Polarimetric SAR Images Classification Based On Deep Spatial Feature Learning

Polarimetric synthetic aperture radar(POLSAR),which is a multi-channel and multiparameter microwave active radar imaging system,can detect hidden targets and observe the true form on the surface of the earth.It has the advantages of all-day and all-weather.Due to its characteristic of multi-polarizations,it can provide rich polarization information.Hence it is widely applied in military and civilian fields,such as disaster relief,vegetation detection,precision agriculture and geological research,etc..The POLSAR image classification as one of the key technologies in the POLSAR image understanding and interpretation,is currently a hot research topic in the field of remote sensing.Classification results provide useful information for the follow-up works such as target recognition and detection.However,because the POLSAR images are difficult to obtain discriminative features,inevitably affected by speckle noises in the process of imaging,and lack of artificial labeled samples,these characteristics add difficulty for Pol AR image classification task.Based on the deep learning and POLSAR image characteristics,the dissertation in view of the problems in POLSAR image classification task,including the difficulties in obtaining discriminative features,the large amount of speckle noises and the small amount of labeled samples,provides a series of robust and accurate classification methods.The contributions are as follows.1.Because the scattering measurements of individual pixels in POLSAR images are affected by the speckle noises,the performance of pixel-based classification approaches would be poor.Since the spatial correlation between adjacent pixels is of great help to overcome the speckle noises,a classification method,which makes use of the stacked sparse autoencoder(SSAE)and the local spatial information of POLSAR image,is provided.This method is based on the perspective of feature learning and can automatically learn useful features layer by layer in an unsupervised manner.Firstly,high order neighborhood window is used to extract the neighborhood region.For the purpose of remaining the spatial relation in the neighborhood window,the weights calculated depending on the spatial Wishart distances from the neighbor pixels to the central pixel are proposed,then the pixels in the neighborhood window are multiplied by their corresponding weights.Finally,input the neighborhood window into SSAE.SSAE learns the deep spatial features of the central pixel in the neighborhood window and accomplishes the classification task.The experimental results show that the features extracted by this method have strong robustness,can effectively overcome speckle noises.It will improve the classification accuracy and obtain better consistency of homogeneous area.2.Most of the POLSAR image classification methods which are based on deep learning are single-scale methods,but there often exists targets with different scales in the image scene.Hence considering that the POLSAR images naturally have multi-scale characteristics,a multi-scale feature extraction and classification framework based on deep learning technology is proposed which can obtain multi-scale deep spatial features.The framework is a hybrid of SSAE,average pooling and softmax classifier.Firstly,SSAE is trained end-by-end in different scales separately,thus the features of different scales are extracted.Then these features are combined in series.In order to improve the classification efficiency and reduce amount of calculation,a 1-D average pooling strategy is proposed for reducing the dimension of the features.Finally,input the multi-scale deep spatial features into the softmax classifier to obtain the final classification result.Since the multi-scale features at the same time has the characteristics of large scale features and small scale features,they can not only improve the consistency of homogeneous area but also classify the details accurately.3.In order to solve the drawback of the method in the chapter 2: though the region-based classification methods can obtain good consistency of homogeneous area,they lost detail information sometimes,a new classification framework is proposed which makes use of the rich structure information of POLSAR image.Firstly,a concise structural representation--the polarimetric sketch map is applied to separate the homogeneous region and detail region.Then specific classification method is applied to each kind of region respectively,thus the characteristics of different regions can be remained.The condition random field(CRF)is used to classify the detail region.Because CRF combines the single pixel characteristics with the spatial context,it can keep the detail classification accuracy meanwhile overcome the speckle noises.For the homogeneous region,a hierarchical deep learning methods is proposed.The first layer is a convolution layer with the kernel of the Gabor wavelet transform which can extract the primary features at multi-scale and multi-orientation.Then SSAE is used to learn more abstract deep features and classify the homogeneous region.Lastly,fuse the classification results of the homogeneous region and the detail region to obtain the final classification result.The experimental results show that the method can improve classification accuracy,obtain better consistency of homogeneous area and keep the image details.4.Aiming at the lack of labeled samples of POLSAR images,a method based on semi-supervised deep integrated learning is proposed.A sample selection strategy is designed which can select highly confident samples in every iteration.This method firstly uses very small amount of labeled training sample to train each classifier respectively,and predict the label of all the unlabeled samples.Then select the unlabeled samples which have the highest prediction probability,and from these selected samples,the ones which have the same prediction label among all the classifier are chosen again.Take these sample as new training samples and put them into the training set.Retrain the classifiers with the new training set.Repeat the process above till the stop condition is reached.The strategy selects high confidence samples and abandon those samples whose predictions are inconsistent,thus reduce the negative effects on classification performance caused by incorrectly predicted unlabeled samples.Convolutional neural network(CNN)and SSAE which are the classical deep learning algorithms are selected as the basic classifier.They are respectively the region-based classification method and the pixel-based classification method,hence they have complementary effects.Experimental results show that this method can achieve satisfactory classification accuracy with only a very small number of labeled training samples.