Adaptive fertilizer management for optimizing nitrogen use efficiency with constrained reinforcement learning

Optimizing nitrogen use efficiency (NUE) in crop production is crucial for sustainable agriculture, balancing the need to maximize yield while minimizing environmental impacts such as nitrogen loss and soil nutrient depletion. Reinforcement learning (RL) emerges as a potent, data-driven approach for achieving optimal farm management decisions, particularly in the context of fertilization, thereby facilitating optimal NUE. Previous literature of RL in crop management have predominantly focused on optimizing yield, profit, or nitrogen loss reduction. However, optimizing NUE has been largely overlooked despite its significance in preventing soil nutrient mining. In this study, we develop an RL environment in various aspects to investigate the capability of RL to optimize NUE through crop growth model simulations. We develop an RL agent with a novel NUE reward function and incorporates action constrains. We compare its performance against baseline methods and other RL agents trained with reward functions from previous literature. Additionally, we evaluate the robustness of our RL agent across various soil conditions, including different initial nitrogen content and drought-(in)sensitive soils. We find that the RL agent trained with our novel reward function is close to the optimal policy, although generalization to different soil texture scenarios prove to be challenging to the RL agent. Further, we identify several open challenges for future work pertaining to RL in crop management.