AI Case Study

University of Melbourne researchers achieve 94% accuracy identifying grape vine varieties using leaf images and neural networks

University of Melbourne researchers use neural networks to determine grape vine variety based only on scanned images of leaves. They achieved 94% accuracy demonstrating an effective, non-destructive way to differentiate between vine types in the field.

Industry

Consumer Goods And Services

Food Beverage And Drugs

Project Overview

Researchers from the University of Melbourne conducted research on 16 different types of grape plant leaves taken from an experimental vineyard in Palma de Mallorca. The leaves were scanned and 12 morpho-colorimetric parameters were determined, including area, perimeter and fractal dimension of each leaf. "The accuracies of the ANN models developed in this study were higher than 87%, with the highest accuracy for the leaf classification model based on morpho-colorimetric parameter (94.2%) compared to the model based on NIR (91.7%). This is highly significant since the morpho-colorimetric model (Model 1) is based on scanned leaves from the different cultivars and automated analysis of images. Compared to the NIR instrument, the morpho-colorimetric method is less expensive and less time-consuming. This method (scanning and ANN Model 1) can also be implemented in-field applications by using inexpensive portable and wireless scanners."

Reported Results

"The ANN Model 1 obtained using the 12 morpho-colorimetric parameters to classify 138 leaves into 16 different cultivars had an overall accuracy of 94.2%".

Technology

"The Matlab Neural Network ToolboxTM 10, which is based on pattern recognition was implemented to develop a model to classify the leaves according to their cultivar by using 13 parameters obtained from their morphometric characteristics... a scaled conjugate gradient algorithm was used for training along with a random division function for training, validation, and testing of data to generate the model using 70% of samples for training, 15% for validation, and 15% for testing. For the construction of the models, different number of neurons in the hidden layer were tested (3, 5, 7 and 10); ten neurons rendered the most accurate models.

The machine learning algorithms reported in this study are artificial neural networks based on pattern recognition with random separation of data for the model development (training), validation and testing. The latter algorithm rendered more accurate models compared to a series of other machine learning classifier algorithms tested for this same study, such as: three decision trees, two discriminant analyses, one logistic regression classifier, six support vector machines, six nearest neighbor classifiers, and five ensemble classifiers".

Function

Operations

Field Services

Background

"Grapevine leaves of different cultivars vary in chemical composition and morphology such as shape, dimension, color and edge shape. These differences in morphometric characteristics have been acquired as evolutionary traits corresponding to specific gene expressions and their interaction with the environment to which each cultivar has been adapted to. Therefore, every leaf morphology and chemical parameter is unique for all cultivars, which allows their classification through a series of different observations or measurements for identification purposes".

Benefits

Data

"A total of 16 different grapevine cultivars were used, from which three mature leaves from each of three different plants for each cultivar were selected, scanned and processed to obtain all the automated measurements." 138 leaf samples in total were used: 84 for the training set, 27 for the validation set, and 27 for the testing set.