Spectral information is essential for objects recognition in remote sensing imagery. However, objects which have particular indices are rather few, and spectra types of spectral library and their universality are limited either. Therefore, an adaptive spectral representation method of remote sensing objects based on endmember matching is proposed. Proper endmember of imagery itself is selected. Spectral angle and distance, which is between the characteristic vectors of spectra of the interested pixel and a specific endmember, are both considered to form a new way for comprehensive spectral matching. Experiments of vegetation and water were adopted in ETM+ (Enhanced Thematic Mapper) images, and were compared to those using USGS (United States Geological Survey) library and normalized difference vegetation index (NDVI) /normalized difference water index(NDWI). Moreover, validations of shadow and bareland images were also carried out to test the effectiveness and universality of the proposed method.