Equity-Constrained Federated Learning for Computer Vision Agriculture Disease Detection Under Village-Level Device Heterogeneity: A Fairness-Accuracy Trade-off Evaluation

Federated learning promises to train crop-disease classifiers across village devices without centralising leaf imagery, but the accuracy it delivers is unevenly distributed when devices differ in camera quality, connectivity and label reliability. the present analysis formalises village-level device heterogeneity as a constrained optimisation problem in which a global model must satisfy per-device accuracy floors while minimising aggregate cross-entropy loss, and it derives the fairness-accuracy frontier that any equity-constrained aggregation rule can attain. Using a theoretical analysis of the constrained empirical risk, the paper shows that the Lagrangian dual of the per-device accuracy constraint yields a device-specific reweighting of local updates, and it establishes a bound on the accuracy gap between the worst-served and best-served village as a function of the heterogeneity in local sample size and label noise. The analysis is illustrated with a meta-analytic calibration in which the pooled association between a village-level infrastructure proxy and child mortality is estimated across 24 African countries (pooled effect -2.19361, standard error 0.26513, 95% confidence interval -2.71326 to -1.67395, p < 0.001, I-squared 99.2%), demonstrating the scale of between-village heterogeneity that equity constraints must absorb. The conclusion is that equity constraints are feasible at modest accuracy cost only when device-level label noise is bounded, and that unconstrained federated averaging systematically under-serves the villages whose data are most valuable for surveillance.