Identification of licorice from different origins using hyperspectral imaging technology

Traditional Chinese medicine identification mainly relies on methods such as observation, touch, smell, taste, water test, and fire test to determine the authenticity, quality, and approximate origin of medicinal materials based on experience. Testing personnel need to have rich experience, but reliability and reproducibility are low. With the advancement of chemical detection technology, thin-layer chromatography, liquid chromatography, and other methods can accurately detect differences between samples. However, pre-processing is time-consuming and labor-intensive, and detection costs are high, which cannot meet the requirements of industrial production for rapid online sorting of licorice. Hyperspectral imaging technology integrates spectral technology and machine vision technology, which can simultaneously study the spectral and image information of object quality attributes. It is a pollution-free, fast, non-destructive, and low-cost detection technology. The application of hyperspectral imaging technology in the field of traditional Chinese medicine is relatively limited. This study aims to obtain the hyperspectral information of licorice from visible near infrared (VNIR) and short wave infrared (SWIR) bands through hyperspectral imaging technology, and use the differences in spectral features to reflect the differences in internal material characteristics of medicinal materials. At the same time, image texture information is fused to quickly and non destructively identify licorice from different origins.

1 Materials and Methods

1.1 Medicinal herbs

The experimental licorice medicinal materials were selected from three representative licorice producing areas, namely Inner Mongolia, Gansu, and Xinjiang. Selecting 30 pieces of each batch of medicinal herbs and arranging them in a matrix, a total of 1620 licorice samples can be collected from 54 batches for hyperspectral image analysis. According to the algorithm, the samples are divided into a training set and a testing set, with 1085 samples in the training set and 535 samples in the testing set.

1.2 Instruments
The hyperspectral camera used in this study can be studied using the FS-15 product from Hangzhou Caipu Technology Co., Ltd. Short wave near-infrared hyperspectral camera, with an acquisition speed of up to 200FPS across the entire spectrum, is widely used in fields such as component recognition, substance identification, machine vision, agricultural product quality, and screen inspection.

1.3 Image acquisition of hyperspectral imaging system

In order to prevent deformation and distortion of the collected images, it is necessary to adjust parameters such as the distance between the sample and the objective lens, camera exposure time, and platform movement speed. To maintain consistent dataset size, the VNIR range length is determined to be 870 pixels and the SWIR range length is 1350 pixels.

2 Results and Analysis

2.1 Analysis of raw spectral curves of VNIR and SWIR

The original spectra in the visible near infrared (VNIR) and shortwave infrared (SWIR) ranges are shown in Figure 1. The signal-to-noise ratio of the 898-1042 nm band in the VNIR range and the 1600-1751 nm band in the SWIR range is low. Therefore, the parts with obvious noise are removed, and the spectra of the first 181 band intervals (435-898 nm) in the VNIR range and the first 421 band intervals (898-1600 nm) in the SWIR range are selected for analysis.

After threshold segmentation and removal of interference background, the regions of interest of licorice samples were extracted. VNIR and SWIR spectra of 1620 samples from 54 batches of licorice were collected, and 1620 spectral curves were analyzed under different bands, including 630 in Inner Mongolia, 510 in Gansu, and 480 in Xinjiang. The average spectra of licorice from different origins were obtained by calculating the spectral curves, and the average spectra of licorice from different origins in VNIR and SWIR are shown in Figure 2. From the graph, it can be seen that licorice in Xinjiang exhibits high reflection intensity in both VNIR and SWIR ranges. The curves corresponding to Inner Mongolia and Gansu have similar trends in SWIR range, only separating after 900-1050 nm and 1500 nm. Overall, the spectral curve trends of samples from different licorice production areas are similar, with no significant differences. At the same time, it was also found that in certain bands, the corresponding reflectivity is different, indicating differences in the chemical composition or physical properties inside, which may be due to differences in soil environment, lighting conditions, and cultivation methods in different production areas.

3 Discussions

The traditional method of classifying the origin of medicinal herbs cannot achieve fast online monitoring, and it is necessary to use spectroscopic and chemometric methods to improve the process control of traditional Chinese medicine and achieve digitalization of its quality. This article uses a three dimensional data fusion method to significantly improve the accuracy of classification models. Spectral and image fusion based on the full band can achieve the best accuracy, while full data fusion in the SPA band can achieve full band classification using only 28 feature wavelengths. With the improvement of instrument accuracy and further optimization of data processing algorithms, hyperspectral imaging technology has great potential and broad prospects in the classification of medicinal origins, identification of base materials, and quality grading.