A Research On Deep En-decoding Structure Based High-resolution Remote Sensing Image Change Detection

With the continuous development of remote sensing image platform and technology,especially the implementation of various high-resolution surface observation projects,satellite remote sensing technology has gradually entered the sub-meter era,also the ground feature information contained in high-resolution remote sensing images is also more sufficient.In recent years,with the continuous maturity of deep learning and big data technology,this has greatly promoted the prosperity and development of remote sensing image technology in various fields,making the use of high-resolution remote sensing images to detect changes in land surface coverage as an efficient and accurate technology means,and have been important applications in many military/civilian fields,such as: agriculture and forestry,water and soil resources,land use,urban expansion,geological environment,coastal areas,marine resources,snow and glaciers,disaster monitoring and mitigation,infrastructure development as well as military use research,etc.Because remote sensing images are extremely susceptible to considerable factors in the imaging process,such as the limited hardware level of the imaging platform,the space/time constraints when acquiring images,the occlusion of weather clouds and smoke,and the existence of various electromagnetic clutter interference,etc.It may cause problems such as low spatial resolution of remote sensing images and the blurred outline of objects in the images.Traditional remote sensing image change detection algorithms have limited performance when processing such remote sensing image data,especially the current amount of remote sensing image data is also increasing.The method based on deep learning breaks through many shortcomings of traditional image processing algorithms,and is widely used in the field of computer vision with its excellent feature learning and representation capabilities,providing a new solution for the research of remote sensing image change detection.This paper is mainly based on the framework of deep learning theory,focusing on high-resolution remote sensing image change detection problems,and building different deep neural network models to flexibly realize problem-driven change detection tasks.The main research content and summary of this article are as follows:1.Combined with the characteristics of the change detection task of high-resolution remote sensing images,and based on the data analysis and model structure,a change detection method based on the deep en-decoding structure built by full convolutional network is proposed.The method is suitable for the change detection task of large-scale remote sensing images.And considering that the remote sensing image usually contains the image content of large scenes,the dilated convolution module is introduced to reduce the number of subsampling,and the details and positioning information in the feature map are retained while enhancing the network receptive field,so as to alleviate the problem of the loss of small targets and details information caused by excessive pooling.2.Aiming at the problem that the segmentation edge of remote sensing image change detection results is not fine enough,a change detection method based on attention mechanism is proposed.In this method,the remote context information in the feature is captured effectively by introducing the visual attention mechanism,and the similar features in the feature graph are aggregated to enhance the feature representation of the changing region,which can help the network to better detect and segment the potential changing region,and improve the precision of edge segmentation of the changing region.At the same time,in order to make full use of the training data,a variety of online enhancement strategies are used and the diversity of the data is increased to improve the robustness of the network.Finally,considering the large difference in the number of positive and negative samples in the change detection task of remote sensing images,the positive samples were amplified for the original data,and the loss of Dice coefficient was introduced to alleviate the imbalance of positive and negative samples in the data,thus improving the final generalization performance of the model.3.Considering the practical value of building change extraction in urban remote sensing images,by combining the characteristics of urban high-resolution remote sensing images,a change detection model for specific target/region categories based on siamese network structure is designed to extract building changes.And in view of the problem that remote sensing images often contain objects of different sizes,especially small targets,multi-scale feature pooling is introduced to help obtain multi-scale context information of target features,and the effective field of view of targets of different sizes is obtained.Without significantly increasing the amount of network parameters,the effect of detecting changes in the building area in the two high-resolution remote sensing images is effectively improved.