基于CBAM-Res-HybridSN的高光谱图像分类研究

A Study of Small Sample Hyperspectral Image Classification Based on CBAM-Res-HybridSN

为了充分利用高光谱图像的“空间-光谱”信息,提高小样本训练数据下的分类精度,文章提出了一种新型卷积注意力(Convolutional Block Attention Module,CBAM)残差单元的混合卷积神经网络(CBAM-Res-HybridSN),以解决高光谱图像小样本分类问题。该模型通过深度可分离卷积层和残差结构来构建深层混合卷积神经网络,在不增加计算机开销的同时,增强对“空间-光谱”鉴别性特征的提取能力;模型还引入了卷积注意力模块,既实现了突出重要特征,同时对冗余和噪声信息也进行了抑制,在小样本数据下提高了分类精度。为了验证方法的有效性,选择雄安新区(马蹄湾村)和DFC2018Houston两组公开高光谱数据集进行了对比试验,当选择标记样本的5%作为训练样本时,分类总体精度分别为99.34%和96.14%。结果表明,所提方法在小样本数据下保证了更高的分类精度。

In order to make full use of the"spatial-spectral"information of hyperspectral images and improve the classification accuracy under small-sample training data.In this paper,a new hybrid convolutional neural network(CBAM-Res-HybridSN)with residual units of Convolutional Block Attention Module(CBAM)is proposed to solve the problem of classifying small samples of hyperspectral images.The model constructs a deep hybrid convolutional neural network with deeply separable convolutional layers and residual structures to enhance the"spatial-spectral"discriminative feature extraction without increasing the computer overhead.The convolutional attention module is also introduced to highlight important features while suppressing redundant and noisy information,thus improving the classification accuracy under small sample data.To verify the effectiveness of the method,two public hyperspectral datasets,Xiong'an New Area(Matiwan Village)and DFC2018 Houston,were selected for comparison experiments,and the overall classification accuracy was 99.34%and 96.14%,respectively,when 5%of the labeled samples were selected as training samples.The results show that the proposed method ensures higher classification accuracy with small sample data.