| Remote sensing image change detection refers to the process of using a certain technology to compare and analyze two remote sensing images taken at different times in the same area and identify the changes of ground objects in the area.The change detection results can provide basis for land and resources management,disaster monitoring,urban planning,vegetation change and so on.With the improvement of remote sensing image acquisition technology and the increase of acquisition channels,remote sensing image data is becoming more and more abundant.In practical application,various fields put forward higher requirements for the accuracy and efficiency of remote sensing image change detection.With the rapid development of deep learning technology in recent years,its superior performance in the field of image processing provides a new research direction for remote sensing image change detection.Therefore,this paper carries out the research on remote sensing image change detection based on deep learning,improves UNET and RESNET,and applies them to solve the problem of remote sensing image change detection.The main work includes:(1)Aiming at the problems of low detection accuracy and inaccurate target boundary segmentation when UNET depth convolution neural network is used for remote sensing image change detection,an AS-UNET model is proposed and used for remote sensing image change detection.The AS-UNET model replaces part of the convolution layer and pooling layer of the encoder in the UNET model with the Shufflenet V2 module,which reduces the information loss caused by the down sampling process;The feature map generated by the encoder is input into the spatial cavity convolution pool pyramid module,and the features extracted under different receptive fields are fused to generate multi-scale feature map,which fully extracts the information of ground objects with different sizes in remote sensing images;The double attention mechanism is introduced into the level communication path of encoder and decoder to enhance the attention of the model to changing pixels and reduce the detection error caused by data sample imbalance.The experimental results show that compared with UNET model and other change detection models,the accuracy of AS-UNET model in remote sensing image change detection is significantly improved,and it is verified that AS-UNET model has certain generalization on other data sets.(2)Aiming at the problems of small target detection,false detection or missing detection when RESNET deep convolution neural network is used in remote sensing image change detection,a twin semantic segmentation model SF-ResNet is proposed and applied to remote sensing image change detection.In order to avoid the loss of information caused by the channel merging process,the encoder part of SF-ResNet network is composed of two networks with shared parameters,which are used to extract the features of dual phase images respectively;Because the traditional semantic segmentation network usually obtains the change map through the top-level network,ignoring the details of the feature map of other layers,it is not conducive to the feature extraction of small targets.Therefore,a feature pyramid network(FPN)is proposed to fuse the shallow and deep features,so that the shallow feature map also has rich semantic information and improve the accuracy of the model for small target change detection.The test results show that the change detection accuracy of SF-ResNet model is higher than that of traditional twin network model,and has certain generalization. |
