The Bayesian backtracking problem in oceanic drift modelling

Backtracking the drift of particles and substances is central to a range of studies in oceanography as well as in law enforcement, search and rescue and the mapping and investigation of marine pollution. Here we demonstrate how a Lagrangian particle model can be used in a forward mode with a Bayesian prior estimate on the release location of the object of interest. We show that for well-behaved drifters, forward and backward (reverse modelling) yield similar results over short periods, if the currents are only weakly divergent. However, for drifters undergoing discontinuous state changes, such as stranding, or objects abruptly and irreversibly changing their drift properties, or for buoyant drifters in strongly convergent flows, backward drift can yield wrongful search areas. We demonstrate this for a case where a liferaft is assigned a wind-speed dependent probability of capsizing, leading to an instantaneous change in drift properties. We also demonstrate the forward and backward methods for a drifter release experiment in the Agulhas current where we also assess the challenges of biases in the current fields. Finally, a method for incorporating multiple observations of debris with a forward model in the Bayesian posterior estimate of the initial location is outlined. • Bayesian backtracking with forward Lagrangian models is presented. • Naive backtracking is compared against forward modelling. • We present the limitations of naive backtracking with discontinuous state changes. • A method for estimating the origin of multiple observations is presented.