[Paper] From Forecast to Action: Uncertainty-Aware UAV Deployment for Ocean Drifter Recovery

Source: arXiv - 2512.09260v1

Overview

The paper proposes a predict‑then‑optimize pipeline that turns ocean‑drifter trajectory forecasts into concrete UAV (drone) deployment plans for maritime rescue. By coupling a large‑language‑model (LLM) based predictor with a spatial‑uncertainty model and a meta‑heuristic optimizer, the authors move beyond static “drop‑a‑drone‑anywhere” strategies and demonstrate measurable gains on real‑world Korean coastline data.

Key Contributions

• End‑to‑end framework that links trajectory forecasting directly to UAV placement, a combination not explored in prior maritime‑search literature.
• LLM‑driven trajectory prediction that ingests heterogeneous sensor logs (e.g., GPS, wind, currents) and outputs probabilistic future paths.
• Gaussian particle sampling to translate forecast uncertainty into a set of plausible drifter locations.
• Dynamic detection radius for each UAV that shrinks with distance, reflecting realistic sensor performance.
• Meta‑heuristic optimization (repair‑aware GA/PSO) that iteratively refines UAV locations while respecting the dynamic radii and coverage constraints.
• Empirical validation on a curated dataset of drifter releases along the Korean coast, showing a 30‑40 % reduction in missed drifters compared with random search baselines.

Methodology

1. Data ingestion & preprocessing – Historical drifter tracks, oceanographic variables, and weather reports are aligned in time‑space windows.

2. Trajectory forecasting – A fine‑tuned large language model (e.g., GPT‑Neo) treats the time‑series as a “story” and predicts the next N hours of latitude/longitude. The model outputs a mean path plus a covariance matrix.

3. Uncertainty modeling – The covariance is used to draw M Gaussian particles around each predicted point, forming a cloud of possible drifter positions.

4. Dynamic UAV detection model – For a UAV at location u and a candidate drifter particle p, the detection probability follows a radial decay function:

   $$
   P_{\text{detect}}(u,p) = \exp!\bigl(-\alpha , |u-p| \bigr)
   $$

   where α is calibrated from field tests.

5. Optimization problem – Choose K UAV launch points to maximize the expected coverage of the particle cloud, subject to flight‑time and no‑fly‑zone constraints.

6. Meta‑heuristic solver – A genetic algorithm with a custom repair operator (to re‑inject infeasible UAVs back into feasible airspace) iteratively improves the placement. The fitness function aggregates detection probabilities across all particles.

Results & Findings

Key takeaways

• The repair‑aware GA consistently outperformed plain random search and a naïve greedy placement.
• Even when the LLM forecast deviated significantly, the Gaussian particle cloud preserved enough diversity for the optimizer to find resilient UAV spots.
• Runtime remains within operational limits for on‑board or edge‑device execution, enabling near‑real‑time mission replanning.

Practical Implications

• Maritime rescue agencies can integrate this pipeline into existing UAV command‑and‑control consoles, turning raw sensor feeds into actionable flight plans within minutes.
• Coastal monitoring firms can automate routine drifter (or buoy) retrieval, reducing manpower and fuel costs.
• The dynamic detection radius model aligns with real sensor characteristics (camera resolution, lidar range), making the solution portable to different UAV platforms.
• Because the framework is modular (forecast → uncertainty → optimizer), developers can swap in domain‑specific predictors (e.g., physics‑based ocean models) or alternative meta‑heuristics without redesigning the whole system.

Limitations & Future Work

• The LLM predictor, while flexible, is data‑hungry; performance degrades in regions with sparse historical drifter releases.
• Gaussian sampling assumes elliptical uncertainty, which may not capture multimodal drift patterns caused by complex currents.
• The current study focuses on a single‑day horizon; extending to multi‑day, multi‑UAV coordination remains an open challenge.
• Future research directions include: integrating physics‑based ocean circulation models, exploring reinforcement‑learning for adaptive UAV routing, and field‑testing the pipeline in harsher weather conditions.

Authors

• Jingeun Kim
• Yong-Hyuk Kim
• Yourim Yoon

Paper Information

• arXiv ID: 2512.09260v1
• Categories: cs.NE
• Published: December 10, 2025
• PDF: Download PDF