Terrain Classification Of Polarimetric SAR Image With Limited Labeled Samples

Polarimetric synthetic aperture radar(Pol SAR)is an actively remote-sensed imaging system,which transmits and receives orthogonally polarized electromagnetic waves to carry out a full range of measurements on targets.Thus,it's widely used in remote sensing area.Terrain classification is the key point in Pol SAR image understanding and interpretation,which is defined as assigning a predefined label for every pixel in the image.Recently,deep learning techniques represented by convolutional neural networks have achieved remarkable achievements in the terrain classification of Pol SAR image at the cost of a huge number of labeled samples.However,due to the complexity of the imaging mechanism,Pol SAR image understanding often requires expert knowledge.Thus,the labeled samples are scarce in Pol SAR.In such a case,this paper carries out three methods using deep learning techniques as following:1.Terrain classification of Pol SAR image based on labels' semantic priors.Based on the prior information that labels in one region are consistent and boundaries between regions are obvious,this method introduces the region consistency constrain and the boundary kept constrain into the training of convolutional neural network to assist the neural network to learn more discriminative features.In addition,the final classification map shows favorable spatial consistency and aligned boundaries.2.Terrain classification of Pol SAR image based on progressively deep seeded region growing and conditional random field.Although the deep features of convolutional neural networks have high level semantic information,extending a labeled training set directly using the seeded region growing is easy to add too many noisy samples at one time,which easily leads to the collapse at training.In this method,during the network training process,unlabeled samples in the neighborhood of labeled samples are selected to extend the labeled training set in each epoch.Thus,the set of labeled training samples is slowly expanded to reduce the probability of collapse during training.At the same time,the prediction of the classification network is smoothed using Pauli pseudo-color image and conditional random field.Moreover,the processed information is fed back to the network during training,so that the prediction of neural network shows great spatial consistency,and the edges between the regions are clear and aligned.3.Terrain classification of Pol SAR image based on inter-class misclassification experience and conditional random field.This method adds residual loss on the basis of the cross entropy loss function,so that the convolutional neural network not only pays attention to the classification of the real category of each sample during training,but also learns the relationship among classes.Therefore,features learned by the convolutional neural network are more discriminative.To reduce the influence of coherent speckle noise,the Pauli pseudocolor image and conditional random field are used to smooth the output of the network,and the result is fed back to the convolutional neural network during training.So the classification result has good spatial consistency and neat edges.