Distributionally Robust Conformal Calibration for Artificial Intelligence Reference-Evapotranspiration Forecasting Under Sensor Degradation: Coverage Gaps and Irrigation Advisory Reliability

Reference evapotranspiration (ETo) forecasts underpin irrigation scheduling, yet the conformal calibration that supplies their prediction intervals assumes exchangeable residuals and a fixed sensor regime. When radiation, humidity or wind sensors degrade, the residual distribution shifts in ways that marginal coverage cannot detect, and advisory reliability falls silently. The framework is methodological: it specifies a calibration procedure, a degradation-aware nonconformity score, and a finite-sample coverage guarantee, and it is illustrated on NASA POWER monthly reanalysis for Kampala, Uganda (2000-2023, 312 monthly observations, mean temperature 22.443 C, warming trend -0.1564 C per decade). Marginal conformal coverage can remain nominal while conditional coverage on degraded-sensor strata collapses, and that the robust calibration restores a distribution-free lower bound on advisory reliability at a quantifiable width cost. The contribution is a calibration protocol that makes sensor degradation an explicit object of inference rather than an unmodelled nuisance, together with the coverage diagnostics that make its failure modes visible.