Machine learning to realize phosphate equilibrium at field level in dairy farming

An important factor in circular agriculture is efficient application of animal manure. Therefore, input and output of nutrients, like phosphorus (P), need to be balanced. Currently, manure application is regulated with rather fixed P application norms as a generic translation of P yields of grassland and maize. Predicting P yields based on field specific, historical data could be an important step to better balance P input and output. This study's objective was to predict P yields based on field and weather data, using machine learning. The dataset contained 640 records of yearly crop yields per field between 1993-2016 with information on P input and output, irrigation, and soil status at field level as well as local weather data. Generalized boosted regression (GBR) was used to predict P yields for the last five years based on information from all previous years. Model performance was evaluated per year as well as together by plotting observed versus predicted values of all five years in one plot. This final plot was compared to a plot with the currently used generic application norms. Model performance per year showed that GBR could predict the trend from low to high rather well (correlations of ~0.8). Results of the five years together showed that GBR performance was better than the generic application norms (correlation 0.68 vs 0.59; RMSE 7.3 vs 8.2). In conclusion, GBR contributed to defining more flexible P application norms with the aim to realize a phosphate equilibrium.