Research On Classification Method Of Few Shot PolSAR Image Based On Relation Network

Polarimetric Synthetic Aperture Radar(Pol SAR)is a powerful microwave imaging technology that can provide all-weather target information on the earth’s surface.Compared with other remote sensing images,polarimetric SAR images can transmit and receive electromagnetic waves in four polarization combinations(HH,HV,VH and VV),thus providing richer information.Due to these characteristics,polarimetric SAR technology has high practical application value in image interpretation,such as image classification,target recognition and detection tasks.Among them,polarimetric SAR image classification has been widely studied as the basis for understanding and interpreting remote sensing.With the continuous improvement of remote sensing technology,it is easy to obtain unmarked polarimetric SAR data.However,the acquisition of labeled polarimetric SAR data is limited by many conditions.Therefore,the classification of polarimetric SAR images based on limited labeled samples has attracted more and more attention of researchers.Aiming at the problem of small samples,combining depth learning and spatial neighborhood features of polarimetric SAR images,the following work is carried out in this paper:(1)The polarimetric SAR marker sample enhancement method based on super-pixel is an image segmentation technology.For polarimetric SAR images,the pixels in the same area usually have high similarity,so they are more likely to belong to the same category.Therefore,aiming at the problem of few labeled samples in polarimetric SAR images,combined with the statistical characteristics and spatial information of the image itself,this paper proposes a super-pixel-based labeled sample enhancement method,which uses the information of very few labeled samples and a large number of unlabeled samples to expand the number of labeled samples,thus improving the classification results of polarimetric SAR images.(2)The polarimetric SAR image classification method based on the integrated dualbranch CNN(EDb-CNN)can obtain different scale features from the polarimetric SAR image,including two parallel CNN structures.Then,the feature fusion model is used to combine the two depth features,and the weighted loss function is used to improve the learning process.Next,use integrated learning algorithm for each CNN branch and Db-CNN network to obtain better classification results.Finally,a post-processing method based on super-pixel is proposed to improve the overall accuracy of classification results.The experimental results of two real polarimetric SAR data sets show that this method is an effective polarimetric SAR image classification method.(3)The 3D residual relationship network(3D-ARRN)method based on SWANet attention mechanism uses a multi-layer CNN with residual structure to extract the depth polarization feature of the image.In order to extract more important feature information and improve the accuracy of classification results,the spatial weighted attention mechanism network(SWANet)is introduced to concentrate important feature information,which is more beneficial to classification tasks.Then,combining the characteristic information of training samples and test samples,the similarity score of training samples and test samples is calculated using 3D-CNN.Finally,the category of the test sample is determined by the calculated similarity score.The effectiveness of this method is verified by three real data sets.