Knowledge-based Remotely Sensed Imagery Classification Method Research

This dissertation concerns with the knowledge-based remotely sensed imagery classification in order to improve the accuracy of landuse/cover classification in arid and semiarid region. At first, the research establishes the landuse/cover classification schemes of the study area. Then according to the principles of interpretation, it analyzes the characteristics of landuse/cover types, such as spectrum, texture, spatial distribution and times. In succession, it discusses the expression of knowledge and establishment of knowledge base. At last, based on research before, we perform the knowledge-based remotely sensed imagery classification and evaluate the result of classification. The conclusion of this paper is as followings:(1) Knowledge-based remotely sensed imagery classification method is the one approach of improving the classification accuracy of landuse/cover in arid and semiarid region. Although, the landuse/cover in study area is complicated, and being lack of validated data as the same time, there is a litter warp in accuracy evaluation, as a whole the accuracy of knowledge-based remotely sensed imagery classification method is improved compared to traditional supervised classification method.(2) In arid and semiarid region, the texture, spatial distribution and times features can improve the accuracy of classification. Sand and desert that are distinct in modality can be differentiated by texture. Shrub and forest, grass and infield which are distinct in spatial distribution, can be differentiated by spatial characteristics. Infield which is uncovered can be differentiated from other deserted glebe by times characteristics.(3) The knowledge-based classification method can effectively synthesize all kinds of assistant information to improve the accuracy. But the quality of assistant data and the match precision between the data also may affect the accuracy. So we must pay attention to the assistant data selection and the rigorous match between the different data in the practical application.