判别流形学习算法的高光谱数据降维与树种识别

Hyperspectral Image Dimension Reduction Based on Manifold Learning Approach DLA and Tree Species Classification

高光谱遥感影像的波段光谱特征是各类地物内在物理化学性质的反映,在对不同地物进行分类与识别时具有巨大潜能,但由于其波段多造成的信息冗余,需要对高光谱数据进行有效降维,以提高高光谱影像的分类准确度。本文提出了基于判别局部片排列的流形学习算法(DLA)对Hypersion高光谱数据进行降维,通过对局部样本数据进行流形学习框架内的优化训练,将原始光谱特征空间转换为低维的最优判别流形子空间,然后在该子空间内利用最大似然分类器对Hypersion影像中的每个像素进行分类,并与主成分分析(PCA)、原始光谱特征(spectral)降维方法的分类效果进行比较。结果表明,DLA能够有效提高高光谱数据的分类准确度,对不同树种分类取得了满意效果。

Hyperspectral image has great potential in the classification and recognization of different objects, whose inherent physical and chemical properties can be reflected by the spectral features of the image bands. In order to overcome the high redundancy among the large number of bands of hyperspectral image, efficient dimensional reduction algorithms should be applied to improve the performance of image classification. In this paper, we present a modified manifold learning algorithm termed discriminative locality alignment （DLA） for the dimensional reduction of Hypersion image data. The proposed method transformed the original spectral feature space into the optimal low dimensional subspace by imposing discriminative information which from given raining samples in the manifold learning framework. In this subspace, the maximum likelihood classifier was then used to classify each pixel of the Hypersion image. Meanwhile, the classifyeation results based on the dimensional reduction algorithms of principle component analysis （PCA） and original spectral were compared with the performance of classification based on DLA. The experiments showed that DLA can effectively improve the classification accuracy of hyperspectral image data, and obtained satisfactory classification results for tree species.