基于加权原型空间特征提取的高光谱图像分类
首发时间:2012-11-09
摘要:针对高光谱数据在原型空间上的表示,在原型空间特征提取(PSFE)方法上提出了一种基于加权原型空间特征提取的高光谱图像数据分类方法,简称w-PSFE。原型空间是一种新的空间中表示高光谱数据,空间中的一个点是所有高光谱数据类在某个波段上对应的光谱信息(光谱的辐射值),即:此空间的维度为数据的类的个数;PSFE方法在原型空间通过模糊c均值聚类(FCM)法,把高度相关的特征与独立的特征分割开,从而实现特征提取用于分类。 PSFE未考虑高光谱数据不同波段(特征)用于分类时的信息量,且未考虑所提取的特征对高光谱图像的分类是否有效。相对地,w-PSFE迭代过程中,计算每个特征的信息量,在特征提取时,根据每个特征的信息量对特征加以不同的权重;为了使具有高有效性的提取后的特征具有高的比率,w-PSFE在每步迭代中计算各个提取后的特征的方差来表示特征有效值,而在下一次迭代中,计算出的有效值用于计算原始特征的类归属度。本文中同时证明了w-PSFE的局部收敛性;实验表明,在提取少量的特征用于高光谱图像数据分类时,w-PSFE的分类精度高于PSFE的分类精度。
For information in English, please click here
A Weighted Prototype Space Feature Extraction Method for Classification of Hyperspectral Data
Abstract:In this paper, a method called weighted prototype space(PS) feature extraction (w-PSFE) is proposed for feature extraction of hyperspectral data. The approach is an extension of previous approach-prototype space feature extraction (PSFE). In PSFE, it represents channels in a new space called PS, where they are characterized in terms of reflection properties of classes; it clusters the channels in PS by fuzzy c-means clustering (FCM), highly correlated and isolated channels are separated by an uncertainty measure. But in PSFE, the FCM does not consider the characteristics of each feature and the contribution rate to clustering analysis when calculating the clusters , and dose not take the validity of the extracted features into consideration, these obviously affect the authenticity and accuracy of the classification. Relatively, in the proposed method, different feature with different importance(information) for classification will be given different weight in clustering algorithm; with the purpose that the extracted features could do well in classification, the effectiveness of extracted features in reduced feature space are also considered in FCM. In the iterative w-PSFE process, to identify the information of features, the weight for feature is computed based on the validity for classification of each feature; to make sure that the extracted feature with higher effectiveness has higher ratio, the validity of extracted feature is computed dynamically based on the variance of the extracted feature vector, and the new validity is used to calculate the cluster memberships of features in next iteration effectively. Moreover, the proposed w-PSFE retains the advantages of PSFE and its local convergence is proofed. Experiments results on hyperspectral data sets show that significant improvements are achieved in w-PSFE in terms of accuracies when compared to results obtained from approach PSFE.
Keywords: feature extraction hyperspectral data weighted fuzzy c-means classifier
基金:
论文图表:
引用
No.****
同行评议
勘误表
基于加权原型空间特征提取的高光谱图像分类
评论
全部评论0/1000