稻瘟病是世界公认的水稻重大病害之一。 实现稻瘟病害的早期分级检测, 对水稻病害早期防治及精准用药具有重要意义。 以大田自然发病水稻为研究对象, 采集稻瘟病发病早期染病叶片和健康叶片, 获取所有叶片样本在400~1 000 nm波段内的高光谱图像并提取光谱数据。 水稻在染病之初不会立刻出现病斑, 无法识别采集到的无斑叶片是否染病。 为实现对自然染病叶片早期无病斑状态的识别, 提出取染病叶片贴近病斑的非病斑区域高光谱数据作为染病等级中的1级样本进行检测分析。 按照病斑面积将样本划分为4个等级: 健康叶片为0级(109片)、 染病无病斑为1级(116片)、 病斑面积小于10%为2级(107片)、 病斑面积小于25%为3级(101片)。 运用主成分分析(PCA)和竞争性自适应重加权(CARS)算法进行特征变量选取, CARS提取的特征波段较多, 利用PCA算法对其进一步降维。 分别以全谱数据、 PCA提取的4个、 8个、 CARS选择的21个、 CARS-PCA提取的6个特征变量为输入, 建立水稻稻瘟病早期高光谱支持向量机(SVM)、 PCA4-SVM、 PCA8-SVM、 CARS-SVM和CARS-PCA-SVM检测模型。 结果显示, 所有模型对各级样本的检测准确率均较高, 其中, 对1级样本的检测准确率与其他级别相当, 识别效果较好; 所有模型的样本总体准确率均大于94.6%, CARS-SVM模型的总体准确率最高为97.29%, CARS-PCA-SVM模型为96.61%略低于CARS-SVM模型, 但其输入变量仅为6个, 较CARS-SVM的21个减少71.43%, 模型更为简洁、 更利于提高检测速度。 因此, 综合评价CARS-PCA-SVM模型最优, 各级准确率分别为97.30%, 94.87%, 94.29%和100.00%。 结果表明, 所建模型检测准确度较高, 可以实现对大田自然发病的稻瘟病早期分级检测, 为稻瘟病染病之初无病斑叶片的检测提供新思路, 为水稻稻瘟病早期防治、 精准施药及检测仪器开发提供理论依据。

Rice blast is a worldwide destructive rice disease. It is of great significance for rice disease control and precision spraying to detect rice blast early and identify the severity of the disease. Based on field experiment and natural infection of rice blast, infected leaves and healthy leaves were collected in the early stage of leaf blast. Hyperspectral images in the spectral range of 400~1 000 nm were captured and the spectral data were extracted. Rice leaves will not immediately show lesions at the beginning of the disease, so it is impossible to identify and collect samples of infected leaves without lesions. In order to realize the early detection of infected leaves without visible lesion, this study proposed to take hyperspectral data of lesion-free areas adjacent to the lesioned areas on the infected leaves as level 1 samples. According to the area of the lesion, the samples were divided into four levels: level 0 (109 pieces) for healthy leaves, level 1 (116 pieces) for infected leaves without visible lesion, level 2 (107 pieces) for leaves with lesion area <10%, and level 3 (101 pieces) for leaves with lesion area <25%. Principal component analysis (PCA) and competitive adaptive reweighting sampling (CARS) were used to extract feature variables; PCA algorithm was used to reduce further the dimension of the bands extracted by CARS. The support vector machine (SVM), PCA4-SVM, PCA8-SVM, CARS-SVM and CARS-PCA-SVM models for early detection of rice blast were build based on the full spectral variables and extracted feature variables, respectively. In this study, all models had high detection accuracy for all levels of samples. Level 1 had good detection accuracy, similar to other levels. All models had an overall accuracy rate above 94.6%. The highest was the CARS-SVM model at 97.29%, and the CARS-PCA-SVM model at 96.61% was slightly lower, but its number of input variables was only 6, which was 71.43% less than that of 21 in the CARS-SVM model. It further reduced the complexity of CARS-SVM model and improved the operation speed. So, the comprehensive evaluation of CARS-PCA-SVM model was optimal, with the identification accuracy of 97.30%, 94.87%, 94.29% and 100.00% for each level, respectively. Therefore, it is feasible to use hyperspectral imaging technology to detect the early stage of rice blast. The results presented in this paper can provide new ideas for the detection of infected leaves without lesions at the beginning of rice blast, and provide a theoretical basis for the early control of rice blast, precision spraying of pesticide and the development of detection instruments.