SAR Image Change Detection Based On Metric Learning And Multi-objective Immune Algorithm

Change detection refers to changes in the same area over a period of time.SAR image refers to the image obtained by synthetic aperture radar.Compared with other imaging methods,SAR images are not limited by external conditions such as weather,and have all-weather and all-time working ability.Therefore,the change detection based on SAR image has become a hot topic of research.For SAR image change detection,how to find the difference between two samples is very important.Because of the unique imaging technology,SAR image has speckle noise.If the traditional Euclidean distance is used,it is easy to be affected by speckle noise.Based on the shortcomings of traditional distance measurement,two SAR image change detection algorithms based on metric learning are proposed in this paper.Meanwhile,Because of the imbalance of samples in SAR image change detection and the need to manually adjust parameters in building models,a SAR image change detection algorithm based on iterative multi-objective immune method is proposed.The main contents of this paper are as follows: 1.In view of the fact that traditional distance measurement is not robust to noise and can't measure the difference between samples well,this paper proposes SAR image change detection based on spatial context metric learning.In SAR image change detection,there is a problem that the samples of the change and unchanged boundary areas are easy to be misclassified,so this method uses the ecological method to obtain the boundary areas,and takes the samples of the boundary areas as training samples;moreover,the method takes full account of the spatial neighborhood information when constructing the constraint pair,and to a certain extent,suppresses the influence of speckle noise and registration error.Finally,when this method constructs the model,the traditional subtraction is transformed into a more robust LR operator.The method is tested on five standard data sets and achieves good results.2.Aiming at the problem that the first single-mode metric learning can't make good use of multi-mode data information,which leads to low accuracy,this paper proposes a SAR image change detection algorithm based on spatial priori and multi-mode metric learning.Similar to the first method,the method takes the samples of the boundary area as training samples,which solves the problem that these samples are easy to misclassified;moreover,the method takes full account of spatial priori in constructing constraint pairs and finds the same kind of constraint pairs in the neighborhood;finally,a multi-mode metric learning model is constructed,which learns two kinds of mapping matrices,i.e.,the specific mapping matrix for each model and the shared mapping matrix for all models.The original samples are mapped to feature spaces by mapping matrices,and finally measures the distance in the mapping space.At the same time,this method is applied to heterogeneous data.First,the change detection results of each channel are obtained,and then the final change detection results are obtained by voting method.The method is tested on four homogeneous datasets and two heterogeneous datasets,and good results are achieved.3.Aiming at the problem of sample imbalance in SAR image and the need for manual adjustment of model parameters,this paper also proposes a SAR image change detection algorithm based on iterative multi-objective immune method.This method regards sensitivity and specificity as objective function,which can overcome the problem of poor performance of single objective function such as accuracy.Then the method uses iterative multi-objective immune method to optimize the objective function.In this process,the model parameters are optimized iteratively,avoiding the problem of manual adjustment of parameters.This method generates the optimal Pareto solution set by iteration optimization.Then it selects the solution with the highest AUC value in the solution set as the final optimal solution.The method is tested on four standard data sets and achieves good results.