PREDICTIVE, SENSOR-INFORMED NITROGEN MANAGEMENT FOR RAINFED LOWLAND RICE USING MACHINE LEARNING AND DSSAT-CERES-RICE IN AFIKPO, EBONYI STATE, NIGERIA

Smallholder lowland rice in south-east Nigeria receives small, late, fixed nitrogen (N) doses, and it is unclear whether “smart” in-season N control adds value over a well-designed fixed schedule. DSSAT-CERES-Rice v4.8 was coupled with external controllers used to simulate rainfed, bunded FARO 44 rice (literature-derived cultivar coefficients, not field-calibrated) at Afikpo (5.89°N, 7.92°E) over 33 seasons (1993–2025), using CHIRPS rainfall and NASA POWER temperature and radiation. The study compared four kinds of management: farmer practice (46 kg N ha⁻¹ at 21 days after transplanting); 30/40/30 split schedules at 40–120 kg N ha⁻¹; a reactive controller emulating a canopy N-stress sensor; and a three-stage gradient-boosting controller; trained on 2,970 simulated dose combinations and tested leave-one-year-out. The reactive controller out-yielded farmer practice by 0.39 t ha⁻¹ (p = 0.002), but only by applying 8 kg N ha⁻¹ more; at equal total N it yielded 0.44 t ha⁻¹ less than a fixed 30/40/30 split (p < 0.001). With N priced at 4 kg paddy per kg N, perfect in-season information was worth only 93 kg ha⁻¹ over the best history-based fixed schedule. The predictive controller captured little of this (+6 kg ha⁻¹, p = 0.72) for 1 July transplanting, against +90 kg ha⁻¹ (p < 0.001) for 1 June. Correcting N rate and timing relative to farmer practice was worth about 1,700 kg ha⁻¹. NASA POWER rainfall showed a strong drying trend, much weaker in CHIRPS, that changed one conclusion; all results are therefore reported with CHIRPS. In-season prediction adds value, but rate and timing dominate.