Biomass is one of the important indicators reflecting the growth status of winter wheat. In order to achieve efficient and non-destructive monitoring of winter wheat biomass, field experiments under different nitrogen conditions and varieties were carried out in this study. DJI Phantom 4 Multispectral (P4M) UAV was used to obtain the multispectral images of six growth stages during the 2020-2021 growth season (including jointing, booting, heading, flowering, filling and maturity stage), and the corresponding vegetation indices and textures were extracted to correlate with the field measured data. The optimal indices and feature features were selected as the input independent variables of regression model. Linear regression, partial least square (PLSR) and random forest (RF) methods were used to assessment biomass of different growth stages, and the determination coefficient (R2), root mean square error (RMSE) and mean absolute error (MAE) were used to evaluate the models. The results showed that: (1) The prediction accuracy of the regression model considering texture features is higher than that of the model using only vegetation indices. (2) When modeling based on optimal vegetation indices or optimal vegetation indices and optimal texture features, the overall performance of linear model was better than that of RF model. When modeling based on total vegetation index and texture feature, the stability of PLSR and RF models was poor, and the estimation accuracy fluctuated greatly in different growth periods. (3) Booting stage was the best choice for biomass estimation by linear model, while the heading stage was the best satge for PLSR and RF models. In conclusion, the coupled texture features of vegetation index extracted by P4M can effectively improve the estimation effect of winter wheat biomass, and provide a reference for the estimation of wheat biomass on a small scale by UAV technology.