Grapevine variety identification using “Big Data” collected with miniaturized spectrometer combined with support vector machines and convolutional neural networks

Highlights

• Data gathered for an unprecedented number of varieties, 64.

• Use of unprecedented number of samples, 35833.

• First use of Convolutional Neural Networks in grapevine variety identification.

• Test AUROC for Touriga Nacional identification was 0.7922.

• Test AUROC for Touriga Franca identification was 0.9803.

Abstract

Several experiments have been previously reported suggesting that the application of spectroscopy and machine learning allows the identification of grapevine varieties, however, up to now, the maximum number of varieties separated was twenty and the total number of sample spectra used does not go beyond the few hundreds. The present work aim is to answer the question: Is it possible to separate one variety from an enlarged group of other varieties when the number of samples is also significantly increased? With this in mind, a total of 35,833 spectra from leaves of 626 plants from 64 varieties were gathered for the study. This is a non-trivial evolution from previous works because it originates an increase in the variability of spectra which brings in a higher risk that a significant percentage of spectra of different varieties are equal and cannot be separated. Simultaneously, it was studied if a miniaturized and easy to use spectrometer could deliver data whose quality was enough to allow varieties separation even with data being collected in the field, non-destructively, and under uncontrolled solar lighting. This data was used to build support vector machines and convolutional neural networks for separating Touriga Nacional from 63 other varieties (including Touriga Franca) or Touriga Franca from 63 varieties (including Touriga Nacional), and the classification efficiencies are analysed.

Introduction

Variety identification is an important topic in viticulture because the wine quality potential is variety dependent and also because the consumers know and want wines of certain varieties which affects the price of grapes. It is therefore important to have methods that ensure trueness-to-type of plants that come out of nurseries. Conventionally, this variety identification is done using ampelography and ampelometry (Tomic et al., 2013) where an expert analyses and measures tens of grapevine features. However, the large number of features to analyse and the similarities between varieties make this a hard and laborious process that cannot be applied to hundreds of plants in a short time period. In addition, training a good ampelographer can take years. Even though ampelography and ampelometry are widely accepted and reliable methods there have been famous cases where producers thought that they were producing a certain variety but were in fact producing another (Tassie, 2010). This can have high costs for these producers due to the influence of the grapes variety on their commercial value. More recently, new DNA based methods (Tomic et al., 2013) have been developed that in spite of being highly reliable are still slow and expensive which prevents their extensive use. With the objective of creating simpler methods, in the last few years, spectroscopic methods have been combined with machine learning methods with promising results (Gutiérrez et al., 2015a, Gutiérrez et al., 2016, Gutiérrez et al., 2015b, Cao et al., 2010, Arana et al., 2005, Diago et al., 2013, Yang et al., 2012). Spectroscopy and machine learning has also been applied to grapevine clone identification but this topic is beyond the scope of the present article (Fernandes et al., 2014).

Spectroscopy measures how electromagnetic radiation interacts with matter, i.e., how this radiation is absorbed or not depending on its wavelength. The ratio between the amount of light incident on a material and the amount of light coming from the material is called reflectance, and its plot versus the wavelength is called a reflectance spectrum. The necessity to use machine learning algorithms to process spectroscopic data comes from the large amount of information that this data contains. These algorithms learn from the spectral data to distinguish between different varieties. The current state-of-the-art in variety identification using spectroscopy and machine learning is described in Fernandes et al. (2018). Gutiérrez et al. (2015a) which separated 20 varieties and Cao et al. (2010) that used 197 samples for a single variety and 439 samples in total set the state-of-the-art in number of separated varieties and samples employed. The present work boosts these values by distinguishing samples of a certain variety from those of 63 other varieties; each variety to separate has more than 3000 samples and the total number of samples available for all varieties is 35833. This means a three-fold increase in the number of varieties, a 17-fold increase in number of samples in a variety and an 80-fold increase in the total number of samples employed. This leads to a more realistic and harder to solve problem than those in previously reported works. The reason is that there is an increased probability that the used dataset contains equal spectra from different varieties. This increase in number of varieties and samples also means an important step towards the creation of a robust grapevine variety identification system that can be commercialized. The present work reports the construction of machine learning classifiers capable of separating Touriga Franca (TF_var) or Touriga Nacional (TN_var) from the remaining varieties. When separating TF_var, TN_var was added to the set of remaining varieties and vice-versa. The classifiers used were Support Vector Machines (SVM) and Convolutional Neural Networks (CNN) and their results will be compared. To the best of the authors’ knowledge this is the first time that CNN are being employed in grapevine variety identification even though they have been used once in rice variety identification (Qiu et al., 2018). The built classifiers were of the one-vs.-all type meaning that they are binary and indicate if a spectrum belongs to a certain variety or not. The classifiers were tested with data gathered in a different day of the training and validation data to minimize the influence, on the spectra separation, of environmental or biological parameters specific to a certain day. The choice of TF_var and TN_var as the main varieties to be separated has to do with their importance in Portugal in the production of the worldwide famous Port wine. In fact, TF_var and TN_var have each 7% (Ranking de castas, 2017), of the total grapevine area planted in Portugal making them the two most planted Portuguese autochthonous varieties in the country. Portugal which has one of the largest pools of autochthonous grapevine varieties in the world, 239 (Cunha et al., 2016), is actively working towards their preservation and dissemination.