Semi-supervised PolSAR Image Classification Based On Fuzzy Superpixel

Polarized Synthetic Aperture Radar(Pol SAR)is capable of all-time and all-weather imaging,and has microwave penetration.Compared with using a single polarization channel,Pol SAR can provide richer information when sending and receiving electromagnetic signals with various polarization states.Because of these characteristics,Pol SAR is widely used in various remote sensing tasks,such as disaster monitoring,crop estimation,and resource exploration.Pol SAR image classification is an important research field for understanding and interpreting remote sensing images or preprocessing for further applications.In recent years,deep learning has attracted great attention and achieved outstanding performance in most computer vision scenarios.Inspired by the successful application of optical image classification,deep learning has been considered an effective feature extraction algorithm for Pol SAR image classification.The success of the Pol SAR image classification method based on deep learning depends on the appropriate labeled data set.With the development of imaging technology,it has become easier to acquire a large number of unlabeled Pol SAR images.However,the annotation of Pol SAR images is much more expensive than the annotation of optical images,which has attracted the attention of recent research on Pol SAR image classification based on semi-supervised deep learning.This paper takes Pol SAR images as the research object,based on fuzzy superpixels,and launches research on the application of deep learning in semi-supervised Pol SAR image classification.The research results are as follows:(1)In view of the difficulty in obtaining labeled samples in Pol SAR image classification and the direct use of pixel classification will be affected by speckle noise,we proposed fuzzy superpixels based semi-supervised similarity-constrained CNN for PolSAR image classification.This method can obtain higher classification accuracy with fewer labeled samples.The algorithm first applies fuzzy superpixel algorithm to generate superpixels and uncertain pixels.Second,the labeled and unlabeled sample sets are constructed based on superpixels and undetermined pixels.Then,we propose a similarity-constrained convolutional neural network(SCNN)model and a two-step label propagation strategy for assigning pseudo-labels to unlabeled data.Finally,bothlabeled data and pseudo-labeled data are used to train SCNN to get the final classification result.(2)The performance of the pseudo-labeled-based semi-supervised Pol SAR image classification algorithm is largely affected by the accuracy of the pseudo-labeling.In order to obtain higher pseudo-labeling accuracy,we propose a semi-supervised Pol SAR image classification algorithm based on fuzzy superpixels and fully convolutional neural network(FS-FCN).The algorithm uses superpixels as the basic unit of label propagation,uses FCN to extract the features between labeled superpixels and unlabeled superpixels,and uses these features to establish a combined optimization problem.Solving this combinatorial optimization problem realizes label propagation.The expanded training set is used to train FCN,and the trained FCN is used as the Pol SAR image classifier.(3)The pseudo-label-based semi-supervised Pol SAR image classification algorithm often requires multiple iterations in the two steps of training the network and obtaining the pseudo-label,which has high time complexity,and the performance of the algorithm is affected by the accuracy of the pseudo-label.For this reason,we propose a semisupervised Pol SAR image classification algorithm based on full convolutional neural network and deep mutual learning.The algorithm uses multiple networks for training at the same time,and is a semi-supervised Pol SAR image classification algorithm that does not rely on pseudo-labels.The algorithm first uses the fuzzy superpixel algorithm to expand the initial set of labeled samples,and then uses deep mutual learning to train three FCN networks at the same time.One FCN uses the supervised classification loss for training,and the remaining two FCNs use the unsupervised classification loss for training.Finally,the prediction results of the two FCN models trained with unsupervised classification loss are weighted as the final classification result.