Evaluation of Integrated-Gradients Attribution Stability for Artificial Intelligence Agriculture Drought-Stress Classifiers Under Multispectral Band Corruption: An Interpretability Sensitivity Bound for Smallholder Irrigation Triggers

Attribution maps produced by Integrated Gradients (IG) are increasingly used to justify irrigation-trigger decisions in smallholder agriculture, yet the stability of these maps under realistic multispectral sensor degradation remains unquantified. The bound is derived from the path-integral definition of IG and the Lipschitz continuity of the classifier gradient, yielding an explicit dependence on per-band perturbation magnitude and on the condition number of the band-covariance matrix. A controlled corruption protocol is specified across four degradation regimes (gain drift, additive noise, band dropout, spectral smearing) and applied to a computed drought-stress classification task calibrated to the Nairobi record. IG attribution rank stability degrades predictably as a function of the bound, and that the bound provides a conservative but actionable trigger-reliability certificate for irrigation advisory systems. The contribution is a methodological framework that converts an interpretability claim into a testable numerical contract suitable for deployment auditing in data-scarce smallholder contexts.