Multi-channel spectral sensor-based instrument as an alternative to the Agtron for predicting coffee roast degree using classical chemometric and machine learning approaches.

This study developed a new alternative Agtron detection system using a low-cost multichannel spectral sensor (410-940 nm) combined with classical and modern chemometric regression methods, and implemented the selected model in an embedded system to produce a portable, affordable, and field-ready device. Spectral data were collected from Arabica and Robusta samples (from different origins) at various roast degrees. Modelling was performed using quantitative chemometrics ranging from classical (Multiple Linear Regression-MLR and Partial Least Squares Regression-PLSR) to modern, including three machine learning (ML) algorithms (Random Forest-RF, Support Vector Regression-SVR, Artificial Neural Networks-ANN). All approaches and scenarios yielded high prediction accuracy (R2v 0.983-0.994, RMSEV <4.7, RPD > 7.7), even with 50 % of the selected wavelengths. Furthermore, the selected model was successfully embedded into a stand-alone microcontroller device and delivered reliable results (R2 0.965-0.967 and RMSE <3), even when tested on new samples from different coffee types (Liberica and Excelsa).