Weakly-Supervised Millet Rust Severity Estimation for Artificial Intelligence Agriculture Under Sparse Plant Pathologist Scoring: A Computer Calibration and Fairness Evaluation

Millet rust severity scoring by plant pathologists is expensive, slow and unevenly distributed across agro-ecological zones, which constrains the supervision available to computer-vision models intended for artificial intelligence agriculture. The problem is framed as a design-characterisation question rather than an empirical field study: given a declared variance structure with between-cluster standard deviation of 6.0, residual standard deviation of 4.5 and intraclass correlation of 0.64, the operating characteristics of a before-after-control-impact estimator were characterised across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. Power curves, minimum detectable effect curves, a cluster-count scatter, a kernel density of estimated power, a coefficient plot with 95% confidence intervals and a residual-versus-fitted diagnostic were derived from the declared structure. The design required 54 clusters to detect a 5 percentage point effect at 80 per cent power, a requirement that exceeds the largest grid value of 32 and therefore exposes a calibration gap between what sparse pathologist scoring can support and what fairness evaluation demands. The contribution is a formal statement of the identifiability and power conditions under which weakly-supervised severity estimates can be defended, together with a fairness-evaluation protocol that treats cluster-level heterogeneity as a first-class object rather than a nuisance.