Research On Polarimetric SAR Image Terrain Classification With Few Labels

Polarimetric Synthetic Aperture Radar can obtain abundant target scattering information,and its terrains classification method has gradually become a research hotspot in radar image interpretation in recent years.At present,a large number of supervised and unsupervised polarimetric SAR classification methods have been developed.Among them,the unsupervised learning method does not require labeled samples,but usually requires strong expert knowledge to design discriminative polarimetric scattering features,and the classification accuracy is not ideal.Supervised learning uses a large number of labeled samples to train models,which can get relatively accurate classification results,but the cost of labeling samples is high and the practicability is limited.Aiming at the low accuracy of unsupervised methods,and the high cost of traditional supervised methods,this paper proposes a series of polarimetric SAR terrain classification methods using small-scale labeled samples.The main research work and innovations are as follows:A kind of FSW-LapSVM classification method for polarimetric SAR terrains in case of small samples is proposed.Firstly,this method extracts features from the polarimetric covariance matrix,and design the Spatial-Wishart similarity and corresponding regular graph for a large number of unlabeled samples based on the complex Wishart distribution characteristics and spatial relationship.Second,the introduction of pixel fuzzy labels reduces the impact of noise on classification and improves model robustness.This method uses a large number of inexpensive unlabeled samples to assist in classification,which can reduce the cost of manual labeling samples and still obtain better classification performance under the condition of small labeled samples.The experimental results on multiple polarimetric SAR datasets show that this method can obtain high classification accuracy under a condition of a few 0.5%~1% of labeled samples.A method for classification of polarimetric SAR terrains based on deep convolutional siamese network in case of small samples is proposed.This method introduces spatial neighborhood information based on spatial consistency assumptions,and uses deep neural networks to automatically extract features.For the problem that the deep network heavily depends on a large number of samples,this method constructs a new type of deep network model with supervised pre-training and supervised fine-tuning,to achieve end-to-end accurate classification by deep learning under small sample conditions.Firstly,a weight-sharing convolutional siamese network is constructed to greatly expand the small sample dataset and to extract differentiated features that are easier to classify.Subsequently,a fully-connected classification network is added on top of siamese network to form deep convolutional siamese network,and the supervised information is used to fine tune the whole network to achieve classification.The experimental results on multiple polarimetric SAR datasets show that this method can obtain better classification effect under the condition that only 10 samples are labeled in each class.A method for classification of polarimetric SAR terrains based on deep convolutional Bi-LSTM siamese network in case of small samples is proposed.On the basis of deep convolutional siamese network,this method bidirectionally serializes the spatial neighborhood information in order to make more reasonable use of the neighborhood information.The bidirectional LSTM can also extract more complete sample information,making the classification result more accurate.In addition,the introduction of convolutions can characterize the local features of the data and reduce data redundancy,preventing overfitting of the model.Under small sample conditions,the method can use sample information more fully and reasonably to improve classification performance.Finally,the experimental results on multiple sets of polarimetric SAR datasets show that the method can further improve the classification accuracy under the condition that only 10 samples are labeled in each class.