番茄早疫病感染性强、 破坏性大, 潜育期症前特征的检测识别是番茄早疫病监测预警和科学防治的关键。 在实验室以离体番茄叶片作为研究对象, 利用高光谱图像监测番茄叶片早疫病的病程演变情况, 结合可见光图像和光谱特征进行数据分析。 实验发现, 番茄叶片感染早疫病后其近红外光谱平均值和红边反射率随着时间不断降低, 且在接种36 h时已出现潜育期病症信息。 选择接种36 h的光谱数据作为番茄早疫病潜育期的建模数据, 分别利用了主成分(PCA)变换、 多元散射校正(MSC)对建模数据进行光谱降维或降噪处理, 进而建立梯度提升决策树(GBDT)和支持向量机(SVM)识别模型, 并导入数据进行训练识别。 讨论了PCA和MSC的预处理方法对梯度提升决策树(GBDT)和支持向量机(SVM) 模型识别效果的影响; 进一步讨论常见核函数对SVM识别模型的影响, 优选出预处理方法和识别模型的组合算法。 结果发现, PCA-GBDT、 PCA-SVM(高斯核)、 PCA-SVM(线性核)、 MSC-GBDT、 MSC-SVM(多项式核)这几类组合算法准确率均为95%以上, 能很好的实现番茄早疫病潜育期的光谱识别; 其中MSC-GBDT的识别召回率和准确率最好, 而PCA-SVM(高斯核)识别效率最高。 研究表明, 通过降噪处理后的番茄早疫病潜育期高光谱数据减少了噪声、 更加符合真实的分布、 具有较大的可信数据量, 配合简单的识别模型会导致识别能力不足, 而配合复杂的识别模型可达到一个较可靠的测试结果; 通过降维算法能使番茄早疫病潜育期高光谱数据的维度降低、 数据量减少; 降维后的特征能够表达出病变信息, 配合简单识别模型时识别效果好, 而配合过于复杂的识别模型会导致识别模型的过拟合。

Tomato early blight is highly infectious and destructive. The detection and identification of pre-symptom characteristics in the incubation period is the key to Tomato Early Blight monitoring, early warning and scientific control. In this paper, the evolution of early tomato blight was monitored by hyperspectral images, and the data were analyzed combined with visible light images and spectral characteristics. The results showed that the average value of near-infrared spectrum and red edge reflectance of tomatoes infected with early blight decreased with time, and the disease information of the incubation period appeared 36 hours after inoculation. This paper selected the spectral data of 36 h inoculation as the modeling data of Tomato Early Blight incubation period. The principal component (PCA) transformation and multivariate scattering correction (MSC) were used to reduce the spectral dimension or noise of the modeling data. Then the gradient lifting decision tree (GBDT) and support vector machine (SVM) recognition models were established, and the data were imported for training and recognition. The influence of PCA and MSC preprocessing methods on the recognition effect of gradient lifting decision tree (GBDT) and support vector machine (SVM) models is discussed. The influence of common kernel functions on SVM recognition models is further discussed, and the combination algorithm of preprocessing method and recognition model is optimized. The results showed that the accuracy of PCA-GBDT, PCA-SVM (Gaussian kernel), PCA-SVM (linear kernel), MSC-GBDT and MSC-SVM (polynomial kernel) was more than 95%, which could well realize the spectral recognition of Tomato Early Blight incubation period; Among them, MSC-GBDT has the best recognition recall and accuracy, while PCA-SVM (Gaussian kernel) has the highest recognition efficiency. The research shows that the hyperspectral data of the Tomato Early Blight incubation period after noise reduction reduces the noise is more in line with the real distribution, and has a large amount of data. The recognition ability will be insufficient, while combined with a complex recognition model, a higher test result can be achieved; The dimension reduction algorithm can reduce the dimension and amount of hyperspectral data in the incubation period of early tomato blight, and the features after dimension reduction can express the lesion information. When combined with a simple recognition model, the recognition effect is good, while with an overly complex recognition model, it will lead to over fitting the recognition model.