Reconstructing flood-induced human mobility using LLM agent with disaster context adaptation

Abstract

Understanding human mobility during extreme flood events is critical for emergency evacuation and resilient transport planning, yet flood-induced behavioral heterogeneity remains difficult to characterize. Progress has been constrained by data scarcity and the inherent complexity of emergency decision-making, which often induces abnormal travel behaviors and anomalous mobility fluxes. We propose the Disaster-aware Trajectory Generator (DTG), integrating Large Language Model (LLM) agents into an agent-based framework to generate individual-level mobility trajectories under evolving urban flood conditions. Validated against 1.32 billion mobility records from the 2021 “7.20” Zhengzhou flood in China, DTG outperforms all traditional and LLM-enabled baselines, achieving a flood-induced inflow-shift prediction accuracy of 0.713 with outflow and OD distribution JSD of 0.069 and 0.047 during the disaster period. By combining accessibility-constrained disaster context reasoning with the representation of heterogeneous socio-demographic responses, DTG reveals critical infrastructure vulnerabilities and abnormal behavioral patterns, offering actionable insights for flood-resilient transport planning.