Research Of Remote Sensing Image Classification Methods Based On Generative Adversarial Network

In recent years,the classification method of remote sensing image based on neural network has become an important research direction of remote sensing image classification.At the same time,more and more research proves that the Generative Adversarial Network(GAN)has great potential in image classification and feature extraction.Because of the existence of the “adversarial” mechanism,GAN can carry out feature learning and feature representation more quickly and effectively than other methods.In this thesis,inspired by GAN model,we propose two novel models for solving two kinds of remote sensing image classification problems.1.To propose Enhancing Pix2Pix(named ePix2Pix)based on Pix2 Pix for RGB remote sensing image classification.In this model,the classification of remote sensing images is considered as an image generation or image segmentationproblem,and improvements are made based on the state-of-the-art model.On the basis of the original model structure,we add one Controller.The function of the controller is to reconstruct the input of the model according to the output of the generator.The reconstruction error is added to the optimization target of the model,which can effectively improve the classification performance and stability of the model.2.To propose a Discriminative GAN(named DGAN)classification model for hyperspectral remote sensing image classification.In order to solve the phenomenon of Different Objects with Same Spectrum(DOSS)and Same Object with Different Spectrum(SODS),we put forward the concepts of IntraclassPair,Interclass Pair and True-fakePair.Through the training process of the whole network,DGAN learns the characteristic similarity of the IntraclassPair,the characteristic distance of the Interclass Pair and the characteristic distance of the True-fakePair.Taking advantage of GAN's adversarial characteristics and advantages in feature extraction,the similarity between Interclass Pair becomes larger and the similarity of IntraclassPair becomes smallerin the process of model training.So the classification performance of DGAN is better thanother models.DGAN learns the relationship between feature pairs,rather than just feature mapping.A series of comparative experiments are done to verify the superiority of our models.The classification accuracy of ePix2 Pix on multiple datasets is more than 3%higher than that of Pix2 Pix,and the classification accuracy of DGAN on multiple datasets is more than 2% higher than that of other existing methods.