Authors:
Mark Benjamin Kuhne, Jamila Shaaban Ebrahim, Jouma Ali Al-Mohamad, Ennouni Assia, Mihiro Otaki, Cheang Chi Seng

Addresses:
Department of Computer Science, Secure and Trustworthy Systems Group, ETH Zurich, Zürich, Switzerland. Department of Computer Engineering, Faculty of Electrical and Electronic Engineering, Aleppo University, Aleppo, Syria. Department of Computer and Mobile Communication Engineering, Faculty of Information Engineering, Al-Shahbaa Private University, Aleppo, Syria. Khenifra Higher School of Technology, Sultan Moulay Slimane University, Béni-Mellal, Beni Mellal-Khenifra, Morocco. Department of Information and Communication Engineering, University of Tokyo, Bunkyo, Tokyo, Japan. Department of Computing and Information Systems, Singapore Management University, 81 Victoria Street, Singapore.

Abstract:

Syria faces one of the most severe landmine and explosive ordnance contamination crises globally, with over 15.4 million people at immediate risk of injury or death from explosive hazards following nearly 15 years of continuous conflict. Manual demining is successful but slow, laborious, and risky, with survey teams covering just 50–100 square meters per day. This research provides a framework for developing machine learning algorithms to interpret magnetic survey data for landmine detection, shifting from raw magnetic field measurements to predictive categorization of buried explosive hazards. This study shows that machine learning can transform magnetometry-based detection by analyzing the physical principles of magnetometry-based detection, assessing the algorithmic landscape, including Support Vector Machines (SVM), Random Forests, Convolutional Neural Networks (CNN), and unsupervised anomaly detection methods like RUDE, and evaluating Syria's pioneering drone-based magnetometer trials in November 2025. In controlled trials, SVM classifiers achieved macro-F1 scores of 0.9586 ± 0.0372, CNN-based models achieved UXO classification accuracy above 95.8%, and deep learning approaches for magnetic data reconstruction outperformed state-of-the-art algorithms by 3–5% while operating up to 80× faster. Data scarcity for training robust models, the need for large labeled field datasets, field-deployable device computation constraints, and the near-zero false-negative requirement in safety-critical applications remain hurdles. A workflow for data collection, preprocessing, feature extraction, model training, and deployment with Syrian tweaks is presented in the paper. 

Keywords: Machine Learning (ML); Magnetic Survey; Landmine Detection; UXO Classification; Convolutional Neural Networks (CNNs); Support Vector Machines (SVMs); Humanitarian Demining.

Received on: 09/07/2025, Revised on: 12/09/2025, Accepted on: 09/11/2025, Published on: 07/09/2026

DOI: 10.69888/FTSESS.2026.000788

FMDB Transactions on Sustainable Environmental Sciences, 2026 Vol. 3 No. 3, Pages: 165-183

• Views : 38
• Downloads : 9

Download PDF