Semi-Supervised Scene Classification Of Remote Sensing Images Based On GAN

Remote sensing image scene classification is based on the content of remote sensing images,which is automatically recognized and classified by specific high-level semantic labels.It is widely used in smart city construction,transportation and tourism planning,natural disaster monitoring and national defense resource management.Due to the progress of earth observation technology,the number of massive unlabeled remote sensing images is increasing rapidly.How to accurately and effectively classify remote sensing images becomes particularly important.However,the annotation of remote sensing images requires extensive engineering skills and expertise.Therefore,how to use a small number of labeled remote sensing images and a large number of unlabeled remote sensing images for scene classification has become a research hotspot.Generative adversarial network is the most promising semi-supervised learning method in recent years,and its application in the field of remote sensing can solve the problem that a large number of unlabeled remote sensing images cannot be effectively utilized.In summary,based on the related models of generative adversarial networks,this thesis is committed to enhancing the stability of generative adversarial training,improving the ability of the discriminator to extract robust features and the generalization adaptability of the model,thereby improving the performance of scene classification.The main research contents are as follows:1.Aiming at the problems of unstable generative adversarial training,unlabeled data cannot be used in supervised scene classification algorithm and the weak feature extraction ability of traditional generative adversarial network,a spectral residual gated attention generative adversarial network was proposed based on semi-supervised theory.Firstly,a spectral normalized residual block is introduced to replace the two-dimensional convolution of discriminator to enhance the stability of generative adversarial training.Secondly,a multi-branch feature fusion module is introduced to fuse low-level features,high-level features extracted by spectral normalized residual block and features extracted by external network Inception V3 to reduce feature loss.Finally,an attentional module combined with gating is added to the discriminator of generating adversarial network to enhance the representation ability of the model.The proposed method is validated on Euro SAT and UC Merced datasets,and the results show that the proposed method can effectively extract features with strong discrimination and improve the performance of semi-supervised classification.2.Aiming at the problems of few unlabeled samples in remote sensing images and the loss of local feature information caused by traditional region discarding algorithms,a balanced data enhancement classification method based on Cut Mix is proposed.The basic theory and sample balance strategy of Cutmix are introduced in detail.Through the visual analysis of the feature heat map and the comparative experiment of the UC Merced dataset,it is shown that Cut Mix can effectively improve the effect of data augmentation and improve the generalization ability of the model.3.Design and implement a remote sensing image scene classification system based on B/S structure,including system requirement analysis,overall architecture design,browser design,server design,network model design and database design,etc.The test shows that the system has certain convenience and usability.