The over-application of fertilizers is harmful to the environment, therefore the European Union’s “Farm to Fork strategy” stipulates that fertilizer use should be reduced by 20% by 2030. The variable rate technique has demonstrated the capacity to reduce the use of fertilizers by adapting the fertilizer dose to the crop needs. However, transferring technologies tested in small and highly controlled plots is not easy to implement in farms. The aim of this study was to evaluate the ability of different machine learning (ML) algorithms trained in one farmer’s plots with a yield monitor data to make an early yield prediction of another farmer plot. Based on this map and working together with the farmer, a fertilizer prescription map was created. This on-farm experiment was car- ried out on a 12-ha wheat plot located in the Basque Country (North of Spain). To obtain the map with the lowest prediction error, 15 ML models were trained and tested. The auxiliary information used was the previous season NDVI obtained from Sentinel-2 and the elevation. The algorithm that best result obtained was the Random Forest (RF) with an R2 of 0.90 and RMSE of 0.29. Before using the prescription map, this was checked with a map elaborated by the farmer and consequently slightly modified. Finally, the adoption of this technique allowed the fertilizer input to be reduced by 19% without decreasing yield. The novelty of the work lies in the collaborative work between farmers and researchers and the use of machine learning techniques to solve the problem of the lack of yield maps.