AgWise based site-specific and climate smart fertilizer advisory for long rains maize in Western Kenya

Smallholder maize systems in Western Kenya face chronic nutrient depletion, widespread soil acidity, and increasing seasonal rainfall variability. These interacting constraints make fertilizer decisions economically risky, particularly when recommendations are generic and do not account for local soil and climate conditions. This report documents a machine-learning (ML) based pipeline that generates site-specific nitrogen (N) and phosphorus (P) fertilizer recommendations for the long rainy season. The pipeline integrates geo-referenced agronomic trial data with spatial covariates (soil, topography, and climate), trains predictive models of maize grain yield, and then searches over feasible N and P application rates to identify locally appropriate recommendations.
The approach supports risk-aware guidance through rainfall scenario analysis. Integration with operational seasonal forecast products will be added in the next phase. Model diagnostics (held-out validation and test splits) indicate good predictive skill for advisory: R2 is 0.74 at training and 0.75 at testing, and prediction performance is consistent across a wide yield range, with some residual dispersion at the upper tail. Interpretation tools (permutation importance, SHAP summaries, and partial dependence plots) confirm that fertilizer N and P rates, rainfall amount/variability, and key soil properties are among the strongest drivers of predicted
yield response. The current outputs include (i) summary statistics on spatial and temporal coverage of the trial dataset, (ii) model performance and interpretability plots, and (iii) gridded maps of best N, best P, and expected yield under rainfall tercile scenarios. These outputs are intended to support iSPARK Work Group 3 (WG3) evaluation of the Agwise soil nutrient management innovation
bundle and to provide a transparent technical basis for advisory deployment and iterative improvement.