Remote Sensing Change Detection Based On Siamese Network And Attention Mechanism

Satellite remote sensing technology is capable of obtaining the information on the earth surface periodically and extracting the dynamic changes of the earth surface rapidly,and is widely utilized in the field of land resources survey,urban building extraction,disaster prevention and mitigation,etc.How to identify the land surface changes accurately has become one of the essential researches in remote sensing.In recent years,with the development of artificial intelligence theories,deep learning algorithms have made many breakthroughs in the application of remote sensing change detection.Deep learning is of great ability on feature mining and feature representation,and can extract the effective characteristics of remote sensing images.Compared with traditional change detection methods,the deep learning algorithm can better capture complex surface conditions and achieve higher change detection accuracy.However,the issues of remote sensing data redundancy and the diversity of scale features of ground objects are still the crucial barrier of the application and development of deep learning-based change detection algorithm.This paper summarized the problems existing in the current deep learning change detection methods and explored the optimization of the structure of deep neural network.In view of the above issues,two change detection models based on Siamese network and attention mechanism are proposed towards two different change detection labelling methods on pixel level and image level,respectively.The main research contents of this paper are as follows:(1)For pixel-level labelled samples,a point-to-point change detection method has been proposed,which is called Siamese difference-learning attention network(SDANet).First of all,Siamese structure based on convolutional neural network(CNN)is constructed for changed feature learning.To enhance the representation of change information,the difference learning module is explored.A spatio-channel self-attention mechanism is then proposed considering the correlation between each pixel and channel in the image,which enables the network to capture the internal relationship on both spatial and channel dimensions and reinforces the perception ability towards the image characteristics.The result demonstrates the superiority of the proposed method on change detection accuracy compared with other comparison methods.(2)For image-level samples,an image-to-image change detection method has been proposed,which is called Siamese high resolution network(Siam?HRNet).To handle with the loss of spatial information caused by multiple successive down-sampling operations in current fully convolutional networks,the multi-resolution parallel structure was introduced and the image information with different resolutions is comprehensively employed without any spatial information loss caused by multiple down-sampling.The result shows that Siam?HRNet can extract the feature of two periods images accurately and achieve high change detection accuracy.(3)To tackle the flaw that the existing attention mechanism can not emphasize the difference information in image-to-image change detection,a novel change detection algorithm named high-resolution-feature difference attention network(HDANet)has been proposed.Moreover,atrous spatial pyramid pooling(ASPP)has been formed by combining a group of convolution kernels in parallel with different dilation rates to capture the multi-scale features in the same image.The result proved that HDANet could assign more attention weight to the regions with high change magnitude and is more sensitive to the difference information between two images,and the detection results outperformed other count parts in terms of the integrity and edge details of ground objects.