基于K-means聚类改进的光谱降维方法

Improved Spectral Dimension Reduction Method Based on K-means Clustering

光谱降维是解决光谱数据高冗余度问题的重要方式,针对光谱降维中常用的主成分分析法(PCA)在主成分个数较少时重构数据与原始数据误差较大的缺点,本研究提出了一种基于预先聚类减少光谱数据降维压缩后信息损失的方法。首先用K-means算法分别在光谱空间和颜色空间将光谱数据聚类,然后对每子类利用PCA法进行降维并重构。结果表明,相比于非聚类的整体降维,本研究改进方法的重构数据在光谱精度和色度精度方面均提升明显,在仅3个主成分的前提下即可获得较高的重构精度,并且光谱空间中的聚类结果优于颜色空间聚类。实验证明,对光谱数据的预先聚类处理对光谱数据的高保真降维压缩具有重要作用。

Spectral dimension reduction is an important way to solve the problem of high redundancy of spectral data.Aiming at the disadvantage of PCA that the reconstructed data has errors from the original data when the number of principal components is small,a method based on clustering pretreatment to reduce the information loss after dimension reduction compression of spectral data was proposed in this study.First,K-means algorithm was used to cluster spectral data in spectral space and chromaticity space respectively,and then the PCA method was used to reduce dimensions and reconstruct each subcategory data.The results showed that compared with the overall dimension reduction of non clustering,the reconstructed data based on the improved clustering method had significantly improved in spectral accuracy and chromaticity accuracy.With only three principal components,higher reconstruction accuracy could be obtained,and the clustering results in the spectral space were better than those in the chromaticity space.The experiment proved that the clustering pretreatment of spectral data plays an important role in the high fidelity dimension reduction compression of spectral data.