Counterfactual sensitivity analysis of reinforcement-learning irrigation schedulers under unobserved soil-moisture states: a climate-analytics evaluation of water-allocation regret

Reinforcement-learning (RL) irrigation schedulers are trained and evaluated against soil-moisture trajectories that, in field deployment, are only partially observed. The framework is methodological and theoretical; it does not report a field trial. Using a partially observable Markov decision process (POMDP) formulation, the paper defines counterfactual regret as the expected shortfall between the return of the deployed belief-based policy and the return of a clairvoyant policy that observes the latent state, and it shows that this regret decomposes into an information term and an approximation term. A sensitivity coefficient is introduced to quantify how regret responds to perturbations in the observation model, and a bound is proved that relates this coefficient to the contraction properties of the belief update. The framework is illustrated with a climate-analytics case study of Lagos, Nigeria, using NASA POWER monthly reanalysis air temperature at 2 m for 2000–2023 (312 monthly observations; mean 26.414 °C; standard deviation 0.943 °C; warming trend 0.1553 °C per decade). The case study is used only to parameterise the climate forcing that drives irrigation demand; it is not an empirical evaluation of any deployed scheduler. Counterfactual sensitivity analysis of this kind is a necessary precondition for credible climate-analytics claims about RL irrigation schedulers, because regret under unobserved soil moisture is not identifiable from observational evaluation alone.