On-The-Go Hyperspectral Imaging Under Field Conditions and Machine Learning for the Classification of Grapevine...

Area of Science:

• Agricultural Science
• Plant Biology
• Remote Sensing

Background:

• Grapevine varietal classification is crucial for viticulture and the wine industry.
• Traditional methods like ampelography and DNA analysis are destructive and laboratory-bound.
• There is a need for efficient, non-destructive field-based phenotyping tools.

Purpose of the Study:

• To develop and evaluate an on-the-go hyperspectral imaging system for classifying grapevine varieties under field conditions.
• To compare the performance of different machine learning algorithms for this classification task.
• To establish a non-destructive method for plant phenotyping in commercial vineyards.

Main Methods:

• On-the-go hyperspectral imaging was conducted using a camera mounted on an all-terrain vehicle at 5 km/h.
• Spectral data were collected from 30 grapevine varieties across two phenological stages in a commercial vineyard.
• Support Vector Machines (SVM) and Multilayer Perceptrons (MLP) were employed for classification model development.

Main Results:

• Both SVM and MLP models achieved high performance, with recall F1 scores and AUC up to 0.99.
• The best SVM kernel was linear, and the optimal MLP activation function was hyperbolic tangent.
• MLP achieved individual variety prediction performance ranging from 0.94 to 0.99, while SVM ranged from 0.83 to 0.97.

Conclusions:

• On-the-go hyperspectral imaging coupled with machine learning is a viable and effective method for grapevine varietal classification in the field.
• This non-destructive technology offers a promising new tool for plant phenotyping, supporting grape growing and the wine industry.
• The system demonstrates the potential for practical deployment in commercial vineyards for accurate variety identification.