Spatially constrained technology applications in information extraction from remote sensing images
The problem of high intraclass variance is apparent in Very High spatial Resolution(VHR)remote sensing images.This problem limits the performance of many remote sensing information extraction methods.Consequently,Spatial Constraints(SCs)within image pixels have become a hot topic,resulting in many research results,but they lack associations and systems orientation from a general perspective.This study reviews and summarizes more than 100 related studies published in the past two decades to provide references for further research on information extraction in VHRs.This study has four sections:In the first section,the three stages of the SCs process(mining and expression of spatial information and construction of the SCs)are described in detail.The primary sources of spatial information were the neighborhood of pixels,imaging relations,and prior knowledge.The spatial information included the mean,median,extreme,and azimuth order.The SCs construction methods included objective functions,energy functions,and discriminant functions.In the second section,the SCs applications are divided into six scenarios(image matching,image segmentation,target detection,image classification,change detection,and others),and the implementation methods and characteristics of the main application scenarios are summarized.The SCs method is closely related to the specific application of the material.For example,SCs is mainly used to build descriptors and perform transformations in image matching;is implemented by model constraints,graph construction in space,and objective functions in image segmentation,target detection and image classification;and emphasizes the neighborhood between pixels and prior knowledge in change detection.The common feature of these scenarios is the development of a robust,unique,and representative descriptor via geometric space information,which can solve specific problems in images.In the third section,the SCs methods are divided into six types according to their implementation and principles(local templates,auxiliary references,spatial graph construction,model constraints,rule constraints,and others),and the advantages and disadvantages of the first five methods are compared.The results showed that the different SCs methods exhibited varying usability across application scenarios.(1)A local template uses the spatial information of the neighborhood and obtains more instances of stable information expression;thus,this approach is suitable for many application scenarios,especially image classification.(2)The point constraint in the auxiliary reference method relies on the spatial relations between feature points and often appears in image matching,while line constraints focus on the connection between the target and the linear object.Thus,this approach is suitable for extracting anthropogenic objects.Furthermore,surface constraints are spatially extensible and suitable for target detection.(3)Graph construction in space can intuitively and effectively extract multidimensional spatial information and is suitable for classifying hyperspectral images.(4)Model constraints are generalized in practical applications but rely on specific mathematical expressions.(5)Rule constraints can specify professional applications and are often used in image classification and change detection.Fully analyzing and considering application scenarios and specific problems are necessary for ensuring the effectiveness of SCs tools.In the fourth section,the development trends and possible shortcomings of SCs research are discussed.Specific suggestions for future work are also provided.
NETL
NSTL
万方数据
