Study On Unsupervised SAR Image Change Detection Method Based On Supervised Strategy

In the field of SAR image processing,change detection is a research topic of great practical significance.It distinguishes whether the features have changed by analyzing two images of the same region obtained at different times.At present,all kinds of change detection methods emerge in an endless stream,which can be roughly divided into two categories: supervised methods and unsupervised methods.The supervised method requires some known information of the change features and can obtain relatively high detection accuracy.However,it is difficult to obtain the real information on the ground surface.Therefore,the unsupervised method is more widely applicable in practical applications.Based on the National Natural Science Foundation of China(Pol SAR image classification based on generative adversarial network,No.61771379)and the National High Level Talent Special Support Program(SAR Image Interpretation and Target Identification),this paper attempts to obtain a preliminary detection result through an unsupervised method in which some samples with high confidence are selected as training sets to train a classifier with better performance and finally achieve unsupervised SAR image change detection with supervised strategy.This paper mainly start from this idea and conducts study on the problem of change detection,and the specific work is as follows:1.A multi-level clustering method for SAR image change detection based on superresolution is proposed.This method begins with unsupervised change detection.Aiming at the problem that the spatial resolution of the filtered image is reduced and the detail is partially blurred,super-resolution technology is used to improve image resolution and enhance detail information,thereby improving detection accuracy.Although the clustering algorithm can look for the inherent distribution structure of the data,the noise in the image can lead to a larger pseudo-change region in the change detection result which is obtained by clustering individual pixel points directly.This method constructs difference maps for two images after super-resolution and clusters to generate multiple categories.Based on this,the hierarchical clustering algorithm is used to merge categories,which can effectively mitigate the effect of noise on the results of change detection.2.A detection method for SAR image changes based on Ada Boost-MLP is proposed.This method combines unsupervised change detection methods and multi-layer perceptron to complete unsupervised change detection under supervised strategies.Using our proposed multi-hierarchy clustering method to obtain some high-confidence samples and combine them into a training set.Because there are relatively few samples in the training set,resulting in over-learning of unchanging samples,we have constructed a multi-layer perceptron model that is sensitive to the weights of the samples.Firstly,a multi-layer perceptron is trained according to the training set.The Ada Boost algorithm is used to adjust the sample weight according to the classification error rate of the current classifier,and then the second classifier is trained.Finally,all the classifiers are combined to output the change detection result.3.A novel SAR image change detection method based on multi-scale skip convolutional network is proposed.Based on CNN,the method has realized the unsupervised change detection.The inputs of the convolutional layer and the pooling layer in the network model are composed of the original data and all the feature maps extracted before.The feature reuse is achieved through the jump connection.This makes the results of change detection not only rely on abstract features extracted by high-level,but also effectively utilized shallow structural information such as edges.In addition,by using different convolutional kernels to observe data from different scales,a richer feature representation of the input image can be obtained.The hopping connection of pooling layer downsamples the outputs of different convolutional layers,which can reduce the possibility of over-fitting of the model.Compared with CNN,this method can effectively improve the accuracy of the change detection results.