Research On The Theory And Application Method Of Knowledge Transfer In Homogeneous Target Recognition In Multi-source Remote Sensing Data

As the most important carrier of human activities at sea,ships play an important role in a variety of maritime matters including border control,environmental protection,traffic surveillance,maritime search and rescue,and so on.Space-borne synthetic aperture radar(SAR)has been widely used for maritime ship monitoring,since space-borne SAR is able to provide wide-area,day-and-night and weather-independent images.The increase of the number of moderate-and high-resolution SAR images with the advent of the new generation of satellite missions makes it possible to further identify the type of ship beyond normally provide geographic location of the ship target.For the moment,there are still many challenges to be resolved on the task of ship classification in SAR images.On the one hand,SAR images can only provide limited discriminative information due to the limitation of SAR imaging mechanism,which leads to the widespread fine-grained classification problem with small inter-class variance and large intra-class variance is severe on the task of ship classification in SAR images.The specific performance is that ships only have subtle visual appearance variations among different subclasses but the appearance of ships shows many changes in the same subclass.On the other hand,labeling ships in SAR images is often a time-consuming and costly process that requires laborious expert interpretation and expensive ground campaigns,and is even not possible in many real-world application scenarios.This leads to the labeled data scarcity problem,which seriously restricts the performance of supervised learning methods on the task of ship classification in SAR images.Therefore,this paper focuses on designing targeted machine learning methods for the above problems.The novel contributions are summarized as follows.1.For addressing the fine-grained classification problem,a distribution shift metric learning method is proposed,which jointly optimizes pairwise constraints,inter-class distribution shift and manifold regularization,to learn the distance metrics of adapting training data.In addition,two kinds of inter-class distribution shift regularizations(local and global)are designed to improve the discriminative ability of distance metrics by widening the distribution discrepancy among different subclasses.2.For addressing the labeled data scarcity problem,a discriminative joint adaptation regularization framework is proposed,which jointly optimizes structural risk function,joint distribution adaptation,manifold consistency alignment and source discriminative information preservation,to learn the domain-invariant classifier.In addition,two kinds of source discriminative information preservation regularizations are designed based on within-class scatter matrix and pairwise constraints,which can transfer the source discriminative information to the target domain while implementing domain adaptation,thereby improving the discriminative ability of the domain-invariant classifier.3.For addressing the above two problems simultaneously,a geometric transfer metric learning method is proposed,which jointly optimizes pairwise constraints,joint distribution adaptation and manifold regularization,to learn the distance metrics of adapting target domain.In addition,two practical application scenarios of transfer learning,which are zero labeled sample task and scarce labeled samples task,are resolved based on different training strategies.Systematic experiments and in-depth analysis on SAR ship databases demonstrated that the proposed methods not only outperform most existing methods and achieve good classification results,but also effectively solve the fine-grained classification problem and labeled data scarcity problem.