Generative Adversarial Networks Based Attention Mechanism For Change Detection In Remote Sensing Images

With the rapid development of aerospace,there are massive amounts of remote sensing images every day.These data is widely used for city planning,cartography,resources survey,land using,environmental protection and so on.Change detection based remote sensing images is gradually a kind of research field.Recently,deep learning has the advantage of self-learning features compared with traditional change detection models and become a new trend.Change areas are based on texture features,geometric features,spectral features of a pair remote sensing images.So this paper focuses on introducing attention-based generative adversarial network(GAN)model and it can capture the relationship between the whole and the part of images and improve the ability of feature extraction.The main work and results are as follows:(1)Construct samples: This paper uses remote sensing images with 30 m resolution and 8 bands and get 3600 samples that include many kinds of change types by visual interpretation.Based these data,this paper designs a processing and making tool of samples.At the same time,we use data augmentation to enlarge samples and finally get 10800 samples.(2)Design change detection model: For improving accuracy,we propose an attention-based generative adversarial network model for change detection.At first,we use a convolution neural network(CNN)model based on attention mechanism for feature extraction.Attention mechanism makes image key location and features can be highlighted.Then we should combine these feature representation on two images and construct CNN based attention mechanism model for generating change areas.At last,a classification model is constructed in order to distinguish images.(3)Experiment: Dataset is divided by a ratio of 10:1 and this experiment uses change detection model based on U-Net,change detection model based on GAN,change detection model based on attention-based GAN to analysis results.By evaluation criteria of overall accuracy(OA)and Kappa coefficient,the experiment shows attention-based GAN has an improvement of 0.02 and 0.1.