Feature Fusion Algorithm For High Resolution Remote Sensing Image Classification Based On Canonical Correlation Analysis

High-resolution remote sensing image information extraction and intelligent interpretation has always been one of the hot topics in the field of remote sensing image processing. The characteristics of color, structure and texture of images which reflect the essence of objects from different angles are widely used in various classification algorithms. How to eliminate the irrelevant and redundant information among the multiple features has attracted the attention of multinational scholars. In general, the use of multiple features of remote sensing images for scene classification will inevitably lead to a dramatic increase in redundant information, irrelevant information, and feature dimensions, which will inevitably lead to a large number of irrelevant or redundant In addition, a large number of ineffective features can be used to engulf the main classification features with strong discriminative characteristics, which will have a great impact on the classification performance. Therefore, multiple features fusion is an important research area in the field of remote sensing image scene classification. The purpose of feature fusion is to obtain meaningful low-dimensional image representation and establish multiple mappings in high-dimensional multi-feature space. The basic idea is to combine the different features of the same remote sensing image in order to construct a new feature space, and eliminate the invalid information in the new feature space by the method of correlation analysis, so as to promote the subsequent classification work smoothly. By using the complex vectors to fuse two sets of features, or connecting feature vectors end to end to get a new set of high-dimensional vector directly, although these methods can improve the classification accuracy to some extent. But the multi-representations of the data and the intrinsic correlation between the features have been greatly damaged. Hence, it is difficult to effectively fuse multiple features. For this reason, how to construct the multiple features fusion algorithm by using the canonical correlation analysis is an urgent problem to be solved in the high-resolution remote sensing image scene classification. In this thesis, we study the multiple features fusion algorithm based on canonical correlation analysis. We will introduce effective correlation criterion function, further optimize the projection bases and find a new model learning method as the starting point of the research work. A kind of universal high-resolution remote sensing image feature fusion classification framework is designed, such that the effective delivery and coordination can be achieved between the different types of remote sensing image features, so as to achieve automatic scene classification and interpretation of high-resolution remote sensing imagery, and then contributing to the smooth transition between the remote sensing data and the specific application field. At the same time, avoiding the problem that traditional methods only aim at one specific application scenario. In this way, the application scope of the algorithms will be widened in a significant degree, meanwhile the value of practical application of the algorithms will also be improved.For maximizing the preservation of the global geometric structure of the remote sensing image data, this thesis introduces an effective method to establish the correlation criterion function. In the meanwhile, we have implemented the scene classification experiment based on the multiple feature fusion framework. The feature fusion framework can solve the problem of multiple feature fusion of high-resolution remote sensing data. It should be noted that it is a new idea of multiple feature fusion in remote sensing data processing domain. From the point of view of data mining, the framework is not only able to express the correlation information between different remote sensing features, but also have the ability to explore the intrinstic geometry information of the data. In the high-resolution remote sensing image data adaptation stage we found some problems existing in this framework, and for these problems give the corresponding solutions in the subsequent research.This thesis presents a new method combined with canonical correlation analysis and a bayesian information criterion based smoothly clipped absolute deviation penalty function. The method exploits smoothly clipped absolute deviation penalty function to optimize canonical correlation analysis, and then puts the bayesian information criterion to use for obtaining sparsing projection bases. It is important to note that this method provides a new direction to sparse the projection bases of canonical correlation analysis. It is important to note that this approach can not only directly control the sparsity of features, but also can preserve Oracle Property to some extent. In the experimental phase, we verified the effectiveness of the algorithm.For unsupervised property of canonical correlation analysis algorithm, this thesis introduces a semi-supervised classification method, which the main idea is to use label propagation algorithm for establishing the graphical model of the sample data to infer label information of unlabeled data by utilizing given label information. Meanwhile, the human intervention is reduced and the efficiency and accuracy are enhanced in the task of remote sensing image scene classification after introducing the thought of semi-supervised learning.