基于数字脊波和样条权神经网络的高光谱图像融合分类  

Fusion Classification of Hyperspectral Remote Sensing Images based on Digital Ridgelet and Sample Weight Neural Network

作　　者：赵春晖[1] 尤佳 

机构地区：[1]哈尔滨工程大学信息与通信工程学院,哈尔滨150001

出　　处：《黑龙江大学工程学报》2010年第4期70-77,共8页Journal of Engineering of Heilongjiang University

基　　金：国家自然科学基金项目(61077079);高等学校博士学科点基金(20060217021);哈尔滨市优秀学科带头人基金(2009RFXXG034)

摘　　要：提出了一种基于数字脊波和样条权神经网络的高光谱图像融合分类新算法。在特征级融合中,针对数字脊波的特点,对不同分辨率的脊波系数采用局部信息熵进行融合,并运用样条权神经网络实现了分类。在决策级融合中,提出了先用样条权神经网络进行预分类,然后用主体投票法进行决策融合的算法,为避免作为局部分类器的神经网络结构过于复杂,对输入数据先进行了像素层的融合实现数据降维,这实质上体现了一种多层次融合的思想。实验结果表明,这两种方法都能有效的实现高光谱图像的融合及分类,在较少的训练样本下分类精度能达到92%以上,其中特征级融合可达到95.87%。A novel fusion classification algorithm of hyperspectral remote sensing images based on digital ridgelet and sample weight neural network was introduced.On feature level fusion,the local information entropy was calculated for ridgelet coefficients on diverse resolution level by fully considering the characteristics of this kind of ridgelet transform,then the sample weight neural network was chosen for classifying the fused data.On decision level fusion,a new method was proposed,i.e.the local classification based on sample weight neural network was executed previously,and then all the results of the local classifiers were sent to decision fusion center which based on majority voting rules.By avoiding the complexity of neural networks owing to too many inputs,a pixel level fusion aimed at reducing dimensionality of input data was applied.The idea of multi-level fusion was implemented.Experimental results showed that both of the methods were effective on fusion and classification of hyperspectral data： the classification accuracy（CA） based on a few training samples was over 92%,and 95.87% of CA was obtained in feature fusion level fusion.

关 键 词：数字脊波变换 快速Slant Stack算法 样条权神经网络 主体投票规则 融合分类 

分 类 号：TN911.73[电子电信—通信与信息系统]