Investigating data-driven approaches to optimize nitrogen recommendations for winter wheat

Abstract

The optimal nitrogen (N) application rate is an important concept to guide the strategic/static N management of crops. However, there is still a knowledge gap on how to combine the optimal N application rate, crop growth status, meteorological, and soil multi-source information to make tactic/dynamic N recommendation. For investigating the data-driven approaches to optimize nitrogen recommendations for winter wheat, this study proposed the concept of optimal topdressing N rate that coupling ‘strategic’ with ‘tactic’ ideas. Then a direct N fertilizer recommendation algorithm based on the optimal topdressing N rate was constructed integrating multi-source data with Random Forest (RF). The agronomic, economic, and environmental benefits of the algorithm were further evaluated based on a large-scale field-level validation experiments, and compared with the indirect N recommendation algorithm of RF-based yield prediction, farmers’ conventional N application rate, and regional optimal N application rate. Finally, multi-objective evolutionary optimization was used to verify the reasonability of the N fertilizer recommendation algorithms. The results showed that the optimal topdressing N algorithm had good model performance (Random Forest prediction R2 = 0.59 ∼ 0.72, RMSE = 39.63 ∼ 40.48 kg/ha), and vegetation indices, maximum temperature, and solar radiation were important features. The validation experiment results showed that compared to conventional management measures, the optimal topdressing N algorithm could reduce N fertilizer application by 18.9 % ∼ 48.4 % with intelligent adjustment, improve partial factor productivity by 12.99 % to 46.09 %, and reduce reactive N losses and greenhouse gas emissions by 20.56 % ∼ 47.24 % and 16.34 % ∼ 40.45 %, respectively. The rationality verification of solving nitrogen topdressing rate through multi-objective evolutionary algorithm proved that both indirect and direct N recommendation algorithms could make decisions within a reasonable threshold. The study provided insights for the data-driven based N management, strategy selection, and benefit evaluation process of N recommendation.