Classification of Arabica and Robusta coffee samples subjected to different technological treatments using various image analysis methods

Coffee varietal differentiation based on NIR spectroscopy has been widely investigated in the last 20 years [1-3]. In this work, we have applied hyperspectral imaging in the NIR range (900-1700 nm) for the classification of Arabica and Robusta coffee varieties, considering coffee beans subjected to different processing methods, i.e., the so-called dry method (to produce natural coffee), wet method (to produce washed coffee) and a somewhat intermediate processing method, referred to as polishing method (to produce polished coffee). PCA has been used as an exploratory technique both on the image mean spectra and on the hyperspectrograms obtained from the images. The hyperspectrograms are built by compressing the useful information contained in each hyperspectral image into a signal composed by the frequency distribution curves of quantities calculated by PCA [4]. This procedure allows to compress the information conveyed by the hyperspectral images, maintaining at the same time both spatial- and spectral-related features. The PCA models obtained showed a clear clustering of Arabica and Robusta samples, whereas, considering the technological treatment, the polished coffee samples are clearly distinguishable from the others, while natural and washed coffee samples are quite superimposed. Image mean spectra and hyperspectrograms were then subjected to PLS-DA classification after preprocessing using SNV followed by meancentering or meancentering only. Concerning the discrimination of coffee samples between Arabica and Robusta categories, the same value of classification efficiency in prediction (EFFPRED = 86.3%) has been obtained considering both the mean spectra and the hyperspectrograms. After forward iPLS-DA variable selection, EFFPRED increased up to 98.6% for models calculated using the mean spectra and up to 100% for the models calculated using the hyperspectrograms. As for the discrimination of the coffee samples into the three natural, polished and washed processing categories, the PLS-DA models calculated using mean spectra led to EFFPRED values equal to 81.1%, 95.7% and 49.8%, respectively, while the PLS-DA models calculated using hyperspectrograms led to EFFPRED values equal to 94.7%, 100% and 92.4%, respectively. In this case, iPLS-DA variable selection led to an increase of the performances of the model calculated on mean spectra (EFFPRED equal to 82.9%, 98.6% and 86.5%, respectively) and to a decrease of the performances of the model calculated using hyperspectrograms (EFFPRED equal to 82.9%, 89.3% and 86.5%, respectively).