Optimising wilderness missing person search through the application of machine learning for location probability prediction and mission planning

Ewers, Jan-Hendrik (2026) Optimising wilderness missing person search through the application of machine learning for location probability prediction and mission planning. PhD thesis, University of Glasgow.

Abstract

Wilderness search and rescue operations present a critical challenge where swift decision making, and efficient resource allocation can significantly affect the likelihood of saving a lost person’s life. Traditional search planning methods rely heavily on manual coordination, and are constrained by limited human resources. However, in recent years, the expanding geospatial data, and advancing unmanned aerial vehicle capabilities have opened new possibilities for more adaptive, and data-driven search systems.

This work improves the search planning aspect of the wilderness search and rescue operations through the development of two components for an autonomous unmanned aerial vehicle. The first component is a probabilistic lost person behaviour model which leverages psychological profiles established from historical data to generate probability distribution maps through Monte-Carlo methods. The second component is a continuous-space reinforcement learning search planner which ingests the probabilistic location information over a given area to produce a search path subject to real-world operational constraints. By integrating behavioural modelling with heuristic-free path planning, this framework transforms raw search parameters into executable autonomous flight trajectories.

Results show that the probability distribution map generation algorithm closely matches historical datasets without requiring per-location retraining, which is the case with other methods from the literature. The search planning algorithm is then able to leverage the probability distribution map effectively whilst outperforming benchmark algorithms like the local hill climbing based algorithm LHC_GW_CONV by over 250%. Furthermore, by training the reinforcement learning policy in an environment with a mid-fidelity six degrees-of-freedom dynamic simulation controlled by a full navigation stack, it was shown that the resultant search paths were able to better leverage the search platform dynamics resulting in higher find rates than when a lower fidelity environment was used.

The proposed modular approach of generating a probability distribution map, which is then ingested by a search algorithm, has provided insights into the strengths and limitations of each part. Future work will focus on the combination of the two components into a full end-to-end algorithm capable of fully autonomous wilderness search and rescue using unmanned aerial vehicles.

This research provides a validated foundation for autonomous wilderness search and rescue systems, reducing operational burden and improving real-world rescue outcomes.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Subjects: T Technology > T Technology (General)
• Colleges/Schools: College of Science and Engineering > School of Engineering
• Funder's Name: Engineering and Physical Sciences Research Council (EPSRC)
• Supervisor's Name: Anderson, Dr. David and Thomson, Dr. Douglas
• Date of Award: 2026
• Depositing User: Theses Team
• Unique ID: glathesis:2026-86000
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 12 Jun 2026 10:19
• Last Modified: 12 Jun 2026 14:54
• Thesis DOI: 10.5525/gla.thesis.86000
• URI: https://theses.gla.ac.uk/id/eprint/86000