Early detection of pre-symptomatic downy mildew (Pseudoperonospora cubensis) infectionin cucumbers by using deep learning tools based on hyperspectral imagery

Abstract

Downy mildew is one of the most severe diseases affecting cucumber cultivation in both field and greenhouse environments. Early detection is crucial for implementing effective management strategies to minimize or prevent disease spread. Hyperspectral imaging (HSI) offers a powerful non-invasive approach for plant disease detection by capturing detailed spectral information. In this study, a deep learning-based detection system was developed for the pre-symptomatic identification of downy mildew in cucumber. Hyperspectral images were acquired using a Specim IQ camera at 24- and 48-hours post-inoculation, including both control (healthy) and infected samples to construct the dataset. Preprocessing steps included Minimum Noise Fraction (MNF) transformation, Principal Component Analysis (PCA), and data augmentation. The Hyper3DNet deep neural network (DNN) architecture was employed for classification, and cross-validation (CV) was conducted to ensure model reliability. Model performance was evaluated using subsets of spectral bands derived from PCA. The coefficient of variation was calculated to support the selection of the optimal number of PCA bands, with the lowest value obtained at 12 bands. PCA loadings were used to assess the contribution of the original bands to the principal components, highlighting influential wavelengths in the 750–1000 nm range, as well as around 400 nm and 560 nm. The results demonstrate robust classification performance, with the mean accuracy, precision, recall, and F1- score values for 24-hours, 48-hours and combined (24h + 48h) groups being above 81.50, 71.61 and 73.28, respectively, across the evaluated metrics. This highlighted the potential of deep learning and hyperspectral imaging for early disease detection in cucumbe