DFlee: An Agent-Based Model for Flood-Induced Population Displacement

Each year, natural disasters displace millions globally, highlighting the need for accurate population movement predictions to improve disaster response. This study presents DFlee, an agent-based model for simulating flood-induced evacuations. Built on the Flee toolkit, originally developed for conflict-driven displacement, DFlee has been adapted for flood scenarios through revised rule sets, refined parameters, and new algorithms. The model integrates flood data and agent behaviour to simulate evacuation and return decisions across diverse hydrological contexts. We ground synthetic scenarios within U.S. Geological Survey (USGS) flood frequency classifications, testing DFlee across riverine minor floods (10--25 year recurrence), riverine major floods (50--100 year), coastal storm surges (25--50 year), and compound flooding (100+ year). Given the limited availability of empirical flood-displacement datasets, validation is conducted using stylised facts, robust behavioural regularities reported in the literature. We focus on three well-documented facts: (i) evacuees prefer closer safe locations, (ii) most households return promptly after floodwaters recede, and (iii) evacuation timing is heterogeneous, with delayed responses influenced by social observation. DFlee reproduces these dynamics, providing confidence in its behavioural realism. The model also demonstrates near-linear scalability, completing simulations with up to 160,000 agents in under 10 minutes, making it a practical tool for forecasting displacement trends and informing humanitarian decisions in flood-prone regions.