Coffee is one of the most highly valued agricultural products worldwide, and accurate bean selection is essential for commercialization. This study evaluates the acoustic properties of Coffea arabica beans to detect the presence of parchment coffee. A total of 4,000 healthy beans were analyzed: 3,000 green beans (hulled and polished) classified into three commercial sizes (14/64”, 17/64”, and 20/64”) and 1,000 parchment beans. Acoustic responses were obtained by recording the impact sound produced when each bean collided with plates of stainless steel, polyvinyl chloride, and glass inside an acoustically isolated chamber. Impact signals were acquired at 51.2 kHz and processed to extract maximum amplitude and power spectral density features across the 0–25.6 kHz range. These features were used to train several classification algorithms to discriminate between parchment and green coffee. Three feature-reduction approaches—maximum relevance–minimum redundancy, principal component analysis, and supervised variable selection—were applied to optimize model performance. Among the materials tested, glass provided the most discriminative acoustic response. Overall, artificial neural networks achieved the highest classification accuracy, reaching 98.89% in validation and 99.17% in testing using a reduced set of 62 key features derived from the supervised variable selection approach.
Chaves et al. (Fri,) studied this question.