摘要
提出一种处理AVIRIS高光谱图像数据的计算机分类算法。首先采用投影梯度(ProjectedGradient)改进的非负矩阵分解(NMF)方法对高光谱数据进行特征提取,大大降低了分解过程中两个子迭代问题的时间复杂度,而后利用径向基函数神经网络(RBFNN)分类器对提取结果进行分类。结果表明,与传统NMF和主成分分析相比,PGNMF-RBF算法消耗时间最少,分类精度最高,6类地物的分类精度达到83.34%。该算法在保留非负矩阵分解明确物理意义的基础上,获得了更快的分解速度和更高的分类精度,在高光谱图像分类领域具有较大的应用潜力。
A new method combined Non-negative Matrix Factorization （NMF） with Projected Gradient （PG） is proposed for hyper-spectral image classification. Projected Gradient method demonstrates much faster convergence than the popular multiplicative update approach in the iteration process of two subproblems from NMF thus effectively maintains higher classification accuracy than traditional methods; RBF neural network achieves higher accuracy and faster classification process compared to BP network. The new method combines the advantages of the above two, applying PGNMF for feature extraction and RBFNN as classifier. The experiment shows that compared to traditional NMF and PCA,PGNMF-RBF has higher accuracy for classification and less time consumption. The classification accuracy for 6 classes reaches 83.34%. This paper demonstrates PGNMF-RBF an effective and promising method in hyper-spectral image classification.
出处
《遥感技术与应用》
CSCD
北大核心
2009年第3期385-390,共6页
Remote Sensing Technology and Application
关键词
投影梯度
非负矩阵分解
RBF神经网络
图像分类
Projected gradient
Non-negative matrix factorization
RBF neural network
Image classification
作者简介
狄文羽（1984-，女，硕士研究生，主要从事遥感图像处理与分析研究。Email：djwenyu@gmail．com。
