Terrain Classification For Polarimetric SAR Images Based On Small Samples And Spatial Polarimetric Information Cooperation

Polarimetric Synthetic Aperture Radar(Pol SAR)that can measure the target scattering echo coefficients without delay and that has high multi-channel and multi-parameter is a radar imaging system.As a result,the polarimetric Synthetic Aperture Radar(SAR)images which is obtained by the Pol SAR can bring people a variety kind of information of the target.As one of the important research contents of radar images interpretation,polarimetric SAR images terrain classification technology has been widely used in many aspects such as target detection and disaster monitoring.Improving the accuracy rate of the terrain classification is one of the most popular topic both at home and abroad.In recent years,many well performed methods for the polarimetric SAR images terrain classification problems are mainly rely on the big size of the training data to train an excellent classifiers model and finally achieve the goal of the high accurate in this problem.However,to obtain the large amount of training data need to pay a high price in the resource of the labor,which is becoming more and more expensive nowadays.What's more,polarimetric SAR images not only contains measured values of echo information but also has a spatially organized structure.Based on the above discussion,this paper mainly studies the terrain classification techniques for polarimetric SAR images with small sample cooperating spatial polarimetric information.The main contents are as follows:1.An incremental and small samples classification method based on spatial point domain Knearest neighbors is designed.The existing supervised methods of the polarimetric SAR images classification problem mostly use a large number of marked samples for batch processing training.This traditional scheme not only has high cost of data but also cannot fully utilize historical information during the training process.In this paper,the labeled data is gradually predicted through incremental learning under the condition of the small labeled.At the same time,a new spatial point domain distance was designed in the local neighborhood to determine the accuracy of the stage prediction results.The proposed method performs performance verification experiments on five images such as the Flevoland image in the Netherlands,the San Francisco Bay image in American,the ESAR image in Germany and so on.The experimental results show that the method can achieve a classification accuracy higher than 84% under the condition of 0.1% labeled sample.2.A small samples classification method based on dual-channel low-rank and spatial polarimetric distance is designed.This method mainly divides the input polarimetric SAR images into two channels,the real part and the imaginary part.The speckle noise in the polarimetric SAR images can be removed by low rank matrix decomposition.A weight calculation formula is designed to measure the similarity of the samples.At the same time,the infinite norm is used to construct spatial constraints to improve the accuracy of data classification.Experimental results on five polarimetric SAR images show that the method can achieve classification accuracy of 85%-98% under the condition of 0.1% labeled sample,and it has obvious effectiveness and superiority.3.A small samples classification method about reinforcement learning based on spatial polarimetric reward is designed.According to the defined local neighborhood area,the method computes the indications of actions which labeled samples provided.At the same time,the functions of space and polarization reward are designed in the local neighborhood area to realize the fusion of spatial polarization information.This method introduces reinforcement learning into polarimetric SAR images classification and constructs a dynamic continuous learning classifier.Experimental results on multiple polarimetric SAR images show that this method can achieve a classification accuracy of more than 90% under the condition of 0.1% labeled samples,and the overall accuracy,average accuracy and Kappa coefficients with good performance.