Answers are grounded in this paper and cite their sources.
Blanket fertilizer recommendations ignore the soil variability of the northern Nigerian savanna. This study evaluated a machine-learning-guided site-specific recommendation (MR) against the regional recommendation (RR), farmers’ practice (FP) and an unfertilized control for maize. Multiple linear regression, Random Forest and Gradient Boosting were trained on 2,103 plot records from 177 nutrient-omission trial fields (2015–2016). Random Forest performed best (held-out R² = 0.61 under a random split; 0.32 when whole fields were withheld) and guided site-specific P and K allocation at a fixed N rate of 100 kg ha⁻¹. The recommendations were tested in 2025 at six on-farm locations in the Northern Guinea (NGS) and Sudan (SS) Savanna, with locations as blocks, using a linear mixed model. Strategy affected grain yield ( P = 0.006). MR gave the highest mean yield (3.57 t ha⁻¹), exceeding FP (2.44 t ha⁻¹) and the control (2.01 t ha⁻¹), but its advantage over RR (3.08 t ha⁻¹) was not significant (0.50 t ha⁻¹; 95% CI − 0.32 to 1.31). MR supplied 36% less N + P₂O₅ + K₂O than RR, raising agronomic efficiency of total applied nutrient from 5.3 to 12.6 kg grain kg⁻¹ ( P = 0.014). Yields were much higher in the NGS than the SS (4.57 vs. 0.98 t ha⁻¹). MR gave the highest net benefit (US$641 vs. US$413 ha⁻¹ for RR) and was the only non-dominated fertilized strategy, but no fertilized strategy out-performed the control in the SS. MR matched RR yields with less nutrient and at lower cost; multi-season validation is needed.
No takes yet. Share an insight, caveat, or question.
Aiyelabegan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: