Advancing dynamic phosphorus prescription models for cassava ( Manihot esculenta ) using a random forest approach

Cassava (Manihot esculenta) is a key staple crop supporting global food security, yet its productivity is increasingly constrained by declining soil fertility and climate variability. In many cassava-producing regions, phosphorus (P) recommendations rely on static fertilization grids that do not adequately account for site-specific soil and crop response variability. This study evaluates a dynamic P prescription framework for cassava using a Random Forest (RF) model and compares its performance with a conventional three-step recommendation approach comprising calibration, site-specific optimization, and grid-based clustering. The analysis used an international dataset combining cassava yield response trials with soil and climatic variables, including PBray-1, soil organic matter (SOM), pHwater, and annual rainfall. Results showed that the traditional grid-based approach exhibited a weak correspondence between prescribed rates and site-specific economic optima. In contrast, the RF-based model achieved improved predictive performance, with R2 values of 0.74 for training and 0.59 for testing. The dynamic framework enabled reconstruction of site-specific P response curves by integrating multiple soil and climatic covariates. Overall, the results indicate that data-driven approaches can complement traditional P recommendation methods by better capturing inter-site variability and supporting more adaptive fertilization strategies for cassava production systems.