基于光谱-空间一致性正则化的高光谱图像分类

Hyperspectral Image Classification With Spectral-spatial Consistency Regularization

近年来高光谱遥感技术迅速发展,高光谱图像分类是遥感领域中的热点研究方向。传统的光谱-空间分类框架,将光谱特征提取与空间特征提取分开进行,忽略了二者之间的相关性,导致分类精度不佳。文中提出基于光谱-空间一致性正则化的高光谱图像分类方法,建立长短期记忆神经网络(LSTM)和八度卷积(Octave Convolution)两个支路,分别提取光谱特征和空间特征,应用一致性正则化项建立光谱、空间与光谱-空间三个支路之间的显式相互作用,约束不同的支路对同一像素具有相同的分类结果,最后利用光谱-空间特征用于分类。在Pavia University数据集上证明了本文提出的高光谱图像分类方法的有效性。

With the rapid development of hyperspectral remote sensing technology in recent years,hyperspectral image classification has become a hot research direction in remote sensing. The traditional spectral-spatial classification framework separates spectral feature extraction from spatial feature extraction,and ignores the relationship between them. This leads to classification accuracy is not particularly good. In this paper,a hyperspectral image classification method based on spectral-spatial consistency regularization is proposed. LSTM branch and Octave Convolution branch are used to extract spectral features and spatial features respectively. At the same time,the consistent regularization term is used to establish the explicit interaction among the three branches of spectral,spatial and spectral-spatial. It is applied to constrain different branches to have the same classification result for the same pixel. Finally,the extracted spectral features and spatial features are integrated for classification. Experiments conducted on the Pavia University dataset show the superiority of the proposed method.