Study About Spectral Unmixing Based On Factor Analysis

Remote sensing data records electromagnetic radiation information ofcontinuous distribution things in space. Because of the existence of mixed pixels, thetraditional remote sensing classification is difficult to meet the application needs,spectral unmixing can obtain end member abundance in the mix-pixel, so spectralunmixing has important significance in improving the accuracy and depth of remotesensing application.Factor analysis is a multivariate analysis method which makes compre-hensive analysis on observed data of multiple variables and multiple samples, its basicidea is based on the study about the similarity matrix of variable or sample, and sumsthe perplexing many variable or samples up to a few factors, and makes analysis onthe combination relation of variable of sample under the premise of the least amountof information loss, then gets the leading role essential factors.This paper studies both the model of factor analysis and linear spectral mixingmodel and their relationship, for the ETM data of the region of Xingcheng area, andthe spectrum of the mixture is decomposed by the use of factor analysis model andlinear spectral mixing model, and compares the two results. In the paper, I emphasizesthe geological significance of factor loadings and factor scores, and analysis the factorcomponent of typical ground landscape, at last confirm the mapping method of factorloadings imagine. The main conclusions are as follows:First, the model of factor analysis and linear spectral mixing are most intimatelyassociated with numerical relationship. Among them, factor scores in Q-mode factoranalysis represents the correlation of the variables and the factors, and it correspondsto the end members of linear spectral mixing model; factor loadings in Q-mode factoranalysis represents the correlation of the samples and the factors, and it corresponds tothe abundance.Second, based on ETM data, Q-mode factor analysis should be carried onfollowed the steps below: solve the factor loadings and factor scores using of therepresentative sample-variable data matrix, and calculate the factor loadings of thefull scene according to the got factor score.Third, analysised the two results we can see, the similarity coefficient of allkinds of typical features is0.6396-0.9985, with an average of0.894. It shows that the factor analysis can be used for spectral unmixing, at the same time, after the analysisof the components characteristics of the typical features/landscape, it can realizeDepth information mining based on factor loading, including both the classification ofthe terrain category and the altered information extraction.The last, this study shows the physical significance explanation of the first threefactors which contains99.085%of the total information, and defines as soil、vegetation and water factor. According to the rotation factor loading matrix, water、reservior both have the maximum factor loading in the water factor, and coniferousforest、broad-leaved forest has the maximum factor loading in the vegetation factors,and grass shrubs、farmland、town and roads have the maximum factor loading in thesoil factor.