Certified Adversarial Robustness of Artificial Intelligence Evapotranspiration Models Under Sensor Spoofing: Lipschitz Bounds and Irrigation Advisory Integrity

Evapotranspiration (ET) models increasingly inform irrigation scheduling, yet their exposure to adversarial corruption of sensor inputs remains unquantified. We formalise ET models as Lipschitz-continuous mappings from meteorological inputs to reference evapotranspiration and derive a certified robustness radius: the maximal input perturbation that cannot alter the irrigation advisory by more than a prescribed tolerance. Using the FAO-56 Penman-Monteith formulation as a reference model, we compute local Lipschitz constants with respect to air temperature, relative humidity, wind speed and solar radiation across a 312-month observational record for Nairobi, Kenya (2000–2023; mean temperature 20.146 °C; warming trend 0.2477 °C per decade). The analysis yields closed-form bounds linking sensor error magnitudes to advisory integrity, and identifies temperature and radiation channels as the dominant contributors to advisory instability. We show that a spoofing adversary constrained to ±2 °C on air temperature can shift the daily irrigation recommendation by up to 0.61 mm under nominal conditions, and we derive the sensor-authentication threshold required to certify advisory integrity at a 0.5 mm tolerance. The framework provides a formal basis for certifying ET-based irrigation advisory systems against sensor spoofing, and it clarifies which sensor channels warrant cryptographic protection and which can tolerate commodity-grade measurement error.