Autonomous UAV Subsurface Sensing for Humanitarian Landmine and UXO Survey

1. Problem Statement

Land contaminated by mines and explosive remnants of war cannot be farmed, built on, or walked across, and the party who pays to change that is a donor government buying clearance by the hectare. The price is the problem. The HALO Trust reports that clearing one hectare of Ukrainian land by conventional methods costs $3,000 to $5,000. Against that unit cost sits roughly 132,000 square kilometers of suspected contamination in Ukraine alone, revised down from an original government estimate of 174,000 square kilometers as survey returned land to use. The World Bank puts the total cost of clearing Ukraine at $34.6 billion.

Labor sets the unit cost. A deminer working a metal detector and prodder clears 20 to 50 square meters in a full day, and the rate falls further in scrap-littered ground because a metal detector responds to metal rather than to explosive. In post-battle terrain saturated with shell fragments, a deminer excavates hundreds of signals for every mine found. Mechanical flails and tillers raise throughput but consume the ground they process and cannot work soft, wet, or steep terrain. Neither method is improving; the productivity figure has not materially moved in twenty years.

The buyer is a shrinking pool. International support for mine action fell 5 percent in 2024 to $761 million (Landmine Monitor 2025). In January 2025 the US State Department paused work on Weapons Removal and Abatement grants for an 85 day review under the executive order on foreign assistance; programs were subsequently cleared to resume under existing awards, and a bipartisan group in Congress has requested $271.703 million for Conventional Weapons Destruction in FY2027. The direction is flat-to-down funding against rising need: 6,279 people were killed or injured by mines and explosive remnants of war in 2024, the highest total since 2020, with civilians 90 percent of casualties. At least 57 states and areas remain contaminated, seven massively so (Afghanistan, Bosnia and Herzegovina, Cambodia, Ethiopia, Iraq, Türkiye, Ukraine).

The economic loss is agricultural and measurable. GLOBSEC estimates 5 million hectares, approximately 15.2 percent of Ukraine's farmland, unusable because of mines, ordnance, and hostilities, with a potential envelope of 8 million hectares. Direct losses to Ukrainian farmers average $1,500 per hectare of arable land, and the sector has absorbed at least $40.2 billion in losses and damages since February 2022.

When money is fixed and need is growing, the only lever is cost per hectare, and that lever is pulled in survey rather than clearance. Land release doctrine already accepts that most suspected hazardous area contains no mines; survey exists to find the fraction that does, so manual clearance is spent only where it must be.

2. State of the Art

Aerial detection has split into two tracks that do not meet, and the commercial market occupies only one.

The commercial track detects what is visible on the surface. Safe Pro Group Inc. (NASDAQ: SPAI), through its Safe Pro AI unit, operates SpotlightAI, a machine learning pipeline analyzing RGB imagery collected by commercial off-the-shelf drones. Its dataset exceeds 2.9 million drone images with more than 51,750 confirmed detections across roughly 37,835 acres surveyed in Ukraine, and its software has been integrated with drones selected under a US Army program of record. This is a working, revenue-generating product, and it is by construction restricted to ordnance that reflects light to a camera. The HALO Trust and Amazon Web Services are running a $4 million trial applying AI to war debris in drone imagery over Ukrainian minefields, in the same visible-surface regime. HALO reports one drone surveying up to five hectares during daylight hours, ten in optimal conditions.

Surface detection is valuable and insufficient. Scatterable mines such as the PFM-1 lie on the surface and are addressable this way. Buried anti-personnel and anti-tank mines, and the large inventory of buried unexploded ordnance, are not. They are the population that forces manual excavation, and they sit outside the physics of an RGB camera.

The research track sees beneath the surface and does not leave the laboratory. Four groups have independently built and flown subsurface payloads. The Signal Theory and Communications group at the Universidad de Oviedo (García-Fernández, Álvarez López, Las-Heras) developed a UAV-mounted ultra-wideband GPR with synthetic aperture processing and centimeter-accurate RTK positioning, progressing from a single-channel prototype to full three-dimensional subsurface imaging and then to an antenna array that raises throughput. The University of Maribor (Šipoš, Gleich) built a stepped-frequency GPR light enough to fly as a 780 gram payload at 4.2 watts. At the Technical University of Denmark, Døssing's group demonstrated high-speed UAV magnetometry with a towed sensor bird. A Spanish team led by González-Aguilera flew proton magnetometry over eleven buried UXO at Spain's CENAD San Gregorio range. At Binghamton University, Nikulin and de Smet established UAV thermal and multispectral detection of scatterable mines with deep learning, continued by Ientilucci's group at the Rochester Institute of Technology in hyperspectral imagery.

Three limitations are common to every published system, and none is a sensor problem. All are open-loop: the aircraft flies a preplanned grid, lands, and the data is post-processed offline before an analyst reviews it. The Oviedo group's most recent paper is titled, accurately, "Towards real-time processing." The detectors do not generalize, as Section 3 quantifies. And the payloads are one-off laboratory builds, while useful GPR return requires flying 20 to 45 cm above irregular vegetated ground at roughly 0.6 m/s, which is a terrain-following control and mechanical integration problem.

The gap between the two tracks is the opportunity. One track has a market and cannot see underground. The other sees underground and has no product.

3. Foundational Research

García-Fernández M, Álvarez López Y, Arboleya Arboleya A, González Valdés B, Rodríguez Vaqueiro Y, Las-Heras Andrés F (2018). "Synthetic Aperture Radar Imaging System for Landmine Detection Using a Ground Penetrating Radar on Board a Unmanned Aerial Vehicle." IEEE Access, 6:45100–45112. DOI: 10.1109/ACCESS.2018.2863572. A UAV-mounted radar whose measurements are coherently combined by a SAR algorithm under centimeter-level positioning, with clutter removal suppressing the air-soil interface reflection. Because the sensor is a radar rather than a metal detector it detected both metallic and dielectric targets, validated in controlled and real scenarios. With 154 citations recorded by Semantic Scholar this is the field's anchor paper, and it establishes the core feasibility claim: a flying, non-contact sensor can image buried dielectric objects a metal detector cannot distinguish from scrap.

Šipoš D, Gleich D (2020). "A Lightweight and Low-Power UAV-Borne Ground Penetrating Radar Design for Landmine Detection." Sensors, 20(8):2234. DOI: 10.3390/s20082234. A stepped-frequency continuous-wave GPR from 550 MHz to 2.7 GHz (2.15 GHz maximum bandwidth), 30 grams for the radar board and 780 grams for the complete payload with antennas, drawing 4.2 watts. Flying 10 to 50 cm above ground (20 to 45 cm preferred) at approximately 0.6 m/s, it detected a metallic anti-personnel mine of 8 by 14 cm buried at 20 cm depth and a plastic anti-tank mine of 27 by 13 cm at the surface. The authors state that scanning speed and soil moisture handling still require improvement and that dry warm seasons are most suitable. The result fixes the payload budget: subsurface sensing fits inside the mass and power envelope of a small multirotor, converting the problem from sensor development into flight control and manufacturing.

García-Fernández M, Álvarez-Narciandi G, Laviada J, Álvarez López Y, Las-Heras F (2023). "Array-Based Ground Penetrating Synthetic Aperture Radar on Board an Unmanned Aerial Vehicle for Enhanced Buried Threats Detection." IEEE Transactions on Geoscience and Remote Sensing. DOI: 10.1109/TGRS.2023.3272982. Having established centimeter-level resolution with single-channel GPR-SAR, the group attacked scanning throughput by integrating an antenna array onto the UAV and testing against several classes of buried target in realistic scenarios; coherent combination across all transmit-receive channels raised both throughput and detection. Cited 47 times, this is the work that makes area coverage plausible, because throughput rather than detection physics decides whether aerial subsurface survey competes on cost per hectare.

Kolster ME, Wigh MD, Lima Simões da Silva E, Bjerg Vilhelmsen T, Døssing A (2022). "High-Speed Magnetic Surveying for Unexploded Ordnance Using UAV Systems." Remote Sensing, 14(5):1134. DOI: 10.3390/rs14051134. Three scalar magnetometers towed in an airframe 10 meters beneath a sub-25 kg UAV at approximately 10 m/s acquired 58 minutes of data, each sensor traversing 31.7 km to densely cover a 600 by 100 meter area over disarmed UXO. Apparent noise floors ran to tens of picotesla, precise enough to model and remove high-frequency noise at ±5 picotesla. All gradiometer configurations recovered most targets including every major target, though the horizontal configuration performed significantly worse. This establishes the complementary modality: magnetometry covers ground roughly seventeen times faster than the GPR flight profile and finds ferrous ordnance, at the cost of blindness to non-metallic mines. A credible system needs both, and needs to know when to use which.

Ugarte-Goicuría I, Guerrero-Sevilla D, Carrasco-Garcia P, Carrasco-Garcia J, Gonzalez-Aguilera D (2026). "Aerial Drone Magnetometry for the Detection of Subsurface Unexploded Ordnance (UXO) in the San Gregorio Experimental Site (Zaragoza, Spain)." Drones, 10(2):88. DOI: 10.3390/drones10020088. A GEM GSMP-35U proton magnetometer on a hexacopter surveyed at 7 m and 2 m above ground on 1 m line spacing over eleven UXO buried at known coordinates. It resolved anomalies of 2 to 18 nT against a high ferromagnetic noise background, reaching signal-to-noise ratios above 5 at 2 m altitude with geolocation accuracy near 0.5 m, and covered 0.53 hectares in under one hour of effective flight. This is the most recent controlled-ground-truth result available, and it quantifies the two numbers a survey business runs on: how small an anomaly is recoverable, and how many hectares per flight hour.

Qiu Z, Guo H, Hu J, Jiang H, Luo C (2023). "Joint Fusion and Detection via Deep Learning in UAV-Borne Multispectral Sensing of Scatterable Landmine." Sensors, 23(12):5693. DOI: 10.3390/s23125693. PMID: 37420862. RGB plus four narrow bands (550, 660, 735, 790 nm) flown at roughly 3 m and 8 m altitude produced 789 image pairs across four mine types (M14, T125, B91, M93) and four vegetation scenes, with about 60 percent of mines occluded by vegetation, evaluated on a test set of 991 labeled mines. Fusion raised mAP@0.5 from 0.848 to 0.922, precision from 0.872 to 0.933, and recall from 0.848 to 0.906, cutting false positives from 123 to 64. This is the quantitative case for sensor fusion over any single modality, by controlled ablation on identical scenes. The authors are explicit that simulated settings remain far from a natural minefield and that mine aging, corrosion, and burial conditions are not represented.

Baur J, Steinberg G, Nikulin A, Chiu K, de Smet T (2020). "Applying Deep Learning to Automate UAV-Based Detection of Scatterable Landmines." Remote Sensing, 12(5):859. DOI: 10.3390/rs12050859. Using multispectral and thermal datasets from an automated UAV survey system over scattered PFM-1 landmines, a Faster R-CNN detector achieved 99.3 percent testing accuracy on a partially withheld testing set and 71.5 percent on a completely withheld testing set. Cited 83 times, this is the field's most-used demonstration that automated aerial detection works and simultaneously its first clear measurement of the generalization cliff. A 27.8 point drop between partially and completely withheld data is not a tuning problem; it is a statement about how narrowly these models learn.

Karwandyar S, Pingel T, Nikulin A (2026). "Deep Learning and Multiview-Based Detection of Scatterable PFM-1 Landmines: Performance, Out-of-Sample Evaluation, and Field Readiness." Geomatics, 6(3):54. DOI: 10.3390/geomatics6030054. Six years later, from the same research lineage, a YOLOv11 detector reached 78 to 91 percent precision and 76 to 88 percent recall on validation data, then fell to 74 to 80 percent precision and 14 to 24 percent recall out of sample against 3D-printed paint-matched replicas and an inert PFM-1. The authors conclude that out-of-sample evaluation is critical to field readiness, and they demonstrate edge deployment for locating a minefield through trigonometric and kernel density relationships. For anyone entering this market this is the decisive result: validation-set performance here does not predict field performance, and the replicas needed to test honestly can be manufactured cheaply and safely.

4. Competitive Landscape

Safe Pro Group Inc. (NASDAQ: SPAI) is the only scaled commercial detection player. It raised approximately $22.0 million gross across an August private placement of 2,000,000 shares with three-year warrants (roughly $8.0 million) and an October sale of 2,000,000 shares at $7.00 (roughly $14.0 million), which management stated mitigates prior going-concern conditions. Full-year 2025 revenue was approximately $1.5 million; Q1 2026 revenue exceeded $1.2 million, up roughly 560 percent year over year, and the company projected Q2 2026 revenue above $1.3 million against $92,753 in Q2 2025, driven by US Army subcontract awards. Its limitation is architectural rather than commercial: SpotlightAI analyzes RGB imagery from unmodified commercial drones, which is exactly why it scales and exactly why it cannot address buried ordnance.

findmine gGmbH (Illertissen, Germany) was founded in 2022 with targeted funding from the Urs Endress Foundation and is developing ground-penetrating synthetic aperture radar and metal detection on unmanned aircraft, translating Oviedo-lineage research toward products. It is the closest direct analogue to this opportunity, and it is a non-profit gGmbH operating pre-commercially, which says something about the sector's ability to support a venture-scale private company on humanitarian demand alone.

Adjacent suppliers who are not competitors. Mechanical demining OEMs (DOK-ING, Hydrema, Armtrac) sell machines that destroy ordnance rather than find it; 98 such machines operated in Ukraine as of October 2024 alongside 4,300 sappers. Geophysical instrument makers (Geometrics, GEM Systems, SENSYS, Bartington) sell sensors, not autonomy; in a 2023 comparative test the SENSYS MagDRONE R3 detected a 105 mm shell at all tested heights and a 60 mm shell at most heights.

Why the space has not commoditized. Four reasons, one technical. The physics is hard: useful GPR return needs a 20 to 45 cm standoff at 0.6 m/s over uneven vegetated terrain, a flight control problem sensor vendors do not own. The evidence pathway does not exist: land release under the International Mine Action Standards requires evidence a National Mine Action Authority will accept, and no NMAA currently accredits aerial subsurface detection as a basis for releasing land, so there is no procurement category to sell into. The buyer is donor-funded and correctly conservative, because the cost of a false negative is a civilian casualty. And detectors do not yet generalize across soil, vegetation, and ordnance type, with no shared cross-site benchmark against which a vendor could prove they do. Commoditization requires the accreditation pathway to open, a two to four year sequence, and whoever completes it first defines the category.

5. Total Addressable Market

Market definition. Aerial subsurface survey systems and services for humanitarian mine action and post-conflict land release: UAV-borne GPR and magnetometry payloads with onboard detection, sold or operated to reduce the area requiring manual clearance. This is the survey segment, not clearance, and not the surface-imagery segment Safe Pro occupies.

Bottom-up, anchored on Ukraine. Suspected contaminated area is approximately 132,000 square kilometers, or 13.2 million hectares. Not all is aerially surveyable; forest canopy, urban rubble, and active-conflict areas are excluded. The addressable subset is open agricultural terrain, which GLOBSEC sizes at 5 million hectares of unusable farmland with a potential envelope of 8 million hectares.

Price per hectare is derived from published productivity rather than a quoted rate. HALO reports one drone surveying 5 to 10 hectares in daylight hours; the Spanish magnetometry trial covered 0.53 hectares per effective flight hour on a dense 1 m grid, bracketing fast reconnaissance against high-confidence survey. At a fully loaded field cost of roughly $500 per day for a two-operator team with amortized equipment, aerial survey lands between $50 and $100 per hectare. Taking $75:

• 5 million hectares × $75/ha = $375 million (Ukraine agricultural land, one-time program value)
• 8 million hectares × $75/ha = $600 million (upper contamination envelope)

Ukraine is roughly 40 to 50 percent of currently recorded global contaminated area, scaling global one-time aerial survey TAM to approximately $750 million to $1.5 billion across a 15 to 25 year clearance horizon, or $30 million to $100 million per year.

Sanity check against actual money flow. Total international mine action funding was $761 million in 2024. An annualized aerial survey market of $30 to $100 million is 4 to 13 percent of all money entering the sector, a plausible share for a technology displacing labor at the survey stage, and a hard ceiling no forecast can exceed.

Top-down cross-check, and a caution. Future Market Insights publishes a "Demining and Mine-Clearance Robots Market" report, and several syndicated houses size the adjacent "mine clearance system" market. Their estimates are not usable as stated: Coherent Market Insights values it at $72.9 million in 2026 growing at 5.4 percent CAGR to $105.4 million by 2033, while Verified Market Reports values the same nominal market at $6.5 billion in 2026 growing at 8.9 percent CAGR to $12.86 billion by 2034. A ninety-fold disagreement on the same market name means scope definitions differ by orders of magnitude and none should be load-bearing. The Landmine Monitor donor-funding figure is the reliable anchor.

SAM. Ukraine agricultural land survey at TRL 6 readiness: approximately $375 million one-time, or $25 to $40 million per year across a 10 to 15 year program, of which a technology provider captures the system and software share rather than full service revenue.

Flag: the humanitarian core market is below $100 million per year. That is small for venture funding and must be stated plainly. The path to a fundable business runs through dual use of the same payload and autonomy stack: pre-construction UXO survey (an established commercial service market in Europe, where firms already sell drone magnetometer surveys), Department of Defense range and Formerly Used Defense Site characterization, military route clearance and minefield breaching, and buried utility detection. The humanitarian application is the proving ground and the moral case; the defense and commercial survey applications carry the revenue.

Payment mechanism. There is no reimbursement code analogue here. Survey is procured through per-hectare or per-project contracts let by national mine action authorities, UNMAS, the US State Department Office of Weapons Removal and Abatement, and implementing NGOs, funded by the donor flows above. Contract award depends on operator accreditation by the National Mine Action Authority in the country of operation, which makes accreditation the commercial gate rather than a compliance afterthought.

6. Research Gap and Commercial Opportunity

The components exist and have been separately validated. Nobody has combined them, and that combination is the business.

Open-loop survey is the wrong architecture, and it is what everyone built. The commercial opportunity is the closed loop: run detection onboard, and when the sensor returns an anomaly, have the aircraft autonomously descend, slow, re-scan at higher spatial density, and confirm with a second modality before committing a mark. This converts a fixed-cost grid survey into an adaptive one that spends flight time in proportion to information, which is precisely how cost per hectare falls.

Generalization is unsolved and measurable. The 99.3 to 71.5 percent drop in 2020 and the collapse to 14 to 24 percent out-of-sample recall in 2026 are the same failure six years apart. No shared cross-site benchmark exists, no domain-adaptation method has been validated across soil and vegetation shift, and no vendor can currently prove generalization to a procurement officer. Whoever builds the benchmark and the evaluation protocol defines how the category is judged, which is a durable position independent of any single model.

Manufacturability separates a paper from a fleet. Payloads in the literature are laboratory builds. Fielding requires hundreds of units at a price a donor-funded operator can buy, calibration that survives transport and non-specialist handling, field-replaceable modules, and tolerance control tight enough that antenna geometry, which directly sets GPR image quality, does not drift unit to unit.

Why the incumbents have not closed this. Safe Pro's economics depend on running inference over imagery from unmodified commercial drones with no payload integration, the source of its scaling advantage and a structural exclusion of subsurface sensing; adding a GPR payload would make it a hardware company competing against its own margin profile. The academic groups are organized around publications, not fielded systems: Oviedo and Maribor are electromagnetics and antenna groups whose deliverable is a payload and a processing chain, DTU is a geophysics group, Binghamton and RIT are remote sensing groups, and their funding periods end with a paper. The instrument makers sell sensors into a market that buys sensors, and building an autonomous aircraft around one would put them in competition with their own customers. Each actor behaves rationally, and the integration nobody's incentives reward is the integration the field needs.

7. Comparable Funded Projects

Government funders across three countries are spending on this exact technology stack, which validates both feasibility and procurement appetite.

Agile Electromagnetics Inc. | Canada, Department of National Defence | Innovation for Defence Excellence and Security (IDEaS), Concept Development | "Unmanned Aerial Vehicle Polarimetric Ground Penetrating Radar" | $999,895 | 15 November 2024 to 30 May 2026. Funded against the stated challenge "Safer passage during minefield breaching operations in Ukraine," seeking TRL 1 to 6 solutions for detecting mines in modern minefields. A near-million-dollar award for UAV-borne GPR against the same targets, on a timeline concurrent with this analysis.

Vishal Monga, Ph.D. | The Pennsylvania State University | SERDP, Munitions Response | MR21-1330, "Prior Guided, Informed Deep Learning for Detection and Classification of Underwater Military Munitions." Physics-informed neural networks that embed the resonant acoustic behavior of ordnance into the architecture rather than treating detection as generic image classification, targeting accuracy under limited training data. The award amount is not published on the project record. The methodological thesis, that domain physics must be built into the model when data is scarce, transfers directly to GPR and magnetic signatures.

Laurens Beran, Ph.D. | Black Tusk Geophysics, Vancouver | SERDP, Munitions Response | MR-2226 and MR-1629. Algorithms and software supporting the classification decisions geophysical data analysts and project managers must make during a munitions response project. The closest existing precedent for the decision-support layer between a raw anomaly and an actionable mark.

The HALO Trust with Amazon Web Services | $4 million | 2025. Investment to trial AI detection of the debris of war in drone imagery across existing Ukrainian minefields and battlefields. The largest operator in the sector is committing to machine detection, in the surface-imagery regime.

US Army C5ISR/NVESD, Fort Belvoir | DoD Humanitarian Demining Research and Development Program (HD R&D) | established 1994. A standing program for rapid development, testing, and validation of demining technology, with requirements set at a biennial workshop convened by the Office of the Assistant Secretary of Defense for Special Operations and Low Intensity Conflict. US Government investment in its HALO Trust partnership rose from several million dollars annually in the early 2000s to roughly $33 million in 2024.

The pattern is consistent: defense and environmental funders are paying for sensing and classification, and no funder has yet paid for the closed-loop integration.

8. Opportunity Assessment

TRL 4, with the evidence chain. Multiple independent groups have integrated subsurface payloads onto UAVs and detected buried inert targets under controlled ground truth: a 780 g GPR detecting a metallic AP mine at 20 cm depth (Šipoš and Gleich, 2020), array-based GPR-SAR against multiple buried target classes in realistic scenarios (García-Fernández et al., 2023), and proton magnetometry resolving eleven buried UXO at 2 to 18 nT with 0.5 m geolocation accuracy (Ugarte-Goicuría et al., 2026). No system has been validated inside an accredited operational mine action workflow, and the 2026 out-of-sample results establish that field readiness is unproven. TRL 5 would require validated performance in a relevant operational environment against ordnance the model has not seen.

Risk 1: false negatives are lethal and the sector is correctly conservative. Humanitarian demining is commonly held to a 99.6 percent probability of detection requirement, and no aerial system approaches it. The mitigation is positional rather than technical: the product enters as a survey and triage aid that reduces and prioritizes the area requiring manual clearance, never as a substitute for clearance and never as the sole basis for declaring land safe. Land release doctrine already accepts non-technical survey evidence for area reduction, which is the door this walks through.

Risk 2: domain shift destroys detector performance. The measured collapse to 14 to 24 percent out-of-sample recall is the largest technical risk. Mitigation is an out-of-sample-first evaluation protocol adopted as the primary metric from day one, site-specific calibration flights over emplaced known targets before each survey, and 3D-printed paint-matched replicas as safe, cheap, scalable training and test targets, a method already demonstrated in 2026.

Risk 3: terrain-following flight is the throughput constraint. Holding 20 to 45 cm above uneven vegetated ground at 0.6 m/s limits hectares per hour and therefore cost per hectare. Mitigation is the array-based GPR approach, which raises throughput per pass, combined with a two-tier survey design in which fast magnetometry at 2 to 7 m triages the area and slow GPR is spent only on flagged anomalies.

Regulatory pathway. This is not a medical device; no FDA pathway applies, there is no 510(k), De Novo, or PMA route, and no CPT or HCPCS reimbursement analogue. The governing framework is the International Mine Action Standards plus accreditation of the operator and method by the National Mine Action Authority in each country of operation, civil aviation authorization (FAA Part 107 with a beyond-visual-line-of-sight waiver for US operations, and national equivalents elsewhere), and ITAR and EAR export control on both payload and trained models, a live constraint on any deployment to Ukraine.

One governance question carries over from medical device regulation and is load-bearing. If the detector adapts on device, per site, the detection statistics accreditation was granted against no longer describe the system in operation. The useful precedent is the FDA's Predetermined Change Control Plan framework for AI-enabled devices, which authorizes a bounded envelope of post-market model change declared in advance, alongside the cleared devices that established adaptive and autonomous algorithm governance: the NeuroPace RNS System for closed-loop adaptive therapy, and IDx-DR (now LumineticsCore, De Novo DEN180001) as the first autonomous AI diagnostic authorized to produce a result without physician confirmation. These are governance analogues for adaptive-algorithm accreditation, not a regulatory pathway for this device class. The design consequence is concrete: lock the deployed model per accreditation cycle, and confine adaptation to a declared calibration procedure whose bounds are part of the accreditation package.

Regulatory position as moat. Because no National Mine Action Authority currently accredits aerial subsurface detection for land release, the first organization to complete a statistically designed trial on an accredited test site and publish confusion matrices against known ground truth does not merely earn a permit. It supplies the evidentiary template authorities will use to judge everyone who follows, on a two to four year lag a fast follower cannot compress.

9. Team Requirements

Executing this requires three capabilities that rarely coexist in one organization, which is the structural reason the opportunity remains open.

Sensor physics and airborne system design. Selecting modality per mine class and soil condition (GPR for dielectric targets, magnetometry for ferrous ordnance, multispectral and thermal for surface-laid mines), specifying the terrain-following flight profile GPR return demands, and designing the statistically powered trial on an accredited test site that converts a prototype into evidence a mine action authority will accept.

Machine learning with evaluation discipline. Detector development matters less here than evaluation methodology: a cross-site out-of-sample benchmark, validated domain adaptation across soil and vegetation shift, edge inference within the compute and power budget of a small multirotor, and the adaptive re-survey policy that closes the loop between detection and flight.

Manufacturing engineering. Design for manufacturability on a ruggedized payload produced in the hundreds, tolerance control on antenna geometry tight enough that image quality does not drift between units, calibration surviving field transport and non-specialist operators, and quality systems adequate to an accreditation package. This is the capability every research group in this field lacks, and it determines whether a validated prototype becomes a fleet.