Abstract: Toimprovetheaccuracyofspectralanalysisforsubpixelmaterialsfurther,anonlinearunmixingalgorithmbasedonthekernelpartialnonnegativematrixfactorization(KPNMF)wasproposed.Firstly,anendmemberextractionalgorithmbasedonthetheoryofconvexgeometrywasusedtogenerateacandidatepixelsetofpureendmembers,andthenthepureendmemberwasdeterminedaccordingtothespatialpurityindicesofthecandidatepixels.Giveninformationofpureendmembers,thekernelmethodwasadoptedtoextendpartialnonnegativematrixfactorization(PNMF).Thecorrespondingobjectivefunctionwasconstructed,andtheiterativesolutionwasalsoderivedtoobtainthesubpixelendmembersandabundancesofalltheendmembers.Theexperimentalresultsdemonstratethattheproposedunmixingalgorithmhasgoodnonlinearunmixingability,andtheunmixingresultsarebetterthanthoseoflinearunmixingalgorithms.

Key words: Hyperspectralunmixing, Subpixel, Convexgeometry, Spatialpurityindex, Partialnonnegativematrixfactorization, Spaceremotesensing