EOS-MODIS Pixel Component Unmixing And Its Application

When one pixel is just one component, it is called as pure pixel. Whileone pixel is made up of more than one component, it is called as mixed pixel.The mixed pixel can be unmixed into different components and theirproportions, and these components are called as basic components cells or endmembers. This process is called as pixel unmixing.Up to now, pixel component unmixing models include linear spectralmixture model(LSMM), fuzzy classification model, neural networks modeland Gaussian mixture discriminant analysis model.LSMM is defined as that the reflectivity (or spectral brightness) of apixel is linear combination of their end-members reflectivity, every of whichcoefficient is end-members proportions occupied in pixel separately. Theexpression as followsWhere, R_iλ , is known spectral reflectivity of i pixel onλ spectral band,f_ki is the unknown fraction of k end-member in i pixel, C_kλ is the reflectivityof k end-member on λ spectral band, ε_iλis a zero mean noise term, n is thenumber of end-members, m is the number of available bands( n ≤ m+ 1).Thevector of proportions f_ki is estimated by minimizing the quadratic. ε_iλ orRMS (root-mean-square error) is used to evaluate the model. The expressionof RMS as follows2 1/21[ ( ) / ]nikkRMS εn== ∑ (2)The average spectral signature C kλ of each end-member is extractedfrom the mixed pixel and f ki can be solved by the linear formula (1) usingLSMM. Finally, all of the pixels are unmixed into the fractions of theend-members. The results are the fraction images of end-members and theerror image in the form of root-mean-square error RMS.As the spatial resolution of EOS-MODIS image is lower, the mixedpixels are usual. In this study, Principal Component Analysis (PCA) method isused to define end-member spectrum. The number, the type and the spectralvalue of every end-member are analyzed. Three end-members, includingwater, soil salinity, and vegetation, were obtained firstly. However, bycomparison and analysis the first results, we found that the number ofend-members was not satisfied for the study areas. Then water areas wereeliminated from the image, and then another end-member "forest and grass"was added. Finally, reliable fraction image for four kinds of end-memberswere produced.Taking the fraction images of four end-members (water, soil salinity,plantation, forest and grass) as the nodes of decision tree classification (DTC),the MODIS image is classified as four classes and mapped the thematicinformation of four land cover classes. To evaluate the precision, maximumlikelihood classification (MLC) and decision tree classification was separatelyimplemented using the MODIS bands1-6 images not unmixed. The Kappacoefficients of MLC and DTC results are 0.3476 and 0.4542. Their wholeprecisions are 48.9346% and 64.4610% separately. However, Kappacoefficients and whole precisions of DTC after pixel component unmixing is0.8625 and 90.7290%. Therefore, the DTC results after pixel componentunmixing is effective.