Precision agriculture is revolutionizing modern farming practices through the integration of advanced technologies that enable accurate monitoring and optimized crop management. Among these technologies, the use of multispectral images is particularly promising, as it allows for the collection of detailed data on the health status of plants. These data, acquired by drones and satellites, provide fundamental information for optimizing agricultural productivity and sustainability, supporting agronomic decisions based on scientific evidence. This thesis, in collaboration with AIRFARM S.r.l.s., focuses on developing a convolutional neural network model based on the U-Net architecture, optimized for the automatic segmentation of multispectral images obtained using a multirotor drone equipped with two cameras: one for the visible spectrum and one for the infrared spectral bands. The main objective is to accurately distinguish vine rows from the surrounding soil, improving the accuracy of the NDVI (Normalized Difference Vegetation Index), a key indicator for assessing vegetation health, thereby minimizing the interference caused by soil in the images. The proposed model not only allows for detailed segmentation of the vine rows but also supports dynamic monitoring of the leaf surface, tracking vegetative biomass to promptly detect any stress or disease. Further research development includes the automatic detection of gaps, i.e., areas devoid of plants, which can indicate growth failures or plants that have not taken root, providing strategic information for targeted replanting and management. The results of the thesis demonstrate that the neural network-based approach offers greater precision compared to traditional methodologies, highlighting the advantages of multispectral images acquired by drones over satellite data. This innovative approach provides advanced tools for agricultural monitoring, contributing significantly to more efficient and sustainable vineyard management, with potential large-scale applications to enhance precision farming practices.
L’agricoltura di precisione sta rivoluzionando le pratiche agricole moderne attraverso l’integrazione di tecnologie avanzate che permettono un monitoraggio accurato e una gestione ottimizzata delle colture. Tra queste tecnologie, l’uso di immagini multispettrali è particolarmente promettente, poiché consente di raccogliere dati dettagliati sullo stato di salute delle piante. Questi dati, acquisiti da droni e satelliti, forniscono informazioni fondamentali per ottimizzare la produttività e la sostenibilità agricola, supportando decisioni agronomiche basate su evidenze scientifiche. Questa tesi, in collaborazione con AIRFARM S.r.l.s., si concentra sullo sviluppo di un modello di rete neurale convoluzionale basato sull’architettura U-Net, ottimizzato per la segmentazione automatica delle immagini multispettrali ottenute tramite un drone multielica equipaggiato con due fotocamere: una per lo spettro visibile e una per le bande spettrali nell’infrarosso. L’obiettivo principale è distinguere con precisione i filari di vite dal terreno circostante, migliorando l’accuratezza del calcolo dell’NDVI (Normalized Difference Vegetation Index), un indicatore chiave per valutare la salute della vegetazione, minimizzando così l’interferenza causata dal suolo nelle immagini. Il modello proposto non solo consente una segmentazione dettagliata dei filari, ma supporta anche il monitoraggio dinamico della superficie fogliare, tracciando la biomassa vegetativa per rilevare tempestivamente eventuali stress o patologie. Un ulteriore sviluppo della ricerca riguarda l’individuazione automatica delle zone mancanti di piante, che possono indicare lacune nella crescita o piante non attecchite, fornendo informazioni strategiche per interventi di reimpianto e gestione mirata. I risultati della tesi dimostrano che l’approccio basato su reti neurali offre una maggiore precisione rispetto alle metodologie tradizionali, evidenziando i vantaggi delle immagini multispettrali acquisite da droni rispetto ai dati satellitari. Questo approccio innovativo fornisce strumenti avanzati per il monitoraggio agricolo, contribuendo significativamente a una gestione più efficiente e sostenibile dei vigneti, con potenziali applicazioni su larga scala per migliorare le pratiche di agricoltura di precisione.
Segmentazione di immagini aeree tramite reti neurali per l’agricoltura di precisione
POMARO, DAVIDE
2024/2025
Abstract
Precision agriculture is revolutionizing modern farming practices through the integration of advanced technologies that enable accurate monitoring and optimized crop management. Among these technologies, the use of multispectral images is particularly promising, as it allows for the collection of detailed data on the health status of plants. These data, acquired by drones and satellites, provide fundamental information for optimizing agricultural productivity and sustainability, supporting agronomic decisions based on scientific evidence. This thesis, in collaboration with AIRFARM S.r.l.s., focuses on developing a convolutional neural network model based on the U-Net architecture, optimized for the automatic segmentation of multispectral images obtained using a multirotor drone equipped with two cameras: one for the visible spectrum and one for the infrared spectral bands. The main objective is to accurately distinguish vine rows from the surrounding soil, improving the accuracy of the NDVI (Normalized Difference Vegetation Index), a key indicator for assessing vegetation health, thereby minimizing the interference caused by soil in the images. The proposed model not only allows for detailed segmentation of the vine rows but also supports dynamic monitoring of the leaf surface, tracking vegetative biomass to promptly detect any stress or disease. Further research development includes the automatic detection of gaps, i.e., areas devoid of plants, which can indicate growth failures or plants that have not taken root, providing strategic information for targeted replanting and management. The results of the thesis demonstrate that the neural network-based approach offers greater precision compared to traditional methodologies, highlighting the advantages of multispectral images acquired by drones over satellite data. This innovative approach provides advanced tools for agricultural monitoring, contributing significantly to more efficient and sustainable vineyard management, with potential large-scale applications to enhance precision farming practices.| File | Dimensione | Formato | |
|---|---|---|---|
|
Pomaro_Davide.pdf
Accesso riservato
Dimensione
33.3 MB
Formato
Adobe PDF
|
33.3 MB | Adobe PDF |
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/20.500.12608/82600