e scalabile per il monitoraggio non invasivo delle piante, e ne evidenziano il potenziale per il rilevamento precoce dello stress (early detection), aprendo la strada a una gestione delle colture più tempestiva e a strategie di miglioramento genetico più mirate ed efficienti.
Morphophysiological analysis of plant responses to different environmental conditions and definition of machine learning algorithms for predictive and automated analytics activities
DEL CIOPPO, Giorgia
2025-06-09
Abstract
Abiotic stresses such as drought and salinity pose serious challenges to plant health, leading to impaired morphological development and crop productivity. These stresses affect processes like photosynthesis and water balance, often resulting in visible symptoms such as reduced growth and leaf chlorosis. While traditional laboratory methods for detecting plant stress are accurate, they are typically time-consuming, destructive, and limited in throughput. In contrast, image-based phenotyping offers a faster, non-invasive approach to monitor plant health under changing environmental conditions. This thesis, titled "Morphophysiological analysis of plant responses to different environmental conditions and definition of machine learning algorithms for predictive and automated analytics activities", presents an integrated framework combining laboratory-based biochemical analyses with advanced image processing and machine learning techniques. The study focuses on plants subjected to drought and salt stress, including model and crop species like Arabidopsis thaliana, different landraces of common bean (Fagiolo d’acqua and Fagiolo della levatrice), and two dwarf tomato cultivars (Tiny Tim and Micro-Tom). Visible light (RGB) and chlorophyll fluorescence (ChlFl) images —collected using both custom and automated platforms—were used to extract digital traits linked to stress responses, including color- and geometry-based features. In parallel, physiological and biochemical markers such as fresh and dry weight, electrolyte leakage percentage, as well as pigment, proline, malondialdehyde, and relative water contents, were measured as ground truth indicators of stress. Areas, NPQ (non-photochemical quenching) and several color indices (i.e. Chroma indices) were among the most informative digital traits measured. By analyzing the full pixel distribution on false-colored images, subtle changes in plant appearance and physiology were detected, often preceding the onset of visible symptoms. These parameters showed high effectiveness not only in identifying the presence of stress but also in classifying its intensity, and Random Forest models achieved 76% precision in distinguishing between drought and salt stress. Notably, as the complexity of the classification task increased, image-derived traits proved more informative than laboratory-derived attributes. By bridging plant biology, imaging technologies, and artificial intelligence, this thesis advances the understanding of plant stress response and contributes to the development of a more sustainable and “smart” approach to crop growth and productivity. The findings validate image-based phenotyping as a robust, scalable tool for non-invasive monitoring of plant stress, and the demonstrated potential for early stress detection opens the door to more timely crop management and targeted breeding strategies.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_G_Del-Cioppo.pdf
embargo fino al 09/12/2026
Descrizione: Tesi di Dottorato
Dimensione
5.46 MB
Formato
Adobe PDF
|
5.46 MB | Adobe PDF | Visualizza/Apri Richiedi una copia |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


