Application Of Graph-based Semi-supervised Classification Method In Remote Sensing Images

Remote sensing image contains a high spectral resolution and a great deal of geographic information, which has an important application in the field of ecological, geological exploration and precision agriculture. The classification of remote sensing images is one of the important technologies of remote sensing image analysis processing. Because of convexity, Graph-based semi-supervised classification method which is based on the labeled samples and unlabeled samples classifying data has good classification performance. Therefore, it has obtained many attention and study. However, in real application, for the large data sets, the graphical models which is constructed based on graph theory is so complexity that the computer memory overflows. The appropriate number of nearest neighbors is beneficial to obtain the manifold of data. If the number of nearest neighbors is too small, we can't get more useful information. On the contrary, if the number is too large, there is much interference information. Therefore, it is significant to adaptively define the number of nearest neighbors.To solve the above problems, this paper studies the semi supervised classification algorithm based on the anchor map building, and presents the corresponding improvements. The main work of this paper is as follows:1. The basic theory and the current research situation of semi-supervised classification are given. The theoretical basis of algorithm for semi-supervised classification, the classical algorithm, and the current status of the research are described. At the same time, the disadvantages and difficulties of the semi supervised classification algorithms are presented.2. The multi-spectral remote sensing image is more complex, and the boundary between different objects is not obvious. One object has different spectrum and different objects have the same spectrum. Therefore, it is very difficult to classify the remote sensing image with large size. To solve the problem, this thesis combines the mean shift algorithm and semi supervised classification algorithm based on anchor construction graph and utilizes them to the large-scale remote sensing image classification. The presented algorithm firstly obtains clustering centers to form the anchor set by mean shift clustering and deduces the graph by the anchor and the labeled sample. Then, the anchor classification is obtained by semi supervised classification algorithm. Finally, the anchor sets are mapped to the whole image data sets, and remote sensing image classification results are obtained. The experimental classification results show that the presented method can improve the classification efficiency and classification precise.3. During the cause of anchor construction graph, the improper parameters of neighbor anchor points affect the classification accuracy. To solve the problem, we propose an adaptive selection algorithm for the number of nearest neighbor anchor. Objects in remote sensing images are more complicated, and some categories stagger serious. When the neighbor number is too small, we only obtain a little useful information. On the contrary, the nearest neighbor number is too large, we maybe obtain the interference information, which can affect the remote sensing image classification accuracy. In this paper, based on the relationship between the measuring distance and the Euclidean distance, we present an adaptive selection algorithm for the number of nearest neighbor anchor. Experimental results tell that the proposed algorithm can improve the classification performance of the remote sensing image.