Using　the　significant　spectral　correlation　within　the　hyperspectral　images,　we　present　a　lossless　compression　algorithm　in　this　paper.　By　means　of　hand　ordering　according　to　spectral　correlation　coefficient　and　2D/3D　hybrid　predicton,　which　are　based　on　local　texture　and　neural　networks,　hyperspectral　data　are　decorrelated.　The　prediction　residuals　are　then　entropy　coded　by　context-based　Golomb　coding.　Experimental　results　show　that　this　method　can　remove　the　spatial　and　spectral　redundancy　efficiently　and　outperforms　JPEG-LS　and　3D-APA　on　average　bit　rate　obviously.