Studies On Endmember Extraction For Hyperspectral Image Based On TF-IDF Model

With the development of hyperspectral technology,spectral resolution and spatial resolution of hyperspectral image are constantly improving,which lay the foundation for the better separation and identification of object features.On the other hand,with the demand for the target object recognition accuracy,the problem of mixed pixel decomposition need to be solved urgently.While the endmember extraction technology is the key step to solve the problem.At present,most of the existing endmember extraction algorithm is based on convex geometry theory.The mechanism of these algorithms are complex and the algorithms are very sensitive to noise,while the algorithm based on statistical model starts late and number is alse less.In this paper,the term frequency inverse document frequency(TF-IDF)model is applied to the endmember extraction field,then a spectral matching based,endmember extraction method on account of feature selection is proposed.The algorithm according to the spectral matching degree of the target pixel to its classification and other classification,the importance of the target pixel to its classification is measured.The weight value of each pixel in each classification is calculated by TF-IDF model,and the pixel with maximum weight value is chosen as the endmember of the class.The algorithm breaks through the dependence of convex geometry theory and explore the endmember extraction method from the view of statistical characteristic.Moreover,the algorithm considers the spatial information of pixels and better use local spectral features of various features by deviding multiple class subspace.It makes the endmember selection has a good class represents.In this paper,the simulated hyperspectral image and ROSIS?AVIRIS hyperspectral remote sensing image are used to carry out the experiment.The endmember extraction algorithm based on TF-IDF model and PPI(Piwel Purity Index)?VCA(Vertex Component Analysis)?SGA(Samplex Growing Algorithm)and MVC-NMF(Minimum Volume Constrained Nonnegative Matrix Factorization)algorithm are compared for the precision of the endmember extraction.the experimental results are as follows:(1)Integrated two simulated hyperspectral image data experiments,it is found that when the number of endmembe in image increased from 3 to 5,the accuracy of endmember extraction are overall improved and the ability against noise are enhanced too.Two groups experiments of the endmember extraction algorithms'accuracy ranking are basically the same,in descending order is:VCA>TF-IDF>PPI>SGA>MVC-NMF.In the image with less noise,the VCA algorithm endmember extraction accuracy is the highest,TF-IDF algorithm follows.But,with the enhancement of the noise signal,TF-IDF algorithm endmember extraction accuracy constantly outperforms other algorithms,because the endmember extraction algorithms based on convex geometry theory are very sensitive to noise.(2)Comprehened ROSIS benchmark hyperspectral image and AVIRIS real hyperspectral image data experimental results,with the based ues the accuracy evaluation index of SAM and SID,the AVIRIS image data experiment also combines the feature abundance map getted in linear decomposition way and RMSE error map to carry out the accuracy evaluation of the endmember extraction algorithms synthetically.The algorithms'accuracy from high to low order is:TF-IDF>PPI>SGA>MVC-NMF>VCA,and TF-IDF algorithm has the optimal results.Comprehened 4 groups of experimental results' analysis reveals that the TF-IDF model applied to endmember extraction work obtains good results.Especially for the hyperspectral image with high noise impact,which means TF-IDF algorithm has good robustness to noise.Moreover,the accuracy of the endmember extraction algorithm is much better when spectral feature signal is strong which means the objects with a wide range of reflectance values and reflectance values floating large than the weak one.