POLSAR Image Change Detection Based On Curvelet DSN

The technology of Pol SAR image change detection is based on two Pol SAR images from different phases in the same area,it is to detect changed regions in the area.This technology is a key one for monitoring military intelligence,detecting natural disasters and so on.Traditional technologies of Pol SAR image change detection have some disadvantages: they need to extract features manually,so the processing procedure is cumbersome,and makes the amount of calculation large;they also have strict requirements for the resolution,polarization characteristics,distribution characteristics etc.of Pol SAR image data.In order to overcome these shortcomings,this paper proposes a new method for Pol SAR image change detection based on deep learning.It takes full use of the outstanding capabilities of deep learning to extract features and the high resolution and polarization characteristics of Pol SAR images to deal with change detection problems.The main works of this paper are as follows:(1)It proposes a Pol SAR images change detection method based on deep stacking network(DSN).This algorithm overcomes the shortages of traditional methods,such as tedious calculation and the high requirements for Pol SAR image.It also takes use of the capabilities of deep learning to extract features.The main idea of this algorithm is to take change detection as a binary classification problem and build a DSN,then use the two preprocessed Pol SAR images from different phases in the same area to train the DSN,finally we can classify the changed and unchanged samples and obtain the final change detection result.(2)It proposes a Pol SAR images change detection method based on curvelet deep convolution stacking network(DCSN).This algorithm takes advantage of curvelet to extract contour features and filter the images,it also combines the ability of feature extracting of convolution neural network with the special stacking structure of DSN,so that it can enhance the performance of the network as well as improve the accuracy of change detection.This algorithm uses curvelet to decompose the Pol SAR images to get coefficients,and applies the inverse curvelet transform to some coefficients to reconstruct the images and get the training data set.Then it replaces the full connection layer of DSN with convolution layer and builds a binary classification DCSN.Finally the DCSN is trained by the training data set to classify the samples to get the final change detection results.(3)It proposes a Pol SAR images change detection method based on SOM_Kmeans and Wishart-DCSN.This algorithm introduces the idea of unsupervised clustering and solves the problems of lacking of real Pol SAR image labels.It also uses the characteristic that Pol SAR image obeys Wishart distribution,and uses Wishart distance to choose samples.This algorithm first applies SOM_Kmeans to the difference map of Pol SAR images to get a pseudo label,and chooses initial samples according to the pseudo label,then the Wishart distances of the selected samples from the two Pol SAR images are calculated,and precise samples are chose by the distances.Then a binary classification DCSN is trained by the precise samples to classify the Pol SAR images and get the final change detection results.