Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (4): 169-175.DOI: 10.3778/j.issn.1002-8331.2001-0150

Previous Articles     Next Articles

3D-CNN with Standard Score Dimensionality Reduction for Hyperspectral Remote Sensing Images Classification

SHE Hailong, XIE Shanjuan, ZOU Jingjie   

  1. Research Academy of Remote Sensing and Earth Sciences, Hangzhou Normal University, Hangzhou 311121, China
  • Online:2021-02-15 Published:2021-02-06

标准分数降维的3D-CNN高光谱遥感图像分类

佘海龙,解山娟,邹静洁   

  1. 杭州师范大学 遥感与地球科学研究院,杭州 311121

Abstract:

Aiming at the problems of Hughes phenomenon and low utilization efficiency of spatial and spectral features in hyperspectral images, a hyperspectral image classification algorithm combining standard score dimensionality reduction and deep learning is proposed. Standard scores are used to evaluate the band quality of hyperspectral data to eliminate redundant bands in hyperspectral remote sensing images. Combined with the optimized 3D Convolutional Neural Network(3D-CNN) classification method, a large stride convolution layer is used to replace the pooling layer, and a series of tricks such as L2 regularization, Batch Normalization(BN) and Dropout are introduced to reduce network parameters and effectively prevent overfitting. Through the experimental tests of two hyperspectral datasets of Pavia Centre and Pavia University, the algorithm greatly reduces the parameters and computational costs of the network model, and achieves classification accuracy of 99.01% and 95.99%.

Key words: convolutional neural network, deep learning, standard scores, hyperspectral remote sensing images, classification

摘要:

针对高光谱图像存在Hughes现象,以及空间和光谱特征利用效率低的问题,提出了一种结合标准分数降维和深度学习的高光谱图像分类算法。利用标准分数对高光谱数据的波段质量进行评价以剔除高光谱遥感图像中的冗余波段,结合优化过的3D-CNN(3D Convolutional Neural Network)分类方法,通过使用大步距卷积层替代池化层,引入L2正则化、批量归一化(Batch Normalization,BN)、Dropout等一系列策略,在减少网络参数的同时有效防止过拟合现象。通过Pavia Centre和Pavia University两个公开高光谱数据集的实验测试,该算法大幅度降低了网络模型的参数和计算量,取得了99.01%和95.99%的分类精度。

关键词: 卷积神经网络, 深度学习, 标准分数, 高光谱遥感图像, 分类