Research On Feature Learning Base Semi-supervised Terrain Classification For PolSAR

Polarimetric Synthetic Aperture Radar(Pol SAR)is one of the most advanced sensors in remote sensing.It has unique imaging characteristics such as all-weather,all-day,multi-band,multi-polarization,and can provide images with high resolution.Therefore,a large amount of valuable information can be obtained in the interpretation of the Pol SAR data.In particular,the classification of polarimetric SAR images,as an important area of polarimetric SAR interpretation,is widely used in earth resource exploration and military systems.For the classification of polarimetric SAR images,the machine learning algorithms can be divided into unsupervised,supervised and semi-supervised classification algorithms.The unsupervised classification algorithm does not use any label information,the model is simple,and the accuracy tend to be low.The supervised classification algorithm requires a large number of labeled samples,and the labeling work is time-consuming and labor-intensive.Semi-supervised methods combine the advantages of unsupervised and supervised learning,and can obtain higher classification performance by using only a small number of labeled samples.Although some conventional semi-supervised methods have good classification performance,they are in low efficiency,which is harmful to practical applications.And because the required features by manual design,so that the classification results are largely dependent on the extracted features.In recent years,deep learning,as an important branch of machine learning,can automatically extract the intrinsic features of target objects,and is widely applied in polarimetric SAR data.However,if there are no sufficient labeled samples,most supervised neural network models do not train sufficient,and it is difficult to obtain better classification results.Thus,based on the feature learning of polarimetric SAR data,three novel semi-supervised classification methods are proposed to improve the efficiency and classification accuracy.1.A fast semi-supervised Pol SAR classification method based on histogram density estimation is proposed,which aims to address the low classification efficiency of the conventional semi-supervised methods.The main idea of the method is as follows: firstly,a non-iterative co-training algorithm is proposed to extend the labeled training samples.Then,the histogram density estimation method is used to quickly project the original features and learn the relationship between the original data;next,selecting the optimal sub-space features by the sub-module optimization algorithm to reduce the correlation among the mapped features.And then the classifier is trained and the feature mapping table is generated.Finally,according to the feature mapping table,the corresponding features are extracted from the test samples,which reduces the time consumed by the feature mapping and achieves a fast classification task.The experimental results show that the method can not only greatly reduce the classification time,but also effectively improve the classification accuracy.2.A new semi-supervised generative adversarial networks(SGAN)based on manifold regulation constraints is proposed.The main idea of the method is: based on SGAN,adding manifold regulation constraints to the generated samples,there will be a large number of unlabeled generated samples to mine the inherent geometric structure in the data,which enable the discriminant network to learn a smooth decision function,and improve classification accuracy,at the meanwhile,feature matching is introduced to stabilize training.The experimental results show that compared with the traditional algorithm,the classification accuracy of the Pol SAR data is dramatically improved.3.A new SGAN based on self-attention model is proposed.The main idea of the method is that the SGAN model is constructed using a convolutional layer,and the convolution process the information in the local space neighborhood.Adding a self-attention model to the convolution can help model the global and multi-level dependencies of the image region.Due to the self-attention model,the image generated by generative network's details are related in each position and the distant part of the image,and at the meanwhile,the discriminative network can also accurately force geometric constraints on the global image structure.In addition,the spectral normalization technique is introduced to stabilize the training of SGAN,so that the training process is more stable and easier to converge.The experimental results show that the classification accuracy is improved compared with stateof-the-art methods.