摘要
使用高光谱成像技术在主成分空间的距离作为样本相似性的判断依据,选择出训练子集,实现了一种基于主成分空间样品光谱特征分类的局部建模方法,并将这种方法与BP神经网络结合,用于苹果糖度的高光谱图像技术的定量分析,增强了检测效果。该方法首先提取高光谱图像的光谱信息并划分为训练集、验证集和测试集,对训练集的光谱进行主成分分析;然后在主成分空间根据欧氏距离和马氏距离选择训练子集,并建立基于BP神经网络的局部回归模型对验证集进行预测;使用全光谱+BP神经网络、全光谱+PCA+BP神经网络、欧氏距离+PCA+BP神经网络和马氏距离+PCA+BP神经网络,选取出的训练子集建立的模型对验证集糖度进行预测。结果表明,与全光谱相比,局部变量建模相关系数r提高,RMSEP降低,提高了检测效果;局部变量建模中,马氏距离+PCA+BP神经网络建立的局部预测模型预测能力更强。将这种建模方法应用于测试集的糖度预测,均方根误差为0.10661,相关系数r为0.89081,可以较好的实现苹果中糖度含量高光谱定量分析。因此马氏距离+PCA+BP神经网络模型有望成为一种有潜力的苹果糖度检测方法,对提高模型的精度有重要的意义。
The distance of hyperspectral imaging technology in the principal component space is used as the judgment basis of sample similarity,and the training subset is selected to realize a local modeling method based on the classification of spectral characteristics of samples in the principal component space.This method is combined with BP neural network for the quantitative analysis of hyperspectral image technology of apple fructose,which enhances the detection effect.Firstly,the spectral information of hyperspectral images is extracted,and the training set,validation set and test set are divided,and the spectral of training set is analyzed by principal component.Secondly,Euclidean distance and Mahalanobis distance of the principal component space are adopted to construct a prediction model based on BP neural network.The model established by the training subset selected by the full spectrum+BP neural network,the full spectrum+PCA+BP neural network,the Euclidean distance+PCA+BP neural network and the Mahalanobis distance+PCA+BP neural network was used to predict the validation set sugar content.The results show that compared with the full spectrum,the local variable modeling correlation coefficient r is increased and RMSEP is decreased,which improves the detection effect.In local variable modeling,the local prediction model established by Mahalanovic distance+PCA+BP neural network has better prediction ability.When testing,the root mean square error is 0.10661 and the correlation coefficient is 0.89081,which can achieve better prediction results of sugar content in apple.Therefore,the modeling of Mahalanobis distance+PCA+BP neural network is more suitable for the detection of apple sugar content.
作者
陈杰
姚娜
武宁
吕海芳
CHEN Jie;YAO Na;WU Ning;LYU Haifang(College of Information Engineering,Tarim University,Alar,Xinjiang 843300)
出处
《塔里木大学学报》
2022年第4期69-76,共8页
Journal of Tarim University
基金
塔里木大学校长基金一般项目“苹果糖分含量高光谱检测方法与定量模型”(TDZKYB202002)
塔里木大学校长基金硕士项目“基于小波变换及智能优化的阿克苏苹果表面损伤检测图像处理方法研究”(TDZKSS202130)。
关键词
高光谱
糖度
主成分空间
欧氏距离
马氏距离
hyperspectral
sugar
principal component space
Euclidean distance
Mahalanobis distance
