Change Detection For Polarimetric SAR Images Based On Contourlet DBN

The change detection of Polarimetric Synthetic Aperture Radar(SAR)images is to detect the changed regions of two polarimetric SAR images acquired at different times in the same area.The polarimetric SAR images have been applied widely in various fields such as military,agriculture and urban planning,so the research on polarimetric SAR images has important practical significance.The traditional polarization SAR image change detection methods are usually based on pixel level,which may result in the low accuracy of the change detection results easily.This paper makes use of the spatial information and scale information between the pixels,and combines those information with the advantages of deep learning to extract features automatically to achieve the final change detection result.The main works of this paper are as follows: 1.A method of polarimetric SAR image change detection based on scattering characteristics and Supervised Deep Belief Network(SDBN)is proposed.This method first uses the Wishart distance between two polarimetric SAR images for clustering to achieve initial change detection result,then selects the labeled training samples according to the initial change detection results,and finally trains the SDBN by the labeled training samples to obtain the final change detection result.This method makes use of the characteristics of complex Wishart distribution of polarimetric SAR image data,which can improve the detection accuracy of the initial change detection result map effectively,so that we can select more reasonable and effective training samples.At the same time,the initial change detection result map eliminates the process of labeling images manually.The advantage of the SDBN is that it introduces the supervised mechanism based on the DBN to improve the accuracy of the final change detection result.In addition,the deep learning algorithm implements change detection,which can form a universal unsupervised change system to solve the problems of difficult acquisition of difference maps and thresholds.The method is tested on multiple different datasets,and it can perform well on all of these datasets.2.A method of polarimetric SAR images change detection based on NSCT_SDBN model is proposed from the perspective of model optimization based on the first work.This method mainly combines the non-subsampled contourlet transform with SDBN to construct NSCT layer,which is used to extract the multi-scale features of polarimetric SAR images to improve the performance of the neural network model as well as the detection accuracy and edge sharpness of the change detection result.Experiments are performed on multiple different datasets,and the results show that the proposed method achieves better results of change detection.3.A polarimetric SAR images change detection method based on selective ensemble NSCT_SDBN is proposed.The method firstly selects a small number of training samples according to the label and then uses the selected training samples to train the NSCT_SDBN by the means of selective ensemble learning to obtain the result of the change detection.Selecting a small number of samples can reduce the workload of labeling image manually,and avoid the dependence on the initial change detection result.The training subset obtained from the self-sampling process in ensemble learning can fit the distribution of real data well.At the same time,the proposed selective mechanism can ensemble models with good performance to improve the generalization performance of the network and the change detection accuracy.Experiments have shown that the selective ensemble learning performs more effectively under the training situations with a small number of samples.