.png){kind=small}
VE3 AI framework that identifies buried ferromagnetic objects from magnetometry data - using magnetic dipole modeling and DBSCAN clustering - with zero ground-truth labels required at any stage.
Buried ferromagnetic objects - from unexploded ordnance to ageing infrastructure - pose persistent risks across marine, terrestrial, and archaeological environments. Yet reliable automated detection remains elusive for three core reasons.
Labeled data scarcity
Supervised AI models require large geophysical datasets - rarely available in remote or marine survey environments.
Signal noise
Background geological variation and sensor interference create false positives that undermine detection confidence.
Poor generalisation
Most approaches fail to adapt across varying object depth, orientation, and burial conditions.
Full methodology for simulating realistic Total Magnetic Intensity (TMI) signals using a magnetic dipole model - covering object orientation, burial depth, sensor altitude, and Gaussian noise conditions.
How maximum, minimum, mean, and peak-to-peak amplitude are extracted per sensor channel and combined with survey-specific parameters to build a compact, clustering-ready feature vector.
Systematic sensitivity analysis across epsilon (0.5 to 1.5) and MinPts (3 to 9) with a complete results table for 20 to 300 samples. Includes actionable guidance on adaptive parameter tuning as dataset size scales.
Empirical evidence that removing standard deviation from the feature vector reduces noise points and improves cluster separation - a counterintuitive but validated result explained in full.
2D PCA projections demonstrating anomaly vs non-anomaly cluster compactness, inter-cluster separation, and the distribution of outlier points across all tested parameter configurations.
As sample count grows from 40 to 300, optimal epsilon decreases from 1.0 to 0.6 while MinPts increases from 3 to 9. The paper documents exactly why and provides configuration guidance for practitioners.
Discussion of how the framework extends to UXO clearance, marine seabed surveying, pipeline inspection, and mineral exploration - plus a roadmap for hybrid machine learning and deep learning integration.
Epsilon values of 1.04 to 1.05 with MinPts = 3 produce stable two-cluster formation with minimal noise points for 40-sample datasets.
As dataset size grows, optimal epsilon decreases from 1.0 to 0.6 and MinPts increases from 3 to 9 - a clear, actionable scaling rule for practitioners.
Removing standard deviation from the feature vector reduces outlier contamination and improves cluster separation. Counterintuitive and empirically confirmed.
The entire pipeline - from data generation through cluster identification - operates with zero labeled training samples.
Improved deterministic + probabilistic logic delivers stronger reconciliation across BFSI, Government, and Healthcare use cases.
Download the complete paper - including full experimental results, DBSCAN parameter tables, and PCA cluster visualisations across 20 to 300 sample configurations.
Download Free Research Paper.png)
