Remote sensing image land classification applications based on deep learning require massive data sets as training samples,and the image label data sets are often difficult to meet the training requirements due to the small number.Using existing samples to increase is an effective technical method.The traditional data augmen-tation technology only changes the color and sharpness of the image,and the amount of augmentation has a cer-tain limit.In order to automate the augmentation of more diverse samples,a remote sensing image sample aug-mentation method based on Pix2pix network is designed in this paper.Pix2pix network generator is used to gen-erate virtual images according to unmanned aerial vehicle and Google image tags,and the discriminator com-pares the virtual images with the real images.After generating adversarial training for many times,the sample pairs are output to achieve augmentation.The results show that the visual contrast similarity of the generated re-sults is high and the average cosine similarity of the unmanned aerial vehicle image and Google image is 0.85 and 0.96,respectively,and the average histogram similarity is 0.50 and 0.61.It is an effective method for re-mote sensing image sample augmentation.
关键词
样本扩增/Pix2pix网络/深度学习/遥感影像
Key words
Data augmentation/Pix2pix network/Deep learning/Remote sensing image
维普
万方数据