Application of Artificial Neural Networks for Multi-Criteria Yield Prediction of Winter Wheat

Three independent models were constructed for the prediction of yields of winter wheat. The models were designed to enable the prediction of yield at three dates: 15th April, 31st May, and 30th June. The models were built using artificial neural networks with MLP (multilayer perceptron) topology, based on meteorological data (air temperature and precipitation) and information on applications of mineral fertilizer. Data were collected in the 2008–2015 from 301 crop fields in the Wielkopolska region of Poland. The evaluation of the quality of predictions made using the neural models was verified by determination of prediction errors using the RAE, RMS, MAE and MAPE measures. An important feature of the constructed predictive models is the ability to make a forecast in the current agricultural year based on up-to-date weather and fertilization information. The lowest MAPE error values were obtained for the neural model WW30_06 (30th June) based on an MLP network with the structure 19:19-15-13-1:1, the error was 8.85%. Sensitivity analysis revealed which factors had the greatest impact on winter wheat yield. The highest rank (1) was obtained by all networks for the same independent variable, namely, the mean air temperature in the period from 1st September to 31st December of the previous year (T9-12_LY).