The present study piled a cube array with the winter wheat's cap reflection hyperspectra sampled from 35 experimental districts during 8 growth periods, and then used a multiway partial least squares (NPLS) algorithm to establish a model to assess the per-mu-yield of winter wheat. The model was used to predict the per-mu-yields of other 13 experimental districts. The correlation coefficient of predicted and actual values was 0.9366, and the mean relative error was 4.44%. These results showed a good prediction of the model. The study has guiding meaning to actual yield assessment.