The yield estimation of winter wheat is of great reference significance for the country to guide agricultural production and make accurate management decisions for the whole growth period of winter wheat. The moderate resolution imaging spectroradiometer (MODIS) was taken as the data source to extract the information of visible and near-infrared bands at different periods and selected seven remote sensing features of vegetation, including normalized difference vegetation index (NDVI), normalized difference water index (NDWI), soil adjusted vegetation index (SAVI), optimal soil adjusted vegetation index (OSAVI), green-normalized difference vegetation index (GNDVI), modified soil-adjusted vegetation index (MSAVI) and green red vegetation index (GRVI). Using its histogram distribution information as an input variable, a convolutional neural network (CNN) was used to predict winter wheat yield, comparatively analyze the performance of NDWI in winter wheat yield estimation, and explore its accuracy changes under the influence of frost damage. The results showed that compared with NDVI, SAVI, OSAVI, GNDVI, MSAVI and GRVI, NDWI had a better prediction effect on the yield prediction of winter wheat in the early growth stage, the prediction accuracy of NDWI was higher than that of NDVI, SAVI and other vegetation indices before and after yield detrending, the coefficient of determination (R2) was up to 0.79, and it can still maintain a good prediction effect under the influence of frost damage. NDWI had the greatest influence on the final yield of winter wheat at the stage of heading and grouting. From April 23th to April 30th, the R2 of NDWI can reach 0.72. In terms of spatial distribution, the winter wheat in the study area had the highest yield in the east, followed by the middle, and the lowest yield in the west, and the western and northern mountains and the eastern plains at the junction of large error. The results could provide scientific reference for early growth yield prediction of winter wheat.