基于深度卷积神经网络的水稻穗瘟病检测方法

发布时间：2018-03-04 23:19

本文选题：病害　切入点：模型　出处：《农业工程学报》2017年20期 　论文类型：期刊论文

【摘要】：穗瘟是一种严重影响水稻产量及品质的多发病害,有效地检测穗瘟是水稻病害防治的重要任务。该文提出基于深度卷积神经网络GoogLeNet模型的水稻穗瘟病检测方法,该方法利用Inception基本模块重复堆叠构建主体网络。Inception模块利用多尺度卷积核提取不同尺度穗瘟病斑分布式特征并进行级联融合。GoogLeNet利用其结构深度和宽度,学习复杂噪声高光谱图像的隐高维特征表达,并在统一框架中训练Softmax分类器,实现穗瘟病害预测建模。为验证该研究所提方法的有效性,以1 467株田间采集的穗株为试验对象,采用便携式户外高光谱成像仪Gaia Field-F-V10在自然光照条件下拍摄穗株高光谱图像,并由植保专家根据穗瘟病害描述确定其穗瘟标签。所有高光谱图像-标签数据对构成GoogLeNet模型训练和验证的原始数据集。该文采用随机梯度下降算法(SGD,stochastic gradient descent)优化GoogLeNet模型,提出随机扔弃1个波段图像和随机平移平均谱图像亮度的2种数据增强策略,增加训练数据规模,防止模型过拟合并改善其泛化性能。经测试,验证集上穗瘟病害预测最高准确率为92.0%。试验结果表明,利用GoogLeNet建立的深度卷积模型,可以很好地实现水稻穗瘟病害的精准检测,克服室外自然光条件下利用光谱图像进行病害预测面临的困难,将该类研究往实际生产应用推进一大步。
[Abstract]:Ear blast is one of the most frequent diseases affecting the yield and quality of rice. The effective detection of ear blast is an important task of rice disease control. A method of rice ear blast detection based on deep convolution neural network (GoogLeNet) model is proposed in this paper. In this method, the main body network .Inception module is constructed by repeated stacking of Inception basic modules. The multi-scale convolution kernel is used to extract the distributed features of ear blast disease in different scales, and cascaded fusion. Google LeNet uses its structure depth and width. In order to verify the effectiveness of the proposed method, we studied the hidden high dimension feature expression of complex noise hyperspectral images and trained Softmax classifier in a unified framework to predict and model panicle blast disease. In order to verify the effectiveness of the proposed method, 1 467 panicle plants collected in the field were selected as experimental objects. Gaia Field-F-V10, a portable outdoor hyperspectral imager, was used to photograph the hyperspectral images of panicle plants under natural illumination. According to the description of panicle blast disease, plant protection experts determine the panicle blast label. All hyperspectral image-label data pairs constitute the original data set for training and verification of GoogLeNet model. In this paper, the stochastic gradient descent algorithm is used to optimize the GoogLeNet model. Two kinds of data enhancement strategies are proposed to increase the size of training data, to prevent the model from overfitting and to improve its generalization performance. The experimental results show that the depth convolution model established by GoogLeNet can be used to detect rice panicle blast disease accurately. In order to overcome the difficulties of disease prediction by using spectral images under outdoor natural light conditions, the study of this kind of disease can be advanced to practical application.
【作者单位】：华南理工大学电子与信息学院;华南农业大学工程学院;广东省农业科学院植物保护研究所;
【基金】：广东省科技计划公益研究与能力建设专项(2014A020208112) 现代农业产业技术体系建设专项资金资助(CASRS-01-33)资助
【分类号】：S435.111.41;TP391.41

【相似文献】