番茄叶片在感染病害后首先发生的是内在生理反应, 肉眼无法观察到。 叶片从被感染到出现肉眼可见病斑期间, 称为叶片病害潜育期。 为了实现番茄叶片表面未见明显病斑的灰霉病潜育期诊断, 对接种样本进行叶片编码、 跟踪、 采集所有编码叶片样本1~8 d连续高光谱图像数据, 建立番茄叶片样本时序高光谱数据集。 采用跟踪的叶片样本出现肉眼可见病斑前几天同一位置区域的高光谱数据作为潜育期感兴趣区域进行检测分析。 为了建立番茄叶片灰霉病潜育期诊断和不同病斑等级分类模型, 采用基于教学优化算法(TLBO)优化极限学习机(ELM)的分类模型进行建模。 通过TLBO算法优化ELM的输入权值和隐藏层的偏差, 提高模型分类性能。 利用高光谱成像系统在近红外高光谱波段388~1 006 nm波段获取五个等级的感兴趣区域进行数据建模, 共采样213个高光谱数据, 其中, 健康类(56个)、 潜育期类(42个)、 小病斑类(43个)、 大病斑类(39个)和严重类(33个)。 通过对比不同的光谱预处理方法, 采用效果最好的小波滤波变换(DWT)对样本数据中每类数据分别滤波。 DWT滤波后, 在610~840 nm波段间五个等级光谱曲线能区分明显, 共包含91个波长, 波长数量较多。 因此, 采用竞争性自适应重加权抽样法(CARS)对采用DWT预处理后的光谱数据在610~840 nm波段重复3次优选特征波长, 合并去除重复项后得到9个特征波段: 694, 696, 765, 767, 769, 772, 778, 838和840 nm。 最后分别选取全波段FC、 610~840 nm波段、 CARS提取的9个特征波段建立3个分类模型FC-TLBO-ELM, DWT-TLBO-ELM, DWT-CARS-TLBO-ELM进行对比, 其中DWT-CARS-TLBO-ELM检测精确度最高达100%, 潜育期召回率100%, 利用时间最短为0.068 9 s, 表明该模型可以实现番茄灰霉病潜育期高精度诊断和灰霉病病害程度高精度分类, 为番茄灰霉病早期防治、 精准施药提供理论依据。

Tomato leaves in the infection of disease occurred after the first internal physiological reaction, the naked eye can not observe, from the blade infection to the appearance of visible disease spots, for the left disease latent period. In order to achieve the tomato leaf surface did not see obvious disease spots of gray mold latent period diagnosis. This paper is for the inoculation samples for leaf coding, daily tracking, and collection of all encoded leaf sample hyperspectral image data, the establishment of tomato leaf sample sequence hyperspectral data set. Based on the tracked leaf samples, hyperspectral data from the same location area a few days before the appearance of visible spots with the naked eye were used for detection and analysis as latent period data. In order to establish the diagnosis of the latent period of tomato leaf gray mold disease and the classification model of different disease plaque levels, the classification model based on the teaching learning-based optimization algorithm (TLBO) optimization extreme learning machine (ELM) is used to model. The input weight and hidden layer deviation of ELM are optimized by the TLBO algorithm, and the model classification performance is improved. Data modeling was obtained in the near-infrared hyperspectral band 388-1007nm band to obtain 5 levels of interest, and 213 hyperspectral data were sampled, including health (56), latent (42), small disease plaque (43), major disease plaque (39) and severe disease (33). The best-performing wavelet filtering transformations (Discrete Wavelet Transform, DWT) filter each type of data in the sample data separately by comparing different spectral preprocessing methods. After DWT filtering, the five class spectral curves between the 610 and 840 nm bands can be distinguished significantly, containing 91 wavelengths and a larger wavelength. Therefore, competitive adaptive reweighted sampling is used (Competitive Adaptive Reweighted Sampling, CARS) to repeated the preferred feature wavelength 3 times in the 610~840 nm band using DWT pre-treated spectral data and combined to remove duplicates to obtain 9 feature bands: 694, 696, 765, 767, 769, 772, 778, 838 and 840 nm. Finally, three classification models FC-TLBO-ELM, DWT-TLBO-ELM, DWT-CARS-TLBO-ELM, were selected for experimental comparison, in which DWT-CARS-TLBO-ELM detection accuracy was up to 100%, and the potential recall rate was 100%. Using the minimum time of 0.068 9 s, it is shown that the model can realize the high-precision diagnosis and high-precision classification of the disease degree of gray mold disease during the latent breeding period of tomato ash mold, and provide a theoretical basis for the early prevention and treatment of tomato ash mold disease and the precise application of medicine.