基于高光谱的水稻稻曲病早期监测研究

为了快速、精准地感知水稻稻曲病的发生，实现稻曲病大面积早期监测，利用机载UHD185高光谱仪采集带有发病区域的多组水稻冠层高光谱图像数据，对图像数据进行预处理并建立数据集。对健康区域和发病区域进行分类训练，建立支持向量机(SVM)识别模型和主成分分析(PCA)加人工神经网络(ANN)的识别模型，通过验证样本来检验识别模型的准确性，达到识别发病水稻的目的。支持向量机识别模型选用两组特征波长下的假彩色图像：第1组波长组合(TZH1)为654、838、898 nm;第2组波长组合(TZH2)为630、762、806 nm,两组数据的错分误差/漏分误差总体分别达到4.24%和5.41%;其中S型核函数的SVM模型诊断性能最好，总体分类精度最高可达到95.64%,Kappa系数可达到0.94,基本达到了准确识别水稻稻曲病的目的。主成分分析加人工神经网络的识别模型选用前3个主成分，贡献率分别为93.67%、2.80%、1.24%,作为最优波长建立人工神经网络识别模型；其中非线性分类的效果优于线性分类的效果，总体分类精度达到了96.41%,Kappa系数可达到0.95。通过两个实验组数据的支持向量机...

Early Monitoring of Rice Koji Disease Based on Hyperspectroscopy

In order to detect the occurrence of rice koji disease quickly and accurately, and realize the early monitoring of rice koji disease in a large area, the airborne UHD185 hyperspectrometer was used to collect multiple sets of rice canopy hyperspectral image data with the disease area, and the image data was preprocessed to establish data sets. The classification training of healthy and diseased areas was carried out, and the recognition model of support vector machine (SVM) and principal component analysis (PCA) plus artificial neural network (ANN) was established to identify diseased rice, and the accuracy of the recognition model was verified by validating the samples. The support vector machine recognition model selected false color images under two sets of feature wavelengths. The first group of wavelength combination (TZH1) was 654nm, 838nm and 898nm, and the second wavelength combination (TZH2) was 630nm, 762nm and 806nm. The total commission error/omission error of the two sets of data reached 4.24% and 5.41%, respectively. Among them, the SVM model of the S-type kernel function had the best diagnostic performance, and the overall classification accuracy could reach 95.64% and the Kappa coefficient was 0.94, which basically achieved the purpose of accurately identifying rice disease areas. The recognition model of principal component analysis plus artificial neural network used the first three principal components, and the contribution rates were 93.67%, 2.80% and 1.24%, respectively, which were used as the optimal wavelength to establish the ANN recognition model. In the classification results, the nonlinear classification was better than the linear classification, the overall classification accuracy was 96.41% and the Kappa coefficient was 0.95. The results showed that through the diagnostic results of the support vector machine in the data of the two experimental groups, it can be seen that the classification accuracy of the recognition model using the support vector machine was stable overall, and there was no obvious difference in the diagnostic effect of the four kernel functions. In terms of overall classification accuracy, the nonlinear classification in the principal component analysis plus artificial neural network recognition model was 0.77 percentage points higher than that of the S-type kernel function classification of the support vector machine recognition model. Therefore, the nonlinear classification model in principal component analysis plus artificial neural network model was more suitable for early monitoring of rice koji disease.