Spectral discrimination of tea plant varieties by statistical, machine learning and spectral similarity methods

摘要

Remote discrimination and mapping of tea plantations is a valuable tool for efficient management of inventory and optimization of resources. Apart from the planting multiple tea varieties, growth of natural vegetation species is common scenario in plantations. The objective of this research was spectral discrimination of nine important tea varieties in the presence of six natural vegetation species in Munnar, Western Ghat highlands of south India. Discrimination models using six methods: k-nearest neighborhood 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) were applied on the hyperspectral reflectance measurements collected at canopy level. The existence and statistical significance of spectral differences of the tea and natural vegetation species were assessed by MANOVA. Results indicate that six out of 9 tea varieties could be discriminated with best accuracies 75 to 80%. While a closer spectral similarity is observed in few tea varieties, the presence of natural vegetation species has decreased inter species variability for few tea varieties while enhancing the same for few other tea varieties at the cost of reducing spectral separability among the vegetation species.