Abstract: Because of high spectral and lempora rsoluionus arge coverage ,and low cosl,MODIS(Moderale Resoution Imaging Serordioionere da has ben widely lused lo quickly extral information of for est types aul regional ,naina and global scales. However is coase spaial resouio ofen leads lo mised pixels and low cssisatio acuracy of forest types. Using seta ummixing can,to some extent,incrase the acurce of casisaio But, how 1lo acurately entract pure endmembers for a study area ofenei an great callnge. The seleion of liner or non-linor sectra unmixing algoritim is anoher callnge. In this study ,a merhod 1 extraet endmembers from MoODIS images was developed. In this mehod the time sries of MODIS derived vegetation index was fist derived and the phenologca variaio of forest trpes were analyed. Decisio treee casificsit ion was then conducled and the obuaine resuls were used lo ex-.trnct endmembers. In adino fo comparson, the casictio was also made using a widely lused clasi fier - maximum ielihoo. Theee impie that liner spetal umming was the best reanlss of wih and without cosrainsns then maximum ikelihoo casicati and nm-liner speral ummixing.

Key words: Remole sensing, Deision tree, Endmember extraction, Mixed pixels, MODIS, Forest