为了实现对茶叶病害的准确预测, 避免病害特征提取过程中对茶叶的二次破坏, 利用荧光透射技术对茶叶赤叶病叶片的荧光透射光谱特性展开研究。 实验采集了健康茶叶叶片样本45个、 赤叶病初期叶片样本60个及赤叶病中期叶片样本60个, 并按照2∶1的比例划分成训练集和预测集样本数, 通过荧光透射手段利用高光谱仪器采集这些叶片的原始荧光透射光谱。 通过对这3组叶片样本平均光谱强度曲线的分析, 证实了利用荧光透射光谱信息对这3种病害类型叶片进行分类的可行性。 然后使用多项式平滑(savitzky-golay, S-G)方法对原始光谱进行平滑和降噪处理。 最后采用竞争性自适应重加权抽样法(competitive adaptive reweighted sampling, CARS)对预处理后的光谱数据进行特征波长的选取。 经过50次加权采样后, 最终选取出4个特征波长, 分别为: 463, 512, 586和613 nm。 为了最大化提取样本的病害特征信息, 强化分类器输入病害特征值的典型性, 使用高光谱反射技术, 采集4个特征波长下的高光谱图像, 分别使用2种不同的纹理提取算法提取病害叶片图像的纹理信息进行对比分析。 首先利用灰度共生矩阵(GLCM)提取4幅图像的纹理信息, 分别计算4个方向的灰度共生矩阵(0°, 45°, 90°及135°), 然后计算5个共生矩阵的均值和方差。 为了提高鲁棒性, 取4幅图像纹理信息的平均值作为该叶片的纹理特征值, 最终得到10个特征值。 利用LBP(local binary patterns)算法获取特征波长下高光谱图像的纹理信息, 并使用Uniform模式对LBP模型进行降维, 最终每幅图像得到944个维度的LBP特征值, 同样取4幅图像的平均值作为该叶片的LBP纹理特征值。 最后通过极限学习机(ELM)分别建立特征光谱联合灰度共生矩阵纹理信息及LBP算子纹理信息的预测模型, 由于模型的输入特征值不在一个量纲, 首先对输入特征值进行归一化处理, 然后再定义模型的输出标签, 即健康叶片的预测模型输出为1, 赤叶病早期为2, 中期为3。 实验测得基于CARS-GLCM-ELM模型的预测准确率为81.82%, 基于CARS-LBP-ELM模型的预测准确率为85.45%, 说明利用荧光透射光谱联合LBP算子纹理信息预测效果更好。 由于没有达到预期效果, 利用Softplus函数对ELM的隐含层激活函数进行了优化, 替换掉原来的Sigmod函数, 优化后的模型预测分类正确率达到92.73%, 基本达到了预期效果。 该研究将病害叶片的荧光光谱信息和对应特征波长下高光谱图像的纹理信息进行了融合, 研究结果可为茶叶病害的快速、 准确预测提供一定的参考价值。

In order to realize accurate prediction of tea disease and avoid secondary damage in the process of disease feature extraction, the fluorescence transmission technology was used to study the spectrum characteristics of tea red leaf disease. The total of 45 samples of healthy tea leaves, 60 samples of early stage of red leaf disease and 60 samples of intermediate stage of red leaf disease were collected in the experiment, and were divided to training set and prediction set according to the proportion of 2∶1 for each kind. The original fluorescence transmission spectra of these leaves were collected using hyperspectral instrument by fluorescence transmission. Through the analysis of average spectral intensity curves of the three groups of leaves, the feasibility of using fluorescence transmission spectral information to classify the three types of leaves was confirmed. Then the polynomial smoothing (Savitzky-Golay, S-G) method was carried out for smoothing and noise reduction on the original spectral. Finally, competitive adaptive reweighted sampling (CARS) algorithm was used to select the characteristic wavelengths of the preprocessed spectral data. After 50 weighted samples, 4 characteristic wavelengths were selected finally, which were 463, 512, 586 and 613 nm respectively. In order to maximize the disease feature information of the samples and strengthen the typification of the classifier input value of disease feature, hyperspectral images were collected on 4 characteristic wavelengths respectively. Gray level co-occurrence matrix (GLCM) algorithm was used to extract image texture information, and 0°, 45°, 90°and 135° direction of the four gray level co-occurrence matrix were calculated. Then, the mean value and square error of the five symbiotic matrices were calculated, and the average value of the four image texture information was taken as the texture feature value of the leaf in order to enhance the recklessness. Finally, 10 feature values were obtained. The LBP (Local binary patterns) algorithm was used to extract the texture information from spectral image, and the uniform mode was used to reduce the dimension of LBP mode. Eventually, 944 dimension characteristic values of LBP were got from each image, similarly, the average value of the four images was taken as the characteristic value of LBP texture. Finally, the LBP eigenvalues of 944 dimensions were obtained for each image, and the average value of 4 images was also taken as the LBP texture feature value of the leaf. Finally, the prediction model was established under characteristic spectrum associated with the gray level co-occurrence matrix and the LBP operator respectively by using the extreme learning machine (ELM). As the input eigenvalues of the model were not in the same dimension, the input eigenvalues were normalized firstly, and then the output labels of the model were defined, that is, the output of the prediction model of healthy leaves was 1, the early stage of red leaf disease was 2, and the intermediate stage of red leaf disease was 3. The prediction accuracy based on CARS-GLCM-ELM model was 81.82%, and the prediction accuracy of CARS-LBP-ELM model was 85.45%. It showed that the effect of combining fluorescence transmission spectrum with LBP operator texture information was better. Due to the undesired results, the hidden layer activation function in ELM was optimized by using Softplus function instead of Sigmod function. The prediction accuracy of the optimized model was 92.73%. In this study, the fluorescence spectrum information of diseased leaves and texture information of hyperspectral images at corresponding characteristic wavelengths were fused, and the results can provide some reference for rapid and accurate prediction of tea diseases.