The Technology Of Geo-ontology Modeling Driven Object Classification For Remote Sensing Image

Remote sensing image classification has been a hot and difficult problem in the field of international remote sensing research. It experienced from visual classification to semi-automatic classification, now it is developing towards automation and intelligent direction. It has the problem of "attaches great importance to technology, ignores knowledge, and lacks of ontology", which reduces its degree of automation and intelligent, and limits its engineering application.In order to solve the problem, this study takes the mainstream GEOBIA (Geographic Object-based Image Analysis) as example, and builds a geo-ontology classification framework for remote sensing image from the perspective of geo-ontology theory.GEOBIA has been the main technology of remote sensing image classification. It is not only a hot topic of current remote sensing and geographical research, but also becomes a new development paradigm in the field. It also has the problem of "attaches great importance to technology, ignores knowledge, and lacks of ontology". Specifically, the lack of a systematic approach designed to conceptualize and formalize the classification elements makes it a highly subjective and difficult method to reproduce. The introduction of geo-ontology provides a new idea and brings a real revolution for remote sensing image classification. Geo-ontology has obvious advantage to solve the problem with the characters of conceptualization, explicit, formalization, space-time and sharing. The image features and expert knowledge could be expressed formally using geo-ontology, which could reduce the semantic gap between low-level features and high-level semantic features, and solve inconsistency caused by different expert knowledge. Ontology model is established by simulating human perception and in the form of operable computer language, thus achieving automatic classification of remote sensing images.Therefore, this research puts forward a geo-ontology framework of "geographical entity concept ontology description - image classification geo-ontology modelling - geo-ontology driven object classification" for remote sensing image classification. Under the guidance of this framework, the knowledge of geographic entities is described, and the geo-ontology classification model is constructed, and the geo-ontology driven object classification methods are realized, which provides theoretical basis, model and method for remote sensing image classification. The main contents and conclusions of this thesis are as followings:1. A geo-ontology framework of "geographical entity concept ontology description - image classification geo-ontology modelling - geo-ontology driven object classification" is put forward for image classification. In view of the problem of "attaches great importance to technology, ignores knowledge, and lacks of ontology". Specifically, image classification is highly subjective and hard to repeat as the lack of the modelling of every element. The framework uses the geo-ontology to link subjective and objective knowledge to realize a unified cognition and objective description of geographical entities, which avoids inconsistent problem caused by different interpretation experience of different experts. All kinds of knowledge are organized organically and expressed explicitly using semantic network model, and semantic relations are expressed in the formal language that the computer could operate. Image objects are classified based on geo-ontology and using data-driven machine learning methods and knowledge-driven expert rules. It could supply unified, standard and scientific overall framework, objective model and new methods for remote sensing image classification, and change uncertainty to certainty, and raise the scientificity and unity of image classification, and promote its automation development and engineering application.2. The concept and knowledge system of geographical entity are built to meet the needs of geo-ontology modeling. In the face of the requirements of the objective modelling of remote sensing image classifcation, the domain knowledge of geographical entity is summarized from four aspects of geographical knowledge, image characteristics, object features, and expert knowledge. The conceptual ontology framework of domain knowledge is built, which could improve the objectivity of geographical entity, and ensure the consistency of geographical entity knowledge, and avoid inconsistent results caused by different experts' knowledge. Take land cover entity for example, whose domain knowledge is summed up, and whose conceptual ontology expression pattern is built, which lays knowledge foundation for geo-ontology modeling of remote sensing image classification.3. The image classification geo-ontology models are built. On the basis of constructing concept system and knowledge system of geographical entity, the ontology model of remote sensing imagery, image object and classifier are built by use of OWL (Web Ontology Language). The ontology models of two typical classifiers such as decision tree and expert rule are built. The models are expressed and edited with Protege software which is developed by Stanford University. At last the entire semantic network model is built, which lays the model foundation for object classification based on geo-ontology.4. Four levels of image object classification method based on geo-ontology is proposed.1) The construction of geo-ontology model for image classification.2) A parallel segmentation method based on Graph theory and Fractal Net Evolution Algorithm (FNEA) is proposed.3) A feature selection method based on Random Forest is proposed.4) An object-based semantic classification method based on semantic network model is proposed. Experiments show that this method could not only obtain the classification result and semantic information of geographical entity, but also realize the reuse of the domain knowledge in the field of geographical entity, and promote the fundamental change of data driven approach to knowledge driven approach.5. The land cover classification tests for Geography Census are carried out. As for fieldland, orchardland, woodland, grassland, building, road, water, bareland of land-cover, take RuiLi city of YunNan and LinTong city of ShanXi as study areas, ZY-3 and WorldView-2 high resolution imagery are the main data source, FeatureStation_GeoEX software developed by Chinese Academy of Surveying and Mapping (CASM) and Protege software developed by Stanford University and are the main test platform. Using the framework, models and methods for land cover classification test, the theory and results show that, this study realized objective modeling of various elements, image semantic classification and sharing of domain knowledge. The proposed framework has the characteristics of objectivity, clarity, extendibility, scalability and so on. It is able to provide unified, standard and scientific overall framework, objective model and new methods for remote sensing image classification.