With the aim of meeting a constantly growing food demand while reducing its environmental impact, a more efficient and sustainable agriculture is becoming increasingly necessary globally. For this reason, precision agriculture presents itself as a crucial response to the challenges of the agricultural sector today and even more so in the coming decades. Precision agriculture is based on steps common to various applications and involves an initial collection of data using various types of sensors mounted on the most diverse platforms such as drones. Subsequently, this data is analyzed by specific software or programs. Based on the information obtained, farmers can make more precise decisions and therefore carry out more targeted actions; hence precision agriculture. Artificial intelligence, AI, with its ability to analyze large amounts of data and make autonomous decisions, offers enormous potential to transform the agricultural sector in this sense. This work examines how AI can be used to analyze grapevines, Vitis vinifera, and in particular their woody components. In the spring of 2024, in seven temporal repetitions spaced about a week apart, images were collected of several sections of vine rows, each containing two to six grapevines. The images were collected in four different vineyards, all with white grapes, two of the Solaris B. variety aged three and six years and two of Pinot Bianco B., also aged three and six years, present on the hill of Cart, in the municipality of Feltre, province of Belluno, Italy. The collected images were processed using a program called Roboflow which allows you to outline the contour of the desired objects. Three classes of objects were then identified, namely the woody components of the vine: rootstock, trunk, and shoots. The images thus modified were then analyzed with YOLOv8, an artificial intelligence algorithm for instant segmentation. This deep learning algorithm developed by Ultralytics LLC uses a convolutional neural network (CNN) that is able to identify and locate objects within images and videos in real time and classify them into groups. The trained model, therefore, produced various results with these images. In the test part, in fact, the effectiveness of the algorithm useful for field monitoring of vines is demonstrated. From the analysis of 108 images, an F1 score, or the harmonic mean of precision and recall, of 0.64 with a confidence of 0.446 is obtained. This data provides a balanced evaluation of the performance of a model considering both false positives and false negatives. Furthermore, in the Mask(P, R, mAP50, mAP50-95) metric, the model test provided the following results: a model precision "P" of 0.711 out of 1, a recall "R" or the model's ability to identify all object instances in the images of 0.589 out of 1 and a mean precision calculated with an intersection over union (IOU) threshold of 0.50 "mAP50" of 0.608 out of 1. Furthermore, the processing speed conducted by the YOLOv8 model was 36.76 FPS.
Con lo scopo di soddisfare una domanda alimentare in costante crescita, riducendo al contempo il proprio impatto ambientale, un'agricoltura più efficiente e sostenibile si fa sempre più necessaria a livello globale. Per questo motivo l’agricoltura di precisione si presenta come risposta cruciale alle sfide del settore agricolo oggi e ancor di più nei prossimi decenni. L’agricoltura di precisione si basa su dei passaggi comuni alle varie applicazioni e cioè prevede un’iniziale raccolta di dati tramite sensori di vario genere montati sulle piattaforme più diverse come droni. Successivamente tali dati vengono analizzati da specifici software o programmi. In base alle informazioni ottenute, gli agricoltori possono prendere decisioni più precise e svolgere quindi azioni più mirate; da qui agricoltura di precisione. L'intelligenza artificiale, AI, con la sua capacità di analizzare grandi quantità di dati e di prendere decisioni autonome, offre un potenziale enorme per trasformare il settore agricolo in tale senso. Questo lavoro esamina come l'AI possa essere utilizzata per analizzare le piante di vite, Vitis vinifera, e in particolare le loro componenti legnose. Nella primavera 2024, in sette ripetizioni temporali distanti circa una settimana l’una dall’altra, sono state raccolte le immagini di diversi spezzoni di filari di vite contenenti ognuno dalle due alle sei piante di vite. Le immagini sono state raccolte in quattro vigneti diversi, tutti a bacca bianca, due a varietà Solaris B. di tre e sei anni e due a Pinot Bianco B., sempre di tre e sei anni, presenti sul colle di Cart, nel comune di Feltre, provincia di Belluno, Italia. Le immagini raccolte sono state elaborate utilizzando un programma chiamato Roboflow che permette di delineare il contorno degli oggetti desiderati. Sono quindi state identificate tre classi di oggetti, ovvero le componenti legnose della vite: portainnesto, tronco e tralci. Le immagini così modificate sono poi state analizzate con YOLOv8, un algoritmo di intelligenza artificiale per la segmentazione istantanea. Tale algoritmo di deep learning sviluppato da Ultralytics LLC utilizza una rete neurale convoluzionale (CNN) che è in grado di identificare e localizzare oggetti all'interno di immagini e video in tempo reale e di classificarli in gruppi. Il modello allenato, quindi, con queste immagini ha prodotto vari risultati. Nella parte di test viene, infatti, dimostrata l’efficacia dell’algoritmo utile per il monitoraggio in campo delle viti. Dall’analisi di 108 immagini risulta un F1, ovvero la media armonica di precisione e richiamo, di 0,64 con una confidenza di 0,446. Tale dato fornisce una valutazione equilibrata delle prestazioni di un modello considerando sia i falsi positivi che i falsi negativi. Inoltre nella metrica Mask (P, R, mAP50, mAP50-95) il test del modello ha fornito i seguenti risultati: una precisione del modello “P” di 0,711 su 1, un richiamo “R” ovvero la capacità del modello di identificare tutte le istanze di oggetti nelle immagini di 0,589 su 1 e una precisione media calcolata con una soglia di intersezione su unione (IOU) di 0,50 “mAP50” di 0,608 su 1. Inoltre la velocità dell’elaborazione condotta dal modello YOLOv8 è risultata essere di 36,76 FPS.
Utilizzo dell'algoritmo YOLO per il riconoscimento delle componenti legnose in Vitis vinifera
SASSO, TOMMASO
2023/2024
Abstract
With the aim of meeting a constantly growing food demand while reducing its environmental impact, a more efficient and sustainable agriculture is becoming increasingly necessary globally. For this reason, precision agriculture presents itself as a crucial response to the challenges of the agricultural sector today and even more so in the coming decades. Precision agriculture is based on steps common to various applications and involves an initial collection of data using various types of sensors mounted on the most diverse platforms such as drones. Subsequently, this data is analyzed by specific software or programs. Based on the information obtained, farmers can make more precise decisions and therefore carry out more targeted actions; hence precision agriculture. Artificial intelligence, AI, with its ability to analyze large amounts of data and make autonomous decisions, offers enormous potential to transform the agricultural sector in this sense. This work examines how AI can be used to analyze grapevines, Vitis vinifera, and in particular their woody components. In the spring of 2024, in seven temporal repetitions spaced about a week apart, images were collected of several sections of vine rows, each containing two to six grapevines. The images were collected in four different vineyards, all with white grapes, two of the Solaris B. variety aged three and six years and two of Pinot Bianco B., also aged three and six years, present on the hill of Cart, in the municipality of Feltre, province of Belluno, Italy. The collected images were processed using a program called Roboflow which allows you to outline the contour of the desired objects. Three classes of objects were then identified, namely the woody components of the vine: rootstock, trunk, and shoots. The images thus modified were then analyzed with YOLOv8, an artificial intelligence algorithm for instant segmentation. This deep learning algorithm developed by Ultralytics LLC uses a convolutional neural network (CNN) that is able to identify and locate objects within images and videos in real time and classify them into groups. The trained model, therefore, produced various results with these images. In the test part, in fact, the effectiveness of the algorithm useful for field monitoring of vines is demonstrated. From the analysis of 108 images, an F1 score, or the harmonic mean of precision and recall, of 0.64 with a confidence of 0.446 is obtained. This data provides a balanced evaluation of the performance of a model considering both false positives and false negatives. Furthermore, in the Mask(P, R, mAP50, mAP50-95) metric, the model test provided the following results: a model precision "P" of 0.711 out of 1, a recall "R" or the model's ability to identify all object instances in the images of 0.589 out of 1 and a mean precision calculated with an intersection over union (IOU) threshold of 0.50 "mAP50" of 0.608 out of 1. Furthermore, the processing speed conducted by the YOLOv8 model was 36.76 FPS.| File | Dimensione | Formato | |
|---|---|---|---|
|
Sasso_Tommaso.pdf
accesso aperto
Dimensione
7.49 MB
Formato
Adobe PDF
|
7.49 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/77888