Research On Sample Distribution Matching Based Transfer Learning For Polarimetric SAR Data

Synthetic Aperture Radar(SAR)records the target scattering information through active transmitting and receiving electromagnetic microwaves.The all-time and all-weather imaging capability enables it to establish a relatively complete Earth Observation system,together with optical sensors.Compared with single-polarization SAR,which means data is acquired under the single polarization state combination,Polarimetric Synthetic Aperture Radar(PolSAR)can further acquire the complete backscattering information of targets under a certain polarization basis.Therefore,it has a natural advantage in the analysis,extraction,and inversion of object's attributes and parameters.With the rapid development of sensor technology,more and more air-and space-borne radar systems have emerged in recent years.In particular,China's self-developed Gaofen-3 radar satellite was successfully launched in 2016,marking a stable domestic data source of SAR and even PolSAR research and applications.However,the scattering characteristics are not only the integrated reflection of target's own size,structure,and dielectric properties,but also affected by imaging parameters,such as microwave frequency and incident angle,etc.Therefore,similar targets often have inconsistent scattering performances on different images.This point has led to the difficulty of reusing historical archive data to solve recent data tasks in practice.With much more abundant information,PolSAR data has especially been more severely affected.The fragmentation phenomenon between radar data greatly reduces the usability and value of historical data.Besides,the labeling work of ground object in recent data not only takes time and effort,but also reduces the timeliness.As one of the research hotspots in the field of machine learning and artificial intelligence,transfer learning aims to explore the application of knowledge accumulated in a certain field to another field related to it.Its fundamental purpose is to solve the adaptation problems of previous domain data in recent domain tasks.This provides possibility for reusing historical data and relieving label information dependence of recent data.Therefore,this dissertation focuses on the transfer learning problem of PolSAR data,which are acquired under different time,space or imaging conditions,and attempts to improve the matching degree of sample distribution between historical data and recent data under a series of hypothesis and application conditions,in order to achieve cross-domain information transfer.The main content of this article includes:The basic statistical distribution form of PolSAR data is verified.Two potential problems existed in the classical classification model based on complex Wishart distribution are analyzed,and the corresponding solutions are proposed.Thus using statistical hypothesis testing,a novel Wishart distribution based model is established.The new model can obtain some representative sub-category centers while effectively reducing the adverse effects of interference samples.It improves the performance of the classification model based on the basic distribution form,and later,the precondition and applicability for maintaining the reasonableness of complex Wishart distribution are summarized.Aimed at matching inter-domain conditional probability distributions,several instance-based transfer learning methods were investigated.In order to cope with the problem that the instance-based transfer learning methods cannot work when the labeled data is very rare in target domain,this article proposes a low-cost approach to expand the labeled sample set in target domain and thus enriches the diversity of target data in some extent,by the use of statistical hypothesis testing theory and region growing technique.After the comparative study of two Bagging based transfer learning algorithms,a novel instance-based transfer method is proposed.This method filters the samples in source domain that are not related to the target domain task,by the use of Wishart distribution,and reduces the negative transfer effect by the use of fallback classifier.Finally,it matches the inter-domain sample conditional distribution via collaboratively updating the evaluation set and weak classifier set,and maintains the stability of target recognition results.Aimed at matching inter-domain marginal probability distributions,the transition process from conventional subspace learning to transfer subspace learning is analyzed systematically and the principles of information retention for transfer subspace learning are summarized.And then,in order to keep label information in source domain,data structural information in both domains,and match inter-domain sample marginal distributions,a subspace learning criterion is proposed and thus a new feature representation based transfer learning model is constructed.Furthermore,it can employ the kernel mapping function based on Wishart hypothesis testing and take the statistical characteristics of PolSAR data into account,which helps reduce the difference between inter-domain marginal distributions and upgrade the precisions in object recognition tasks without using label information in target domain.As for the matching of mixture probability distributions,the effect of equal-dimensional projection transformation on the multivariate complex Gaussian distribution,the complex Wishart distribution,and their relevant finite mixture distribution models are analyzed,and the changes of transformed probability density functions are discussed together.Based on the previous work,Bregman divergence is introduced to evaluate the differences between inter-domain mixture distributions,and thus this article proposes an equal-dimensional transfer transformation function that simultaneously keeps the intra-domain sample distributions and matches the inter-domain sample distributions.Finally,by introducing the polarization basis theory,the difficulty of solving this objective function is reduced and the interpretability of transfer relationship between two domains is improved.A variety of processing tasks can benefit from it,including target recognition,change detection,etc.