5 °C). Diffuse reflectance (DR) measurements were performed in diffuse reflection mode with a Shimadzu sampling accessory (DRS8000A). The ground coffee sample was mixed with KBr check details (100 mg) and then 23 mg of this mixture was placed inside the sample port. Pure KBr was employed as reference material (background spectrum). All spectra were recorded within a range

of 4000–400 cm−1 with a 4 cm−1 resolution and 20 scans, and submitted to background subtraction. The spectra were also truncated to 2500 data points in the range of 3100–600 cm−1, in order to eliminate noise readings present in the upper and lower ends of the spectra. Preliminary tests were performed in order to evaluate the effect of particle size (0.39 mm < D < 0.5 mm; 0.25 mm < D < 0.39 mm; 0.15 mm < D < 0.25 mm; and D < 0.15 mm) and coffee/KBr mass ratio (2, 5, 10, 20, 30, 40 and 50%) on the quality of the obtained spectra. The conditions that provided the best quality spectra (higher intensity and lower noise interference) were D < 0.15 mm and 10% coffee/KBr mass ratio. In order to improve performance of prediction models, the following data pretreatment techniques were evaluated: (0) no additional processing

(raw data), (1) mean centering, (2) normalization, (3) baseline correction, (4) first derivatives PS341 and (5) second derivatives. Mathematical treatments such as mean centering and normalization are commonly applied to data in order to remove

redundant information and enhance sample-to-sample differences ( Wang et al., 2009). Mean centering corresponds to subtraction of the average absorbance value of a given spectrum from each data point. Normalization is calculated by dividing the difference between the response at each data point and the minimum absorbance value by the difference between the maximum and minimum absorbance values. Baseline correction and derivative transformations are usually performed in order to compensate for baseline offset between samples and also to reduce instrument variations ( Esteban-Díez, González-Sáiz, Sáenz-González, & Sitaxentan Pizarro, 2007). The statistical software XLSTAT Sensory 2010 (Addinsoft, New York) was employed for all the chemometric calculations. Average spectra obtained for defective and non-defective roasted coffee samples are shown in Fig. 1. A comparative evaluation of these spectra indicates that they are quite similar, although variations in band intensity are perceived, with absorbance values being higher for non-defective and light sour beans and lower for black beans. The two sharp bands at 2920 and 2850 cm−1 have been previously identified in Arabica and Robusta roasted coffee samples (Kemsley et al., 1995) and also on Arabica green coffee samples (Craig et al., 2011 and Craig et al., 2012), in association to asymmetric and symmetric stretching of C–H bonds.