基于高光谱成像的蓝莓内部品质检测特征波长选择方法研究

Characteristic Wavelength Selection Method of the Detection of Internal Quality of Blueberry Based on Hyperspectral Image

为了实现蓝莓内部品质快速、准确检测,采用高光谱成像技术对蓝莓的糖度和硬度多指标同时进行检测研究。提出多阶段特征波长选择方法,即采用连续投影法(SPA)和逐步多元线性回归(SMLR)等特征波长选择方法同时将糖度和硬度的特征波长选择出来。通过高光谱成像系统(400~1000nm)采集了200幅蓝莓图像,首先对高光谱图像进行多元散射校正、标准正态变量变换和Savitzky-Golay平滑等光谱预处理,选取最优的预处理方法。然后利用SPA或者SMLR选择出糖度的几个特征波长,在此基础上再利用SPA或者SMLR选择出硬度的几个特征波长,从而形成四个特征波长选择方法 (SPA-SPA、SMLR-SMLR、SPA-SMLR和SMLR-SPA),采用4种多阶段特征波长选择方法提取同时反映蓝莓糖度和硬度的特征波长的组合。最后以全波长光谱信息(FS)和4种多阶段特征波长选择方法得出的光谱信息作为BP神经网络模型的输入矢量,建立了蓝莓糖度和硬度的预测模型。结果表明:Savitzky-Golay平滑为最优的预处理方法 ,结合BP神经网络,采用SPA-SPA多阶段特征波长选择方法所得的预测性能最优,糖度校正集的相关系数(Rc)和校正均方根误差(RMSEC)分别达到0.959和0.318°Brix,硬度校正集的相关系数(Rc)和校正均方根误差(RMSEC)分别达到0.956和0.153°Brix。糖度预测集的相关系数(Rp)和预测均方根误差(RMSEP)分别达到0.952和0.391°Brix,硬度预测集的相关系数(Rp)和预测均方根误差(RMSEP)分别达到0.953和0.234°Brix。该研究表明,应用高光谱成像技术可以对蓝莓糖度和硬度多指标同时进行检测研究,所获得的特征波长可为开发多光谱成像的蓝莓品质检测和分级系统提供参考。

In order to realize the rapid and accurate detection of blueberry quality,multi-stage characteristic wavelength selection method of hyperspectral image was proposed to detect the soluble solid content and firmness simultaneously.Successive projections algorithm(SPA)or stepwise multiple linear regression(SMLR)was used to select several characteristic wavelengths of the soluble solid content and firmness.The hyperspectral images of blueberry over the spectral region between400nm and1000nm were acquired for200blueberry samples.Firstly,the hyperspectral image was preprocessed by MSC,SNV and Savitzky-Golay,and the optimal preprocessing method was selected.Then successive projections algorithm(SPA)or stepwise multiple linear regression(SMLR)was used to select several characteristic wavelengths of the soluble solid content.And on the basis of the firmness spectral information of these characteristic wavelengths,some characteristic wavelengths were selected by SPA or SMLR to form four characteristic wavelength selection methods,i.e.SPA-SPA,SMLR-SMLR,SPA-SMLR and SMLR-SPA.Four kinds of multi-stage characteristic wavelength selection methods were used to extract the combination of characteristic wavelengths which simultaneously reflect the soluble solid content and firmness of blueberry.Finally,the spectral information selected from full wave wavelength and the spectral information selected from four multi-stage characteristic wavelength selection methods were taken as input vector to build BP neural network model to predict soluble solid content and firmness.The results suggested that Savitzky-Golay Smoothing method was the optimal preprocessing method.The prediction result of SPA-SPA was the best among the multi-stage characteristic wavelength selection methods using BP neural network prediction model.The calibration results of soluble solid content were correlation coefficient(Rc)0.959,root mean square error(RMSEC)0.318°Brix.The prediction results of firmness were correlation coefficient(Rp)0.956and square mean error(RMSEP)0.153°Brix.The prediction results of soluble solid content were correlation coefficient(Rc)0.952and root mean square error(RMSEC)0.391°Brix.The prediction results of firmness were correlation coefficient(Rp)0.953and mean square error(RMSEP)0.234°Brix.The study demonstrated that hyperspectral image technique can simultaneously detect soluble solid content and firmness of blueberry,and the characteristic wavelengths obtained by multi-stage characteristic wavelength selection method can provide the reference for the development of multispectral image of blueberry quality detection and classification system.