The reflectivity characteristic wavebands were explored to identify different early blight disease stages of tomato leaves based on feature ranking (FR).Spectral reflectance information at the wavelengths of 345.75-1 042.25 nm were acquired for healthy and infected samples at both early stage and advanced stage.The spectral reflectance was considered as x variable, and the dependent variables of healthy, early stage and advanced stage samples were set as 0, 1 and 2, respectively.The Naive Bayes (NB) classification models were established to identify the different stages of diseased samples.The wavelengths at the beginning and end were cut due to the noise, resulting in spectral reflectance at 479.69~920.38 nm were used.The corresponding results were 85.71%, 90.91% and 100% for 400.09~1 000.08 nm, 78.57%, 63.64% and 81.82% for 479.69~920.38 nm as well as 92.86%, 63.64% and 63.64% for selected wavebands (658.73, 654.19, 642.33 and 689.46 nm).Although the recognition ability of FR-NB classification model decreased, fewer input variables simplified the model and improved the computational efficiency, which provides a basis for the development of the multispectral detection sensor.%利用特征排序(feature ranking,FR)提取反射率特征波段识别轻度和严重染病的番茄早疫病样本.依次采集样本在345.75~1 042.25 nm波长范围内的反射率信息,将光谱反射率作为x变量,健康、轻度和严重染病样本的因变量设为0、1和2,建立朴素贝叶斯(Naive Bayes,NB)分类模型,识别不同病害程度的样本.由于全波段首尾段含有噪声,切除噪声后进一步研究479.69~920.38 nm波段范围内的光谱反射率信息.在基于400.09~1 000.08 nm波段范围的分类模型(NB)中,验证集识别率分别为85.71%、90.91%和100%;在479.69~920.38 nm得到的验证集识别率分别为78.57%、63.64%和81.82%;在基于特征波段(658.73、654.19、642.33、689.46 nm)的分类模型(FR-NB)中,验证集识别率分别为92.86%、63.64%和63.64%.结果表明,基于光谱反射率特性可识别番茄叶片早疫病害,虽然FR-NB分类模型的识别效果有所降低,但较少的输入变量简化了模型,提高了计算效率,为病害检测多光谱传感器的开发提供了依据.
展开▼
