Linear pixel unmixing is one of the key technologies for hyperspectral image application.However

there are two problems for the hyperspectral decomposition in operational cases.One is the endmembers of an image can't be extracted automatically with traditional supervised ways;the other is unmixing hundreds of spectral bands directly may reduce accuracies due to the high correlation between bands.To mitigate the problems

we proposed a method for abundance estimation from spectral domain wavelet features.We utilized the discrete wavelet transform (DWT) as a preprocessing step for the feature extraction

then selected endmembers with projective iterative algorithm in an unsupervised fashion based on the features.In the end

we performed a constrained least square method for the abundance estimation.Algorithm validation and comparison were done with real PHI data.Experimental results show that the use of DWT-based features can improve the abundance estimation

as compared to those of original hyperspectral signals or conventional PCA-based features.