基于半监督局部保持投影的高光谱遥感影像分类方法研究

发布时间：2018-08-03 14:46

【摘要】：高光谱遥感影像具有所含光谱信息量大、相关性强的大数据量等特点，若用传统分类算法对其进行分类易产生“维数灾难”，因此对高维数据进行降维处理则显得尤为重要。在诸多降维算法中，如主成分分析(PCA)算法、线性判别分析(LDA)算法等，它们或是不能有效利用数据中的类别信息，或是对数据的类别信息要求严格。针对这些问题，论文提出一种半监督局部保持投影(SSLPP)算法。论文首先对高光谱图像及其自身特点作简单介绍，并结合监督学习与非监督学习对高维数据的特征提取方法进行总结和分析，提出SSLPP算法；其次从SSLPP算法的原理、算法流程等方面，对算法进行详细介绍；为验证SSLPP算法的有效性，与目前几种主流特征提取算法进行对比性实验，如主成分分析(PCA)算法、局部保持投影(LPP)算法、监督局部保持投(SLPP)算法。实验中对两种实际情况下的高光谱遥感图像数据进行分类实验，，首先用各种算法对原数据集进行降维处理，然后使用K近邻分类器对低维数据进行判类识别，计算出各个算法的总体分类精度，由此实验结果对SSLPP算法的有效性进行验证；最后为了探究SSLPP算法与各种分类器的融合性，分别用三种分类器与之结合对四种遥感图像进行地物分类实验，结果表明在该算法下各分类器均获得较高识别率，由此验证SSLPP算法具有较好的融合性。经过实验分析，论文所提SSLPP算法相比较于其他特征提取算法具有以下几点优势：①SSLPP算法相对于非监督降维算法，它充分利用了数据中的类别信息，使高维数据经过低维映射后具有较好的可分性；②SSLPP算法相对于监督降维算法，其不仅利用了数据中的标记样本并同时充分利用大量的未标记样本，使得在进行低维投影时更好的把握原始数据的整体性；③在对高光谱数据进行分类处理，SSLPP保证较高分类精度的同时，又避免了对原始数据的全类别标定工作，从而很好的提高数据计算处理效率。综上所述，论文主要研究了高光谱遥感图像基于半监督学习的特征提取与分类方法，提出一种半监督数据特征提取算法，通过对几种实际高光谱遥感图像的分类识别实验证明了论文算法的有效性。
[Abstract]:Hyperspectral remote sensing images are characterized by large amount of spectral information and large amount of data with strong correlation. If the traditional classification algorithm is used to classify hyperspectral remote sensing images, it is easy to produce "dimensionality disaster", so it is very important to reduce the dimension of high-dimensional data. In many dimensionality reduction algorithms, such as principal component analysis (PCA) algorithm, linear discriminant analysis (LDA) algorithm and so on, they either can not effectively use the category information in the data or require strictly the data category information. In order to solve these problems, this paper presents a semi-supervision department preserving projection (SSLPP) algorithm. Firstly, the hyperspectral image and its own characteristics are briefly introduced, and the feature extraction methods of high-dimensional data are summarized and analyzed by combining supervised learning and unsupervised learning, and then the SSLPP algorithm is proposed, and then the principle of SSLPP algorithm is introduced. In order to verify the effectiveness of the SSLPP algorithm, a comparative experiment is carried out with several popular feature extraction algorithms, such as principal component analysis (PCA) (PCA) algorithm, local preserving projection (LPP) algorithm, and so on. The Supervisory Department maintains the (SLPP) algorithm. In the experiment, two kinds of hyperspectral remote sensing image data are classified. Firstly, the original data set is reduced by various algorithms, and then the low-dimensional data is identified by K-nearest neighbor classifier. The overall classification accuracy of each algorithm is calculated, and the validity of SSLPP algorithm is verified by the experimental results. Finally, in order to explore the fusion of SSLPP algorithm with various classifiers, Three classifiers are used to classify the ground objects of four remote sensing images respectively. The results show that each classifier has a high recognition rate under this algorithm, which verifies that the SSLPP algorithm has a better fusion performance. After experimental analysis, compared with other feature extraction algorithms, the proposed SSLPP algorithm has the following advantages over the unsupervised dimensionality reduction algorithm, which makes full use of the class information in the data. Compared with the supervised dimensionality reduction algorithm, the high-dimensional data has good separability after low-dimensional mapping. It not only makes use of the labeled samples in the data, but also makes full use of a large number of unlabeled samples at the same time. In order to better grasp the integrity of the original data in the low-dimensional projection, the classification of the hyperspectral data can be processed by SSLPP to ensure a higher classification accuracy, and at the same time, the whole classification of the original data can be avoided. In order to improve the efficiency of data calculation and processing. To sum up, this paper mainly studies the feature extraction and classification method of hyperspectral remote sensing image based on semi-supervised learning, and proposes a semi-supervised data feature extraction algorithm. The effectiveness of the algorithm is proved by the classification and recognition experiments of several real hyperspectral remote sensing images.
【学位授予单位】：重庆大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】