Classification of Imbalanced Near-infrared Spectroscopy Data

Due to the imbalanced distribution of real near-infrared spectroscopy data, it is difficult for traditional machine learning methods to correctly classify samples during the modeling process. In general, near-infrared spectral data sets are high-dimensional and have few samples. In order to enhance the classification accuracy of machine learning, here we propose an ensemble-based learning approach. Specifically, the proposed method first generates a number of samples using a variational autoencoder (VAE) network, and merges these with the original data to form a new balanced data set. Then a classification model is built using the multi-feature fusion cascade forest (FCForest) method. We verified and evaluated our approach using an imbalanced near-infrared spectroscopy data set from citrus greening. The experimental results showed that using VAE to generate the samples improved the classification accuracy for the imbalanced data. Furthermore, by using the FCForest method on the new, balanced data set, the performance of the classifier was further improved.