Machine learning applied to canopy hyperspectral image data to support biological control of soil-borne fungal diseases in baby leaf vegetables

Abstract

Baby leaf vegetables constitute a significant segment of the convenience fresh food market. Due to cultivation conditions under plastic tunnels favourable to pathogen and to restrictions about synthetic fungicide applications, these crops are prone to soil-borne diseases and need effective biological management. Real-time tracking by digital means of the performances of antagonistic agents against plant pathogens may be a great opportunity to optimize field practices and increase disease biocontrol efficacy. In this study, a non-linear machine learning approach, based on Artificial Neural Networks technique, was used to assess the biocontrol efficacy of Trichoderma spp. from VIS-NIR spectral reflectance, estimating disease severity in baby leaf young plants during specific plant-pathogen-antagonist interactions. The most successful accurate model architecture achieved a predictive success rate of 74%. The variable impact analysis on the 207 variables considered showed that the 20 most important frequencies lie in the intervals Δω1 = [426,460] nm Δω2 = [495,530] nm Δω3 = [570,667] nm Δω4 = [770,880] nm Δω5 = [940,1000] nm. The model with improved classification accuracy is highly suitable for the automated detection of healthy status considering a wide spectrum of crop/pathogen targets under Trichoderma spp. beneficial dealings. Findings indicate that hyperspectral image-derived features could be used as proxy for disease level tracking under biological control of telluric pathogens in baby leaf vegetable cultivations.