基于松驰聚类假设的高光谱数据分类研究
发布时间：2018-11-23 09:52
【摘要】：近年来，随着遥感技术的迅猛发展，人们可以获取大量的高光谱数据，，如何根据这些数据进行准确的地物分类，是遥感数据应用水平提高的关键。目前已有大量机器学习算法用于高光谱数据的分类，这些方法基本上都基于聚类假设，即光谱相同/相似的数据具有相同/相似的标签。但是，由于传感器噪声以及成像设备分辨率的限制，实际高光谱数据中存在严重的“同物异谱”和“同谱异物”现象，严重制约了聚类假设下分类器的性能提高。针对这一问题，本文将松弛聚类假设思想引入到高光谱数据分类中，与稀疏编码分类器和SVM(Support Vector Machine)分类器结合，发展了松弛聚类假设下的高光谱数据分类方法。主要研究工作和创新包括以下内容： (1)设计了一种松弛聚类假设下的高光谱数据稀疏编码分类算法。将松弛聚类假设思想以概率向量的形式应用到稀疏编码模型下，设计了相应的优化目标函数与优化算法。在实际高光谱数据集上进行的仿真实验结果表明：松弛聚类假设能够大大改善由于混合像元引起的数据分类正确率不高的情况。 (2)设计了一种松弛聚类假设下的半监督高光谱数据分类算法(RCA-SLR-SSC)。该算法是将松弛聚类假设思想应用到SVM分类器下，定义了松弛聚类假设下的半监督图Laplacian正则，以及根据高光谱数据的空间约束正则。松弛聚类假设思想降低了像元误分的可能性，空间图正则的加入增强了样本标记的平滑性，减少了样本标记中的奇异点。所以，该算法能够得到较高的分类正确率。在实际高光谱数据集上进行的仿真实验验证了该算法的性能。与同类算法相比，该算法能够在较少样本情况下得到较好的分类结果。 (3)设计了一种空-谱松弛聚类假设下的半监督高光谱图像分类方法。该算法是对RCA-SLR-SSC算法在空间上进行了进一步松弛。即利用sketch算法提取出高光谱数据中背景像元与待分类像元之间边界处的像元，并减小这些像元对应的空间约束矩阵中的权值，构造一个新的拉普拉斯图矩阵。再将这个新的拉普拉斯图矩阵应用到RCA-SLR-SSC算法中。松弛聚类假设思想在空间上的应用大大降低了因空间约束而造成的边界处像元的误分，提高了分类正确率。高光谱数据集上的仿真实验也表明：与RCA-SLR-SSC算法相比，该算法在边界处像元的分类上具有明显的优越性。
[Abstract]:In recent years, with the rapid development of remote sensing technology, people can obtain a large number of hyperspectral data. How to classify objects accurately based on these data is the key to improve the application level of remote sensing data. At present, a large number of machine learning algorithms have been applied to the classification of hyperspectral data. These methods are based on the clustering assumption that the data with the same spectrum / similar spectrum have the same / similar labels. However, due to the limitation of sensor noise and resolution of imaging equipment, there are serious phenomena of "isomorphism" and "isospectral foreign body" in the actual hyperspectral data, which seriously restrict the performance improvement of the classifier under clustering assumption. To solve this problem, this paper introduces the idea of relaxation clustering hypothesis into hyperspectral data classification, and combines with sparse coding classifier and SVM (Support Vector Machine) classifier to develop the hyperspectral data classification method under relaxed clustering assumption. The main research work and innovations are as follows: (1) A sparse coding algorithm for hyperspectral data under relaxed clustering assumption is designed. The relaxation clustering hypothesis is applied to the sparse coding model in the form of probability vector, and the corresponding optimization objective function and optimization algorithm are designed. The simulation results on the actual hyperspectral data set show that the relaxed clustering assumption can greatly improve the low accuracy of data classification caused by mixed pixels. (2) A semi-supervised hyperspectral data classification algorithm (RCA-SLR-SSC) under relaxed clustering assumption is designed. In this algorithm, the relaxation clustering hypothesis is applied to the SVM classifier, and the semi-supervised graph Laplacian canonical under the relaxed clustering assumption is defined, as well as the spatially constrained regularization based on hyperspectral data. The loose clustering hypothesis reduces the possibility of pixel misdivision, and the addition of regular space graph enhances the smoothness of the sample marking and reduces the singularity in the sample marking. Therefore, the algorithm can achieve a higher classification accuracy. The performance of the algorithm is verified by simulation on the actual hyperspectral data set. Compared with similar algorithms, this algorithm can obtain better classification results with fewer samples. (3) A semi-supervised hyperspectral image classification method under the assumption of space-spectrum relaxation clustering is designed. The algorithm further relaxes the RCA-SLR-SSC algorithm in space. The algorithm of sketch is used to extract the pixels at the boundary between the background pixel and the pixel to be classified in hyperspectral data, and to reduce the weights in the spatial constraint matrix corresponding to these pixels, and to construct a new Laplace map matrix. Then the new Laplace matrix is applied to RCA-SLR-SSC algorithm. The application of relaxation clustering hypothesis in space greatly reduces the misclassification of pixels at the boundary caused by spatial constraints and improves the classification accuracy. The simulation results on the hyperspectral data set also show that the proposed algorithm is superior to the RCA-SLR-SSC algorithm in the classification of pixels at the boundary.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751
本文编号：2351149
[Abstract]:In recent years, with the rapid development of remote sensing technology, people can obtain a large number of hyperspectral data. How to classify objects accurately based on these data is the key to improve the application level of remote sensing data. At present, a large number of machine learning algorithms have been applied to the classification of hyperspectral data. These methods are based on the clustering assumption that the data with the same spectrum / similar spectrum have the same / similar labels. However, due to the limitation of sensor noise and resolution of imaging equipment, there are serious phenomena of "isomorphism" and "isospectral foreign body" in the actual hyperspectral data, which seriously restrict the performance improvement of the classifier under clustering assumption. To solve this problem, this paper introduces the idea of relaxation clustering hypothesis into hyperspectral data classification, and combines with sparse coding classifier and SVM (Support Vector Machine) classifier to develop the hyperspectral data classification method under relaxed clustering assumption. The main research work and innovations are as follows: (1) A sparse coding algorithm for hyperspectral data under relaxed clustering assumption is designed. The relaxation clustering hypothesis is applied to the sparse coding model in the form of probability vector, and the corresponding optimization objective function and optimization algorithm are designed. The simulation results on the actual hyperspectral data set show that the relaxed clustering assumption can greatly improve the low accuracy of data classification caused by mixed pixels. (2) A semi-supervised hyperspectral data classification algorithm (RCA-SLR-SSC) under relaxed clustering assumption is designed. In this algorithm, the relaxation clustering hypothesis is applied to the SVM classifier, and the semi-supervised graph Laplacian canonical under the relaxed clustering assumption is defined, as well as the spatially constrained regularization based on hyperspectral data. The loose clustering hypothesis reduces the possibility of pixel misdivision, and the addition of regular space graph enhances the smoothness of the sample marking and reduces the singularity in the sample marking. Therefore, the algorithm can achieve a higher classification accuracy. The performance of the algorithm is verified by simulation on the actual hyperspectral data set. Compared with similar algorithms, this algorithm can obtain better classification results with fewer samples. (3) A semi-supervised hyperspectral image classification method under the assumption of space-spectrum relaxation clustering is designed. The algorithm further relaxes the RCA-SLR-SSC algorithm in space. The algorithm of sketch is used to extract the pixels at the boundary between the background pixel and the pixel to be classified in hyperspectral data, and to reduce the weights in the spatial constraint matrix corresponding to these pixels, and to construct a new Laplace map matrix. Then the new Laplace matrix is applied to RCA-SLR-SSC algorithm. The application of relaxation clustering hypothesis in space greatly reduces the misclassification of pixels at the boundary caused by spatial constraints and improves the classification accuracy. The simulation results on the hyperspectral data set also show that the proposed algorithm is superior to the RCA-SLR-SSC algorithm in the classification of pixels at the boundary.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751
【参考文献】
相关期刊论文 前3条
1 杨哲海,韩建峰,宫大鹏,李之歆;高光谱遥感技术的发展与应用[J];海洋测绘;2003年06期
2 陈进;王润生;;高斯最大似然分类在高光谱分类中的应用研究[J];计算机应用;2006年08期
3 高恒振;万建伟;徐湛;钱林杰;;基于光谱加权直推式支持向量机的高光谱图像半监督分类[J];信号处理;2011年01期
相关博士学位论文 前2条
1 陈进;高光谱图像分类方法研究[D];国防科学技术大学;2010年
2 崔燕;光谱成像仪定标技术研究[D];中国科学院研究生院（西安光学精密机械研究所）;2009年
本文编号：2351149
本文链接：https://www.wllwen.com/guanlilunwen/gongchengguanli/2351149.html
