Research On Remote Sensing Image Classification Based On Adversarial Unsupervised Domain Adaptation

In recent years,remote sensing image classification has attracted plenty of attention.A type of method called scene classification,determining which scene the image belongs to,has been applied to lots of fields,including land use,forest monitoring,urban planning,and vegetation management.Although deep learning-based scene classification of remote sensing images has gained a lot of attention and become an active research area,these methods require sufficient annotated data to extract effective features.The reality is that the remote sensing images we acquire through sensors are often unannotated.Labeling data is time-consuming and expensive.The specialized personnel is even required in order to obtain high-quality annotated data.At the same time,due to different sensors,illumination,reflectance conditions,and topographic features,when it turns to a new environment,the accuracy of the supervised classification model will drop dramatically.The unsupervised domain adaptation can effectively take advantage of the existing knowledge to overcome the problem of lacking labels and weak generalization.This paper proposes a remote sensing image classification method based on adversarial unsupervised domain adaptation.Firstly,during the adversarial adaptation,the decision boundary of the classifier has the problem of noise.To overcome the problem,we construct the inter-domain sample alignment loss function.In order to construct the loss function,the target domain samples are assigned false labels by kNN algorithm,and the target domain samples with reliable false labels are selected.Then,the similarity matrix is constructed by using the source domain samples and the target domain samples with reliable pseudo labels.The similarity matrix records the similarity between the target domain samples and the source domain samples.With the similarity matrix,we can construct the set of positive and negative sample pairs of target domain samples.With the set of positive and negative sample pairs,we can construct the inter domain sample alignment loss function.Minimizing the inter domain sample alignment loss function can pull the similar samples closer and push away dissimilar samples in the source domain and the target domain during the adaptation process so that the aligned samples are far away from the decision boundary.Secondly,in view of the phenomenon of high intra class difference and low inter class difference in remote sensing image datasets,this paper solves this problem by constructing the intra-domain sample alignment loss function.The intra-domain sample alignment loss function shortens the distance of similar samples in the same domain and widens the distance of different classes of samples in the same domain,so as to alleviate the problem of large intra class differences and small inter class differences in the same domain.In this way,the negative transfer during adaptation can be reduced during the adaptation process.Twelve transfer adaptations are carried out on four remote sensing image data sets:UCM,WHU,AID and RSSCN7.The proposed method achieves 82.96% average accuracy on the twelve transfer adaptations,which verifies the effectiveness of the proposed method.At the same time,ablation experiments are also done to verify the effectiveness of inter domain sample alignment loss function and intra domain sample alignment loss function.