Quantifying the dynamics of collective human decision making is essential to optimizing communication systems, transportation networks, and action protocols for ensuring public safety in uncertain and risky situations. Representing human factors has been elusive since decision making is driven by a host of interacting factors such as time pressure, perceived risks, and social influence. We develop two complementary data-driven models to describe decision making in natural disaster scenarios based upon the observed behavior of subjects in a controlled behavioral experiment where participants must decide whether and when to evacuate from a virtual natural disaster. We first develop a rate-based model that quantifies an individual’s decision-making strategy as a function of the reported disaster threat level, which we observe to be a primary influence of collective behavior, and use this model to simulate and predict evacuation at the population level. We investigate the effect of social influence on strategy adjustments that emerge in group contexts, comparing evacuation decisions made individually with those that are arrived at via group consensus or vote. We find that the accurate characterization of group behavior mandates the consideration of individual heterogeneity, leading to a second approach using an artificial neural network which incorporates more experimental parameters, including a measure of individual variation, to predict precise evacuation times. An alternative method of quantifying individual differences in the neural net- work uses participants’ social media data as input to the model and achieves comparable prediction accuracy.