高光谱成像的柑橘病虫害叶片识别方法

为监测柑橘生长状况，实现病虫害无损识别，利用高光谱成像技术和机器学习方法进行柑橘病叶分类研究。使用高光谱成像仪采集46片柑橘正常叶、46片溃疡病叶、80片除草剂危害叶、51片红蜘蛛叶和98片煤烟病叶的高光谱图像，在478～900 nm光谱范围内对每个叶片一个或多个发病区提取5×5的感兴趣区域（ROI），将ROI内每个像素的反射率值作为光谱信息，则一个ROI得到25个光谱信息样本，最终五类叶片共得到13250个光谱样本。利用随机法将全部样本划分为9 938个训练集和3 312个测试集。分别采用一阶求导（1stDer）、多元散射校正（MSC）和标准正态变换（SNV）三种方法对原始光谱信息进行预处理，对不同预处理方法后的数据采用主成分分析法（PCA）提取特征波长。1st Der预处理后得到7个特征波长，分别是520.2，689.0，704.8，715.4，731.2，741.8和757.6 nm；MSC和SNV预处理后得到7个相同的特征波长，分别是551.9，678.5，704.8，710.1，725.9，731.2和757.6 nm；原始光谱得到7个特征波长，分别是525.5，678.5，710.1，720.7，725.9，757.6和762.9 nm。分析PCA后的样本分布散点图可知，正常叶片、溃疡病叶片和红蜘蛛叶片样本有一定程度聚类，除草剂叶片和煤烟病叶片样本有大量重叠，仅依据PCA不能完成病虫害叶片的识别。对全波段（FS）和PCA特征波长数据在不同预处理方法下进行支持向量机（SVM）和随机森林（RF）建模，结果表明：数据在1stDer预处理方法下识别效果最佳，1st Der-FS-SVM模型总分类精度（OA）为95.98%，Kappa系数为0.948 2，1st Der-FS-RF模型OA为91.42%，Kappa系数为0.889 2，1stDer-PCA-SVM模型OA为90.82%，Kappa系数为0.881 6，1stDer-PCA-RF模型的OA为91.79%，Kappa系数为0.894；对PCA选择的特征波长数据建模，SVM和RF模型下识别率均达到84%，全波段下模型识别率在88%以上，FS数据建模效果优于PCA特征波长。研究结果表明，高光谱成像技术结合机器学习方法进行柑橘叶片分类是可行且有效的，为柑橘病虫害的无损准确识别提供理论根据。

Study on the Identification Method of Citrus Leaves Based on Hyperspectral Imaging Technique

To monitor citrus growth and realize nondestructive identification of pests and diseases, the leaf classification of citrus diseases was studied using hyperspectral imaging technology and machine learning method. Using hyperspectral imager to collect hyperspectral images of 46 normal citrus leaves, 46 canker leaves, 80 herbicide-damaged leaves, 51 red spider diseased leaves, and 98 soot diseased leaves. A 5×5 regions of interest (ROI) were extracted from one or more diseased areas of each leaf in the 478~900 nm spectral range. Taking the reflectance value of each pixel in the ROI as the spectral information, one ROI would get 25 spectral information samples, and finally the five types of leaves get a total of 13 250 spectral samples. The samples were divided into 9938 training sets and 3 312 test sets by random method. The first derivative (1st Der), multiple scattering correction (MSC) and standard normal transformation (SNV) were used to preprocess the original spectral information, and principal component analysis (PCA) was used to extract the characteristic wavelength of the data after different preprocessing methods. After 1st Der pretreatment, 7 characteristic wavelengths were obtained, which were 520.2, 689,704.83, 715.38, 731.2, 741.75 and 757.58nm respectively. After MSC and SNV pretreatment, 7 identical characteristic wavelengths were obtained, which were 551.85, 678.45, 704.83, 710.1, 725.93, 731.2 and 757.58 nm, respectively. The original spectrum obtained seven characteristic wavelengths, which were 525.48, 678.45, 710.1, 720.65, 725.93, 757.58 and 762.85 nm, respectively. The scatter plot of sample distribution after PCA analysis showed that there was a certain degree of clustering of normal leaves, canker leaves and starscream leaves, and a large amount of overlap between herbicide leaves and soot leaves, so the identification of pest and disease leaves could not be completed only based on PCA. Support vector machine (SVM) and random forest (RF) were used to model the all-band spectrum (FS) and PCA characteristic wavelength data under different pretreatment methods, and the results showed that: The OA of 1st Der-FS-SVM model was 95.98%, the Kappa coefficient was 0.948 2, the OA of 1st Der-FS-RF model was 91.42%, the Kappa coefficient was 0.889 2, the OA of 1st Der-FS-SVM model was 90.82%, and the Kappa coefficient was 0.881 6, OA and Kappa coefficient in 1st Der-PA-RF model was 91.79% and 0.894 respectively. For PCA characteristic wavelength data modeling, the recognition rate of SVM and RF models reached 84%, and the recognition rate of the full-band spectrum model was above 88%. The FS data modeling effect was better than that of PCA characteristic wavelength. The results show that it is feasible and effective to classify citrus leaves by hyperspectral imaging technique combined with machine learning method, which provides a theoretical basis for the accurate and nondestructive identification of citrus pests and diseases.