A Comparison between the Post- and Pre-dispersive Near Infrared Spectroscopy in Non-Destructive Brix Prediction Using Artificial Neural Network

Abstract

Even though near infrared (NIR) spectroscopy have been implemented in determining the Brix of pineapples, no traceable study compares the effects of different acquisition designs. Thus, this study aims to evaluate the prediction performance of both pre- and post-dispersive NIR sensing devices in non-destructive Brix prediction using artificial neural network (ANN). The pre-dispersive device has five narrowband light emitting diodes (LEDs) with different wavelengths and a photodiode detector, whereas the post-dispersive device has a bifurcated fiber optic, a broadband LED, and a spectral sensor. First, the NIR diffuse reflectance was non-destructively collected using both NIR devices. Then, the collected diffuse reflectance was calibrated with the white and dark references, and then pre-processed using normalization and standard normal variate methods. After that, ANNs were built for both devices using the pre-processed data. Results show both devices are suitable for sample screening application with range error ratio (RER) of more than seven. Nevertheless, the ANN that trained using the post-dispersive device outperformed that trained using the pre-dispersive device with an 8.1% improvement of correlation coefficient of prediction (i.e. from 0.6853 to 0.7408), and a 5.7% improvement of root mean square error of prediction (i.e. from 1.3918 to 1.313°Brix).