Assessment of grapevine variety discrimination using stem hyperspectral data and AdaBoost of random weight neural networks

Grapevine variety discrimination is important for growers to be able to certify plants, however, this is not simple to do using conventional ampelometric methods or DNA based methods. The present work proposes, to the best of the authors' knowledge, for the first time, to differentiate grapevine varieties using only spectroscopic stem information. A total of 1200 measurements were gathered non-destructively in the field. The spectroscopic data was processed with a combination of AdaBoost and random weight neural networks (RWNN) which is still rare in scientific literature. Ten grapevine varieties, 5 red and 5 white, were used: Cabernet Sauvignon (CSvar), one of the most widely planted varieties worldwide, and nine Portuguese autochthonous varieties, including Touriga Franca and Touriga Nacional (TNvar) that are very important in Port wine production. Portugal is known to be one of the countries with the largest number of autochthonous varieties in the world. In general, the overall correct classification percentages were significantly better than the 10% level of random classification. True positive rates for a classifier separating ten varieties varied between 41.7 and 70.8%, depending on the variety. The false positive rate (FPR) varied between 2.5 and 7.3%. TNvar and CSvar measurements were correctly classified in 70.8 and 70% of cases with FPR of 5.3 and 3.3%, respectively. The algorithms AdaBoost of extreme learning machines, random vector functional link with pseudoinverse and no regularization, support vector machines and random forest did not provide better results than the AdaBoost of RWNN.