光学学报, 2020, 40(21): 2128002, 网络出版: 2020-11-01

基于注意力机制的多目标优化高光谱波段选择

Multi-Objective Optimization of Hyperspectral Band Selection Based on Attention Mechanism
作者单位

1中国人民解放军空军航空大学航空作战勤务学院, 吉林 长春 130022

2东北师范大学地理科学学院, 吉林 长春 130024

摘要
神经网络的注意力机制可以从数据中提取关键信息,将这一特性运用在高光谱波段选择上有助于充分学习波段之间的相互依赖和非线性关系,提取更重要的波段。提出了一种基于注意力机制的多目标优化高光谱波段选择算法。首先,利用注意力模块和自编码器构建网络;然后,将一维光谱数据作为网络输入,采用两种损失函数并结合多目标优化方法对输入数据进行训练,使嵌入在网络中的注意力模块充分学习各波段之间的非线性关系,对信息量大和易于分类的波段赋予更大的权重,以实现波段选择;最后,利用支持向量机分类器和平均光谱散度验证波段子集的性能。实验结果表明:相比于其他算法,所提算法在Botswana与Indian Pines数据集上提取的波段子集的分类精度更高,信息量更大,由此证明了所提算法对高光谱波段选择的有效性。
Abstract
The attention mechanism of neural networks can extract key information from data, and the application of this feature in the selection of hyperspectral bands can help fully learn the interdependence and nonlinear relations between bands and extract more important bands. This paper presents a multi-objective optimization method for hyperspectral band selection based on the attention mechanism. First, the attention module and autoencoder are used to construct the network. Then, one-dimensional spectral data is provided as input to the network; two loss functions are used and combined with the multi-objective optimization method for training. Therefore, the attention module embedded in the network learns the nonlinear relationship between different bands and assigns more weight to the bands with a large amount of information and easy classification, thereby realizing band selection. Finally, the support vector machine classifier and mean spectral divergence are used to validate the performance of the band subset. The experimental results show that the band subset extracted using this method from the Botswana and Indian Pines datasets is more accurate and informative than the subsets extracted using other algorithms. Thus, it is demonstrated that this algorithm is more effective in selecting hyperspectral bands.
00 11