Label-Shift Evaluation for Computer Agriculture Rangeland-Carrying-Capacity Estimators Under Pastoral Herd-Mobility Reclassification: An Artificial Intelligence Importance-Weighted Risk Bound

Rangeland-carrying-capacity estimators trained on historical herd-mobility labels are increasingly deployed in computer agriculture settings where pastoral reclassification has altered the joint distribution of covariates and mobility outcomes. The bound is derived from a density-ratio correction between source and target mobility regimes, with a Rademacher complexity term controlling the weighted hypothesis class. A replication study of 600 data sets of N = 500 drawn from a declared data-generating process (intercept 12.0, exposure 0.42, group -3.1, severity 1.35, error SD 4.2) was used to characterise the estimator's behaviour under the stated process. The estimated bias was -0.00523 for the intercept and 0.0005 for the exposure coefficient; empirical coverage of the nominal 95% confidence intervals was 1.0 for both parameters; and the root mean squared error was 0.84673 and 0.01525 respectively. The bound is shown to be tight enough to detect label-shift-induced risk inflation at the sample sizes typical of rangeland monitoring programmes, and the paper argues that importance-weighted risk should replace unweighted validation error as the primary evaluation criterion for carrying-capacity estimators under pastoral reclassification.