Comparing Random Forest to Bayesian Networks as nitrogen management decision support systems

Nitrogen (N) is notoriously difficult to manage and there are many approaches for fertilizer N rate recommendations. Existing fertilizer N rate recommendation systems can be improved by incorporating the effects of weather on sidedress economicoptimum N rates (EONR). In this study, we evaluated the performance of machine learning methods, a Bayesian Network (BN) and a Random Forest (RF) for estimating EONR for corn. BN draws relationships between variables based on assumptions about conditional independence, where the model is structured by an algorithm or, in this case, expert opinion. In contrast, RF determines model structure based on the input variables and model output. The models were trained and validated using a large database (n = 324) of corn yield response to N fertilizer collected across southern Ontario. Sixty‐six of the 324 site‐years were used for validation with success assessed by the frequency that N rate predictions that produced net returns were within CAN$25 ha−1 of the observed EONR. The success rate was 64% and 48% for the BN and RF, respectively. Both models incorporated weather from planting to sidedress and outperformed a benchmark provincial N recommendation system. We argue that BN has advantages when some input variables are unknown or uncertain and for improving model structure with stakeholder feedback. Moreover, RF is easy to implement but the model structure must use point estimates instead of probabilities for uncertain parameter values such as future weather. BN represents a more flexible modeling approach than RF for incorporating both modeling and stakeholder input. Core Ideas Two machine learning models were benchmarked against current provincial N recommendations for corn in Ontario. Machine learning derived N recommendations were financially sound more frequently than current recommendations. Bayesian Network models are more difficult to implement than Random Forest models. Adjustments for weather and SOM improved N recommendations derived from machine learning.