SAR Image Semi-supervised Learning Classification Based On Superpixels And Samples Selective Strategies

As an active ground observation system,the Synthetic Aperture Radar(SAR)system can perform all-day and all-weather ground observation.For its imaging is based on electromagnetic waves,it possesses the capability to penetrate surface,so the system is widely used in military and civilian fields.SAR image classification is one of the hot issues in recent years.Due to the different imaging mechanisms,SAR images have inherent coherent speckle noise,which makes the method of processing optical images is no longer suitable for SAR images.In addition,due to the hardware conditions of the SAR image,the number of samples and the number of labeled samples are very rare compared with optical image,which also brings challenges to SAR image processing.In this paper,considering the current semi-supervised co-training classification algorithm for SAR images has problems such as slow processing,and the classification effect is disturbed by coherent speckle noise.We study the semi-supervised classification algorithm of SAR images in depth.The specific research contents include:(1)A method for texture feature extraction based on superpixels is proposed.Through the depiction and analysis of the characteristics of SAR images and the characteristics of superpixels,superpixels have a good set of characteristics,and the pixels within the same superpixel are homogeneous.Therefore,it is intended to use superpixels as the basis for feature extraction,model training,and classification and marking element.However,the current SAR low-level feature extraction mechanism cannot be directly applied to superpixels.Therefore,this article expands superpixels into regular regions without introducing extraneous noise,then calculate its texture feature parameters.The validity of the method is verified by experiments,and the time complexity is significantly reduced.(2)A semi-supervised classification method for SAR images based on selection strategy is proposed.The selection strategy is reflected in two aspects of the co-training process to obtain pseudo samples.On the one hand,the canonical correlation analysis is used as a supplementary discriminator for SAR image classification sample classification,so as to make fuller utilization of hidden information in unlabeled samples and labeled samples.On the other hand,for low-confidence samples,a method of sample augmentation based on superpixels and active learning is proposed,which pertinently selects samples with considerable information to expand,thereby compensating for the problem of insufficient samples for SAR image.Through theoretical analysis and experimental verification,this method achieves better results and accuracy compared with other semi-supervised co-training related methods with less labeled samples.And effectively reduce the effect of coherent speckle noise on the classification effect.