Evaluation of Causal-Attribution Bounds for Computer Agriculture Yield Anomaly Detectors Under Concurrent Drought and Pest Outbreak Confounding: An Artificial Intelligence Shapley-Decomposition Audit

Yield anomaly detectors in computer agriculture routinely attribute harvest shortfalls to a single dominant stressor, yet drought and pest outbreaks frequently co-occur and confound one another's signatures in remotely sensed and agronomic covariates. The framework is methodological and theoretical: it defines the attribution game, states the confounding conditions under which Shapley values are point-identified, and derives upper and lower bounds when concurrent drought and pest outbreak render the attribution non-identified. A worked illustration uses NASA POWER monthly reanalysis 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 warming trend of 0.1622 C per decade, to show how a single covariate channel can be bounded before attribution is attempted. The contribution is a bound-based audit protocol that replaces point attribution with interval attribution and specifies the evidence required to narrow those intervals in operational detector pipelines.