Polarimetric SAR Deep Feature Learning And Terrain Classification With Data Characteristics And Scattering Mechanisms

Synthetic aperture radar(SAR)is an active microwave remote sensing sensor.Its imaging mode is not disturbed by external weather,and has the advantage of obtaining data in all weather and all time.It is an ideal "clairvoyant" medium for human earth observation.In recent years,deep learning technology has played an important role in human production and life.It is a learning algorithm that simulates the cognitive mechanism of human brain.Deep learning obtains hidden layer features from input layer by layer through complex and deep model structure,and then abstracts them into task results.Thanks to the strong ability of fitting,reasoning,and prediction,deep learning has set off a revolutionary wave in the fields of pattern recognition,fuzzy control,text analysis and so on.However,it has a general effect in the field of SAR image processing.The crux lies in the characteristics of SAR image different from optical image:statistical distribution,small samples,complex-valued data,etc.Compared with SAR image,polarimetric SAR(PolSAR)image also contains rich and complex polarimetric scattering information,which further increases the difficulty of interpretation.This article aims to explore the new ideas brought by deep learning for PolSAR image processing tasks,and then carries out relevant tasks with the carrier of deep network.This article aims to explore the new ideas brought by deep learning for PolSAR image processing tasks,using the deep network model as a carrier and drawing on its advantages of deep feature extraction,deep data fitting,and small sample learning to carry out interpretation tasks.In this way,the disadvantages of traditional methods,such as failure to fully exploit polarization scattering characteristics,being susceptible to data distribution interference,and large demand for training samples,are overcome.Firstly,the researches based on second-order polarimetric expression is demonstrated,then the context from the data to task is systematically combed.Secondly,the deep learning theory and models are summarized,and the characteristics of various models are analyzed,then the internal relationship between deep network and PolSAR tasks is excavated.Finally,task oriented,the deep networks associated with characteristics of PolSAR data and the scattering mechanisms is designed,which provides a new solution for the traditional tasks such as polarimetric target decomposition,decomposition feature selection,terrain classification,image segmentation,change detection.The specific innovations and contributions are as follows:·For the second-order polarimetric expression data,most decomposition algorithms design predetermined decomposition basis or use fixed decomposition mode to analyze original data,so the obtained decomposition features do not reflect the scattering characteristics of the current terrains to the greatest extent.Therefore,this article proposes a polarimetric target decomposition feature learning framework,which can learn the decomposition features suitable for the current data according to its scattering characteristics;however,the scattering information of a pixel to be decomposed is not only determined by itself,but also affected by the scattering information of its neighborhood pixels.Therefore,this article proposes a convolution target decomposition feature learning framework with spatial information,which takes into account the scattering information of the central pixel and its neighborhood pixels in the decomposition process.·For the features after polarimetric target decomposition,the interpretation results of PolSAR image can be obtained by analyzing the data statistics and physical scattering mechanisms.However,making full use of decomposition features will increase the computational complexity,and some invalid features will have a negative impact on final task.Using features of partial algorithms will limit the ability of data to characterize the scattering characteristics of terrains.Therefore,this article is committed to selecting representative feature subsets with appropriate dimensions from a large number of features.For terrain classification,this article proposes a one-dimensional convolutional neural network feature selection model,which integrates the deep network with the feature subset decision-making process,exerts its deep information mining ability,and selects the optimal feature subset with the classification index as the evaluation standard;In order to make the selected feature subset not limited by specific tasks or label information,this article proposes a sparse variational autoencoder feature selection framework from the characteristics of feature decomposition itself,so that the model captures the data distribution characteristics and highlights the representative feature.Taking the reconstruction error without label information as the standard,the feature subset that can reconstruct the original data to the greatest extent is selected,on which the tasks of terrain classification,image segmentation,and change detection are implemented.·For the interpretation tasks of PolSAR image,establishing the mapping between the second-order polarimetric expression data and the specific task is the most direct way.This article uses the deep network to carry out an end-to-end analysis between PolSAR data and terrain classification results.Concretely,it aims at the problem of large intra-class differences and large inter-class similarity in complex areas of PolSAR image,which leads to poor classification effect of inconsistent areas.Combined with the characteristics of less data categories and large amount of data of PollSAR image,a deep metric learning framework based on N-cluster loss is proposed,then the overall discrimination of feature space is improved.Meanwhile,adversarial learning mechanism is adopted to accelerate the convergence speed of the model on complex samples,so as to improve the classification performance in complex regions.