Early Detection of Plant Viral Disease Using Hyperspectral Imaging and Deep Learning

Area of Science:

• Plant Pathology
• Remote Sensing
• Machine Learning

Background:

• Early detection of grapevine viral diseases is crucial for vineyard management and preventing spread.
• Hyperspectral remote sensing offers a non-destructive method for identifying plant diseases.
• The newly discovered grapevine vein-clearing virus (GVCV) requires effective detection strategies.

Purpose of the Study:

• To identify and classify GVCV-infected grapevines at early asymptomatic stages using hyperspectral imagery.
• To explore the utility of various vegetation indices (VIs) and machine learning algorithms for disease classification.
• To evaluate the performance of different deep learning architectures in analyzing hyperspectral data cubes.

Main Methods:

• Hyperspectral images (400-1000 nm) of healthy and GVCV-infected grapevines were acquired.
• Data preprocessing focused on isolating grapevine pixels, followed by statistical analysis of spectral patterns.
• Pixel-wise and image-wise classifications were performed using traditional machine learning (SVM, RF) and deep learning (2D-CNN, 3D-CNN) with selected vegetation indices and spectral features.

Main Results:

• Discriminative spectral regions were identified in the NIR (900-940 nm) at 30 DAS and VIS (400-700 nm) at 90 DAS.
• Key vegetation indices for discrimination included NPQI, FRI1, PSRI, AntGitelson, and WSCT.
• Random Forest (RF) outperformed Support Vector Machine (SVM) in larger feature spaces, while 3D-CNN showed promise for feature extraction from hyperspectral data cubes.

Conclusions:

• Hyperspectral remote sensing combined with machine learning can effectively detect early-stage GVCV infections in grapevines.
• Specific vegetation indices and wavelength regions are critical for accurate disease classification.
• Advanced deep learning models like 3D-CNN offer potential for improved feature learning in hyperspectral plant disease analysis.