It is still not clear that how high the upper limit of land

u

se classification accuracy can be reached using TM imagery in southeast p la in areas of China featuring mass irrigation networks and small tracts. In this p aper a result for the upper limit

65%

is presented based on the GIS layer （as gro und truth） of the detailed land

u

se inventory map at 1∶10000 scale being well o verlaid with a TM image of the same period

according to the national land

u

se c lassification system by taking all measures that may make the accuracy higher wi t hin the framework of conventional image classification mathods. The results show that the selection of training samples and reference pixels for accuracy assess ment in a limited part of test area has significant influence on accuracy assess me nt. The limitation of quality and quantity of reference pixels is probably the m a in reason why some early researches of this kind reported higher classification accuracy.