基于生成对抗网络的高光谱影像解译算法

Hyperspectral image interpretation based on generative adversarial network

高光谱遥感能够获取蕴含着丰富地表覆盖信息的高光谱影像,在国民经济建设和国防军事信息化领域都有着广阔且深远的发展潜力。高光谱影像解译是其应用的核心技术之一。近年来,高光谱遥感系统不断提高的光谱分辨率和不断增加的成像设备量化深度使得获取的影像越来越精细,但同时也给解译任务带来诸多挑战。为解决基于深度学习的解译方法在小样本条件下容易过拟合及泛化能力差的问题,本文基于生成对抗网络,设计了一种用于高光谱影像解译的深度卷积生成对抗网络模型。该网络去除了全连接的隐藏层和池化层,其中,生成器采取微步卷积作为上采样策略,以随机噪声和样本的类别标签作为输入得到伪样本。判别器和解译器均采取步长卷积作为下采样策略,以生成的伪样本和真实的训练样本作为输入,以判别真伪和类别标签。训练中根据多解译损失优化参数,相较传统生成对抗网络,其能够更合理地优化损失函数,通过对抗性训练模型可学习到高光谱影像数据集的真实概率分布,并可根据学习到的代表性空谱特征对高光谱影像进行解译,解译精度高。

Hyperspectral remote sensing can obtain hyperspectral images with abundant surface coverage information,and it has broad and far-reaching potential in the fields of national economic construction and national defense and military information.As one of the core technologies,hyperspectral image interpretation is facing great challenges with the improvement of spectral resolution and the refinement of image owing to the increase of quantization of imaging equipment.Considering that the interpretation method based on deep learning is prone to overfitting and the generalization ability is poor with small samples,a deep convolutional generation confrontation network model is designed for hyperspectral image interpretation based on the generation confrontation network,which completely removes the connected hidden layer and pooling layer.The generator adopts the micro-step convolution as ihe upsampling strategy,and takes random noise and category label of the sample as the input to obtain pseudo samples.The discriminator and the interpreter adopt step-size convolution as the down-sampling strategy,and use the generated pseudo samples and real training samples as the input to distinguish authenticity and category labels.The parameters are optimized based on multiple interpretation losses in training.Compared with the traditional generative confrontation network,the loss function can be optimized in a more reasonable way.Through the confrontation training model,the true probability distribution of the hyperspectral image data set can be obtained,and the hyperspectral images can he accurately interpreted according to the acquired representative spatial spectmm feature.