Deep Learning Polarimetric SAR Classification Based On Scattering Energy And Wishart Distribution

Compared to synthetic aperture radar(SAR)images,polarimetric SAR can transmit and receive polarimetric electromagnetic wave energy in different channels with all-weather.So the polarimetric SAR images contain more terrain information.However,since the polarimetric SAR data is usually presented in complex matrix,traditional image classification methods are difficult to deal with the polarimetric SAR image.In another way,the exit classification algorithms mostly focus on statistical characteristics or scattering characteristics of a single pixel.In order to extract the image feature from the polarimetric SAR image and combine with the scattering feature,this paper mainly combines the traditional methods with deep learning models.The main works are summarized as follows:(1)A new algorithm based on stacked autoencoder and scattering energy is proposed.In this method,orientation compensation is used to compensate the polarization orientation angle,reducing the impact of the polarimetric angle noise.Then,the Freeman-Durden decomposition is adopted to extract the three basic scattering powers—surface,double bounce and volume,as the new features of the polarimetric SAR image.For the image space feature,we use the pixel block centered at the pixel point as the stacked sparse autoencoder input and the stacked sparse autoencoder network as the classification algorithm for classification.This method uses the scattering energy as the feature of polarimetric SAR pixels and reduces the complexity of the data.With pixel blocks as input,the neural network can extract the local space information simultaneously.The stacked sparse autoencoder is used for classification.The algorithm proposed in this chapter can combine polarization scattering features with the spatial information of the image and obtain better results than other algorithms in experiments..(2)A new polarimetric SAR image classification method based on Wishart distance and convolution network(CNN)is proposed.In this chapter,since the coherence matrix and the covariance matrix are both Hermitian matrices,a new method of Wishart distance is proposed,which named faster Wishart.The faster Wishart greatly reduces the computation time.In another way,the faster Wishart is helpful for the optimization of the weights.Based on the new algorithm formula,a new architecture of the convolution layer based on Wishart distance is proposed,which named Wishart CNN.The Wishart CNN has a kernel with the size of 3x3,which can extract the Wishart distance.Wishart CNN can not only extract the statistical information of polarimetric SAR,but also extract the image features of polarimetric SAR based on the CNN.Finally,a new neural network(W-CNN)for polarimetric SAR image classification is proposed,which combine with Wishart CNNs.The experiments show that this model outperforms the state-of-art algorithms.(3)A new method for polarimetric SAR image classification based on the fully convolution network(FCN)and Wishart classifier is put forward.Firstly,a new semantic segmentation network is proposed,which named Close-U-net.The polarimetric SAR image is classified by the Close-U-net and we get a rough segmentation map.Then the indefinite pixels in the segmentation map are classified by the Wishart classifier and we get the final result.The Close-U-net can extract the semantic features of polarimetric SAR images with less training samples.Secondly,combining the statistical characteristic information by Wishart classifier,the results have a higher accuracy in experiments.