Federated Learning for Computer Vision Cassava Disease Triage Under Clinic-Level Label Shift: An Agricultural Fairness Evaluation

Cassava disease triage models trained on federated clinic data face a structural fairness problem that centralised evaluation cannot detect: the label distribution differs across clinics because diagnostic capacity, referral pathways and local pathogen pressure differ, so a model that performs well on pooled data can systematically fail the clinics that most need it. The framework defines clinic-conditional risk, decomposes it into a pooled-calibration term and a shift-amplification term, and derives a fairness bound that relates the worst-clinic risk to the pooled risk through the total variation distance between clinic label priors. A replication-based evaluation of the before-after-control-impact estimator was conducted under a declared variance structure with between-cluster SD of 6.0, residual SD of 4.5 and intraclass correlation of 0.64, using 2000 replications per design cell across cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The analysis indicates that 54 clusters are required to detect a 5-percentage-point effect at 80 per cent power under this variance structure, and that the shift-amplification term dominates the fairness bound when clinic priors diverge by more than the pooled calibration error. The contribution is a shift-aware evaluation protocol that replaces pooled accuracy with clinic-conditional risk reporting, together with a design-power characterisation that tells implementers how many clinics a federated triage evaluation must recruit before its fairness claims become estimable.