ICA与PCA在高光谱数据降维分类中的对比研究

Comparative study on descending dimension classification of hyperspectral data between ICA algorithm and PCA algorithm

利用独立主成分算法与主成分分析法分别对原始数据及3种预处理数据进行降维运算,再利用常用的4种分类算法分类,对比分类结果发现,独立主成分算法与主成分分析算法在乔木树种高光谱数据降维中并不具有非常明显的优势,且独立分量分析(ICA)算法提取用于分类的数据不如PCA算法稳定;从计算机的运行成本上来看,PCA算法优于ICA算法,基本上ICA算法平均成本是PCA算法的4到5倍;不同的数据预处理及降维方法组合对分类结果影响明显,d(log(R))结合PCA算法,log(R)结合ICA算法降维结果最理想;通过比较乔木树种高光谱数据的分类结果发现,Fisher判别法最适合对PCA和ICA降维结果进行分类。

The three kinds of pre-processed hyper-spectra data（first-derivative（d（R））,logarithms（log（R））,logarithms-first-derivative（d（log（R）））） and original data were reduced in dimension by ICA and PCA algorithm,and then classified by Support Vector Machine（SVM）-Gaussian Raial Basis Function（RBF）,Support Vector Machine（SVM）-Liner,Back Propagation（BP）neural network and Fisher classification method.The results show that compared with PCA,ICA did not have obvious advantage in dimension reduction for hyperspectral data of trees.ICA was less stable than PCA.Judging by the running cost of computer,PCA algorithm was better than ICA,ICA＇s average cost was 4 to 5 times more than PCA＇s.Various pre-processing methods and various classification methods showed the different influences on classification results.By comparing several combination methods,the study has found that d（log（R））-PCA combination and log（R）-ICA combination were usually considered to be the most ideal.According to the results of the classification,it is found that Fisher classification method is suitable for the classification of the data preprocessed by PCA and ICA.