Task-Oriented Remote Sensing Images Retrieval Based On Case-based Reasoning

Remote sensing information as an important strategic resource of protecting state security and promoting national economy construction, plays an irreplaceable role in macro decision of agriculture, disaster reduction and many other problems. With the rapid development of earth observation technology, remote sensing data sources will be continuously enriched, one of the key issues is how to access remote sensing data quickly to meet the different needs of various users. At present, the remote sensing image acquisition depends mainly on professional query subscription service or spatial information portal, which require user to submit query that contains the professional remote sensing information qualitatively or quantitatively on different detailed levels. However, the lack of semantic query ability of different specific application field is still an obstacal in remote sensing information retrieval.Therefore, the task-oriented remote sensing image retrieval is put forward, aimed at retrieving remote sensing images by task, which simplify the acquisition mode, and improve the service level. However, the relationship between the task and remote sensing images is hard to abstract into a general rule, due to the complexity of the spatial and temporal geographic environment between them. Thus, this paper introduces the case based reasoning technique (Case-based, Reasoning, CBR), which hidden the complex space-time relationship between the task and remote sensing images into cases, and mining the relationship by analogy reasoning. The main research work and results are as follows:(1) The task-oriented remote sensing image retrieval method based on CBR is proposed. This method use case to express the complex space-time relationship between task and remote sensing images, which is difficult to abstract into a rule. Analogical reasoning is induced to obtain the relationship in cases. And the case adaption is involved to adjust retrieval cases to meet the user query. This method is better than the traditional method that based on rules and ontologies, as the knowledge source is more abundant and easy to obtain, and more easily to express special exceptions.(2) The semantic remote sensing image application case representation model is introduced, whch mainly contains four elements that are time, spatial, task and remote sensing image that satisfiy the task. The elements in model are defined and described by ontology, and the feature of each type of the elements is modelled. The spatial-temproal relationship between cases and elements are also establised. The concept model and the description model of different two levels are defined in the model, while the conceptual model is used to describe the nature of the concept of object properties and relationships, and description model is used to decribe attributes and relationships of concept on the natural language level. Description model can be transformed to the conceptual model by semantic reasoning, which facilitates the case aquicisition and ensure the sufficient expression and reasoning ability.(3) The case retrieving model based on the deep semantic understanding is adopted, which including the similarity measure model and indexing model. The local similarity model of time, spatial, task elements, and the gloable case similarity measure model are established. The time semantic similarity measure model is based on time structure and task characteristics. The spatial semantic similarity measure model is based on the spatial relationship contact strength of geospatial objects and feature vector of different fetaure types, and the spatial environment. The task semantic similarity measure model is based on attributes. An extending case retrieval network that considering spatial characteristics is also proposed, which involves virtual space and time index node. The spatial index node is implemented by a two-level R tree and time index node is implemented by the inverted structure. The vitrual case node is also extened in CRN, which enables performance on top-k retrievl, and supports different users'background of dynamic weighting of case elements.(4) The case adaption model based on the knowledge is produced. The model use existing knowledge on one hand, such as time, space and sensor ontology, on the other hand the case base itself is a training set, which is used to mining difference adjustment rules and high-frequency rules for case adaption. The paper mainly fouces on the expression and learning process of difference adjustment rules. A differential informaiton expression method based on cocept features is provided. And a rule generalization criterion is establisted considering spatial and semantic features. Meanwhile, a rule network based generalization algorithm is introuced. According to the remote sensing image query requirements, the object-oriented application of adaption knowledge which takes query and sort into account is established.(5) The task-oriented remote sensing images retrieval prototype system--iGeoPortal is developed based on previous research. The validation is carried out based on iGeoPortal to verify the method proposed in this paper. At the same time, the natural language processing and semantic inference model is established to realize the transformation from description model to concept model. A query driven method for placename description and a fast intersection method for fuzzy spatial descrition is also developed. Preliminary experiments show that the method proposed in this paper is effective and feasible.