PolSAR Image Terrain Classification Based On Deep Semi-Supervised Learning

Polarimetric SAR terrain classification,as the basis of remote sensing image interpretation,has become an important research direction in the field of remote sensing.In recent years,the deep learning has attracted the attention of many scholars and has also achieved great success in different fields.Therefore,this paper hopes to introduce the advantages of deep learning in polarimetric SAR terrain classification.However,deep learning requires a large number of training samples to ensure the performance of the network,and the lack of labeled samples in the field of polarimetric SAR limits the application of deep learning in this field.Semi-supervised learning can use a large number of unlabeled samples to obtain good results in the case of a small number of labeled samples.Therefore,it is of great significance to study how to combine the advantages of deep learning with semi-supervised learning to achieve the terrain classification of polarimetric SAR.Supported by the National Natural Science Foundation of China(PolSAR image classification based on co-training and sparse representation,No.61173092),and the National Natural Science Foundation of China(PolSAR image classification based on generative adversarial network.No.61771379).Aiming at the problem of small samples in polarimetric SAR terrain classification,this paper proposes a series of improved methods based on semi-supervised ladder networks.The main research results are: A polarimetric SAR terrain classification method based on the Robust Ladder Network(RLN)is proposed.Using the advantages of ladder network in the network structure,supervised learning and unsupervised learning are combined at the theoretical level to make better use of the unlabeled samples to assist model learning.At the same time,in order to obtain a more robust model,we designed a stability loss function from the perspective of supervised learning to reduce the difference in output vectors when the same sample passes through the network multiple times.And from the perspective of supervised learning,we design the mutual exclusion loss function of the output vector,thus constructing a ladder network based on a robust strategy.This not only solves the small sample problem in polarimetric SAR terrain classification,but also effectively enhances the robustness of the model.Finally,the feasibility and effectiveness of this method are verified by experiments.A method for terrain classification of polarimetric SAR based on Wishart Ladder Network(WLN)is proposed.In the first place,the Wishart metric distance is derived from the Wishart distribution.Based on the semi-supervised learning manifold hypothesis and Wishart distance,we construct the neighboring graph on the training set,and define the objective function of the Wishart ladder network.Then,we can optimize the model by the use of the smoothness of decision function on neighboring graph.Finally,according to the assumption of spatial coherence,the neighborhood information is introduced for polarimetric SAR pixels,which can reduce the influence of speckle noise to a certain extent,and effectively improves the classification effect.A method for terrain classification of polarimetric SAR based on attention mechanism and Ladder Network(ALN)is proposed.Although the introduction of the neighbor information in the existing method will improve the classification effect to a certain extent,it also brings about redundancy of the data information and inevitably introduces noise interference.To solve the above problem,this method designed the attention encoder and attention decoder in ladder network based on the attention mechanism,which can make the model automatically select useful information for the current task,and remove the interference of redundant information during the learning process.In this way,attention-aware features are extracted to improve the learning efficiency of the model.Experiments show that this method can effectively improve the convergence of the model.