High Resolution Remote Sensing Image Change Detection Based On Convolutional Neural Network

With the rapid development of space and aviation technology,the ability and quality of earth observation system to acquire space data are constantly improving,and satellite remote sensing technology has entered the sub-meter era.High-resolution remote sensing images contain more complex shapes,textures,structures and spatial information,so people have higher requirements for interpretation accuracy while obtaining high-precision remote sensing images.As an important part of remote sensing image processing,high resolution remote sensing image change detection has important applications in environmental monitoring,geographical database update,disaster relief and urban planning.In recent years,deep learning has been widely applied in various fields by virtue of its excellent regression performance.It breaks through the constraints of traditional algorithms and has a strong generalization ability.The application of deep learning theory to high resolution remote sensing image change detection is a hotspot in remote sensing image processing.This paper firstly points out the research background and significance of the change detection of high-resolution remote sensing image,and expounds the current research status at home and abroad,analyzes the research basis of this topic--convolutional neural network,and introduces the idea of semantic segmentation of convolutional neural network into the change detection of remote sensing image.According to the characteristics of ASPP in extracting the context information of different sensory fields,and combined with the advantages of extraction and fusion of the multi-scale features of Inception structure,ASPP and Inception structure are integrated into the Unet network model,and ASPP Inception-Unet is proposed.Unet code path models in each stage of the first convolution layer replacement for Inception structure,strengthened the feature extraction of code paths and context information ability,on to sample before,will contain the characteristics of high-level semantic information input to ASPP module,used to obtain the characteristics of the different scale,through the above improvements,improve network change detection performance.According to the residual structure feature extraction ability,combining with the characteristics of FPN multi-scale prediction,the residual structure and FPN into Unet model,establish FPN Res-Unet model,the model on the basis of the Unet,introducing Resnet18 residual structure as feature extraction,in the process of expansion path sampling in each level,expand the path to FPN fusion model in the network backbone,fully residual structure,Unet and characteristics of the pyramid network advantages of mutual confluence,It enhances the feature extraction ability of the network and makes up for the lack of semantic segmentation network's ability to detect small targets and suppress noises,which makes the network pay attention to details while acquiring highlevel semantic information and improves the accuracy of change detection.Based on the multi-temporal image of GF-2,a set of GF-2 building change detection data set was made.At the same time,another 3 sets of heterogeneous open source change detection data sets were used to jointly serve as the experimental data in this paper.Will ASPP Inception-Unet,FPN Res-Unet,FCN and Unet in change detection experiment with four sets of data set respectively,the results show that ASPP Inception-Unet and FPN Res-compare Unet FCN and Unet shows better performance,the change detection task average F1 value increased by 3.7%,5.5% and 5.5%,7.3%,and has better generalization ability and application value,to the land resource management,geographic database updates,such as urban planning provides a more efficient method.