Development of a calibration model for near infrared spectroscopy using a convolutional neural network

Development of qualitative or quantitative models is essential to exploit the full potential of near infrared (NIR) spectroscopy. In tandem with one-dimensional convolutional neural network (1D-CNN), a data-driven model is developed using NIR spectroscopy to estimate organic contents. First, the 1D-CNN model is designed to capture the features of the NIR spectra by means of several convolutional and pooling operations. Then, the suitable hyper-parameters of 1D-CNN are obtained by using the grid search algorithm to achieve the optimal performance. Furthermore, the dropout operation is added into the 1D-CNN to suppress the overfitting problem by means of removing some neurons, and the probability distribution of throwing follows the Bernoulli distribution. The developed framework is validated by the application in the sugar content estimation of Huangshan Maofeng tea. The experimental results demonstrate that the key features of the NIR spectra are successfully extracted by the proposed strategy; thereby, a new effective scheme for analyzing NIR spectra is provided for food analysis.