An　adaptive　dimensionality　reduction　method　to　conduct　classification　of　hyper-spectral　imagery　using　optimal　segmentation　of　spectral　signature　is　proposed.　The　method　partitions　the　spectral　signals　into　a　fixed　number　of　contiguous　intervals　with　constant　intensities　in　terms　of　minimizing　the　mean　square　error.　To　automatically　obtain　the　best　number　of　the　segments,　a　quantitative　indictor　based　on　variables　correlation　between　original　and　the　reconstructed　spectral　approximation　is　designed,　and　the　best　segments　can　be　adaptively　determined　by　a　user　specified　threshold.　To　validate　the　method,　an　experiment　with　aerial　push-broom　hyper-spectral　imagery　(PHI)　is　conducted,　and　the　results　demonstrate　that　the　spectra　data　reduction　using　adaptive　optimally　segmentation　can　preserve　the　distinctions　among　spectral　signatures　and　can　improve　the　classification　accuracy　significantly.　Comparisons　with　principal　component　analysis　(PCA)　and　discrete　wavelet　transform　(DWT)　are　also　done,　and　the　proposed　method　can　achieve　better　classification　accuracy　with　overall　accuracy　and　kappa　coefficient.　©2010　IEEE.