激光技术, 2020, 44(4): 485, 网络出版: 2020-04-01

3维卷积递归神经网络的高光谱图像分类方法

Hyperspectral image classification based on 3-D convolutional recurrent neural network
作者单位

空军航空大学,长春 130022

摘要
为了针对高光谱图像中空间信息与光谱信息的不同特性进行特征提取,提出一种3维卷积递归神经网络(3-D-CRNN)的高光谱图像分类方法。首先采用3维卷积神经网络提取目标像元的局部空间特征信息,然后利用双向循环神经网络对融合了局部空间信息的光谱数据进行训练,提取空谱联合特征,最后使用Softmax损失函数训练分类器实现分类。3-D-CRNN模型无需对高光谱图像进行复杂的预处理和后处理,可以实现端到端的训练,并且能够充分提取空间与光谱数据中的语义信息。结果表明,与其它基于深度学习的分类方法相比,本文中的方法在Pavia University与Indian Pines数据集上分别取得了99.94%和98.81%的总体分类精度,有效地提高了高光谱图像的分类精度与分类效果。该方法对高光谱图像的特征提取具有一定的启发意义。
Abstract
In order to extract the features of spatial information and spectral information in hyperspectral image, a 3-D convolutional recursive neural network (3-D-CRNN) hyperspectral image classification method was proposed. Firstly, 3-D convolutional neural network was used to extract local spatial feature information of target pixel, then bidirectional circular neural network was used to train spectral data fused with local spatial information, and joint features of spatial spectrum were extracted. Finally, Softmax loss function was used to train classifier to realize classification. The 3-D-CRNN model did not require complex pre-processing and post-processing of hyperspectral image, which can realize end-to-end training and fully extract semantic information in spatial and spectral data. Experimental results show that compared with other deep learning-based classification methods, the overall classification accuracy of the method in this paper is 99.94% and 98.81% respectively in Pavia University and Indian Pines data set, effectively improving the classification accuracy and efficiency of hyperspectral image. This method has some enlightening significance for feature extraction of hyperspectral image.
00 11