Uncertainty-Aware Evidential Deep Learning for Artificial Intelligence Crop Yield Estimation Under Adversarial Weather Perturbation: A Computer Risk-Calibration Evaluation

Crop yield estimation models that report only a point prediction conceal the epistemic risk that matters most to a grower or a reinsurer: whether the model knows when it does not know. The framework treats the evidential output as a Dirichlet concentration over yield outcomes, defines an adversarial perturbation operator on the weather covariate, and derives a calibration bound that links the evidential uncertainty mass to the worst-case perturbation radius. The empirical component uses NASA POWER monthly reanalysis (MERRA-2) air temperature at 2 m for Accra, Ghana (5.6, -0.19) over 2000-2023, comprising 312 monthly observations with a mean of 26.521 C, a standard deviation of 1.24 C and a linear warming trend of 0.1622 C per decade. The evidential model's uncertainty mass rises monotonically with perturbation radius while the deterministic model's residual grows without a corresponding signal, so that the evidential model's risk-coverage curve dominates at every operating point examined. The contribution is a calibration criterion, not a new architecture: it specifies the conditions under which an evidential yield model can be trusted to flag its own failure under weather stress, and it identifies the perturbation regime in which that guarantee breaks down.