Counterfactual-Policy Inference for Computer Agriculture Fertilizer Timing Under Unobserved Weather Confounding: An Artificial Intelligence Doubly-Robust Evaluation

Fertilizer timing recommendations in computer agriculture are typically learned from observational field data in which weather is an unobserved confounder: farmers apply nitrogen when rain is expected, and the same weather that triggers application also drives yield. The framework combines an outcome regression on weather covariates with a propensity model over observed application determinants, and identifies the average treatment effect of a timing policy through an influence-function correction that remains consistent when either nuisance model is correctly specified. A sensitivity bound is derived that quantifies the bias introduced by an unobserved confounder of bounded strength. The analysis is methodological: it specifies the estimator, states its identification assumptions, and derives the finite-sample bias bound, without reporting field-trial outcomes. The contribution is a policy-evaluation protocol that makes the weather-confounder assumption explicit and auditable, and that yields a defensible interval for the effect of shifting fertilizer timing under a warming climate.