基于3D Octave卷积和胶囊网络的高光谱图像分类

Hyperspectral image classification based on 3D Octave convolution and capsule networks

针对当前胶囊网络分类模型在高光谱图像分类中存在空谱联合信息利用不足和收敛较慢的问题,提出一种结合多尺度Octave三维卷积和胶囊网络的分类模型。首先,使用主成分分析(PCA)来降低高维的光谱特征并保留其关键特征;其次,通过多尺度Octave三维卷积模块使模型能够适应当前数据集目标尺寸跨度较大的特点,在减少空间冗余的同时提高高光谱图像的空谱联合信息的利用;最后,对动态路由算法进行改进,使用向量长度的相似性与方向的相似性来衡量两个向量的一致程度,从而解决网络在训练过程中收敛较慢的问题。为了验证改进后模型的有效性,选择Pavia University公开高光谱数据集,并且通过OA、AA以及Kappa系数将分类结果与当前主流的分类模型进行对比实验。结果表明,在Pavia University数据集上,改进后的模型在OA、AA以及Kappa系数上的精度要高于其他模型,因此,该模型能够更好地应用于高光谱遥感图像分类任务中。

A classification model combining multi-scale Octave 3D convolution and capsule network is proposed to solve the problems of insufficient utilization of space-spectral joint information and slow convergence of the current capsule network classification model in hyperspectral image classification. The principal component analysis(PCA) is used to reduce the highdimensional spectral features and retain their key features. The multi-scale Octave 3D convolution module is adopted to make the model adapt to the characteristic of current dataset with a large target size span,which can reduce the spatial redundancy and improve the utilization of the space-spectral joint information of the hyperspectral image. The dynamic routing algorithm is improved,and the similarity of the vector length and the similarity of the direction are used to measure the degree of consistency between the two vectors,so as to solve the problem of the slow network converges in the process of network training. In an experiment,the public hyperspectral dataset of Pavia University was selected,and the classification results of the improved model were compared with those of the current mainstream classification models by means of OA(overall accuracy),AA(average accuracy) and Kappa coefficients to verify the effectiveness of the improved model. The results show that,on the dataset of Pavia University,the accuracy of the improved model is higher than that of other models in terms of OA,AA and Kappa coefficients. Therefore,this model can be better applied to the hyperspectral remote sensing image classification.