In relation to the classification errors, the SPA-LDA results were better than SIMCA, and similar to PLS-DA. The ‘‘AFC parameters’’ model presented a correct classification of 86.7% when using SPA- LDA. The other models correctly classified all the samples in the test set, which indicates that the parameters selected by SPA-LDA convey enough information for tea discrimination. Despite the good performance of all supervised classification techniques using the ‘‘Elements in tea infusions’’ data set, SPA-LDA provided the more significant result, a feature selection based on only three chemical parameters. Fig. 2 shows the Fisher’s discriminant score plot obtained by using the parameters (K, Al and Mg) selected by SPA-LDA for ‘‘Elements in tea infusions’’. As can be seen, the separation between the classes is more apparent than those of the respective PCA score plot presented in Fig. 1d. Moreover, three of the most relevant parameters indicated in the loading plots in Fig. 1d were selected by SPA, which demonstrates its suitability for feature selection.