基于光谱-空间残差网络模型的高光谱遥感图像分类  

Hyperspectral Remote Sensing Image Classification Based on the Spectral-spatial Residual Network Model

作　　者：韦春桃[1] 肖博林 李倩倩 白风 卢志豪 WEI Chuntao;XIAO Bolin;LI Qianqian;BAI Feng;LU Zhihao(School of Civil Engineering,Chongqing Jiaotong University,Chongqing 400074,China)

机构地区：[1]重庆交通大学土木工程学院,重庆400074

出　　处：《地理信息世界》2020年第3期42-48,共7页Geomatics World

基　　金：重庆市基础科学与前沿技术研究专项重点项目(cstc2015jcyjBX0023)资助。

摘　　要：残差网络是近几年提出的一种新型深度卷积网络,通过增加网络深度提高分类的准确率,也解决了网络退化问题。基于残差学习原理,设计了针对高光谱遥感图像分类的光谱-空间残差网络模型。首先,将原始高光谱遥感数据三维立方体输入网络模型,并使用特定的卷积核对光谱特征进行降维;然后,利用光谱残差模块和空间残差按模块分别且连续地学习光谱和空间特征;最后,对提取到的特征进行池化操作并分类。此外,为规范训练数据和防止过拟合,学习过程中使用了批量归一化和dropout的方法。所设计网络模型在Indian Pines和Pavia U数据集上进行了验证实验,结果表明,所提方法有效地缓解了网络退化的问题,且在分类精度上也高于支持向量机、卷积神经网络等现有算法。The residual network is a new kind of deep convolutional networks proposed in recent years.It improves the accuracy of classification by increasing the depth of the network and also solves the problem of network degradation.Based on the principle of residual learning,this paper designs a spectral-spatial residual network model for hyperspectral remote sensing image classification.First,we input the 3D cube of original hyperspectral remote sensing data to the network model.Then we degrade the spectral features by a specific convolution kernel.After that,the spectral residual module and the spatial residual module are used to learn the spectral and spatial features respectively and continuously.Finally,the extracted features are pooled and classified.In order to standardize the training data and prevent overfitting,batch normalization and dropout methods are used in the learning process.The designed network model is validated by the Indian Pines and Pavia U datasets.Results show that the proposed method can effectively mitigate the problem of network degradation and is better than existing algorithms,such as support vector machines,convolutional neural network in terms of classification accuracy.

关 键 词：高光谱遥感图像分类 残差网络模型 特征提取 批量归一化 DROPOUT 

分 类 号：P237[天文地球—摄影测量与遥感]