Study On The Construction Of Geo-ontology For Land Cover Classification Of Remote Sensing Image

Land cover classification of remote sensing images is a hot of international research.Object-Based Image Analysis is the main and potential technology of remote sensing image interpretation.The lack of a systematic approach designed to conceptualize and formalize the interpretation elements makes it a highly subjective and difficult method to reproduce.Geo-ontology has obvious advantage to solve the problem with the characters of conceptualization,explicit,formalization,and sharing.Based on the land cover secondary classes of geographic national conditions,from the perspective of geo-ontology theory,land cover classification rules are constructed by combination of knowledge driven and data driven,and object-based classification method based on geo-ontology is proposed,and land cover classification experiment is carried out.Specific research contents and achievements are as follows:(1)Land cover classification rules are constructed by combination of knowledge driven and data driven.Based on the land cover secondary classes of geographic national conditions,combining with the basic process of the high resolution remote sensing image classification,the land cover classification rule set is constructed by combining the deep knowledge and the shallow statistics.(2)The geo-ontology model of land cover classification is built.The ontology model of land cover type,image object feature and classifier is constructed in protégé software,and the reasoning machine is finally used to reason the geography ontology,testing logical consistency.(3)Object-based classification method based on geo-ontology is proposed.It is divided into four levels: geo-ontology modeling of land cover classification,image segmentation,random forest feature selection,image object classification.(4)Land cover classification experiment is carried out.In this experiment,We have chosen 13 kinds of land cover,which includes artificial grassland,green forest,low independent building,urban road,pond,dry land,graff,multi-storey building,low rise building,clay surface,highway,shrub forest and orchard of Land-cover.Potsdam of German and LinTong city of ShanXi are taked as study areas and orthophoto and WorldView-2 high resolution imagery are the main data source.The software platform for conducting experiment use FeatureStation_GeoEX software,Protégé software and SPM(Salford Predictive Modeler)in which the models and methods for land cover classification are test.According to the experiment and results,we can conclude that the method provided in the thesis has realized accurate modeling of remote sensing image classification elements and sharing of domain knowledge,and the validity and universality of the research method is verified,which can be used to provide a shared and scalable geo-ontology for land cover classification of remote sensing images.