Hyper Spectral Imaging and Optimized Neural Networks for Early Detection of Grapevine Viral Disease

Abstract

The early detection of viral infections in grapevines is crucial to implement timely countermeasures and prevent the spread of disease across vineyards. This study leverages remote sensing via hyper spectral imaging to non-invasively identify and quantify infections caused by the recently discovered grapevine vein-clearing virus (GVCV), primarily during the initial asymptomatic phase. Post-calibration and preprocessing of hyper spectral images, only pixels associated with grapevines were retained. To discern between reflectance spectra profiles of healthy and GVCV-infected vines, an advanced statistical technique was employed. Subsequent to data preprocessing, an artificial hummingbird optimization technique was utilized for feature extraction, ensuring the selection of the most relevant features for enhancing the overall model classification. Furthermore, a non-invasive method was adopted to estimate the total chlorophyll (Chl) content of grape leaves. The study found a correlation between Chl concentration and the red-edge chlorophyll index, with reflectance measurements in the near-infrared (755-765 nm) and red-edge (710-720 nm) spectral ranges. For both pixel-wise and image-wise classification of disease severity, a hybrid of ZfNet+VGG19 was deployed. The proposed method, termed the Artificial Humming Bird Optimized ZfNet+VGG19 neural network (AHB_ZfNet+VGG19), demonstrated a considerable acceleration and an increase in accuracy, primarily attributed to the incorporation of prior training and model deepening. When contrasted with established methodologies, the proposed approach achieved a superior performance with an accuracy of 98.27%, precision of 97.67%, recall of 97.41% and F1-score of 97.74% for the Salinas dataset, and an accuracy of 98.45%, precision of 97.1%, recall of 97.41%, and F1-score of 97.6% for the Indian pine dataset.