多核学习算法及其在高光谱图像分类中的应用研究进展

Multiple Kernel Learning Algorithm and its Application Research Progress in Hyperspectral Image Classification

多核学习算法在高光谱图像分类领域占据着十分重要的地位。与灰度图像、全色图像和多光谱图像等相比,高光谱图像因具有很强的分类识别能力等多方面优势而被广泛应用。为进一步提高高光谱图像的分类精度,促进多核学习算法在高光谱图像分类中的应用,本文对多核学习算法及其在高光谱图像分类中的应用进行了总结。首先在回顾核方法的基础上阐述了多核学习框架,其次对多核学习核函数组合方法进行综述,随后根据求解多核学习组合系数方法的不同将多核学习分为两类:固定规则的多核学习算法和基于优化的多核学习算法,并对两类多核学习算法在高光谱图像分类中的应用进行综述,总结各类算法在高光谱图像分类的应用进展。同时,为了便于研究者对多核学习算法及其在高光谱图像分类问题中的应用研究,本文对常用核函数和高光谱图像数据集进行了整理归纳。最后,讨论了多核学习算法在高光谱图像分类研究方面的不足,并对未来研究方向进行了展望,以期为该领域的研究和应用提供参考。

Hyperspectral images have been widely used in target detection, spectral decomposition, classification,and many other fields. They have higher recognition ability than grayscale images, panchromatic images, and multispectral images. However, how to effectively use hyperspectral images with large number of bands, huge data volume, and increased information redundancy is an important topic. Multiple kernel learning is a typical multi-view learning method that can make different kernel functions according to different feature spaces and group multiple kernel functions into an optimal kernel function for hyperspectral image classification. Compared to single kernel method, multiple kernel learning has unique advantages in solving problems such as the uneven spatial distribution of high-dimensional features, using information more efficiently, and improving classification accuracy greatly. At present, the research difficulty of multiple kernel learning algorithm are the combination of kernel functions and the selection of optimal weight coefficients. In order to improve the classification accuracy and promote the application of multiple kernel learning algorithm in hyperspectral image classification, we review the development history and current research progress of multiple kernel learning algorithms. First, the kernel learning method and the framework of multiple kernel learning algorithm are introduced. The specific methods of kernel function combination used in multiple kernel learning algorithm are summarized. According to many researches, it can be concluded that the linear method has been widely used because of its simplicity and efficiency. Moreover, according to the methods of determining weight coefficients in multiple kernel learning algorithm combination, multiple kernel learning algorithms can be generally divided into two categories: fixedrule multiple kernel learning algorithm and optimization-based multiple kernel learning algorithm. Then, the applications of different multiple kernel learning algorithms from each category in hyperspectral image classification are reviewed. In order to facilitate researchers to discuss the problems of multiple kernel learning algorithm and hyperspectral image classification, the commonly used kernel functions and the widely used data sets in hyperspectral image classification are also reviewed. Finally, we discuss the deficiencies of multiple kernel learning algorithms in the field of hyperspectral image classification and point out the future research direction to help solve practical application problems.