Sensing and Spectroscopy Lab

This work aimed to predict specialty coffee flavors using machine and deep learning methods on near-infrared spectra.

The developed models could provide flavor prediction for seven categories with the performance of recall above 70% and accuracy above 72%.

The method of focusing plot could visualize the attention of the deep convolutional neural network and explain the decisive feature bands for predicting the coffee flavors.

The feasibility of applying a very deep convolutional neural network on NIR spectra and multi-labeled data were demonstrated.