高光谱影像是典型的高维数据,在光谱维和空间维都包含了大量信息。针对高光谱影像分类时光谱维数据量巨大的特点,提出一种基于三维空洞卷积残差神经网络的高光谱影像分类方法。该方法以高光谱像元立方体作为数据输入,使用三维卷积核同时提取高光谱数据的空间维和光谱维特征,并通过在卷积核中引入空洞结构,在不增加网络参数量和不消减数据特征的情况下提高卷积核的感受野,从而提高神经网络的分类的精度。该方法利用残差结构避免了由网络层数加深导致的梯度消失问题,最终使用Softmax分类器完成高光谱像元的分类工作。实验结果表明:所提方法在Indian Pines和Salinas数据集上分别取得了97.303%和97.236%的总体分类精度,与各对照组相比具有更好的分类效果,由此证明所提方法可以提升高光谱影像的分类性能。

Hyperspectral image is typical high-dimensional data which contains abundant information in both spectral and spatial dimensions. In this paper, a hyperspectral image classification method based on three-dimensional dilated convolutional residual neural network is proposed for characterizing large amounts of data in the spectral dimension during hyperspectral image classification. In this method, hyperspectral pixel cubes were applied as input data. Further, a three-dimensional convolutional kernel was used to simultaneously extract the spectral and spatial characteristics of hyperspectral data. Then the receptive field of the convolutional kernel was enhanced without adding network parameters or reducing data features by introducing a dilated structure in the convolutional kernel. Thus, the classification accuracy of the neural network was improved, avoiding the problem of gradient disappearance caused by the deepening of network layers using a residual structure. Finally, the Softmax classifier was used to complete the classification of hyperspectral pixels. Results show that this method obtained an overall classification accuracy of 97.303% and 97.236% on the Indian Pines and Salinas datasets, respectively, exhibiting a better classification effect than other control groups. Thus, the proposed method can improve the classification performance of hyperspectral images.