联合CNN和Gabor滤波器的高光谱影像分类方法  

A hyperspectral image classification method combining CNN and Gabor filter

作　　者：雷彬 秦进春 魏旻 LEI Bin;QIN Jinchun;WEI Min(Army Engineering University,Xuzhou 221000,China;Xi'an Research Institute of Surveying and Mapping,Xi'an 710054,China;State Key Laboratory of Geo-Information Engineering,Xi'an 710054,China;Xi'an Daoda Tianji Information Technology Co.,Ltd.,Xi'an 710054,China)

机构地区：[1]陆军工程大学训练基地,江苏徐州221000 [2]西安测绘研究所,陕西西安710054 [3]地理信息工程国家重点实验室,陕西西安710054 [4]西安道达天际信息技术有限公司,陕西西安710054

出　　处：《测绘科学与工程》2021年第3期35-39,共5页Geomatics Science and Engineering

摘　　要：本文充分利用高光谱影像的空间信息,提出了一种联合卷积神经网络(convolutional neural network,CNN)和Ga-bor特征的高光谱影像分类方法。首先,设计了一个包含四个卷积层的CNN用于高光谱影像分类,该网络以降维后的高光谱局部影像块为输入,因此能够在分类过程中充分考虑邻域空间信息对分类结果的影响。然后,提出利用Gabor滤波器初始化卷积神经网络的卷积核参数,以充分利用Gabor纹理特征的优势来进一步提高分类精度。在两组真实的高光谱影像上进行分类试验,试验结果表明本文设计的CNN能够取得较高的分类精度,且结合Gabor滤波器对卷积核参数进行初始化能够进一步提高分类精度。In order to make full use of the spatial information of hyperspectral images,a hyperspectral image classification method based on convolutional neural network(CNN)and Gabor features is proposed.Firstly,a CNN with four convolution layers is designed for hyperspectral image classification.The network takes the local patches of hyperspectral image as input,so the influence of neighborhood spatial information on the classification results can be fully considered in the classification process.Then the convolution kernel parameters of convolution neural network are initialized by a Gabor filter to utilize the advantages of Gabor texture features to further improve the classification accuracy.Experiments on two real hyperspectral image dataset show that the CNN designed in this paper can achieve higher classification accuracy,and initializing the convolution kernel parameters with a Gabor filter can further improve the classification accuracy.

关 键 词：高光谱影像分类 卷积神经网络 GABOR特征 空间特征 

分 类 号：P231[天文地球—摄影测量与遥感]