Uncertainty-Calibrated Evidential Segmentation for Computer Agriculture Crop-Disease Triage Under Annotator Disagreement: An Artificial Intelligence Calibration-and-Deferral Evaluation

Crop-disease triage from field imagery is routinely evaluated as if each pixel carried a single ground-truth label, yet phytopathologists disagree systematically about lesion boundaries, and that disagreement is not noise to be averaged away but evidence about which regions are genuinely ambiguous. The framework is characterised analytically rather than empirically: no observational data are used or claimed. A before-after-control-impact design is specified under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), and the operating characteristics of the design are derived across 2000 replications per cell for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points.54 clusters are required to detect a 5-percentage-point improvement at 80% power under this variance structure, and it formalises the conditions under which deferral improves expected triage utility relative to forced labelling. The contribution is a decision-theoretic account of when an evidential segmenter should be trusted, deferred or overruled in agricultural triage, together with the design parameters that make such an evaluation identifiable.