Research On Remote Sensing Image Scene Classification Based On Generative Adversarial Networks

Remote sensing image classification has important application requirements and is widely used in natural disaster detection,land resource utilization and coverage management.The method based on deep feature learning does not require a considerable amount of engineering skill and domain expertise.Deep learning features are learned automatically and deep learning models that are composed of multiple processing layers can learn more powerful feature representations of data.So the researchers pay more and more attention to the method based on deep feature learning.The generative adversarial network(GAN)is the most potential deep learning method in recent years,and it is a new idea to introduce it into remote sensing image classification.Using the appropriate data augmentation method and improving the feature extraction ability of the discriminator are the key techniques and difficulties.Based on the GAN theory,this thesis proposes a more generalized classification model for remote sensing image scene classification.The main contents include:The limitations of the existing remote sensing image datasets and the shortcomings of the classification methods are analyzed.It is pointed out that the datasets is small and it is difficult to extract discriminative representations due to its own texture features.Through the analysis of the existing GAN,the limitation of unsupervised feature extraction is presented.And the instability of the training process will affect the classification results.The characteristics of semi-supervised learning in machine learning are analyzed.It is pointed out that semi-supervised learning is suitable for GAN.In combination with semi-supervised learning and GAN,the original GAN which needs unsupervised feature extraction is improved by making full use of label information and generated data.The classification model of GAN based on semi-supervised feature extraction is established.Compared with the existing methods of supervised feature extraction and unsupervised feature extraction,semi-supervised learning can improve the ability of network feature extraction and the classification results.Through the analysis of the problems existing in original GAN,it is pointed out that the original GAN has the problems of unstable training and insufficient diversity of the generated images,and the fundamental reason lies in the unreasonable design of the loss function.Wasserstein distance and gradient penalty are introduced into WGAN-GP to measure the distance between the generated distribution and the real distribution.On the basis of WGAN-GP,the classification model based on WGAN-GP is established.Compared with the original GAN,using a more reasonable loss function can improve the quality and diversity of the generated images,and the training results tend to be stable.By analyzing the empirical risk minimization theory and the vicinal risk minimization theory,the thesis points out the overfitting problem of the former in the case of insufficient data.And the superiority of the mixup data enhancement method based on the vicinal risk minimization theory is presented.A semi-supervised feature extraction network based on mixup data augmentation is proposed.This method can effectively augment the data,stabilize the training process and improve the classification results.