State of the art
Published work on each part of the platform, and the gap this project addresses.
A literature and product search was carried out in September 2026 for each component of the planned platform. Each component has been demonstrated before. We found no published system that combines them, and none that is openly documented or built at low cost.
2.1Software-defined ground-penetrating radar
Stepped-frequency continuous-wave (SFCW) radar implemented on software-defined radios is an established academic approach. A Georgia Tech thesis and the corresponding IEEE conference paper [R1, R2] implemented SFCW GPR on SDR hardware. A 2025 GNU Radio Conference paper [R3] used commercial off-the-shelf SDRs, and a hobby project imaged the subsurface with two HackRF units [R4]. These works confirm that the radar architecture in Chapter 6 is feasible with SDR hardware. None of them identifies the material of a target.
2.2Multi-axis electromagnetic induction for unexploded ordnance
The closest related technology is the family of advanced induction sensors developed for classifying unexploded ordnance in the United States. MetalMapper [R5] uses three orthogonal transmit loops and multi-axis receivers. TEMTADS and ALLTEM follow similar principles [R6]. These systems invert a dipole model to obtain the magnetic polarizability tensor of each target and classify on it. Reported performance at live sites is about 90% probability of detection at about 10% false positives. The mathematics of polarizability-tensor spectra as object signatures is an active research topic [R7, R8]. These instruments work in the time domain and are vehicle-towed or cued systems in the price range above 100,000 USD.
2.3Combined induction and radar detectors
Hand-held detectors that combine induction with radar are standard equipment in humanitarian demining. The US Army's AN/PSS-14 (HSTAMIDS) has been standard issue since 2006 [R9]. The Vallon–Cobham Minehound and Minelab's MDS-10 and MDS-20 [R10, R11] pair simultaneous multi-frequency induction with ultra-wideband radar. Research on combining electromagnetic induction spectroscopy with GPR for anti-personnel mines continues [R12]. The main result of this line of work is that the two sensors fail on different kinds of clutter, so fusion lowers the false-alarm rate. These products are closed: raw data and tensor information are not available to the user. Minelab's consumer detectors (Multi-IQ) also transmit several frequencies simultaneously, which is the excitation principle chosen for LODESTONE.
2.4Machine learning on detector data
Recent work applies deep learning to induction data for classifying mines and small metal content in mineralised soil [R13]. It applies one-dimensional convolutional networks to time-domain polarizability features [R14] and machine learning to pulse-induction data [R15]. Methods that start from physically meaningful features, such as the polarizability tensor, are common, and the pipeline in Chapter 8 follows that direction. A limitation mentioned in almost all of these works is the small size and closed nature of the available labelled datasets.
2.5Comparison
Table 2.1.Capabilities of related systems compared with the planned platform.
| Capability | MetalMapper / TEMTADS | AN/PSS-14, MDS-20 | SDR GPR (academic) | LODESTONE |
|---|---|---|---|---|
| Full-tensor induction (3 TX × 3 RX) | Yes | No | No | Yes |
| Multi-frequency induction spectra | Yes, time domain | Yes, closed | No | Yes, 12 tones |
| Radar | No | Yes | Yes | Yes |
| Centimetre-level positioning | GPS on vehicle | No | No | RTK, IMU, camera |
| Open hardware and data | No | No | Publications only | Yes |
| Approximate cost class | above 100,000 USD | 20,000 to 50,000 USD | 1,000 to 5,000 USD | about 2,300 USD |
2.6Position of this project
We did not find a platform that combines (1) multi-frequency, full-tensor induction on an audio codec, (2) SDR radar on the same time line, (3) centimetre positioning with hardware time synchronisation, (4) classification on orientation-independent tensor spectra, and (5) an open labelled dataset. The individual components are each supported by the literature above. The contribution of LODESTONE is their integration at low cost with full documentation.
From the related work we adopt the dipole-model inversion and tensor-based classification of the ordnance programmes, simultaneous multi-frequency excitation, the principle of fusing induction with radar to reduce false alarms, and the spectral-tensor mathematics of [R7, R8]. The dataset is designed into the project from the first work package in response to the data shortage noted above.
The search covered English-language publications and product literature. Patents and literature in other languages have not yet been reviewed; this is listed as task T6.5.