Abstract: Timely and accurate crop distribution maps derived from satellite observations could assist crop growth monitoring. Although crop mapping methodologies have been widely studied, there are still some drawbacks, such as the limitation of ground reference data and low efficiency of crop type mapping caused by using time series data of the entire year. The objectives of this study are: (1) to develop a new method, which can identify crop types using the crop records of the previous years; (2) to evaluate the performance of the method with different time series length, and try to acquire the crop type map at 30 m spatial resolution. The study area of this paper was the ASD30 of Kansas State, USA. We firstly used the cropland data layer (CDL) data and MODIS EVI(enhanced vegetation index) time series between 2006 and 2013 to generate reference EVI time series with the ABNet algorithm for the major crops in the study area, i.e. alfalfa, corn, sorghum and winter wheat. Then, we acquired the “possible” training samples in 2014 using the CDL records between 2006 and 2013. If a pixel was labeled as “Crop A” more than 4 times among the 8-year CDL records, the pixel was labeled as “possible Crop A” in 2014. Next, we compared the MODIS EVI of the “possible crop A” pixels and the reference EVI time series of Crop A, if the 2 profiles were matched, the “possible Crop A” was confirmed as a training sample of “Crop A”. Finally, we used these training samples and monthly composited Landsat NDVI (normalized differential vegetation index) to identify crop types at 30 m resolution. To analyze the effect of time series length on crop type identification performance, we tried 7 time series lengths (April, April-May, April-June, April-July, April-August, April-September and April-October), used MODIS EVI time series to acquire training samples for each time series length, and then identified crop types using the corresponding training samples and Landsat NDVI time series. Several metrics derived from the confusion matrix, such as overall accuracy, Kappa coefficient, were used to evaluate the classification performance. Results showed that when only time series data in April were used, we acquired 5 088 samples, and 91.86% among these samples had the same crop label with the CDL data. When longer time series data were used, more training samples in 2014 were acquired with higher accuracy. When entire EVI time series data were applied, 10 803 samples were acquired and 10 317 samples had same crop label with CDL data. When using these training samples and monthly composted Landsat NDVI to identify crop types at 30 m resolution, classification accuracies were low if April or April-May time series data were used, and overall accuracies were 66.12% and 52.51%, respectively. When time series length was April-October, overall classification was 94.89%. April-August time series achieved good classification performance, as 10 183 training samples were acquired, 96.32% samples had same label to CDL data, overall classification accuracy was 94.02%, and acreage of major crops was similar to CDL data. Finally, we could conclude: (1) The method proposed in this study can acquire train samples in the classification year when the ground reference data are absent. Using these training samples, we can obtain crop type distribution maps with high accuracy (better than 90%). (2) We can acquire the crop type map of the study area in August with the high classification accuracy which is similar to the result derived from the entire EVI time series, and has the similar crop acreage with CDL data for each crop. In the future, we can enhance this method by improving the previous-year training samples with CDL crop confidence layer.