Predicting site-specific economic optimal nitrogen rate using machine learning methods and on-farm precision experimentation

Abstract

Applying at the economic optimal nitrogen rate (EONR) has the potential to increase nitrogen (N) fertilization efficiency and profits while reducing negative environmental impacts. On-farm precision experimentation (OFPE) provides the opportunity to collect large amounts of data to estimate the EONR. Machine learning (ML) methods such as generalized additive models (GAM) and random forest (RF) are promising methods for estimating yields and EONR. Twenty OFPE N trials in wheat and barley were conducted and analyzed with soil, terrain and remote-sensed variables to address the following objectives: (1) to quantify the spatial variability of winter crops yield and the yield response to N using OFPE, (2) to evaluate and compare the performance of GAM and RF models to predict yield and yield response to N and, (3) to quantify the impact of soil, crop and field characteristics on the EONR estimation. Machine learning techniques were able to model wheat and barley yield with an average error of 13.7% (624 kg ha−1). However, similar yield prediction accuracy from RF and GAM resulted in widely different economic optimal nitrogen rates. Across sites, soil available phosphorus and soil organic matter were the most influential variables; however, the magnitude and direction of the effect varied between fields. These indicate that training a model using data coming from different fields may lead to unreliable site-specific EONR when it is applied to another field. Further evaluation of ML methods is needed to ensure a robust automation of N recommendation while producers transition into the digital ag era.