Hyperspectral Imaging Combined with Machine Learning to Classify Flavescence Dorée Symptoms

One of the main reasons for money loss for European winegrowers is represented by diseases affecting vines. Flavescence dorée (FD) is a phytoplasma transmitted by a leafhopper (Scaphoideus titanus), thus after a few plants are diseased the whole vineyard is quarantined and, in the most severe cases, destroyed. FD symptoms may vary according to the variety, including altered coloration and texture of the leaves. Scientists typically detect the disease using molecular analysis and only recently by leveraging visual inspection techniques, such as hyperspectral imaging (HSI). However, the molecular test is insufficient to detect the disease in asymptomatic samples. This work proposes a novel approach based on HSI and Machine Learning (ML) to increase the success rate of disease detection on asymptomatic samples. We collected the average spectral response of 100 samples (10 of which are healthy), subdivided into 3 classes: Healthy, Asymptomatic, and Diseased. Due to the low number of samples in the dataset, we also propose a synthetic augmentation procedure leveraging a Constrained Generative Adversarial Network (CGAN) to generate 103 spectra per class. From the synthetic spectra, we calculate 10 vegetation indexes: NDVI, RGI, CI, mCARI, REP, CUR, mARI, mARI2 (proposed in this work), ACI, and mACI. These values are used as the input features of an Ensemble of 30 Bagged Trees to classify the samples into the 3 classes. The model was tested on the real 100 samples, achieving a True Positive Rate (TPR) per class equal to 80%, 93%, and 96% respectively. As a further contribution, we evaluated the system's performance on real data using a Bayesian test.