ABSTRACT:

We introduce FLOOD-XML (Flood LOss and Observed Damage using eXplainable Machine Learning), a multi-model ensemble dataset that addresses this gap by developing data-driven flood damage functions and estimating event-level damages globally using a suite of machine learning (ML) models. Long-term flood event records from the Emergency Events Database (EM-DAT) since the 1980s have harmonized with event attributes (location, duration, fatalities, and affected area) to train ML models spanning tree-based methods, support vector regression, deep learning, and non-parametric approaches. Damage functions are calibrated using economic losses reported in U.S. dollars in EM-DAT and validated through leave-one-out and group K-fold cross-validation before being used to estimate flood damage for the Dartmouth Flood Observatory (DFO) dataset. Multiple evaluation metrics are used to assess model performance, and feature contributions are determined through explainable ML techniques.
The FLOOD-XML dataset provides globally consistent, explainable flood damage estimates, offering risk analysts, insurers, and policymakers a reproducible tool for impact-based flood hazard assessment and adaptation planning.

FLOOD-XML: https://github.com/nassernajibi/FLOOD-XML
FLOOD-XML Mapper: https://nassernajibi.github.io/FLOOD-XML-Mapper