Evaluation of Saliency-Map Spatial-Agreement Audits for Computer Vision Agriculture Weed Detectors Under Variable Soil-Tillage Residue: An Interpretability Localization Bound for Precision Herbicide Application

Saliency-map audits are increasingly proposed as a quality gate for computer-vision weed detectors deployed in precision herbicide application, yet the spatial-agreement metrics on which such audits rely have no established sampling theory under the variable soil-tillage residue that dominates conservation-agriculture fields. The audit is modelled as a before-after-control-impact design in which the intervention is a residue-stratified retraining pass and the outcome is the intersection-over-union between detector saliency and agronomist-annotated weed masks. Under a between-cluster standard deviation of 6.0, a residual standard deviation of 4.5 and an intraclass correlation of 0.64, the design was evaluated across cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points, with 2000 replications per cell.54 clusters are required to detect a 5-percentage-point shift in spatial agreement at 80% power, and it derives the corresponding minimum detectable effect curve. The contribution is a localisation bound that converts a saliency-agreement claim into a falsifiable statement about the smallest residue-induced misalignment an audit can reliably detect, together with the cluster budget that statement implies for smallholder herbicide-advisory settings.