Hyperspectral Classification for Chinese Medicinal Herbs via Multi-Perspective Feature Extraction and Information-Theoretic Fusion.

Accurate and efficient classification of the origin of Chinese medicinal herbs is crucial for ensuring their quality, safety, and efficacy. This study takes the Chinese medicinal herb Angelica dahurica as an example for research. Hyperspectral reflectance data, which provides reflectance values of samples at different wavelengths, offers an effective way to characterize Chinese medicinal herbs. Hyperspectral reflectance data can be viewed as tabular data, allowing feature extraction via tabular data feature extraction methods; it can also be viewed as a reflectance sequence, enabling feature extraction by borrowing time-series feature extraction methods. That is, multiview learning can be applied to classify hyperspectral data. In this paper, an autoencoder (AE), a 1D convolutional neural network (1DCNN), and a Gated Recurrent Unit (GRU) are used for feature extraction. Then, a fusion network consisting of a linear transformation, a bilinear transformation, and a compression unit is proposed for feature fusion, and the feature fusion is based on the Information Bottleneck criterion. The fused features are fed into a fully connected neural network (the classifier) to perform the classification of the Chinese medicinal herb. Experimental results on Angelica dahurica data show that integrating features from different views effectively improves classification accuracy. Multiview classification can serve as a method for the classification of Chinese medicinal herbs.