基于TBTA网络的高光谱图像分类

Classification of Hyperspectral Image Based on TBTA Network

近年来,利用深度学习方法对高光谱图像(Hyperspectral Image, HSI)进行分类受到了研究人员和学者的广泛关注。为了充分利用HSI丰富的光谱和空间信息,并考虑到HSI小样本高维度的特点,提出了一种三分支三重注意力机制(Triple-Branch Ternary-Attention, TBTA)网络结构。在TBTA中,基于DenseNet和3D-CNN,设计了三个分支来分别提取高光谱图像中包含的光谱特征、空间x轴方向的特征和空间y轴方向的特征,可以将三种特征分离开。在三个分支中分别引入了其特征方向的注意力机制,针对信息丰富的光谱波段设计了光谱注意块,信息丰富的像素点分别设计了空间X和空间Y注意块,使得TBTA能够对提取的特征进行细化。在4个高光谱数据集上进行了实验,并对比了5种算法:SVM、CDCNN、SSRN、FDSSC、DBMA,实验结果表明本文的算法在OA、AA、KAPPA等评价标准均获得了更好的效果,其中TBTA的OA指标比次优的算法平均提高0.1%-2.08%。

In recent years, the classification of Hyperspectral Image(HSI) using deep learning methods has received a lot of attention from researchers and scholars. In order to make full use of the rich spectral and spatial information of HSI and consider the characteristics of small samples and high dimensionality of HSI,a Triple-Branch Ternary-Attention(TBTA) network structure is proposed in this paper. In TBTA,firstly, based on DenseNet and 3D-CNN,three branches are designed in TBTA to obtain the spectral features, spatial X-axis features and spatial Y-axis features contained in the hyperspectral image, respectively, which can separate the three features. Secondly, attention mechanism is applied in these three branches, and spectral attention blocks are designed for information-rich spectral bands, and spatial X-axis and spatial Y-axis attention blocks are designed for information-rich pixel points, which enable TBTA to refine the extracted features. In this paper, experiments were conducted on four hyperspectral datasets and five algorithms were compared: SVM,CDCNN,SSRN,FDSSC,and DBMA. The experimental results show that the algorithms in this paper obtained better results in OA,AA,KAPPA and other evaluation criteria, in which the OA index of TBTA was improved by 0.1%-2.08% on average over the suboptimal algorithms.