For the planning of site-specific nitrogen fertilization, adequate decision rules are needed. Prerequisite for site specific nitrogen fertilization is the site specific forecast of yield. For this the use of artificial neural networks (ANN) has proven particularly interesting. Therefore, ANN based small-scale yield forecasts are realized in order to deviate the economic optimum of fertilization. The basis of yield forecasts with ANN are different site-specific input variables that have presumable impact on yield expectation. These input variables for instance could be recorded yield, electrical conductivity, relief (e.g. topographic wetness index (TWI)), draft force resistance, vegetation indices like red edge inflection point (REIP), previous fertilizer applications and so on. In many years the economic advantage of using ANN for nitrogen fertilization is approved. The results are largely promising, but not sustainable in every case. The data survey for the training set underlies natural disturbance. So the accuracy of small-scale yield forecasts varies considerably from year to year. Also direct impact like unreliable yield or electric conductivity recording influences the quality of input data. To improve the quality of results, it may be necessary to manipulate existing input and target variables. It has to be tested whether a classification of the input and target variables, in comparison to the metric scaled input and target variables, offer improvement in accuracy. Therefore, different classification systems are examined in this study. Equal intervals as a classical scheme are tested at first by varying the width of intervals. A further focus lies on quantile classification and at least on a standard deviation classification scheme. Initial studies implementing ANN with focus on soil parameters actually show positive effects, regarding to classification. The paper provides information on the extent to which improvements in the small-scale yield forecast results occur and whether the results found can be generalized. The results found in this paper are essential for further work with ANN in site-specific nitrogen application.