FF-IR: An information retrieval system for flash flood events developed by integrating public-domain data and machine learning

Structured databases on flash flood (FF) events have limited information and lack emerging data (e.g., visual media). The web is rich with information that can bridge this gap. However, search engines return long lists of webpages cluttered with commercial and irrelevant information. To address this challenge, we developed a FF information retrieval (IR) system (FF-IR). The system uses machine learning (ML) models in novel ways to automate and enhance this IR process. FF-IR consists of three steps: (1) creates event-specific search queries from the publicly available Storm Events dataset and directs them to Google to collect candidate webpages; (2) transforms the candidate webpages to relevance features; and (3) classifies each candidate webpage as relevant or non-relevant using our ML models. FF-IR outperforms direct Google searches by over 100%, measured by the F2-score. Natural hazard researchers and practitioners can use FF-IR to facilitate FF risk assessments and mitigation planning.