Semi-Supervised Remote Sensing Image Scene Classification Based On Generative Adversarial Network

With the progress of science and technology,the demand of remote sensing image scene is increasing gradually.As a basic image processing method,remote sensing image scene classification has a very important application value.It has been widely used in natural disaster detection,resource survey,land resource utilization,urban management and other fields.In the past few years,people have proposed various methods to classify remote sensing images,but most of these methods are based on manual classification methods,which require a large number of professionals,and the classification process is time-consuming and labor-intensive.At present,deep learning has attracted the attention of many researchers in the fields of machine learning and pattern recognition.More and more scholars apply deep learning to remote sensing image scene classification to obtain better classification effect.However,the existing remote sensing image data sets are small,the amount of labeled data is small.Moreover,the data sets of remote sensing images are characterized by large differences among images of the same class and small differences among images of different classes,which leads to difficulties in feature extraction and inability to obtain higher classification.Therefore,it is the main task of this study to effectively solve the problem of insufficient remote sensing label samples and improve the classification accuracy of remote sensing images.Generative Adversarial Network(Generative Admittedly Network,GAN)is one of the most promising deep learning methods for unsupervised learning in complex distribution in recent years.It is a new idea to introduce GAN into the field of remote sensing image classification.In the case of limited labeled samples,the GAN model can generate a large number of unlabeled samples.Then an excellent semi-supervised network architecture can be obtained by combining a limited number of labeled samples with a large number of unlabeled samples in the classification model.Therefore,based on the GAN theory,this paper proposes a semi-supervised model with more generalization ability for remote sensing scene classification to improve the classification accuracy.The research of this paper mainly includes the following contents:(1)In order to solve the problems of supervised classification requiring a large number of labeled samples and low accuracy of unsupervised classification,the discriminant of deep convolution generation adversation model(Deep Convolution Generation Adversation Network,DCGAN)was changed from binary classifier to multiple classifier,and the improved model was called cDCGAN.Then a semi-supervised classification model(Classification based on Semi-Supervised Learning,CSSL)based on remote sensing images was designed by combining cDCGAN and deep convolution classifier(VggNet-16).The semi-supervised classification of labeled samples and unlabeled samples was performed by multi-training method.In this way,a large number of unlabeled remote sensing data sets are generated and made full use of to achieve the effect of improving the classification accuracy,and solve the problem of using a large number of labeled data for supervised classification.(2)To solve the problem of poor image quality generated by cDCGAN model in CSSL,which affects the classification accuracy,this paper analyzes the reasons and proposes and use Selu activation function and batch normalization to improve the generator of cDCGAN.Compared with the CSSL model,the improved model can retain more image information,so it can extract more image features,make the generated image closer to the real sample,and further improve the classification accuracy of remote sensing image.In order to objectively evaluate the superiority of the proposed semi-supervised classification method and the improved classification model,the semi-supervised method and the improved method were verified in the NWPU-Resisc45 data set and UC-Merced data set respectively.Experimental results show that the proposed method can not only generate unlabeled remote sensing images to solve the shortage of original label image sets,but it also achieves better classification accuracy.The improved semi-supervised classification method based on the original model further improves the quality and classification accuracy of the generated remote sensing images.