Chemometric procedure for data analysis
Tea samples were divided into training (50%), validation (25%), and test (25%) sets by applying the Kennard–Stone (KS) uniform sampling algorithm (Kennard and Stone, 1969). Differing methods of pattern recognition were evaluated: (a) SIMCA, (b) PLS-DA, and (c) SPA-LDA. The variables chosen by SPA were those correspond- ing to the smallest cost function G value. The modeling procedures were accomplished using the training and validation samples (including the variable selection for the SPA-LDA modeling, and the determination of the principal components for each SIMCA modeling). The test samples were only used for the final data evaluation and comparison of the classification models (Soares et al., 2013).