基于高光谱成像技术的小麦籽粒赤霉病识别

利用高光谱成像技术通过光谱分析和图像处理进行小麦赤霉病的识别。采用标准正态变量变换(SNV)和多元散射校正(MSC)方法对光谱进行预处理,分别利用连续投影算法(SPA)和正自适应加权算法(CARS)进行变量筛选提取特征波段,结果表明采用MSC-SPA和SNV-SPA算法时决定系数分别为0.901 9和0.900 6,均方根误差分别为0.223 8和0.223 2,筛选波长个数分别为7个和5个。利用SVM和BP神经网络算法建立的交叉验证模型及验证模型的准确率均达到90%以上。其中,MSC-SPA-SVM和SNV-SPA-SVM方法的建模集准确率分别为97.08%和94.17%;验证集准确率分别为98.33%和97.50%,均优于MSC-SPA-BP和SNV-SPA-BP模型。为了研究染病小麦的高光谱图像信息,利用主成分分析方法,根据权重系数选择最佳特征波长为627.698 nm。利用图像处理方法对特征波长下的特征图像进行预处理、特征提取。分别提取特征波长图像的形态参数特征和纹理特征参数等,根据特征参数相关性分析选择最优的建模特征参数。分别利用10折交叉验证方法建立线性判别分析、支持向量机和BP神经网络识别模型,结果表明3种识别算法识别准确率均在90%以上,具有较好的识别效果。

Identification of Fusarium Head Blight Wheat Based on Hyperspectral Imaging Technology

Fusarium head blight is one of the main infection diseases in wheat, and the infection of wheat has serious impact on food safety. In order to explore the rapid and nondestructive detection of wheat scab, the identification of wheat scab was carried out using spectral analysis and image processing in hyperspectral imaging technology. Standard normal variable transform (SNV) and multiple scatter correction (MSC) methods were used for spectral data pretreatment, and continuous projection algorithm (CARS) and the positive adaptive weighted (SPA) algorithm were used to select wavelength. The results showed that the determination coefficients ( R 2 ) of MSC-SPA and SNV-SPA were 0.901 9 and 0.900 6, respectively, the root mean square errors were 0.223 8 and 0.223 2, respectively, and the numbers of selected wavelength were 7 and 5, respectively. Support vector machine (SVM) and BP neural network algorithms were used for modeling. The results showed that the accuracy of the four models were above 90%. The accuracy rates of MSC-SPA-SVM and SNV-SPA-SVM were 97.08% and 94.17% for model calibration set, respectively, and those for the model validation set were 98.33% and 97.50%, respectively, which were better than those for model calibration set. According to image information analysis of disease wheat in hyperspectral image, the principal component analysis method was applied, and the best wavelength image was chosen at 627.698 nm according to the weight coefficient. Image processing method was used for preprocessing, feature extraction, etc. The morphological parameters and texture feature parameters of the best wavelength image were extracted respectively, and the optimal parameters of the model were selected according to the correlation analysis of the feature parameters. Ten-fold cross-validation method was adopted to establish linear discriminant analysis, support vector machine and BP neural network identification models. The results showed that the recognition accuracy of the three identification algorithms were all above 90%, which indicated that the proposed method were feasible and effective.