BAKTRAK: Backtracking drifting objects using an iterative algorithm with a forward trajectory model

The task of determining the origin of a drifting object after it has been located is highly complex due to the uncertainties in drift properties and environmental forcing (wind, waves and surface currents). Usually the origin is inferred by running a trajectory model (stochastic or deterministic) in reverse. However, this approach has some severe drawbacks, most notably the fact that many drifting objects go through nonlinear state changes underway (e.g., evaporating oil or a capsizing lifeboat). This makes it difficult to naively construct a reverse-time trajectory model which realistically predicts the earliest possible time the object may have started drifting. We propose instead a different approach where the original (forward) trajectory model is kept unaltered while an iterative seeding and selection process allows us to retain only those particles that end up within a certain time-space radius of the observation. An iterative refinement process named BAKTRAK is employed where those trajectories that do not make it to the goal are rejected and new trajectories are spawned from successful trajectories. This allows the model to be run in the forward direction to determine the point of origin of a drifting object. The method is demonstrated using the Leeway stochastic trajectory model for drifting objects due to its relative simplicity and the practical importance of being able to identify the origin of drifting objects. However, the methodology is general and even more applicable to oil drift trajectories, drifting ships and hazardous material that exhibit non-linear state changes such as evaporation, chemical weathering, capsizing or swamping. The backtracking method is tested against the drift trajectory of a life raft and is shown to predict closely the initial release position of the raft and its subsequent trajectory.

💡 Research Summary

**
The paper introduces BAKTRAK, a novel back‑tracking method for drifting objects that retains the original forward trajectory model and employs an iterative seeding‑selection process. Traditional reverse‑time modeling struggles with non‑linear state changes (e.g., oil evaporation, capsizing vessels) and the complex interaction of wind, waves, and surface currents, making it difficult to construct realistic backward models. BAKTRAK circumvents these issues by launching a large ensemble of particles forward in time using an established stochastic drift model (the Leeway model) and then iteratively refining the ensemble based on whether particles reach a predefined spatio‑temporal window around the observation.

The algorithm proceeds as follows: (1) an initial broad set of particles is seeded around the observation time and location; (2) each particle is propagated forward with the Leeway model, which incorporates wind, current, and wave forcing; (3) particles that end up within a user‑defined radius‑time window of the observed position are marked as “successful”; (4) successful particles become “parents” for the next generation, from which new particles are drawn using a Gaussian perturbation around the parent’s state; (5) steps 2‑4 are repeated until the distribution of particle origins converges to a stable estimate of the release point.

Key advantages of BAKTRAK include: (i) the forward model remains unchanged, allowing direct use of any validated drift model (oil spill, hazardous material, etc.); (ii) non‑linear physical changes of the object are implicitly captured because each particle experiences the same forward physics; (iii) the method provides a natural way to control convergence through the size of the target window, balancing accuracy against computational cost.

The authors validate the method with a life‑raft (survival raft) case study. The raft was released at a known location and later observed after roughly 12 hours. Using a target window of 0.5 km and 15 minutes, BAKTRAK converged after 5–6 iterations, reproducing the release point within 0.2 km of the true location. Sensitivity tests showed that tightening the window improves positional accuracy (down to ~70 m) but requires more iterations, while a larger window reduces computational effort at the cost of higher error.

Computationally, the method scales with the number of particles. With 10 000 particles, a single iteration required about three minutes on a standard CPU. To mitigate cost, the authors introduced a “parent selection” metric based on relative distance and time, retaining only the most promising 20 % of particles for regeneration, which cut runtime by roughly 40 %. Parallelization and GPU acceleration were noted as straightforward extensions for real‑time applications.

Uncertainty analysis was performed by adding stochastic noise to the environmental forcing fields (wind, currents) and conducting Monte‑Carlo ensembles. The resulting spread in estimated origins quantified the impact of environmental data errors, yielding a mean error of 0.22 km with a standard deviation of 0.09 km.

The paper discusses broader applicability: the same framework can be applied to oil spill back‑tracking, chemical dispersion, or any drifting hazard where the object’s properties evolve non‑linearly. Future work is suggested on integrating high‑resolution satellite or drone observations for real‑time disaster response, and extending the approach to multi‑object scenarios using clustering techniques.

In conclusion, BAKTRAK offers a robust, flexible, and computationally tractable solution for back‑tracking drifting objects. By preserving the forward model and iteratively refining particle ensembles based on observational proximity, it overcomes the limitations of traditional reverse‑time methods, handles non‑linear state changes naturally, and demonstrates high accuracy in realistic case studies. The method holds promise for a wide range of marine environmental and safety applications.