Reinforcement-learning irrigation scheduling under soil-moisture sensor dropout: a climate-analytics evaluation of yield-regret bounds and groundwater-drawdown equity

Soil-moisture sensor dropout is a routine operational failure in smallholder irrigation schemes, yet reinforcement-learning (RL) schedulers are typically evaluated under the assumption of continuous state observation. The framework yields three results. First, a dropout-adjusted regret bound is derived in which the effective horizon is inflated by the expected dropout duration, so that the per-season yield regret of any no-regret scheduler grows with the square root of the dropout-adjusted horizon rather than the nominal one. Second, the bound is shown to be tight for the class of optimistic least-squares schedulers under a persistence dropout model, and slack for schedulers that maintain an explicit observation model. Third, the drawdown-equity criterion is formalised as a Gini coefficient over per-farm groundwater extraction, and the framework predicts that dropout-induced regret is distributed regressively across farm sizes because small farms have fewer sensors per hectare. The Harare temperature record is used to parameterise the seasonal evapotranspiration demand that enters the regret bound, not to estimate yields. The contribution is a formal bridge between RL regret theory and groundwater-equity accounting that makes the equity cost of sensor dropout an explicit term in scheduler design.