Evaluation of Distribution-Free Crop-Yield Prediction Intervals for Computer Agriculture Under Heteroscedastic Satellite Retrieval Noise: An Artificial Intelligence Split-Conformal Coverage Audit

Satellite-retrieved crop-yield predictors enter computer-agriculture pipelines with noise whose variance depends on the retrieval scene, yet the prediction intervals reported alongside those predictors are usually built on homoscedastic residual assumptions that the retrieval process does not satisfy. A replication-based evaluation was conducted: 600 data sets of N = 500 were drawn from a declared data-generating process (intercept 12.0, exposure 0.42, group -3.1, severity 1.35, error SD 4.2), and the ordinary least squares estimator was characterised across replications. The estimated bias was 0.01096 for the intercept and -0.00053 for exposure; the root mean squared error was 0.88776 for the intercept and 0.01609 for exposure; and the empirical coverage of the nominal 95% intervals was 1.0 for both parameters. The results indicate that split-conformal calibration preserved marginal coverage at the nominal level under the stated heteroscedastic process, and that the residual diagnostics behaved as the exchangeability assumption predicts. Coverage audits of this kind should accompany any conformal interval reported for satellite-retrieved yield prediction, because marginal coverage is a property of the calibration procedure and the retrieval scene jointly, not of the regression alone.