Above ground biomass (AGB) is one of the most commonly used traits indicating the growth of winter wheat at early growth stages. It is of great practical significance to monitor the growth and to estimate the yield. The conventional methods involving destructive sampling and manual calculating of the dry weight to measure AGB are prohibitively time consuming and laborious. The non-destructive approach to estimate AGB of winter wheat is through the estimation of vegetation indices (VIs) and regression analysis, which heavily depends on tools such as Remote Sensing and LiDAR. Therefore, the method is subject to specialized knowledge and high-cost. An estimating method for above ground biomass of wheat winter at early growth stages was proposed by using the digital images of winter wheat canopy. The canopy images were captured by a digital camera. The image segmentation of vegetation was achieved by using Canopeo. Based on the segmented images, eight images features, i. e., CC (Canopy Cover), ExG (Excessgreen), ExR (Excess red), ExGR (ExG-ExR), NGRDI (Normalized Green-Red Difference Index), GLI (Green Leaf Index), RGRI (Red-Green Ratio Index) and RGBVI (RGB Vegetation Index), were extracted. Correlation analysis was conducted between pairs of the images features and the biomass measurements. The feature that were highly correlated to the biomass measurements were used to build the estimation model. Results showed that ExR, GLIand RGBVI were not correlated to biomass, resulting in the elimination in the following experiments. The left five features were highly correlated to the biomass measurements. Among the five selected features, CC, ExG and ExGR were positively correlated to biomass measurements while NGRDI and RGRI were negatively correlated to biomass measurements. Based on the selected features, four model, i. e., Partial least squares regression (PLSR), BP neural network (BPNN), Support vector machine regression (SVR) and Random forest (RF), were built to estimate the above ground biomass. The influences of the number of features and the sowing density on the estimating accuracy were analyzed quantitatively. Results showed PLSR achieved the best accuracy based on the selected five features, the R2 value was 0.801 5, and the RMSE was 0.078 8 kg·m-2, indicating that the PLSE was able to accurately estimate the above ground biomass of winter wheat at early growth stages. Thenumber of feature was proved to be an influence of factor. The accuracy of the model decreased with the reduction of the number of features. Experiments on the models by using different sowing density datasets were conducted as well. The results showed that PLSR outperformed the other three models over all the density datasets, the R2 values were 0.897, 0.827 9 and 0.788 6, and the RMSE values were 0.062, 0.072 and 0.079 1 kg·m-2, indicating that the PLSR the estimated above ground biomass had a goof agreement with the ground truth. With the increase of the sowing density, the accuracy of all the models decreased while the PLSR achieved the minimum reduction. In Summary, the above ground biomass can be estimated by using the digital images, which can provide support to the field management of winter wheat at early growth stages.