高光谱结合机器学习鉴别林下参参龄

doi:10.19894/j.issn.1000-0518.240204

摘要/Abstract

摘要：

建立一种基于高光谱技术结合机器学习快速、无损鉴别市售林下生晒参参龄的方法。以常见市售林下生晒参为研究对象，采集不同年限的林下生晒参高光谱图像并基于阈值分割方法从感兴趣区域提取光谱数据，然后对数据进行多元散射矫正（MSC）-Savitzky-Golay smoothing（S-G smoothing）-一阶导数（FD）和MSC-S-G smoothing-二阶导数（SD）预处理消除干扰，将预处理后的数据进行3种机器学习，包括支持向量机回归（SVR）、主成分回归（PCR）和偏最小二乘回归（PLSR）。高光谱数据存在冗余性，采用竞争性自适应重加权采样法（CARS）、连续投影算法（SPA）和无信息变量消除算法（UVE）3种算法筛选波段去除无用的波长信息。 CARS-MSC-S-G smoothing-SD-SVR有较低的均方根误差RMSEC（0.0027），RMSEP（0.0120），较高的决定系数 $R C 2$ （0.9998）， $R P 2$ （0.9993），以及相对分析误差RPD（38），竞争性自适应重加权采样法（CARS）结合SVR模型效果最好，能够实现对林下生晒参年限的准确分类。基于高光谱成像技术结合机器学习能够实现对林下生晒参参龄的快速、无损鉴别。

关键词: 林下生晒参, 不同年限, 高光谱成像, 鉴别

Abstract:

The study establishes a method for rapid and non-destructive identification of the age of commercially available forest sun-dried ginseng based on hyperspectral technology combined with machine learning. Using common commercially available forest sun-dried ginseng as the research object， hyperspectral images of ginseng of different ages were first collected， and spectral data were extracted from regions of interest using a threshold segmentation method. The data were then preprocessed using Multiplicative Scatter Correction （MSC）-Savitzky-Golay smoothing （S-G smoothing）-First Derivative （FD） and MSC-S-G smoothing-Second Derivative （SD） to eliminate interference. Three machine learning models， including Support Vector Machine Regression （SVR）， Principal Component Regression （PCR）， and Partial Least Squares Regression （PLSR）， were applied to the preprocessed data. Due to the redundancy in hyperspectral data， three algorithms—Competitive Adaptive Reweighted Sampling （CARS）， Successive Projections Algorithm （SPA）， and Uninformative Variable Elimination （UVE） were used to select bands and remove redundant wavelength information. The CARS-MSC-S-G smoothing-SD-SVR model showed the best performance， with lower Root Mean Square Errors （RMSEC： 0.0027， RMSEP： 0.0120）， higher Correlation Coefficients （ $R C 2$ ： 0.9998， $R P 2$ ： 0.9993）， and RPD （38）. This model effectively achieved accurate classification of the age of forest sun-dried ginseng. Combining hyperspectral imaging technology with machine learning enables rapid and non-destructive identification of the age of forest sun-dried ginseng.

Key words: Forest sun-dried ginseng, Different ages, Hyperspectralimaging, Identification

中图分类号:

O657.3

韩路生, 王昱丹, 李宇航, 李根悦, 宋欣宏, 杨凯丽, 谭鑫, 王恩鹏. 高光谱结合机器学习鉴别林下参参龄[J]. 应用化学, 2025, 42(1): 69-77.

Lu-Sheng HAN, Yu-Dan WANG, Yu-Hang LI, Gen-Yue LI, Xin-Hong SONG, Kai-Li YANG, Xin TAN, En-Peng WANG. Identification of Age of Forest Sun-Dried Ginseng Based on Hyperspectral Imaging Technology[J]. Chinese Journal of Applied Chemistry, 2025, 42(1): 69-77.

图/表 8

图1 可见-近红外高光谱成像识别系统结构示意图

Fig.1 Schematic structure of the visible-near-infrared hyperspectral imaging identification system

图2 不同年份林下生晒参样品参A. 10-year forest sun-dried ginseng; B. 15-year old forest sun-dried ginseng; C. 20-year forest sun-dried ginseng

Fig.2 Samples of forest sun-dried ginseng

图3 不同年限林下生晒参的可见光-近红外高光谱采集的平均光谱曲线图AVE-10. average spectra profile of 10-year forest sun-dried ginseng； AVE-15. average spectral profile of 15-year forest sun-dried ginseng； AVE-20. average spectral profile of 20-year forest sun-dried ginseng

Fig.3 Mean spectral profiles of visible-near-infrared hyperspectral acquisition of forest sun-dried ginseng of different ages

图4 不同生长年限林下生晒参的可见光-近红外高光谱数据预处理结果图A. MSC-S-G smoothing-FD plot; B. MSC-S-G smoothing-SD plot

Fig.4 Pre-processing results of visible-near-infrared hyperspectral data of forest sun-dried ginseng of different growth ages

表1 基于PCR建模的高光谱数据预处理结果

Table 1 Preprocessing results of hyperspectral data based on PCR modeling

Models	Training set		Cross-Validation		Prediction set
Models	RMSEC	$R C 2$	RMSECV	$R C V 2$	RMSEP	$R P 2$	RPD
Original data-PCR	0.269 6	0.882 0	0.279 8	0.873 1	0.332 8	0.820 2	2
MSC-S-G smoothing-FD-PCR	0.261 9	0.888 6	0.270 0	0.888 6	0.262 5	0.884 0	3
MSC-S-G smoothing-SD-PCR	0.240 3	0.906 2	0.247 7	0.909 8	0.238 6	0.913 8	3

表1 基于PCR建模的高光谱数据预处理结果

Table 1 Preprocessing results of hyperspectral data based on PCR modeling

Models	Training set		Cross-Validation		Prediction set
Models	RMSEC	$R C 2$	RMSECV	$R C V 2$	RMSEP	$R P 2$	RPD
Original data-PCR	0.269 6	0.882 0	0.279 8	0.873 1	0.332 8	0.820 2	2
MSC-S-G smoothing-FD-PCR	0.261 9	0.888 6	0.270 0	0.888 6	0.262 5	0.884 0	3
MSC-S-G smoothing-SD-PCR	0.240 3	0.906 2	0.247 7	0.909 8	0.238 6	0.913 8	3

表2 基于MSC-S-G smoothing-SD的建模结果

Table 2 Modeling results based on MSC-S-G smoothing-SD

Models	Training set		Cross-Validation		Prediction set
Models	RMSEC	$R C 2$	RMSECV	$R C V 2$	RMSEP	$R P 2$	RPD
MSC-S-G smoothing-SD-PCR	0.240 3	0.906 2	0.247 7	0.909 8	0.238 6	0.913 8	3
MSC-S-G smoothing-SD-PLSR	0.215 3	0.924 8	0.235 2	0.911 5	0.215 4	0.924 5	3
MSC-S-G smoothing-SD-SVR	0.045 5	0.996 7	0.075 7	0.990 8	0.045 6	0.996 3	16

表2 基于MSC-S-G smoothing-SD的建模结果

Table 2 Modeling results based on MSC-S-G smoothing-SD

Models	Training set		Cross-Validation		Prediction set
Models	RMSEC	$R C 2$	RMSECV	$R C V 2$	RMSEP	$R P 2$	RPD
MSC-S-G smoothing-SD-PCR	0.240 3	0.906 2	0.247 7	0.909 8	0.238 6	0.913 8	3
MSC-S-G smoothing-SD-PLSR	0.215 3	0.924 8	0.235 2	0.911 5	0.215 4	0.924 5	3
MSC-S-G smoothing-SD-SVR	0.045 5	0.996 7	0.075 7	0.990 8	0.045 6	0.996 3	16

图5 特征波段算法筛选过程图A. CARS feature band screening process; B. UVE feature band screening process; C. SPA feature band screening process

Fig.5 Feature band algorithm screening process

表3 基于特征提取算法的模型结果

Table 3 Model results based on feature extraction algorithm

Models	Training set		Cross-Validation		Prediction set
Models	RMSEC	$R C 2$	RMSECV	$R C V 2$	RMSEP	$R P 2$	RPD
CARS-MSC-S-G smoothing-SD-SVR	0.002 7	0.999 8	0.049 6	0.996 2	0.012 0	0.999 3	38
SPA-MSC-S-G smoothing-SD-SVR	0.018 1	0.999 4	0.052 5	0.995 7	0.020 1	0.998 9	30
UVE-MSC-S-G smoothing-SD-SVR	0.040 2	0.997 5	0.176 5	0.950 6	0.041 6	0.997 3	19

表3 基于特征提取算法的模型结果

Table 3 Model results based on feature extraction algorithm

Models	Training set		Cross-Validation		Prediction set
Models	RMSEC	$R C 2$	RMSECV	$R C V 2$	RMSEP	$R P 2$	RPD
CARS-MSC-S-G smoothing-SD-SVR	0.002 7	0.999 8	0.049 6	0.996 2	0.012 0	0.999 3	38
SPA-MSC-S-G smoothing-SD-SVR	0.018 1	0.999 4	0.052 5	0.995 7	0.020 1	0.998 9	30
UVE-MSC-S-G smoothing-SD-SVR	0.040 2	0.997 5	0.176 5	0.950 6	0.041 6	0.997 3	19

参考文献 25

1	PAN Y， ZHANG H， CHEN Y， et al. Applications of hyperspectral imaging technology combined with machine learning in quality control of traditional Chinese medicine from the perspective of artificial intelligence： a review［J］. Crit Rev Anal Chem， 2023， 29： 1-15.
2	BAI Z， HU X， TIAN J， et al. Rapid and nondestructive detection of sorghum adulteration using optimization algorithms and hyperspectral imaging［J］.Food Chem， 2020， 30（331）： 127-290.
3	姚坤杉，孙俊，陈晨，等. 基于高光谱技术的三七不同部位粉末的无损鉴别［J］. 光谱学与光谱分析， 2023， 43（7）： 2027-2031.
	YAO K S， SUN J， CHEN C， et al. Nondestructive identification of powder from different parts of Panax pseudoginseng based on hyperspectral technique［J］. Spectrosc Spectral Anal， 2023， 43（7）： 2027-2031.
4	高梁喨. 基于高光谱技术的柑橘品质检测研究［D］. 雅安：四川农业大学， 2023.
	GAO L L. Research on citrus quality detection based on hyperspectral technology［D］. Ya'an： Sichuan Agricultural University， 2023.
5	DE OLIVEIRA ZANUSO B， DE OLIVEIRA DOS SANTOS A R， MIOLA V F B， et al. Panax ginseng and aging related disorders： a systematic review［J］. Exp Gerontol， 2022， 161： 111731.
6	LEE I S， KANG K S， KIM S Y. Panax ginseng pharmacopuncture： current status of the research and future challenges［J］. Biomolecules， 2019， 10（1）： 33.
7	余江锋，罗吉，黄亿健，等. 吉林省不同种类人参UPLC-PDA指纹图谱及化学模式识别研究［J］. 中药材， 2019， 42（4）： 822-827.
	YU J F， LUO J， HUANG Y J， et al. UPLC-PDA fingerprinting and chemical pattern recognition of different species of ginseng in Jilin Province［J］. Chin Mater Med， 2019， 42（4）： 822-827.
8	肖霞，曹俊凯. 人参当归颗粒的薄层鉴别［J］. 中国实用医药， 2012， 7（9）： 249-250.
	XIA X， CAO J K. Thin-layer identification of ginseng and angelica granules［J］. Chin Practical Med， 2012， 7（9）： 249-250.
9	ZHU H， LIN H， TAN J， et al. UPLC-QTOF/MS-based nontargeted metabolomic analysis of mountain-and garden-cultivated ginseng of different ages in northeast China［J］. Molecules， 2018， 24（1）： 33.
10	ZHANG L， WANG P， LI S， et al. Differentiation of mountain- and garden-cultivated ginseng with different growth years using HS-SPME-GC-MS coupled with chemometrics［J］. Molecules， 2023， 28（5）： 2016.
11	LEE B J， KIM H Y， LIM S R， et al. Discrimination and prediction of cultivation age and parts of Panax ginseng by Fourier-transform infrared spectroscopy combined with multivariate statistical analysis［J］. PLoS One， 2017， 12（10）： e 0186664.
12	YANG S， SHIN Y S， HYUN S H， et al. NMR-based metabolic profiling and differentiation of ginseng roots according to cultivation ages［J］.J Pharm Biomed Anal， 2012， 58： 19-26.
13	刘秋安，徐芳芳，张欣，等. 基于近红外光谱技术和分类与回归树算法建立天舒片崩解时间预测模型［J］. 中草药， 2021， 52（16）： 4837-4843.
	LIU Q A， XU F F， ZHANG X， et al. Establishment of a disintegration time prediction model for Tianshu tablets based on near-infrared spectroscopy and classification and regression tree algorithms［J］. Chin Herbal Med， 2021， 52（16）： 4837-4843.
14	LI X， WEI Y， XU J， et al. SSC and pH for sweet assessment and maturity classification of harvested cherry fruit based on NIR hyperspectral imaging technology［J］. Postharvest Biol Technol， 2018， 143： 112-118.
15	ABDI H， WILLIAMS L J. Partial least squares methods： partial least squares correlation and partial least square regression［J］. Methods Mol Biol， 2013， 930： 549-579.
16	HEMMATEENEJAD B， MIRI R， ELYASI M. A segmented principal component analysis--regression approach to QSAR study of peptides［J］. J Theor Biol， 2012， 305： 37-44.
17	MCKEARNAN S B， VOCK D M， MARAI G E， et al. Feature selection for support vector regression using a genetic algorithm［J］. Biostatistics， 2023， 24（2）： 295-308.
18	傅晨，王开军. 决定系数与相关系数相关系数辅助的LASSO回归［J］. 福建师范大学学报（自然科学版）， 2024， 40（2）： 57-63， 89.
	FU C， WANG K J. LASSO regression assisted by coefficient of determination and correlation coefficient［J］. J Fujian Norm Univ （Nat Sci Ed）， 2024， 40（2）： 57-63， 89.
19	徐萍，米琪，罗文秀，等. 滇重楼皂苷近红外光谱快速定量分析模型建立［J/OL］. 中成药， 1-7［2024-05-20］.
	XU P， MI Q， LUO W X， et al. Modeling of near-infrared spectroscopy for rapid quantitative analysis of Dianthus annuus saponins［J/OL］. Proprietary Chin Med， 1-7［2024-05-20］.
20	陈露萍，徐芳芳，张欣，等. 基于偏最小二乘法建立大株红景天片素片硬度近红外光谱预测模型［J］. 中草药， 2023， 54（8）： 2446-2452.
	CHEN L P， XU F F， ZHANG X， et al. Establishment of a near-infrared spectral prediction model for the hardness of Rhodiola rosea tablets based on the partial least squares method［J］. Chin Herbal Med， 2023， 54（8）： 2446-2452.
21	LI H， LIANG Y， XU Q， et al. Key wavelengths screening using competitive adaptive reweighted sampling method for multivariate calibration［J］. Anal Chim Acta， 2009， 648（1）： 77-84.
22	ZHU H， CHU B， ZHANG C， et al. Hyperspectral imaging for presymptomatic detection of tobacco disease with successive projections algorithm and machine-learning classifiers［J］. Sci Rep， 2017， 7（1）： 4125.
23	HAN Q， WU H， CAI C， et al. An ensemble of Monte Carlo uninformative variable elimination for wavelength selection［J］. Anal Chim Acta， 2008， 612（2）： 121-125.
24	YU H， LU J， ZHANG G. Continuous support vector regression for nonstationary streaming data［J］. IEEE Trans Cybern， 2022， 52（5）： 3592-3605.
25	刘子健，顾佳盛，周聪，等. 基于高光谱成像技术的山楂产地判别研究［J］. 食品工业科技， 2024， 45（10）： 282-291.
	LIU Z J， GU J S， ZHOU C， et al. Research on the origin discrimination of hawthorn based on hyperspectral imaging technology［J］. Food Ind Sci Technol， 2024， 45（10）： 282-291.

[1]	闫志慧, 姜珂, 陈立书, 朱仝飞, 陈竹, 杨永建, 王大成. 藏药柳茶叶乙醇水提取物中化学成分的鉴定[J]. 应用化学, 2014, 31(02): 237-241.
[2]	曾志, 谢润乾, 谭丽贤, 张涛. 川芎水蒸气蒸馏和超临界CO₂提取物化学成分的GC-MS分析鉴别[J]. 应用化学, 2011, 28(08): 956-962.
[3]	杨孝容, 向清祥, 熊俊如, 陈刚. 南北五味子荧光光谱差异的原因[J]. 应用化学, 2009, 26(09): 1070-1074.
[4]	张滨, 刘莉丽, 林炳承, 张亚兵, 盛海平. 用毛细管区带电泳法研究水稻种子胚乳贮藏蛋白[J]. 应用化学, 1998, 0(1): 56-58.
[5]	许汉英, 赵晓君. 分光光度法测定白酒中的糠醛[J]. 应用化学, 1997, 0(3): 84-86.