A first step towards data ecosystem design for AI in agriculture: A case study assessing variable importance for optimising nitrogen decision-making in wheat

Designing an Artificial Intelligence (AI) system in agriculture involves several key steps. First, the correct data and surrogate information must be collected to determine feasibility and ensure the system aligns with existing knowledge. Next, the operational AI is built using field-appropriate variables, which might include proxies for critical drivers. These steps form the data ecosystem needed to deploy the AI model. Lastly, the system's evaluation assesses statistical performance and ease of deployment, including data collection and recommendation speed. In nitrogen (N) fertiliser management, agronomic knowledge shows that optimal N rates depend on crop demand and soil nutrient supply. To establish key variables for predicting the optimal N rate and evaluate how the selection method guides the selection of variables in the database, three wheat field seasons (2019–2020) with on-farm strip trials were studied. Forty-three variables, including weather, soil, plant, and reflectance data, were collected. Using AIC stepwise regression with ANOVA and recursive feature elimination and SHAP (Shapley Additive exPlanations) via random forest, variable selection methods were compared. Random forest identified five key variables: NDVI and NDRE ratios between farmer and rich strips, total soil carbon at depth, chlorophyll canopy content index, and available N at mid-season. Stepwise regression selected eight variables, including total N and carbon at various stages and photosynthetically active radiation. The random forest approach, requiring only five variables, offers a simpler, reproducible system for N decision-making in dryland wheat production, highlighting the importance of efficient database design in agricultural AI systems.