From static grids to dynamic response models: Predicting wheat phosphorus needs with Random Forest

Amine Kassam¹Lotfi Khiari^2*Ismail Kouera¹Hamza Jouichat³

• ¹Universite Mohammed VI Polytechnique, Ben Guerir, Morocco
• ²Laval-University Department of soil and agrifoog Eng, Quebec, Canada
• ³Universite Laval, Québec City, Canada

Conventional phosphorus (P) fertilizer recommendations for wheat are typically based on single-factor soil test P (STP) classifications and static fertilization grids, which often fail to capture field-level variability and site-specific crop responses. This study introduces a data-driven, multifactorial approach using machine learning, specifically the Random Forest (RF) algorithm, to model wheat yield response to P across diverse agroecological conditions. A global meta-dataset was assembled from 294 trials and 927 observations from 42 peer-reviewed studies. Yield response curves (∆Y = Yfertilized − Yunfertilized) were modeled using six covariates: applied P rate, soil organic matter, pHwater, annual precipitation, soil texture, P fertility class, and application method. These RF-derived curves support localized microeconomic optimization by adjusting P rates according to market prices and field characteristics. The traditional STP-based method showed poor predictive power (R² = 5–9%) and highly variable recommendations across fertility classes. In contrast, the RF model achieved high accuracy (R² = 89% training, 78% testing) and successfully replicated site-specific response curves in 80% of cases. When optimal rates were recalculated from RF-derived models and compared to observed values, a strong correlation was observed (R² = 93%, slope = 1.02). This approach represents a paradigm shift in P fertilizer recommendations by enabling robust, site-specific, and adaptive decision-making. It replaces fixed-rate prescriptions with dynamic response models, offering a scalable tool aligned with the goals of precision agriculture and sustainable nutrient management.