The aim of the study was to attempt to build and validate the neural model controlling the qualitative selection of the stream of grain mass as early as the stage of combine harvesting of winter wheat. The model uses the highest possible number of data describing locally changeable environmental conditions such as: protein content, moisture and yield of wheat grain, soil abundance in basic nutrients (total Kjeldahl nitrogen, exchangeable phosphorus and potassium, magnesium) and additionally - the pH coefficient, content of organic matter in soil and the relative altitude. The construction of the neural model was preceded with a multiple regression analysis. The results of the analysis (α = 0.05) indicated statistical significance of all of the traits under analysis, which influence grain quality and are defined as the content of protein. The MLP neural network (9-30-1) consisted of one hidden layer containing 30 neurons, one output and nine inputs. The network learning was done with the BFGS (Broyden-Fletcher-Goldfarb-Shanno) algorithm in a single phase during 827 epochs with the SOS error function. The study was a part of the development project No. R12 0073 06 entitled “Development and validation of the technology for separation grain stream during cereals selective harvesting”, financed by the Polish National Centre for Research and Development.