A Geographic Ontology-driven SVM Approach For Multi-source Remote Sensing Image Classification And Change Detection

Recent years,with the rapid development of remote sensing technology and unmanned aerial vehicles(UAV)technology conditions,it is more and more easily to obtain multi source highresolution aerial / Satellite images.In the era of remote sensing image big data,remote sensing image classification and change detection are an important research topic in the field of remote sensing.Remote sensing image classification and change detection have a very practical value in many fields,such as geographical conditions monitoring,National Geographic National Census and related industries updating the database by remote sensing image classification and change detection technology.But at present the application in related field depends on mainly manpower,complete formal description is not unified on the classification of knowledge,and does not realize the knowledge sharing of classification and change detection.Although the object-oriented image analysis technology has become a new paradigm of remote sensing image information extraction,but the lack of conceptual and formal description for the surface elements always exists in objectoriented image classification and change detection process,and the lack of essential objective understanding for the geographic elements.Therefore,it is an important research topic in the field of remote sensing that how to organize well the top-level classification information and the lowlevel feature information.In this paper,using multi-source remote sensing image as the data source,we study the remote sensing image classification and change detection,and the geographic ontology theory is introduced in remote sensing image classification,by means of the ontology we can share the conceptual model of clear formal description.Based on the big data analysis software,the information of the image object features is analyzed,and the feature map is constructed by ontology.Followed by establishing remote sensing image surface information conceptualization and formalized expression,image object classification characteristics and classification rules information are effectively organized and formalized,achieving the high-level semantic information classification and the low-level features of information communication,eliminating the semantic gap between the two.The classification effect of classifier is improved by combining ontology knowledge and machine learning algorithm.In the study area,the ontology driven remote sensing image classification and change detection experiments are carried out,the main contents include the following aspects:(1)Image object feature analysis based on random forest.The random forest has the ability of feature selection,so we comprehensive analysis the image object spectrum,texture and geometric features using the random forest,by mean of the big data analysis and prediction software SPM(Salford Predictive Modeler)calculates the importance score of image object features,the whole forest balance error rate,out-of-bag data error,and other parameters.Then,the parameters are analyzed comprehensively,and the selection of classification features and the threshold of feature separation are achieved.According to the results of feature selection and the analysis of the importance of features,the image classification rules are constructed.(2)With a high spatial resolution and high spectral resolution airborne / satellite image data and LIDAR data as the data source,we create the remote sensing image classification knowledge ontology model in conceptualization and formalization way,image classification ontology model including image object features ontology model,classification rules ontology model and the cover classification ontology model.On this basis,a method of ontology driven SVM(Support Vector Machine)classification is proposed.Compared with the ordinary SVM classification,this method can improve the classification accuracy of SVM classifier by combining the classification ontology knowledge with the SVM classifier.(3)Based on the classification results obtained with high accuracy,combined with the base vector classification data the after-classification object change detection is carried out.The change patch of the study area is obtained,and the transfer matrixes of land type change are calculated.by mean of the transfer matrixes the change flow information can be more intuitive understand and master.Compared with the results of the artificial change detection,the classification object change detection can reach the level of artificial detection,but the speed is faster than the manual detection.Giving some help to update the outdate database.