基于注意力机制辅助的空谱联合残差网络的高光谱图像分类

Spatial-spectral residual network with attention mechanism for hyperspectral image classification

摘要:

卷积神经网络（convolutional neural network，CNN）是高光谱图像分类中一种常用的方法，有着较好的分类表现。然而，CNN不可避免地会提取出一些冗余特征，这对高光谱图像分类的准确率造成干扰。此外，高光谱图像分类还面临着同谱异物、同物异谱问题。为了解决以上这些问题，提出了一种基于注意力机制辅助空谱联合残差网络的高光谱图像分类方法。一方面，通过使用注意力机制辅助的3-D、2-D残差网络，同时从光谱维度和空间维度提取空谱联合特征，克服同谱异物、同物异谱问题；另一方面，引入通道注意力机制和空间注意力机制，有效降低了冗余空谱特征的干扰。在2种高光谱数据集上的实验结果表明，相比同类对比算法，所提出的方法具有更优越的分类性能。

Abstract:

Recently, the convolutional neural network (CNN) has attracted increasing attention in hyperspectral image classification. The CNN is a useful method due to its satisfactory classification performance. However, the useless features extracted by the CNN inevitably have bad influence on the classification. Besides, the phenomena of inter-and intra-class spectral variability has posed a challenge to hyperspectral classification tasks. In order to overcome the above problems, a novel spatial-spectral residual network with attention mechanism was proposed in this study. The spatial-spectral features were effectively extracted from the spatial domain and spectral domain by the attention mechanism-aided 3-D and 2-D residual networks to overcome the problem of inter-and intra-class spectral variability. Besides, the channel attention module and spatial attention module were used to reduce the interference of the useless features on the classification. Experimental results over two hyperspectral datasets indicate that the proposed method achieves superior classification performance over its counterparts.