Urban flash-flood management requires turning probabilistic forecasts into concrete, resource-constrained actions. We develop a forecast-to-decision pipeline that links calibrated incident-level harm prediction with prescriptive optimization. Using a processed DesignSafe-CI Texas flash-flood incident dataset covering 2005–2019 (6,137 incidents; 2.17% harm prevalence), we construct GIS-derived predictors of hazard intensity, exposure, and transportation infrastructure, plus physically motivated interactions (e.g., precipitation × bridge density). An ℓ 1 -regularized logistic model (LASSO) with out-of-fold isotonic calibration provides decision-grade probabilities and automatic feature selection. Temporal generalization is assessed with a strict year holdout (train 2005-2018; test 2019), leave-one-year-out (LOYO), and rolling-origin schemes. On the 2019 holdout, the calibrated model attains ROC-AUC 0.783, PR-AUC 0.127 (about six times prevalence), and Brier 0.021, although the holdout contains only five positive cases and therefore yields wide uncertainty intervals; complementary LOYO validation gives AUC values spanning 0.66–0.99, reflecting interannual variability that random splits obscure. We embed the learned calibrator as a piecewise-linear mapping in a mixed-integer program that selects intervention sets meeting risk-coverage targets at minimum effort. Results show strong spatial concentration of expected harm within the candidate set of historically active locations: intervening at 61 of 220 candidate assets (28%) captures 80% of baseline risk; 98 candidate assets (45%) achieve 90%, with certified optimality gaps < 0.1%. Leave-one-year-out frontier analysis confirms this concentration pattern is stable across years (median: 27% of assets for 80% coverage; IQR < 2 percentage points), indicating that the prescriptive conclusions are not artifacts of the single holdout year. The framework is reproducible, interpretable, and operationally tractable, providing a transparent pathway from geomatics-informed risk to prioritized preparedness and response portfolios for known high-risk locations.
Gupta et al. (Tue,) studied this question.