利用残差生成对抗网络的高光谱图像分类

Hyperspectral Image Classification Based on Residual Generative Adversarial Network

针对高光谱图像分类过程中存在的标记样本需求量大和分类精度要求高等问题,提出了一种利用残差生成对抗网络(GAN)的高光谱图像分类方法。该方法以生成对抗网络为基础,使用包含上采样层和卷积层构成的8层残差网络替换生成器的反卷积层网络结构,提高数据的生成能力,使用34层残差卷积网络替换判别器的卷积层网络结构,提高特征提取能力。以Pavia University、Salinas及Indian Pines数据集为实验数据,将所提方法与GAN、CAE-SVM、2DCNN、3DCNN、ResNet进行了比较。实验结果表明,所提方法在总体分类精度、平均分类精度和Kappa系数上均有显著提高,其中总体分类精度在Indian Pines数据集上达到了98.84%,较对比方法分别提高了2.99个百分点、22.03个百分点、12.91个百分点、4.99个百分点、1.79个百分点。所提方法在网络中加入残差结构,增强了浅层网络与深层网络的信息交流,可提取高光谱图像的深层次特征,提高了高光谱图像分类的精度。

A hyperspectral image classification method based on residual generative adversarial network(GAN)is proposed to address the problems of high demand for labeled samples and high classification accuracy in the process of hyperspectral image classification.The method is based on GAN and includes:replacing the deconvolution layer network structure of the generator with an eight-layer residual network composed of an upsampling layer and a convolution layer to improve data generation ability;improving feature extraction ability,the discriminator’s convolutional layer network structure is replaced with a thirty-four-layer residual convolutional network.The experiment compares the datasets from Indian Pines,Pavia University,and Salinas.The proposed method is compared to GAN,CAE-SVM,2DCNN,3DCNN,and ResNet.The results demonstrate that the proposed method improves overall classification accuracy,average classification accuracy,and Kappa coefficient significantly.Among them,the overall classification accuracy reached 98.84%on the Indian Pines dataset,which is 2.99 percentage points,22.03 percentage points,12.91 percentage points,4.99 percentage points,and 1.79 percentage points higher than the comparison methods.In summary,adding a residual structure to the network improves information exchange between the shallow and deep networks,extracts deep features of the hyperspectral image,and improves hyperspectral image classification accuracy.