Optimization of precision fertilizer management to reduce nitrogen pollution while maintaining agricultural productivity

Cheap and abundant nitrogen (N) fertilizer has driven revolutionary increases in global crop yields. However, N is susceptible to being lost from agricultural fields and transported into water bodies where, in excess, it contributes to drinking water contamination, harmful algal blooms, and hypoxia. This work investigates the potential to improve the Pareto front of crop yield and N loss outcomes through high-frequency, in-season soil sampling and fertilizer application. The current paradigm in agriculture is (1) to apply N either entirely before plant emergence or split between pre-plant and a single “side-dressing” after plant emergence and (2) to measure soil N concentrations yearly or less often. However, advances in agricultural field robotics and remote sensing are making it possible to attain and act on soil data more quickly and precisely and thus to maintain crop yield while decreasing N losses. To date though, few studies have attempted to describe optimal fertilizer strategies that make use of such high-frequency monitoring and actuation tools or how much N loss abatement these strategies could achieve. Indeed, such analysis is challenging due to the complex and high-dimensional nature of the N management problem.

This study provides a model abstraction and discretization scheme that make dynamic programming (DP) for N management computationally feasible while maintaining decision-relevant features of the problem. The DP algorithm computes a strategy on incremental fertilizer application rates and timing based on the N and water content in the soil as well as weather predictions. The strategy is optimized for a maximum expected land profitability minus a fine that is incurred when N losses exceed a statutory threshold. By varying the level of the fine for N loss violations, we map the DP outcomes to a Pareto front that characterizes the maximum land profitability that can be achieved at different likelihoods of violating the N loss threshold. Finally, we compute multiple strategies via DP, each using different frequencies for soil monitoring and fertilizer application, and we show that increasing the frequency shifts the Pareto front toward improved (lower) likelihoods of N loss violation without sacrificing land profitability.