| Remote sensing images are used in many fields,such as monitoring the coverage of land,forest,grassland and wetland,using satellite maps generated by remote sensing information to guide production and life,and easily obtaining geographic information to guide disaster relief in case of disasters.Therefore,the related research of remote sensing image processing has become very important.At present,as a basic and important part of computer vision processing,semantic segmentation plays an important role in many applications of remote sensing image.Most of the mainstream image semantic segmentation models are based on deep learning.In order to get better segmentation results,these models need a lot of labeled data for training.However,remote sensing image itself has the characteristics of large image,large quantity,changing and growing with time,and various data types.Therefore,the annotation of remote sensing image needs a lot of manpower and material resources,so how to effectively train the semantic segmentation model is a key problem in the case of remote sensing image with only a small amount of annotation or no annotation.In order to solve this problem,researchers have proposed many methods,such as semi supervised semantic segmentation and transfer learning.As a sub direction of transfer learning,unsupervised domain adaptation can train the model using labeled and unlabeled data when the image data set does not meet the same distribution condition.For example,simultaneous interpreting of two image datasets is caused by different sensors or different regions.And only one dataset is marked.By using existing annotated source domain images and unlabeled target domain images,we can train a good semantic segmentation model in the target domain.This method is called unsupervised domain adaptive semantic segmentation.Based on the above ideas,this thesis uses unsupervised domain adaptation method to solve the semantic segmentation problem of remote sensing image.Since the adversary learning method can achieve data distribution alignment through the confrontation between discriminator and generator,it can be used in unsupervised domain to adapt to semantic segmentation and effectively reduce domain offset.This thesis focuses on the research of domain adaptation algorithm of remote sensing image semantic segmentation based on adversarial learning,including the following two aspects:（1）A domain adaptive semantic segmentation method of remote sensing image based on entropy enhanced confrontation learning is proposed.Aiming at the problem that the existing domain adaptive semantic segmentation methods are difficult to predict without discrimination,an entropy enhancement method is proposed to combat loss.Aiming at the problem of class error correspondence in domain adaptation,a cooperative training classifier is introduced.At the same time,in order to make full use of the multi-scale information of remote sensing image and reduce the domain offset caused by the resolution of sensors between domains,the classifier adopts the atrous spatial pyramid pooling module.Compared with the existing remote sensing image domain model which needs multi-step training to adapt to the research of semantic segmentation,the model of this method can directly carry out end-to-end training,and the experimental results show that the model can achieve good segmentation effect in the target domain.（2）A domain adaptive semantic segmentation method for attention enhanced remote sensing image is proposed.In order to solve the problem that convolutional network model can’t effectively utilize long-range dependence information,two attention modules are added in the segmentation part of domain adaptation model.The position attention module can effectively use the long-range dependence information of remote sensing image,and the channel attention module can obtain the relationship between feature map channels.The experimental results show that the two attention modules can improve the performance of remote sensing image domain adaptive semantic segmentation.Through experiments,the optimal combination of the two attention modules in the network for domain adaptive semantic segmentation task is determined. |
