In order to further improve the accuracy of UAV remote sensing yield estimation, taking the mulched winter wheat from 2021 to 2022 as the research object, the coating background of the multispectral images at the greening stage, jointing stage, ear pumping stage and filling stage was removed, and the best remote sensing window period was selected, and a mulched winter wheat yield estimation model was constructed based on the optimal vegetation index. The results showed that the canopy reflectivity was closer to the true value after removing the coating background by the support vector machine supervised classification method, and the yield estimation accuracy of the ear stage and the grouting stage was higher. The correlation analysis between vegetation index and yield at different growth stages showed that the best remote sensing window period was the ear extraction period. When the optimal vegetation index was selected based on stepwise regression and full subset regression, it was found that the yield inversion model had the highest accuracy when the screening variables were MCARI, MSR, EVI2, NDRE, VARI, NDGI, NGBDI, ExG based on stepwise regression. In addition, among the yield inversion models constructed by three machine learning methods, partial least squares, artificial neural network and random forest, the random forest model based on stepwise regression method had the highest inversion accuracy, with an R2 of 0.82 and an RMSE of 0.84t/hm2. The research result can provide technical support for improving the accuracy of remote sensing yield estimation and realizing the fine management of agricultural production.