高光谱遥感图像数据降维和分类算法研究

发布时间：2018-01-07 09:33

本文关键词：高光谱遥感图像数据降维和分类算法研究　出处：《北方民族大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着遥感技术的发展,高光谱遥感作为一种新型的遥感方式已经成为人们获取地表信息的重要手段,并在军用和民用领域发挥着极其重要的作用。高光谱图像分类是高光谱信息处理的重要方向,高精度的分类算法是实现各种应用的前提。但是,高光谱图像以其高分辨率、多波段数、大数据量的特点给传统图像分类技术带来了巨大的挑战。针对高光谱遥感图像分类存在的问题,本文主要研究高光谱遥感图像数据降维和分类算法。随机森林是Breiman于2001年提出的分类和预测模型,作为集成学习的一种,随机森林算法充分利用了单个分类器和统计抽样技术,解决了单分类器在性能上无法提升的问题,但其本身也存在一些有待完善的地方和一些应用上的缺陷,在算法处理不同数据集的性能和算法分类精度的提升方面,还存在着一定的空间。根据随机森林中的不足,本文提出改进型算法,着重从数据降维方面入手进一步提高随机森林的分类精度。前期采用高效的降维算法选择出具有信息量大、相关性小和类间可分性强等特点的特征子集,而后在特征子集空间进行分类。这一分类过程不仅可以提高地物识别分类的效率,同时处理数据的速度也大大改善。本文算法具体实施方案分两步,第一步构造特征空间,将投影寻踪和主成分分析相结合得到降维后的特征,形成新的特征组合。主成分分析主要揭示了数据的全局线性分布特征,代表了原始数据的整体分布方向,而投影寻踪则揭示数据的局部分布规律。全局与局部的结合会使高光谱遥感数据的整体分类精度得到提高。第二步是在新的特征空间里利用随机森林进行地物分类。实验证明,与其他分类算法对比,本文算法有明显的优越性。本文通过对随机森林算法的研究,针对该算法中的不足提出了一种改进型算法,不仅降低了样本数据维数、保留了主要的信息量也提升了随机森林的分类精度。最后,文章分析了本文算法在高光谱遥感图像分类上的优势与不足之处,为今后我们进一步的研究打下了坚实的基础。
[Abstract]:With the development of remote sensing technology, hyperspectral remote sensing as a new type of remote sensing has become an important means for people to obtain surface information. Hyperspectral image classification is an important direction of hyperspectral information processing, and high-precision classification algorithm is the premise of various applications. Because of its high resolution, multi-band number and large amount of data, hyperspectral image has brought great challenges to the traditional image classification technology, aiming at the problems of hyperspectral remote sensing image classification. This paper mainly studies the algorithm of reducing dimension and classifying hyperspectral remote sensing image data. Stochastic forest is a classification and prediction model proposed by Breiman in 2001 as a kind of ensemble learning. The stochastic forest algorithm makes full use of single classifier and statistical sampling technology to solve the problem that the performance of single classifier can not be improved, but there are still some defects in its application. There is still some space to improve the performance of different data sets and the classification accuracy of the algorithm. According to the shortcomings of the random forest, this paper proposes an improved algorithm. The classification accuracy of random forest is further improved from the aspect of data dimensionality reduction. In the early stage, the efficient dimensionality reduction algorithm is used to select the feature subset with the characteristics of large amount of information, small correlation and strong separability among classes. The classification process can not only improve the efficiency of object recognition and classification, but also greatly improve the speed of processing data. This algorithm is implemented in two steps. The first step is to construct the feature space, combining projection pursuit and principal component analysis to get the reduced dimension feature, and form a new feature combination. The principal component analysis mainly reveals the global linear distribution characteristics of the data. Represents the overall distribution direction of the raw data. Projection pursuit reveals the local distribution of the data. The combination of global and local will improve the global classification accuracy of hyperspectral remote sensing data. The second step is to use random forest to divide ground objects in a new feature space. Class. Experimental proof. Compared with other classification algorithms, this algorithm has obvious advantages. Based on the research of stochastic forest algorithm, an improved algorithm is proposed, which not only reduces the dimension of sample data, but also reduces the dimension of sample data. The main information is retained and the classification accuracy of random forest is improved. Finally, this paper analyzes the advantages and disadvantages of this algorithm in hyperspectral remote sensing image classification. It has laid a solid foundation for our further research in the future.
【学位授予单位】：北方民族大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】