Grassland shows obvious seasonal patterns and the effect of grassland classification varies in different stage within its life span.Field spectrum data with high resolution of dominant grasslands in the region around Qinghai lake was collected at 16 temporals in 2003 using GER 1500 spectrometer.Analysis of spectrum classification experiments for grasslands shows that spectrum transformation affects the classification accuracy.Classification method combining with certain spectrum transformation can achieve much better result than using the raw spectrum reflectance.Maximum likelihood and support vector machine using moving average spectrum

spectral angle mapping and minimum distance using first-order derivative of spectrum’s logarithm

and artificial neural network using first-order derivative of normalized spectrum can improve classification result.Then the paper carries out classification experiments for each temporal and determines the optimal temporal for grassland spectrum classification.The optimal temporal for natural and artificial grassland classification is at the beginning of grass turning green or in the middle ten days of August

with highest recognition accuracy mounting to 99%.The optimal temporal for artificial grassland classification is in the middle ten days of May

with highest recognition accuracy mounting to 95%

and it is worst for artificial grassland classification in the middle ten days of July.The optimal temporal for natural grassland classification is in the middle ten days of August

with highest recognition accuracy mounting to 87%.Experiment using TM data testifies the result derived from field spectrum data.