A　novel　approach　to　improve　abundance　estimation　of　hyper-spectral　image　using　spectral　piecewise　constant　features　is　presented.　Firstly　it　extracts　the　spectral　features　by　partitioning　the　spectral　signals　into　a　fixed　number　of　contiguous　intervals　with　constant　intensities　in　terms　of　minimizing　the　mean　square　error.　Then　the　multi-scale　features　can　be　determined　by　using　optimal　segmented　scale　or　combination　of　features　with　different　segmentation　scales.　In　the　end,　the　estimation　is　performed　by　unmixing　the　pixel　in　the　feature　space　with　constrained　least　square　algorithm　to　achieve　the　respective　abundance　fractions　of　these　end-members　present　in　the　pixel.　Algorithm　validation　and　comparison　were　done　with　simulated　and　real　data.　Experimental　results　demonstrate　the　proposed　method　can　significantly　improve　the　least　squares　estimation　of　end-member　abundances　using　remotely　sensed　hyper-spectral　signals,　as　compared　to　those　of　original　hyper-spectral　signals　or　discrete　wavelet　transform　based　features.