Machine learning for minefield pattern identification and completion

Research on minefield patterns and layouts lacks public datasets containing locations and metadata, limiting the training and fair evaluation of pattern and layout-level machine-learning models beyond single-mine detection. To address the main research questions: First, can valuable information about patterns and minefield layouts be drawn from only partial position observations using machine learning, and what are the achievable baselines? And second, can patterns be completed from only partial position observations using machine learning, and what are the achievable baselines? Another foundational question must be answered beforehand: Can realistic training data for minefield layout and patterns be created without access to classified operational records, and if so, how? Therefore, in this thesis, the research questions are addressed by developing an end-to-end training data pipeline for machine learning for pattern identification and completion. It does so based on the prerequisites: it creates a catalogue of deployment schemes and methods that motivates a standardised row-laying experiment in which time-ordered Cartesian coordinates of the placements are recorded. A statistical analysis identifies the sequential signatures that matter for realistic simulation. These findings are converted into explicit requirements for a step-wise simulator, which is then implemented. After validation, this simulator is embedded in a data pipeline that produces realistic, incomplete minefield pattern data with controlled augmentation. On this foundation, the identification track begins with a multi-label taxonomy (method, base pattern, specific pattern, region) and progresses to a multi-class cascade focused on base patterns for sharper decision boundaries. The completion track iterates through several reinforcement-learning environments to achieve partial and full completion, but still leaves room for improvement and further research. The results demonstrate that pattern classification is possible on partial or otherwise incomplete position observations, while pattern completion works with reinforcement learning but still has room for improvement and further research. Also realistic, sequentially faithful simulation is a practical prerequisite for training layout-aware models and a viable substitute for sensitive field data, enabling reproducible benchmarking and targeted improvement. The approach is most applicable to hand-laid rows and closely related doctrines, but can be adapted to other procedures due to the simple configurability of the simulation algorithm with measured data.

Citation

Bruckbauer, A. (2025). Machine learning for minefield pattern identification and completion [Unpublished doctoral thesis]. University of Chester.

Publisher

University of Chester

URI

https://chesterrep.openrepository.com/handle/10034/629986

Type

Thesis or dissertation

Language

en