基于流形学习的多光谱优化波段选择算法研究

Algorithms Study for Selecting Few Characteristic Spectral Bands Based on Manifold Learning

为从多光谱图像特征提取的角度进行优化波段选择,在充分描述数据结构特征的同时使提取选择的特征有明确的物理意义,对基于流形学习算法的优化波段选择算法进行了研究。用判别局部排列(DLA)算法对多光谱数据进行预处理,选取正负样本,利用样本信息,以目标分类为目的进行特征提取。利用特征提取的结果,从特征提取的角度分析当前各谱段对所提取的主特征贡献的总信息量和贡献率,给出了基于权值和基于贡献率的两种优化波段选择算法,分别基于权值和贡献率进行特征选择。用正负样本的可分性可快速高效降维,同时又能保留多光谱图像原物理特性。实测数据验证了优化降维后的5个谱段能保留原数据的物理特性,目标识别概率提高约2%,计算复杂度降低约50%。优化选择的谱段有助于新一代多光谱探测器的研制和应用。

To implement the band selection algorithm from the feature extraction of multi-spectral image,and not only describe the data features but also remain the physical meanings of the selected bands,the optimization algorithm was studied based on manifold learning in this paper.Positive and negative samples were selected after a pretreatment on multispectral data by using discriminative locality alignment(DLA)algorithm.On the basis of the sample information,features were extracted to classify the targets.Using the feature extraction transformation matrix,the gross information content and contribution rate by the bands to the most discriminative and significant extracted features were analyzed and evaluated.Then the two algorithms based on weight and contribution rate,in which the features were selected based on weight as well as contribution rate respectively.The divisibility of the positive and negative samples can rapidly reduce dimension and retain the original physical features of multispectral image.The measured data proved that 5spectrums could reserve the physical features of the original data after dimension reduction optimization.Meanwhile the target recognition rate increased by 2% and the calculation complex rate decreased by 50%.The optimization of band selection contributes to the development and application of the new generation multispectral detector.