Explore the effect of data augmentation of spectroscopic data for deep learning models

Abstract

This study aimed to explore the effect of data augmentation techniques to improve the performance of deep learning models on data sets that contain more features than samples. The data set used as an example for this case study was Raman spectroscopic data. Deep learning models have a reputation of requiring large amounts of data and the sample size of data used in this study was small. As a way to study the effect of increasing sample size, three different augmentation techniques based on the principles of linear combinations, partial least squares and extended multiplicative scatter correction were developed and used to increase the sample size. The effect of these three augmentation techniques was studied for a convolutional neural network by training and evaluating the neural network using augmented data. The evaluation of the augmentation methods was based on the performance of the convolutional neural network. In order to study the behavior of the convolutional neural network, the performance of the neural network was compared to partial least squares regression model. Furthermore, learning curves where also used to analyze the performance of the neural network based on the sample size. The augmentation methods were used to artificially increase the sample size and the learning curves were used to see if the increase in sample size lead to improvement. The results of this study showed that using augmentation techniques to increase sample size does improve the performance of model.