Abstract: At present, based on the research of crop yield, estimation of time series of MODIS_NDVI is widely used, but the MODIS data by the spatial resolution of the sensor and features of complex types of limitations has a great influence on the monitoring effect. In this paper, we used Landsat_5_TM and Landsat_7_ETM images of 30 m spatial resolution, constructed a multi day synthetic vegetation index time series data. Two cotton plots in San Joaquin Valley of California were studied to predict cotton yield with the vegetation index. San Joaquin Valley is the Mediterranean climate that belongs to the high yield region of cotton. Landsat_5 and Landsat_7 images combined with the field measured data were used to establish cotton yield prediction model. The tendency of the normalized difference vegetation index (NDVI) and the enhanced vegetation index (EVI) time series and the correlation between vegetation index and cotton yield were analyzed to acquire the optimal cotton yield estimation phase. Then the optimal yield prediction model of cotton was built. Through comparison of the two vegetation indexes respectively as independent variable, the precision of the model can be found. With NDVI as input variable, the decision coefficient and root mean square error were slightly better than EVI. So in this paper, we selected NDVI as yield estimation factor. The results showed that the time series of NDVI based on pure pixels of Landsat images can reveal the period of cotton growing exactly. NDVI of different stages of cotton growth showed differences with time changing, especially in cotton flowering and boll-setting period with 30 m spatial resolution. The correlation coefficient between the flowering and boll-setting period of NDVI and cotton yield was larger than 0.80, and the maximum correlation coefficient was 0.90. During flowering and boll-setting period, NDVI average value determination coefficient of model was 0.82, RMSE was 463.69, thus using flowering and boll-setting stages was more suitable for cotton yield prediction than other growth stages. The NDVI on July 25th was the optimal model input variable for single period, and the average of three growth stage NDVI value before the maximum NDVI was the optimal model input variable for multiple growth stages, and we used the average of NDVI value in the whole flowering and boll-setting stages as input variable for the model. The determination coefficient (R2) reached 0.82; and the R2 was 0.81, RMSE was 477.82 with the model of taking the maximum NDVI value as input variable. The correlation coefficient was large between NDVI of any time in the flowering and boll-setting growth stage and the cotton yield. Thus any growth stage during the flowering and boll-setting period can be used to predict cotton yield. In addition, in the study area, we chose the model of the plot A to verify the plot B model. We concluded that RMSE was 387.53, compared with the common modeling RMSE value, and the test result was ideal. The research results of this paper provided the theoretical basis for the research on the maximum synthesis of MODIS_NDVI. It provided a reference for other crops to establish the yield estimation model.