新东方在线官网Z S T U
加速度学习网英语Zhejiang Sci-Tech University
硕 士 学 位 论 文
一辈子的英文Master’s Thesis
中文论文题目:
基于高光谱成像技术的大豆花叶病早期 检测与分级方法研究 英文论文题目: Early Detection and Grading Method of Soybean
Mosaic Dia Bad on Hyperspectral
Imaging Technology
学科专业： 信号与信息处理macao>relativeto
作者姓名： 吴子娴
指导教师： 桂江生
完成日期： 2018年12月13日
摘要
drown大豆作为一种有着悠久种植历史的农作物，不仅含有丰富的营养物质，还具有抵抗胆固醇的功效。目前，由于轮作周期不断减少，病虫害预防不及时致使大豆产量降低。其中大豆花叶病(SMV)在我国乃至全球大豆种植区均有出现，严重影响大豆产量。因此，大豆花叶病的病情诊断监测显得尤为重要。当前大豆病害检测的研究主要集中在大豆发病中后期，为了实现对大豆花叶病的及时预防，本文基于高光谱成像技术，对大豆花叶病早期检测以及严重程度分级方法进行研究。主要研究内容如下所述：
(1) 为减轻花叶病对大豆产量的影响，实现对大豆花叶病害早期诊断预警，提出了一种基于SPA2-ELM的大豆花叶病早期检测方法。对接种SC3，SC7大豆花叶病毒七天后的大豆叶片以及正常叶片进行高光谱图像采集，分别采用SG平滑处理和分段多元散射校正对光谱数据进行预处理。采用连续投影算法(SPA)选取9个特征波长SPA1(405、461、522、552、626.6、705、743.4、855、947nm)。为了进一步地减少计算量，再次对其进行SPA处理得到4个最佳特征波长SPA2(461、552、705、855nm)，分别采用人工神经网络(ANN)、最小二乘支持向量机(LSSVM)、极限学习机(ELM)模型建立基于全光谱数据和特征波长提取的大豆花叶病早期诊断模型。实验表明，采用SG平滑处理后的模型效果优于采用分段多元散射校正预处理的模型，采用SPA2-ELM模型在去除数据冗余的前提下保持了良好的精度。模型训练集精度达到89.59%，测试集精度达到87.5%。
(2) 为进一步提高大豆花叶病早期诊断模型的精度，提出了一种基于CNN模型的大豆花叶病早期检测方法。其中模型的卷积层为两层，卷积层中加入ReLu非线性激活函数，池化层中使用Max pooling池化函数。对比采用极限学习机、最小二乘支持向量机模型，无论是不同种类大豆样本集的识别率，还是总体数据集的识别率，CNN模型的识别率都比LSSVM和ELM的高，其模型训练集识别率为94.79%，测试集识别率达到92.08%。
(3) 为了更好的实现对大豆花叶病的生长监测，提出了一种基于CNN-SVM模型大豆花叶病分级检测方法。将正常生长以及不同患病程度的大豆叶片分为0级、1级、2级，将卷积神经网络模型的全连接层接入支持向量机，从而解决小样本带来的误差，对比单独采用卷积神经网络模型，检测结果更为准确。其中训练集准确率达到96.67%，测试集准确率达到94.17%。证明了用卷积神经网络与支持向量机相结合对大豆花叶病分级检测的可行性，并为基于高光谱图像的大豆花叶病害检测提供了一种新的方向。
关键字：大豆；高光谱成像技术；花叶病；极限学习机；卷积神经网络计算机专业美国留学
Early Detection and Grading Method of Soybean Mosaic Dia Bad
on Hyperspectral Imaging Technology
ABSTRACT
As a crop with a long planting history, soybean not only has rich nutrients, but also has the effect of resisting cholesterol. At prent, soybean yield is reduced due to the decrea of rotation cycle and the inadequate prevention of dias and inct pests. Soybean mosaic virus (SMV) dia has appeared around the world. That has affected soybean yield riously. Therefore, it is particularly important to realize the detection of soybean mosaic dias. The current rearch on soybean dia detection mainly focus on the middle and late stages of soybean dia, in order to achieve timely prevention of soybean mosaic dia, this paper studied the non-destructive detection method for the initial diagnosis of soybean mosaic dia and the verity of soybean mosaic dia bad on hyperspectral imaging technology. The main rearch contents are as follows:
(1) In order to reduce the impact of mosaic dia on soybean production and explore a theoretical basis for detection and warning of early soybean mosaic dia, this paper provided a method of early detection of soybean mosaic dia bad on SPA2-ELM. Hyperspectral image acquisition of soybean leaves and normal leaves after inoculation of SC3, SC7 soybean mosaic virus for ven days. The spectral was preprocesd by SG smoothing and PMSC. Then, the feature wavelengths ware performed by successive projection algorithm (SPA). There were 9 characteristic wavelengths(
奥巴马最后一次演讲405、461、522、552、626.6、705、743.4、855、947nm) Further, in order to reduce the amount of calculation, SPA processing was performed again to obtain 4 optimal characteristic wavelengths(461、552、705、855nm). Three models of extreme learning machine (ELM), artificial neural network (ANN), least squares support vector machine (LSSVM) classification algorithm were established bad on full-band information and characteristic wavelength information for initial diagnosis of soybean mosaic dia. Experiments showed that the model with SG smoothing was better than the model with piecewi multi-scatter correction preprocessing. The SPA2-ELM model maintained good accuracy under the premi of removing data redundancy. The accuracy of the model training t reached 89.59%. The prediction t accuracy was 87.5%.
gnu
(2) To further improve the accuracy of the initial diagnosis model of soybean mosaic dia, an early detection method of soybean mosaic dia bad on CNN model was propod. The convolutional layer of the model was two layers, the ReLu nonlinear activation function was added
subi
