Research On The Change Detection Method Of High Resolution Remote Sensing Image Combining Pixel Level And Object Level

With the rapid development of aerospace science and technology,high-resolution remote sensing images have become an important data source for earth observation.The results of urban expansion,natural disaster assessment,vegetation monitoring and analysis can be obtained rapidly by using change detection technology.However,the high-resolution remote sensing image contains a large amount of ground feature information,at the same time,the acquisition of high-resolution remote sensing images will be affected by factors such as illumination,angle,and registration,which improves the difficulty of processing and feature extraction of high-resolution remote sensing images,resulting in more challenges for the change detection technology method of high-resolution remote sensing image.In the existing high-resolution remote sensing image change detection methods,there is salt and pepper phenomenon in the pixel-level change detection results,while the object-level change detection method based on image segmentation has the problem that the image segmentation scale is difficult to determine.Therefore,we propose a method that combines pixel-level change detection with object-level change detection.The results of the two methods are fused to obtain the change area of the target change area,so as to improve the accuracy of change detection.The main research contents of this article are as follows:(1)U-net network model structure improvement.In order to make use of the advantages of deep learning in understanding image features,we use U-net network to carry out object-level image segmentation for high-resolution remote sensing images.There are some problems still existing in the original U-net network such as over fitting and slow model training,so it is improved.The method in this paper improves the U-net model structure and adds a batch standardization layer to the U-net network structure at first.Then transfer learning is introduced to initialize the improved U-net model with the parameters of VGG16 network,and Dropout method is added to avoid overfitting.Finally,the cross-entropy loss function used by the original U-net network is improved and the Focal loss function is used instead.Comparative experiments show that the method proposed in this paper effectively improves the semantic segmentation effect of high-resolution remote sensing images.Higher segmentation accuracy is obtained,and it is more stable during model training.Research on high-resolution remote sensing image change detection method combining pixel-level and object-level change detection.In order to solve the problem that only using a single type of change detection method can not make full use of remote sensing image features,resulting in low accuracy of change detection results,we propose a change detection model combining the advantages of pixel-level and object-level.In this method,multi-dimensional features of high-resolution remote sensing images are comprehensively used.First,random forest classifier is used to classify the combined features of two-phase remote sensing images,and the binary classification results of pixel-level change detection are obtained.Then,the improved U-net network is used to conduct semantic segmentation of the post-temporal remote sensing images to obtain the classification results of image objects.Finally,the results of the above two methods are combined with statistical principles to obtain the area of change on the ground object category.The experimental results show that the method presented in this paper has high accuracy of change detection.