Answers are grounded in this paper and cite their sources.
Disaster early-warning studies often report headline accuracy, which can hide class imbalance, temporal drift and spatial extrapolation error. This paper applies a leakage-aware evaluation protocol to two research modules. The first is flood severity classification among 4,548 recorded flood events from 155 gauges, evaluated with a chronological split (training before 2010, validation 2010–2014, test 2015–2020). The second is a heat-event proxy built from IMD gridded daily maximum temperature (1° grid, 1951–2024). A grid-cell day is positive when it belongs to a run of at least three consecutive hot days, where a hot day exceeds the cell's 1961–1990 calendar-day 90th percentile and its baseline climatological Tmax exceeds 35 °C. The proxy covers 338 eligible cells, April–June target dates and 1991–2024. It is not an official IMD declaration. For flood severity, validation PR-AUC selected logistic regression, which achieved test balanced accuracy 0.6445 and ROC-AUC 0.6483 against a majority-class balanced accuracy of 0.5000. For the heat-event proxy, histogram-based gradient boosting was selected by validation PR-AUC (0.5811). On the 2020–2024 test period it achieved balanced accuracy 88.82%, recall 90.36%, ROC-AUC 0.9548 and PR-AUC 0.5813, but produced 18,415 false positives at threshold 0.50. A simple persistence rule had higher F1 (0.591 against 0.457) and fewer false positives. A spatial hold-out over 51 unseen cells gave balanced accuracy 87.33%. Attributions are associations, not causes, and these results do not establish operational early-warning performance.
No takes yet. Share an insight, caveat, or question.
Patil et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: