Flavescence dorèe leaf symptoms - Analysis using hyperspectral imaging and machine learning

Aug 4, 2026

Upload by Quinn

This work introduces a contactless approach for detecting and classifying Flavescence dorèe (FD) leaf symptoms using hyperspectral imaging combined with machine learning. Three ML models classify Pinot Noir leaves into healthy, asymptomatic, and diseased categories, leveraging vegetation indices derived from per-pixel spectra. The dataset includes 201 hyperspectral hypercubes collected in 2023 and 2024 from the same field, balanced across the three classes. Results indicate good performance for a model trained on vegetation-index population features normalized with z-score.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages7

Reading Time18 min

AccessFree to download

Abstract

Introduction

Materials

Instrumentation

Data collection procedure

19. Analysis of Flavescence dorèe leaf symptoms using hyperspectral imaging and machine learning C. Nuzzi 1*, E. Saldi2, I. Negri2 and S. Pasinetti 1 1 Department of Mechanical and Industrial Engineering (DIMI), University of Brescia, Via Branze 38, Brescia, 25123, Italy; * [cristina.nuzzi@unibs.it](cristina.nuzzi@unibs.it) 2 Department of Sustainable Crop Productions (DI.PRO.VES.), Catholic University of the Sacred Heart, Via E. Parmense 84, 29122, Piacenza, Italy Abstract This work presents a novel approach for the detection and classification of Flavescence dorèe disease based on hyperspectral imaging. Three machine learning models were trained to classify leaf samples of Pinot Noir into healthy, asymptomatic, and diseased classes according to a combination of vegetation indices calculated on the per-pixel spectrum of the sample. The dataset used included 201 hypercubes collected from the same field in 2023 and 2024, equally divided into the three classes. Results highlighted good performance ( accuracy) for the model trained on the population of features extracted from the vegetation indices for each leaf and normalized using z-score. Keywords: Flavescence dorèe, hyperspectral imaging, machine learning, spectroscopy Introduction Flavescence dorèe (FD) is a quarantine disease affecting European vines. Known symptoms are evident in several parts of the plant, including wood, fruits, and leaves (European Food and Safety Authority et al., 2020). According to the grape variety, leaf symptoms are different (leaves become red in the case of dark grapes, yellow for white grapes). The variety also influences the plant’s resistance to the disease; however, even if it survives the infection, it cannot fully recover, leading to large cost for winemakers. The phytoplasma is transmitted to the plants by a vector insect, Scaphoideus titanus, and most control strategies focus primarily on the elimination of the insect in the field to prevent or limit the disease transmission. Several southern European countries are affected by FD, including Italy, France, Spain, Portugal, Croatia, and Slovenia. To control the spread of the disease and reduce potential contagion to nearby fields, local regulations force winegrowers to destroy the whole vineyard if at least 20% of it is affected by FD (Boulent et al., 2020) . Since a whole growing season may pass between infection and the appearance of symptoms, it is fundamental to develop strategies for early detection to mitigate the impact of FD. Currently, the only reliable method to certify the presence ofFD (which may also be confused with other diseases with similar symptoms) is to conduct a polymerase chain reaction (PCR) analysis on leaf samples; however, this method is effective only if the leaves exhibit clear symptoms, making early detection challenging (Pelletier et al., 2009). The research community is trying to come up with cost and time effective solutions, especially involving contactless sensors. The research described by Al-Saddik et al. (2017) focused on the analysis of spectral bands that highlight the presence of FD by using a portable spectroradiometer to analyse the leaves of several affected plants of different varieties. Similar experiments were conducted by Junges et al. (2020), enforcing the conclusion that most physiological changes can be detected in the visible spectral range (400–900 nm) and in the range 1600–2200 nm, where water absorption and phenolic compounds accumulation effects can be observed. In addition to contact spectrometers, hyperspectral imaging (HSI) is rapidly gaining momentum as well. The main advantage is that the spectral information is acquired for each pixel of the leaf sample at once, producing a higher amount of data compared to contact spectrometers thanks to higher spatial Precision agriculture ’25 171 resolution. For example, Silva et al. (2022) adopted machine learning (ML) techniques to analyse hyperspectral images, obtaining pro

FAQs

What disease and grape variety are targeted in this study?

The study focuses on Flavescence dorèe (FD) in Pinot Noir vines. The goal is to distinguish healthy, asymptomatic, and diseased leaves.

How does hyperspectral imaging contribute to the classification?

Hyperspectral imaging captures spectral information for each pixel of the leaf, creating a data-rich representation. The work then derives vegetation indices from these spectra to feed machine learning models.

What makes the proposed dataset and task different from earlier approaches?

Unlike approaches relying only on symptomatic leaves, this work adds an “asymptomatic” class that should test positive by PCR but shows no visible symptoms yet. The dataset is expanded to 201 samples and uses three feature-based ML approaches.

Aug 4, 2026

Upload by Quinn