Research On Content-based Image Mining Methods For Remote-sensing Images

With the development of remote sensing technique and storage technique, there is a great number of spatial data, much of which is remote-sensing image data. However, the use efficiency of the huge quantities of remote-sensing image data is still low. It is very difficult for people to process with thousands of image data and find out knowledge from them. So people only can deal with remote-sensing image which is still in a low lever. With the date mining, information retrieval, multimedia database and other correlative field developing, it becomes possible to manage and analyze remote-sensing images and find out useful information to people. The paper puts a concept frame of the remote-sensing image mining for the problem, and proposes two content–based remote-sensing image mining methods: semi-supervised improved fuzzy c-means clustering to remote-sensing image and interactive learning-based image mining in remote sensing. The mostly work is as follow.(1)The paper puts a concept frame of the remote-sensing image mining. The remote-sensing images are different from ecumenical images, so according to these feathers we generalize the concept of remote-sensing image mining and put forward the general process and the hierarchy frame.(2)The paper proposes a semi-supervised improved fuzzy c-means clustering to remote-sensing image. Because there are the characteristics of uncertainty and mixing meta-pixels in the remote-sensing images, the classical fuzzy c-means method has a low accuracy in the remote-sensing image clustering. This paper improves the fuzzy c-means method. And it adds a-priori information into the patterns to change the method as a semi-supervised clustering. In the clustering process, the unlabelled patterns compare similarities with the labeled patterns, and then the accuracy of the algorithm can be increased.(3)The paper proposes an interactive learning-based image mining in remote sensing. We discuss a simple hierarchical modeling in remote sensing image mining according to KIM. It divides image mining to three parts, extraction of primitive image features, unsupervised clustering and interactive learning. It supports the man-machine interaction, from low level of primitive image features to get high level of logical features, and then studies the training data via man-machine interaction, finally mines knowledge and model needed. The simple hierarchical modeling is suitable for different application in remote sensing image and for the development of remote sensing technology.Finally, I have done experiments to validate the remote-sensing image mining method that this paper brings forward, and the result is satisfying.