High-performance hyperspectral remote sensing and machine learning algorithms for detection of blister blight in Camellia sinensis

Abstract

Camellia sinensis is a widely cultivated crop that is harvested for two leaves and a bud. However, these soft tissues are highly susceptible to the infection known as Exobasidium vexans. This fungal disease reduces the quality and quantity of tea produced. The objective of the study was to develop a remote sensing-based model that could be used to predict the severity of blister blight infections. The study was conducted on five tea varieties susceptible to blister blight infections and the hyperspectral data were collected from leaves with a handheld instrument. Spectral preprocessing algorithms that included Puchwein's and Honig's were applied to select calibration sets and perform feature selection, respectively. Four machine learning algorithms that included artificial neural network (ANN), random forest, k-nearest neighbors, and support vector machine were compared. The result indicated that the ANN outperformed other machine learning models, achieving a training accuracy of 83% (kappa coefficient = 0.78) and a testing accuracy of 92% (kappa coefficient = 0.90). The classification model was tested on another set of Kangra Asha tea leaves, resulting in a classification accuracy of 90% (kappa coefficient = 0.86). Thus, machine learning methods provided a novel technique to identify blister blight disease in the tea crop.