Adversarial-robustness certification of flood-susceptibility classifiers under projected land-cover regime shift: a computer-science evaluation of false-negative exposure in municipal zoning analytics

Flood-susceptibility classifiers used in municipal zoning are typically validated on historical land-cover distributions, yet zoning decisions bind future development under land-cover regimes that differ from those on which the classifiers were trained. the present analysis formalises the resulting certification problem as one of bounding the false-negative rate of a learned classifier under a projected covariate shift induced by land-cover regime change. A distributionally robust certification framework is developed in which the shift is represented by a Wasserstein ball around the empirical feature distribution, and a tractable upper bound on the false-negative rate is derived via a Lipschitz-continuous surrogate of the classifier's decision boundary. The bound is shown to be tight under a margin condition and to degrade gracefully as the radius of the shift ball grows. A case study using NASA POWER monthly reanalysis air temperature for Harare, Zimbabwe (2000–2023, 312 monthly observations, mean 18.914 °C, warming trend 0.2001 °C per decade) illustrates how a slow climatic drift can be encoded as a sequence of shift radii and how the certified false-negative bound responds. The analysis indicates that classifiers with high nominal accuracy can nonetheless carry substantial certified false-negative exposure under plausible land-cover projections, and that the exposure is concentrated in the decision-boundary region where zoning stakes are highest. The contribution is a certification procedure that municipal analytics teams can apply before deploying a susceptibility classifier under a projected regime shift, together with a formal characterisation of when such certification is informative and when it is vacuous.