With the development of unmanned aerial vehicle (UAV) and remote sensing technology, crop yield estimation through rapid acquisition of multitemporal and highresolution remote sensing images at field scale has become a research hotspot. In order to determine the optimal growth stage and sampling times for winter wheat yield estimation by UAV multispectral remote sensing, a field experiment on winter wheat in sandy soil was conducted, which was divided into four groups (36 management zones) by irrigation level and five groups (15 management zones) by nitrogen application level. Then the multi-spectral remote sensing images of eight growth stages for winter wheat from rising to late filling were collected by the UAV platform. Additionally, partial least squares (PLS), random forest (RF), least absolute shrinkage and selection operator (LASSO) were used to establish the yield prediction model of winter wheat at each growth stage. Based on the optimal model selected, five yield estimation schemes for the vegetation indices integration during specific growth periods were developed by the cubic B-spline curve and compound trapezoidal formula. The results showed that significant differences were found for estimation accuracy at different growth stages, which was increased with the growth of winter wheat. In single growth period, the optimal growth periods of PLS, RF and LASSO models were early filling, early filling and late filling, respectively. Compared with PLS and LASSO models, RF had the best precision in estimating winter wheat yield except early joint stage. The accuracy of yield estimation in the multi-growth stages was better than that in a single one. The optimal yield estimation scheme was the vegetation indices from rising to the late filling stage for eight sampling times of remote sensing (the determination coefficient R2 of 0.96 and the normalized root mean square error （NRMSE） of 5.39%). Meanwhile, the yield estimation scheme of six sampling times from rising to flowering stage also performed excellently (NRMSE of 9.16%), which meant that it can not only reduce sampling times and remote sensing cost, but also can predict the winter wheat yield in advance. The results were of great significance for the accurate prediction of winter wheat yield by UAV remote sensing.