Straightforward and rapid detection of Canephora coffee adulteration in Arabica Coffee by Near-Infrared Hyperspectral Imaging and Chemometrics

The current Brazilian federal legislation, through ordinance Nº 570 of the Secretariat of Agricultural Defense (SDA/MAPA) on May 9, 2022, mandates the labeling and inspection of coffee species Arabica and Canephora in commercially roasted and ground coffee. However, both regulatory bodies and industry lack official methodologies to enforce this requirement, in addition to the challenges related to sampling solid and heterogeneous samples for methodologies that employ microscopy techniques. Therefore, the present study aimed to develop and validate a simple, fast, and robust analytical method for detecting the adulteration of roasted and ground Arabica coffee with roasted and ground Canephora coffee across different roast levels. This was achieved using near-infrared hyperspectral imaging supported by chemometric tools. The pure and certified samples of each coffee species, were kindly donated by Bourbon Coffee, Brazil, with light, medium, and dark roasts from the 2021, 2022, and 2023 harvests. Adulterated samples were prepared by blending Canephora coffee into Arabica at concentrations of 1% to 20%. Hyperspectral images of all samples were then acquired using a line-scan hyperspectral camera (Sisuchema, Specim). The resulting spectra were pre-processed (Savitzky-Golay 1st derivative, trimming of the ends, and Standard Normal Variate), and a Partial Least Squares Discriminant Analysis (PLS-DA) model was calibrated with spectra of standards. Using five latent variables, the model achieved high accuracy values of 0.999105 for calibration, 0.999140 for internal validation, 0.999510 for validation with the 2022 harvest, and 0.999200 for validation with the 2023 harvest. These expressive results not only demonstrate excellent performance but also the method’s robustness in handling samples from different harvests. With the calibrated PLS-DA model, predictions of the blends were performed. A tolerance limit was established to distinguish pure from adulterated samples based on the pixel count classified as Canephora. In this step, a significant portion (29%) of the 1% adulterated samples were mistakenly classified as pure. As a result, the method was determined to have an accuracy of 0.98, with a specificity of 0.99 and a sensitivity of 0.93 for detecting fraud above 5%. We conclude that the initial proposal of the study was successfully achieved. Despite the 5% detection limit for fraud, the method is an excellent tool for targeted sampling, as the images can pinpoint suspicious particles that could be collected and analyzed by a reference method.