Conformal Prediction Intervals for Artificial Intelligence Sorghum Yield Estimation Under Sparse Weather Station Coverage: A Computer Coverage-Guarantee Evaluation

Sorghum yield estimation in smallholder regions of southern Africa is constrained by sparse weather-station networks, and the machine-learning models used to fill the resulting gaps typically report point predictions without distribution-free guarantees. The framework is formalised for the case of covariate shift induced by interpolation error, and a weighted conformal score is derived that restores marginal coverage under a bounded shift assumption. The method is evaluated on a real reanalysis record for Harare, Zimbabwe: NASA POWER monthly air temperature at 2 m, 312 monthly observations from 2000 to 2023, with a mean of 18.914 °C, a standard deviation of 3.107 °C and a warming trend of 0.2001 °C per decade. Coverage-guarantee properties are established analytically and illustrated against this record. The contribution is a coverage-guarantee evaluation protocol that separates the uncertainty contributed by the yield model from the uncertainty contributed by the weather interpolation, and that reports interval width as a function of station density rather than as a single scalar. The framework is proposed as a methodological basis for yield-estimation systems that must state what their intervals actually cover.