基于MSRA初始化卷积神经网络的草地牧草分类研究  

Study on grassland forage classification based on MSRA initialization convolutio n neural network

作　　者：刘一磊 刘江平[1] 赵烜赫 马玉宝[3] 闫伟红[3] 潘新[1] LIU Yi-lei;LIU Jiang-ping;ZHAO Xuan-he;MA Yu-bao;YAN Wei-hong;PAN Xin(College of Computer and Information Engineering,Inner Mongolia Agricultural University,Hohhot 010010,China;Inner Mongolia Institute of electronic information technology,Hohhot 010004,China;Institute of Grassland Research,Chinese Academy of Agricultural Sciences,Hohhot 010020,China)

机构地区：[1]内蒙古农业大学计算机与信息工程学院,呼和浩特010010 [2]内蒙古电子信息职业技术学院,呼和浩特010004 [3]中国农业科学院草原研究所,呼和浩特010020

出　　处：《光电子．激光》2021年第1期88-95,共8页Journal of Optoelectronics·Laser

基　　金：国家自然基金(61962048,61562067);中央级基本科研业务费(1610332020020)资助项目。

摘　　要：草地牧草的分类与识别是草原研究与监测的重要环节,利用高光谱成像技术和卷积神经网络进行牧草种类的识别判断,为实现草地牧草自动分类与数字化治理提供了新的途径。本文提出了基于MSRA初始化卷积神经网络的草地牧草高光谱图像自动识别与分类的方法。主要过程包括图像预处理、裁剪、特征提取和识别分类四个环节,首先预处理采用改进的自适应波段选择法进行波段提取,然后将提取后的数据压缩成新图像进行裁剪,最后进入MSRA初始化卷积神经网络提取特征并进行识别分类。本文针对卷积神经网络的鲁棒性、稳定性和识别率等问题创新性的提出了MSRA初始化方法,通过初始化设置参数和权值,使网络的性能得到提升,提高识别准确率。本文对实地采集的蒙古冰草、老麦芒、紫羊毛草、燕麦、黄花杂交苜蓿、光穗冰草6种牧草进行识别分类,为保证实验的可靠性与准确性,对训练集和测试集进行多次划分及多次交叉验证实验。实验结果表明,本文提出的MSRA初始化卷积神经网络相比于SVM、KNN、2D-CNN等方法,对草地牧草高光谱图像的识别准确率较高,达到96.50%。实验结果证明本方法具有良好的分类性能和可行性,为草地牧草的识别分类提供了新思路。The classification and identification of grassland forage is an important part of Grassland Researchand monitoring.Usi ng hyperspectral imaging technology and convolution neural network to identify f orage species provides a new way to realize automatic classification and digital management of grassland.In this paper,an automatic recognition and classification method of grassland hyperspectral image s based on MSRA initialization convolutional neural network is proposed.The main process includes image preprocessing,clipping,feature extraction and recognition and classification.Firstly,the improved adaptive band selection method is used for band extraction,and then the extracted data is compressed into a new image for clipping.Finally,MSRA initialization convolution neural network is used to extract features and classi fy.In this paper,aiming at the robustness,stability and recognition rate of convolutional neural network,an inn ovative MSRA initialization method is proposed.By setting parameters and weights through initialization,the performance of the network is improved and the recognition accuracy is improved.In order to ensure the reliab ility and accuracy of the experiment,the training set and test set were divided and cross validated many times.The r ecognition accuracy of MSRA initialization convolution neural network is 96.50%,higher than that of SVM、KN N、2D-CNN and other methods.The experimental results demonstrated the good performance and feasibility of the al gorithm,which provides a new idea for grass recognition and classification of grassland.

关 键 词：高光谱图像 识别分类 卷积神经网络 MSRA初始化 

分 类 号：S-3[农业科学]