Polarimetric SAR Image Change Detection Algorithm And Its Fast Implementation

PolSAR has the advantages of SAR.In addition,the reflection of ground object is measured in a variety of polarization ways to describe the backscattering characteristics of ground object and provide multi-dimensional remote sensing information of the target.PolSAR change detection algorithm determines the real change area caused by the change of scene target through some characteristics of PolSAR data.This paper aims to improve the performance of PolSAR change detection in agriculture,forest,ocean,disaster assessment and better serve the monitoring of geomorphic change,resource change.At the same time,the parallel acceleration of the relevant time-consuming steps can realize the fast implementation of the change detection algorithm.To achieve the above goals,the following difficulties need to be solved:(1)How to make full use of polarization information to construct difference map.(2)How to generate more semantically rich training sets to provide data support for change detection and classification algorithms.(3)How to design a classification algorithm with low noise,high precision,high Kappa coefficient and consideration of details.(4)In engineering,the speed of PolSAR change detection algorithm is guaranteed,and the time required for algorithm time-consuming point is reduced.To solve the above problems,the main work of this paper is as follows:In order to solve the difficulty(1),a multi-channel difference map construction algorithm based on enhanced data is designed in this paper,which can construct multiple difference maps of different types based on different fine-grained enhanced data to make full use of polarization information.The model of data distribution was established,Parameters are obtained by parameter estimation as enhanced data and the multi-channel enhanced difference map was constructed by Renyi Entropy algorithm,HLT algorithm and MLR algorithm.In order to solve the difficulty(2),this paper optimized the FCM clustering label calibration algorithm and design the FCM clustering label calibration algorithm with compensation.The specific compensation method is to carry on the change and non-change estimation to the samples with uncertain change and increase the sample number of the change and nonchange sample set.The improved algorithm not only provides more positive and negative training samples,but also enriches the semantics of the training set.In order to solve the difficulty(3),this paper improved the tree feature based sample constrained kernel SVM algorithm and the multi-criteria shallow depth forest algorithm of multi-scale network.The former can use tree feature to improve change detection accuracy and Kappa coefficient.It can use sample constraint reduce the influence of speckle noise,as well.The latter can suppress the speckle noise with large sample and multi-scale,and increase the detailed detection information.Multiple criteria are used to enhance the difference and diversity of features,thus improving accuracy and Kappa coefficient.The tree feature based sample constrained kernel SVM algorithm extract tree features with XGBoost algorithm.Meanwhile,the samples in the training set are screened to train the kernel SVM classifier.Multi-criteria shallow depth forest algorithm of multi-scale network performs large-scale clipping of samples.Multi-scale large scale samples are multi-scaled by multi-scale network,and sample size is simplified by tree model and pooling.Multicriteria shallow depth forest algorithm is used to determine the change area.In order to solve the problem of(4),this paper accelerates the time-consuming points in the process.Multi-channel difference map construction and XGBoost algorithm are widely used as two directions of acceleration.The former is accelerated by optimizing matrix storage form,CUDA parallel algorithm,GPU deployment calculation,etc.The latter reduces the time consumption of the XGBoost algorithm by starting to turn on GPU acceleration.