Information Mining Of Typical Landscape Types Of Remote Sensing Images Based On Heterogeneous Kernels

The distribution of ground objects has a high degree of complexity and spatial heterogeneity,so it is difficult to quantitatively describe its inherent laws.Remote sensing observation provides basic data and technical support for understanding and recognizing the distribution law of ground objects,especially the finer spatial features and structural information in high spatial resolution remote sensing images provide more detailed and sufficient basis for the extraction and type identification of ground features.At present,remote sensing data processing is more focused on the spectral information features of a single pixel,but the spatial structure information mining carried by the pixel and its neighborhood pixels is not well utilized.Spatial heterogeneity is an important feature of the spatial distribution of ground objects.In contrast,traditional remote sensing image analysis and information extraction methods are not fully considered.In recent years,some studies have tried to define the spatial heterogeneity index and spatial heterogeneity kernel to describe and quantify the spatial heterogeneity of the central pixel relative to its neighborhood.Although the original spatial heterogeneous core can better reflect the local characteristics,it is difficult to carry out effective statistical analysis because of the influence of gray resolution and image amplitude.In this paper,global and local multivalued methods are introduced to limit the number of types that produce spatially heterogeneous cores from the aspects of direct reduction of gray resolution and type resolution of heterogeneous cores.In this study,six landscape types of plain cultivated land landscape(North China Plain),Hengduan Mountains landscape(Guangxi),urban landscape(Xi'an),Loess Plateau landscape(Yan' an),desert landscape(Tarim Basin)and karst landscape(YunGui Plateau)were selected for spatial feature information extraction and comparative analysis.The results show that,while the theoretical value of heteronuclear species is increasing,there is little influence on the number of heteronuclear components and no new heteronuclear pattern appears.Therefore,in the spatial feature extraction of six landscape types,the local multivalued method is better than the global multivalued method.It is true that the larger the number of spatial heterogeneous nuclei,the greater the potential to extract the spatial structure features,but the higher the corresponding computing cost and the more serious the limitation of the image space range.Therefore,this study focuses on the local binary mode LBP.According to the difference between the heterogeneous core category probability of completely random distributed images and the heterogeneous core category probability of specific images,it tries to propose a randomness method to measure different landscape images,and further compares and distinguishes the heterogeneous core at the image edge and the heterogeneous core at the non-edge.The results show that the proportion of heterogeneous nuclei in the non-marginal areas of the images of six typical landscape types is close to the theoretical value of random distribution,which indicates that the heterogeneous nuclei in the non-marginal areas are mostly generated by the gray value of random distribution.On this basis,a local binary mode TLBP method considering the threshold value is further improved and developed in this paper.The results show that,compared with LBP,TLBP method can significantly improve the difference between different landscape heterogeneous-spectrum.The spatial heterogeneous kernel and its related methods can reflect the local spatial heterogeneity to a certain extent,extract the spatial structural features of typical landscapes,and provide more scientific basis for further identification and classification research.