Feature Statistics And Classification Algorithms For Polarimetric SAR Image

As an active microwave imaging system,synthetic aperture radar(SAR)has the ability of capturing high resolution and large observation data with some penetration in both day and night under almost all weather conditions.SAR has received much attention from major countries through the world and plays an important role in military target detection,resources and environmental observation,land planing and land use,and other potential fields.Compared to SAR,polarimetric SAR(PolSAR)can provide more information about terrain targets by emitting and receiving fully polarized waves.PolSAR has more special advantages over SAR.China is also building its own high resolution Earth observation system(HREOS).As the first civilian space-borne PolSAR satellite,Gaofen-3 has been successfully launched on 10 August 2016.It is believed that Gaofen-3 would continually provide a number of sustainable observation data.Due to the coherent imaging mechanism,PolSAR and the polarimetric features present obvious statistics and randomess.Statistical modeling has been a research focus of PolSAR.Moreover,scene classification is critical for PolSAR image interpre-tation.With the improvement of image resolution,the heterogeneous or extremely heterogeneous in the image become prominent.Conventional statistical models devel-oped for scattering matrix,polarimetric covariance matrix or coherence matrix,have become more and more complex.The complicated and diverse statistics of polarimetric features can be hardly described since no analytical expressions are available in most cases,which hinders the development of PolSAR statistical modeling.For the differences in imaging parameters and the scattering mechanism ambiguity induced by target ori-entation diversity effect,the same targets may present different scattering mechanisms while different targets may scatter similarly.This is the so called "spectral confusion"in PolSAR image.It not only causes difficulty in PolSAR image interpretation but also induces that training samples from one image can not be directly applied to another.How to quickly process and classify a whole PolSAR image is the key problem in prac-tical use.Though the developed classification algorithms perform well on local image regions,they may involve many processing steps and increase computational complex-ity.The practicability on a whole image can not be estimated.Supported by two HREOS sub-projects,the paper focuses on polairmetric features and conducts research of PolSAR statistical modeling and scene classification algorithms.Many polarimetric features can be extracted from PolSAR data.The polarimetric features usually reflect the scattering mechanism and physical properties about terrain targets and present more complicated and diverse statistics.It is difficult to theoreti-cally derive the statistical models.In this paper,nine polarimetric features are directly extracted from the polarimetric covariance matrix,the Alpha-stable distribution is in-troduced to describe the statistics of different polarimetric features.Then the joint statistical model for multiple polarimetric features is constructed based on Copula,and we note it as CoAS.Finally,a flexible PolSAR image classification algorithm is de-veloped based on CoAS and the Markov random field(MRF)model.The proposed classification algorithm takes polarimetric information,statistical information,spatial information,and physical understanding into consideration,and performs better on three real PolSAR data than the compared classifiers.With the combinations of differ-ent polarimetric features,CoAS helps find out how the features contribute to the final classification results.The polarimetric features verifying "total power invariance",carry the information about terrain targets and are affected by imaging parameters.Even though the po-larimetric features of same class from different images may be different,they should present similar variation.The paper focuses on polarimetric features that verify "total power invariance".Via computing the ratios between features,the component ratio based distance(CRD)is defined,and hopefully alleviates system differences of two Pol-SAR images.The combinations with L1 distance and ?2 distance can generate L1-CRD and ?2-CRD and then benefit their robustness to small noise.Finally,we replace the distance measurement in k-nearest neighbour algorithm with CRDs for cross-source PolSAR image classification.The experimental results on simulated data and real Pol-SAR images demonstrate the validity of the proposed classification scheme.And CRDs improve the efficiency of training samples.Based on previous research and practical development experiences,the tree-based classifiers like decision tree,random forest,have special advantages in processing a whole PolSAR image and turn out to be quick and accurate.The introduction of spa-tial information is very important for overall classification improvements.This paper analyzes the polarimetric features of typical terrain targets in Gaofen-3 PolSAR images.Based on the gradient boosting decision tree(GBDT)-XGBoost and the incorporation of spatial information via superpixel-based modified majority voting,a fast classification framework is proposed.We design and develop a whole PolSAR image classification workflow,which supports several classification algorithms.The classification module,the basic processing module,the building damage information extraction module and the assistant module compose the developed PolSAR image basic processing and inter-pretation system(POLSAR-PRINT).In summary,we research and develop the joint statistical model for polarimetric features and the three PolSAR image classification algorithms.The developed software POLSAR-PRINT has been applied in HREOS sub-projects.