In　remote　sensing　images,　the　accuracy　of　land　cover　classification　at　pixel　scale　is　affected　by　mixed　pixels　greatly.　Sub-pixel　mapping　tries　to　predict　land-cover　map　at　sub-pixel　scale　according　to　spectral　unmixing　abundances　and　some　constraints　of　land-cover　distribution　patterns.　In　this　paper,　using　both　spatial　and　spectral　information　of　land-　cover,　we　propose　a　new　spectrum　preserving　sub-pixel　mapping　algorithm　based　on　local　connectivity　and　as　a　constraint　similarity.　Spatially,　local　dependence　is　re-modeled　by　the　with-in　class　scatter,　nonlocal　similarity　is　introduced　by　minimizing　the　representation　errors　among　similar　pixels.　Spectrally,　spectrum　preserving　is　realized　by　minimizing　the　spectra　errors　in　sub-pixel　mapping.　Comparative　experiments　with　artificial　and　real　images　show　that　the　proposed　algorithm　achieves　a　higher　accuracy　than　other　related　algorithms,　thus　it　is　more　suitable　for　practical　application.　©　2014　Acta　Automatica　Sinica.　All　rights　reserved.