Using a thermal camera to obtain canopy temperatures for winter wheat, an infrared crop water stress index (ICWSI) was calculated in the main water-requirement stage. The performance of a BP neural network was tested with ICWSI values for three different periods in one irrigation circle as independent input factors and observed winter wheat yield after harvest as the output. The topology of the neural network was 3-5-1, and after data normalization, convergence performance was enhanced. Results showed that the maximum relative error was only 3.42%. To confirm the superiority of this method, a common nonlinear regression model was also built to compare the predictions with ICWSI values and the observed yield of winter wheat, but the maximum relative error of this model was higher (18.87%). Comparison between these two mathematical methods shows that the approach of combining thermal camera technology with a BP neural network prediction model, which is more precise for nonlinear prediction, was sufficiently better than other models to predict the winter wheat yield successfully and accurate enough to meet production requirements.