Semi-Supervised PolSAR Terrain Classification Based On ? Model

Polarimetric Synthetic Aperture Radar(PolSAR)imaging is an important means of remote sensing detection.Classification of terrain using PolSAR images is an important task in SAR image interpretation.In recent years,the use of deep learning methods has made the PolSAR terrain classification more accurate and faster.However,the labeled samples are very scarce,which is a typical small sample problem in PolSAR data.Training of existing deep models usually requires a large number of labeled samples.The limited labeled samples of PolSAR data greatly restrict the application of deep learning methods to this problem.Aiming at the small sample problem in PolSAR classification,this paper applies ? model to PolSAR image terrain classification and makes a study of a series of classification methods based on the ? model.The main work is described as following:1.A PolSAR terrain classification method based on Triplet-? model is proposed.The model adopts convolutional neural network structure.By adding different perturbations to the samples and using different dropout layers in the network structure,the output vectors are different,and the model is encouraged to reduce this difference and give consistent prediction.The training of this model only needs a small number of labeled samples.At the same time,with the aid of a large number of unlabeled samples,it can effectively solve the problem of small samples.However,this model neglects the relationship between samples.Therefore,this paper introduces the triplet loss in metric learning and constructs Triplet-? model,which can restrict the sample features in embedding space and enhance the classification effect.2.PolSAR terrain classification method based on VAT-? model is proposed.Although adding triplet loss can improve the classification accuracy of ? model,it is necessary to mine triplets before calculating the loss.This process is computationally intensive.Therefore,the idea of virtual adversarial training is introduced.By minimizing the distribution difference between the model's output probability of original input and that of the noisy data,virtual adversarial noise is obtained.Adversarial samples are formed by adding this noise to the input samples.This process is equivalent to automatic generation of hard examples.Training with these samples can make the model more robust to noise while improving the classification accuracy.3.A PolSAR terrain classification method based on temporal-? model is proposed.The ? model has one obvious drawback that the network needs to evaluate the same input sample twice in an epoch.It's time complexity is higher.Through temporal ensembling strategy,the prediction results of current epoch are derived from the cumulative results of previous epochs.In this way,the network does not need to be trained for twice in each epoch,which accelerates the convergence of the model.Experiments show that the training time can be shortened effectively on the premise of guaranteeing accuracy.