联合空谱信息的高光谱影像深度Transformer网络分类

Classification of hyperspectral images based on deep Transformer network combined with spatial-spectral information

卷积神经网络中的局部卷积运算无法对高光谱影像中的全局语义信息进行充分学习,因此,基于Transformer模型设计了一种新颖的深度网络模型,以进一步提高高光谱影像分类精度。首先,利用主成分分析方法对高光谱影像进行降维处理,并选取像素周围邻域数据作为输入样本,以充分利用影像中的空谱联合信息;然后,利用卷积层将输入样本转换为序列特征向量;最后,利用构建的深度Transformer网络进行分类。Transformer模型中的多头注意力机制能够充分利用丰富的判别性信息。试验表明,与现有卷积神经网络模型相比,文章方法能够获得更为优异的分类性能。

The local convolution operation in convolutional neural networks cannot fully learn the global semantic information in hyperspectral images.Given this,this study designed a novel deep network model based on Transformer in order to further improve the classification precision of hyperspectral images.Firstly,this study reduced the dimensionality of hyperspectral images using the principal component analysis method and selected the neighborhood data around pixels as input samples to fully utilize the spatial-spectral information in the images.Secondly,the convolutional layer was used to transform the input samples into sequential characteristic vectors.Finally,image classification was conducted using the designed deep Transformer network.The multi-head attention mechanism in the Transformer model can make full use of the rich discriminative information.Experimental results show that the method proposed in this study can achieve better classification performance than the existing convolutional neural network model.