Hyperspectral discrimination of tea plant varieties using machine learning, and spectral matching methods

Remote sensing-based discrimination and mapping of tea (Camellia sinensis) plantations are valuable for efficient management of inventory and optimization of resources by the tea production industry. Apart from the diverse tea plant varieties, growth of natural plant species is a common scenario in tea plantations. The objective of this research is spectral discrimination of nine popular tea plant varieties in the presence of six natural plant species in Munnar, Western Ghats of India. Canopy level hyperspectral reflectance measurements acquired for tea and natural plant species were analyzed using several statistical, and machine learning methods namely, k-nearest neighbourhood classifier (k-NN), linear discriminant analysis (LDA), support vector machines (SVM), normalized spectral similarity score (NS3), maximum likelihood classifier (MLC), and artificial neural networks (ANNs). In addition, the existence and statistical significance of the spectral separability among 15 tea and natural plant species was assessed by non-parametric MANOVA. Results indicate that six out of nine tea plant varieties could be discriminated with accuracies between 75% and 80%. The presence of natural plant species has decreased the inter-species spectral variability for a few tea plant varieties. However, there has been enhanced spectral variability for a few other tea plant varieties. The presence of natural plant species does not need to be disadvantageous to the spectral discrimination of tea species.