Evaluation of Concept-Probe Adversarial-Robustness Audits for Computer Vision Agriculture Disease Classifiers Under Herbicide-Drift Imagery: An Interpretability Safety Bound for Smallholder Spray Regimens

Herbicide drift can mimic the visual signature of foliar disease, and a convolutional classifier that reads drift-bleached tissue as infection will trigger a fungicide spray that the crop does not need. Concept-probe audits are increasingly used to certify that such classifiers attend to disease-relevant features, but their behaviour under adversarial drift imagery has not been characterised as a safety property. The classifier is modelled as a composition of a feature extractor and a concept probe, and the audit is defined as a bounded perturbation problem on the probe's decision margin. Three results follow. First, a Lipschitz condition on the probe yields a certified radius within which the probe's concept attribution cannot be reversed by any admissible drift perturbation. Second, the radius translates into a bound on the expected number of unnecessary spray events per hectare per season, expressed through the drift-plausible perturbation set. Third, the bound is shown to be tight for linear probes and to degrade predictably for probes trained on class-imbalanced concept labels. The framework is illustrated with a design-based power analysis of the before-after-control-impact estimator used to validate the audit in field trials, with an intraclass correlation of 0.64, between-cluster SD of 6.0 and residual SD of 4.5; 56 clusters are required to detect a 5 percentage-point effect at 80% power across 2000 replications per design cell. The contribution is a certified, deployable safety bound that links an interpretability audit to an agronomic decision, rather than a post hoc explanation of classifier behaviour.