Online Continual Learning for Artificial Intelligence Agriculture Weed Detection Under Species-Order Drift: A Computer Forgetting-Robustness Evaluation

Weed detection models deployed in agricultural fields encounter weed species in sequences that are neither random nor stationary, and the order in which species appear can determine how much of an earlier species the model retains. the present analysis treats species-order drift as a formal property of the online continual learning problem and derives a forgetting-robustness bound for gradient-based detectors under a declared variance structure. The analysis characterises the operating characteristics of a before-after-control-impact design applied to continual-learning evaluation, using a between-cluster standard deviation of 6.0, a residual standard deviation of 4.5 and an intraclass correlation of 0.64. Power was estimated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The design required 56 clusters to detect a 5-percentage-point effect at 80% power. The derived bound separates order-induced forgetting from capacity-induced forgetting and shows that the two are not identifiable from aggregate accuracy alone. The contribution is a formal criterion for when species-order drift can be distinguished from ordinary catastrophic forgetting, together with the design parameters needed to evaluate forgetting-robustness in agricultural deployment conditions.