High-Resolution SAR Image Classification Based On Prior Information

Synthetic aperture radar(SAR)image classification is a very important part of SAR image understanding.It plays an important role in a wide range of SAR image applications.In real world,certain patterns and rules apply on the spatial distributions of land covers and targets.These may include clustering effects in space of land covers or targets within the same category,and dependency or exclusion effects in space of land covers or targets in different categories,which are referred to as prior information on categories in this thesis.As there are limitations in expressing the real world with digital images,it is crucial to enhance the performance of digital image interpretation by utilizing supplementary information,such as geographical information,professional intelligence,and human experience,etc.This is especially important for high-resolution SAR images,since the SAR imaging mechanism and environmental effects can lead to serious noises,brightness variations,and even structure deformation and part missing of targets.Such effects may seriously impact the accuracy and effectiveness of SAR image classification and fine-scale object extraction.To this end,this thesis focuses on improving the accuracy and effectiveness of SAR image classification by combining image features and prior information of the concerned categories(such as spatial contexts,target structures and shapes,etc.).Based on the state-of-the-art research in SAR image classification,this thesis aims on several unsolved key problems in this field,including impact of serious noise on categorization,effect of incomplete object structures on fine-scale categorization and target extraction,as well as difficulties in obtaining the ground truth of samples under a large-scale dataset.By introducing label prior information in SAR image classificationin,the classification accuracy is effectively improved and the cost of training set definition reduced.In this thesis,the contents of label prior information in classification and its modeling methods based on conditional random fields(CRFs)is introduced.Then,characterizations and models for prior information with special focus on spatial contexts,object structure priors and label proportions,as well as their applications in SAR image classification are described.The main contents of this thesis cover spatial context-based SAR image classification,object structure prior-based SAR image classification,and label proportion-based SAR image classification.Those methods are expected to provide better supports in large-scale SAR image understanding.Spatial context is the basic form of label prior information.In spatial context-based SAR image classification,this research concentrates on introducing the local and global presentation of label context.A CRF model is built to combine low-level features and context information,and efficient inference and learning methods are proposed to improve the classification results for SAR images with noises.On the object layer,the spatial contexts appear in form of object structure priors.For object structure prior-based SAR image classification,this thesis proposes a joint model for low-level features,spatial contexts,and object structures.CRFs-based inference and learning for object structures are studied.With the proposed method,fine-scale classification and segmentation results are improved for SAR images with noises and object deformation and part missing.Label proportion is an implicit form of label prior information in spatial distrubutions,which can be applied in weak sample labeling and corresponding classification methods.For label proportion-based SAR image classification method,this research introduces the grid labeling method with label proportions and propose the classification framework for SAR images based on grid labeling.A weakly-supervised SAR image classification is implemented based on the weakly-labeled training set from grid labeling,improving the efficiency and cost of sample labeling.The modeling and application methods for prior information improves the performance of high-resolution SAR image classification and solves a few key problems of SAR image classification in practical applications.Still,certain aspects of this field remain to be explored in depth,such as extending the spatial context model to the analysis of object groups and exploring a multiscale model for object structures.