Remote Sensing Image Scene Classification Based On Generative Adversarial Networks

With the development of deep learning,more and more classification algorithms based on supervised learning are introduced into the real remote sensing image scene classification.However,this kind of scene classification method based on supervised learning needs a large number of labeled samples,but it is difficult to get enough labeled samples directly in practical applications.How to solve the problem of lack of labeled samples in remote sensing image effectively,and to propose a more generalized classification method is a main topic of this study.The purpose of this thesis is to introduce the generation adversarial networks into the research of remote sensing image classification model,and to explore the perspective and adaptability of generation adversarial networks for remote sensing image scene classification based on supervised learning and unsupervised representation learning.More advanced deep learning methods and improved theoretical models are used to improve the feature extraction ability and data augmentation ability of the generative adversarial networks,as well as the problem of instability in the training process.The main contents and conclusions of this thesis are as follows:First,in order to solve the problem of gradient vanishing on the basis of deepening the network structure,the residual module is introduced in this thesis.By changing the generator and discriminator to the residual network,the network structure can be deepened so that the generator can generate high-quality false sample images and improve the fitting ability of the model.Finally,the auxiliary classifier is used to realize supervised learning classification,and generate false sample images according to the tag to achieve the purpose of data augmentation.Second,aiming at the problems of feature extraction difficulty and poor robustness of sample generation in unsupervised representational learning classification methods,a new unsupervised representation learning model based on WGAN with weighted penalty term(WGAN-GP)is proposed in this thesis.This model optimizes the objective function of the original generation adversarial networks and adds a multi-feature fusion layer behind the discriminator.This layer not only provides the feature information for the classifier,but also feeds back the gradient signal of feature matching for the generator,so that the generator can generate the false sample images close to the real sample images.Multi-layer perceptron classifier(MLP classifier)which is composed of all connected networks is used to realize the classification function.Through experimental analysis and verification,the two types of classification methods based on supervised learning and unsupervised representation learning proposed in this thesis can achieve the task of remote sensing image scene classification.Compared with other deep learning methods,they have more powerful sample expansion ability,and provide a way of thinking for the semi supervised learning based remote sensing image scene classification method.