Answers are grounded in this paper and cite their sources.
Scarce empirical damage data constrains how well tsunami damage models perform locally and transfer between Pacific Island countries. We quantify these constraints using damage data from the 2009 South Pacific (Samoa 2009, earthquake-generated, n = 85) and 2022 Hunga Tonga-Hunga Ha'apai (Tonga 2022, volcanic-generated, n = 375) tsunamis. Random Forest and Extreme Gradient Boosting (XGB) were trained on 19 hazard, building and environmental variables, evaluated under four-, three- and two-level damage schemes and compared with simpler baseline models at four levels. At four levels, XGB accuracy was high for Tonga 2022 (0.95) but no better than majority-class accuracy for Samoa 2009 (0.54, 95% CI [0.44, 0.64], against 0.45), and transferred models showed no useful skill. A two-level scheme representing repairable from unrepairable damage transferred in both directions, with skill of +0.13 to +0.34. Multinomial logistic regression matched XGB for Samoa 2009, whereas XGB exceeded it by 0.15 accuracy for Tonga 2022, placing the threshold for ensemble learning between 85 and 375 observations. Sample size and class imbalance explained 0.17 of the 0.40 accuracy gap between events, and the remainder was consistent with hazard simulation uncertainty, under which flow depth, the most important predictor, ranked first in only 10 of 40 Samoa 2009 perturbations. Four-level classification needed approximately 180 to 250 training observations, more than twice the Samoa 2009 sample. With small damage samples, reducing the number of damage levels for prediction rather than increasing algorithm complexity, produced more transferable models.
No takes yet. Share an insight, caveat, or question.
Paulik et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: