Computer Vision vs. Ground-Penetrating Radar: What Aerial Inspection Can and Cannot See

Abstract: Two technologies dominate advanced pavement diagnostics: drone-based optical inspection with AI-driven computer vision, and ground-penetrating radar (GPR). They are frequently presented as alternatives. They are not. They address structurally different diagnostic questions: one characterises the surface, the other the subsurface. This article sets out the physical principles behind each technology, defines their respective detection limits, identifies the failure modes in each, and argues that the decision to deploy one without understanding the other is an analytical error with real cost consequences.

1. Two Technologies, Two Diagnostic Questions

Pavement failure is a multi-layer process. A road structure typically comprises a wearing course, a binder course, a base layer, and a subbase, all resting on a subgrade. Failure can initiate at any layer and propagate upward, downward, or laterally depending on loading patterns, drainage conditions, and material properties.

Surface inspection technologies — whether manual, vehicle-mounted camera systems, or drone-based aerial platforms — observe the wearing course. They classify what is visible: cracking morphology, surface deformation, loss of surface texture, edge deterioration, marking condition. This is exactly the information required for routine maintenance planning, contractual performance verification, and condition scoring under frameworks such as PCI or the European Road Assessment Programme (EuroRAP).

Ground-penetrating radar operates on a fundamentally different physical principle. A GPR antenna transmits electromagnetic pulses into the pavement structure and records reflections from dielectric property boundaries — the interfaces between layers of different material composition or moisture content. The resulting radargram reveals layer thicknesses, detects voids or delaminations within the structure, identifies moisture infiltration, and characterises subgrade conditions — none of which are visible at the surface. [1]

The diagnostic question for surface inspection is: what defects exist on this pavement, where are they, and how severe are they? The diagnostic question for GPR is: what is the structural condition of the pavement layers, and what is the likely remaining load-bearing capacity? These are different questions requiring different answers.

2. What Computer Vision Detects — and How

Modern drone-based aerial inspection platforms capture imagery at ground sampling distances (GSD) typically between 0.5 and 2 cm/pixel, depending on flight altitude and sensor specification. At GSD < 1.5 cm/px, the following defect categories are reliably detectable and classifiable by trained convolutional neural network (CNN) models:

• Cracking: Longitudinal, transverse, diagonal, block, and fatigue (alligator) cracking patterns. Detection sensitivity at GSD 1 cm/px reaches crack widths of approximately 2–3 mm. Severity classification (low/medium/high) is based on crack width, pattern density, and evidence of secondary spalling. [2]
• Potholing: Bowl-shaped surface disintegration. Detectable from aerial imagery through shadow geometry and texture discontinuity. Area measurement is reliable; depth estimation requires photogrammetric structure-from-motion (SfM) processing of overlapping imagery.
• Ravelling and surface texture loss: Aggregate loss visible as surface granularity change. Reliably detected at GSD < 1 cm/px; more challenging at GSD 2 cm/px due to texture frequency aliasing.
• Surface deformation: Rutting and shoving detectable through shadow analysis in low-angle illumination. Photogrammetric DEM generation from drone imagery enables rutting depth estimation with accuracy in the range of ±3–5 mm under optimal conditions.
• Road markings, edge condition, signage: Fully classifiable from aerial imagery. Line retroreflectivity cannot be measured optically from nadir-view imagery — this requires ground-based retroreflectometry.

Critical limitation: Computer vision cannot detect subsurface conditions. A pavement surface can appear structurally sound under optical inspection while carrying voids, delaminations, or saturated base layers that will produce sudden catastrophic failure under traffic loading. The 2016 sinkhole collapses on several European urban road networks involved surfaces that showed no visible distress in the weeks prior to failure. [3]

3. What GPR Detects — and Its Own Limits

Ground-penetrating radar used for pavement assessment typically operates in the frequency range of 1–2.4 GHz (air-coupled antennas) or 400 MHz–1 GHz (ground-coupled). Higher frequency provides finer resolution but lower penetration depth; lower frequency penetrates deeper but resolves thinner layers less precisely.

The principal GPR-detectable conditions in road infrastructure are:

• Layer thickness measurement: Wearing course, binder course, and base layer thicknesses measurable with accuracy of ±5–10 mm when dielectric constants are calibrated from core samples. Critical for construction acceptance and structural capacity modelling.
• Delamination detection: Air gaps or moisture films at layer interfaces generate strong dielectric contrasts. GPR reliably identifies delaminations of > 5 mm vertical extent at depth up to approximately 300 mm.
• Void detection: Sub-surface voids (e.g., from drainage pipe failure, sinkhole precursors, compaction defects) are detectable as hyperbolic reflections. Detection reliability diminishes with depth beyond 500 mm and in highly attenuating (moist, clay-rich) subgrades.
• Moisture infiltration: Water significantly reduces dielectric constant contrast and increases signal attenuation. Saturated base or subgrade layers produce characteristic amplitude reduction signatures. Useful for identifying drainage failure before visible surface distress develops.

GPR limitation: Ground-penetrating radar provides no information about surface cracking patterns, marking condition, or geometric compliance. A radargram is not a defect map in the sense that a georeferenced AI classification output is. Interpretation requires specialist expertise; automated GPR analysis is less mature than computer vision for surface defects, and false positive rates for void detection remain higher in field conditions than in controlled trials. [4]

Additionally, air-coupled GPR is typically deployed from vehicle-mounted systems at traffic speed, which limits spatial resolution relative to drone-based imagery. Ground-coupled systems require traffic management for deployment on open roads.

The decision between optical inspection and GPR is not a technology preference. It is a question of which failure mode you are trying to detect — and which comes first in your asset management decision chain.

4. Failure Mode Sequencing and the Inspection Decision

The practical decision framework for inspection technology selection should be based on failure mode probability and consequence, not on technology availability or unit cost.

For routine network-level maintenance planning — pothole repair scheduling, surface treatment prioritisation, marking renewal, drainage clearance — drone-based computer vision inspection is the appropriate primary tool. It delivers spatial coverage, consistent classification, and georeferenced output at network scale that no GPR deployment can replicate economically.

For structural assessment — construction acceptance testing, post-event investigation (after flooding, heavy freeze-thaw cycles, or abnormal loading events), investigation of suspected void formation, or pre-resurfacing structural adequacy assessment — GPR is the appropriate primary tool. It answers questions that no optical system can.

In practice, the most analytically complete pavement management approach combines both. A routine aerial survey identifies surface distress locations and severities. Where that survey flags accelerated deterioration, unusual cracking morphology (e.g., circular or radial patterns suggestive of sub-surface void), or sections with unexpectedly poor condition relative to traffic loading, targeted GPR investigation of those specific sections provides the subsurface context needed to distinguish between a surface maintenance requirement and a structural rehabilitation requirement.

This sequential approach — surface inspection triggering targeted GPR — is materially more cost-efficient than blanket GPR coverage of an entire network, and more diagnostically complete than surface inspection alone. Several European road agencies, including the Slovak Road Administration (NDS) and the Czech Road Directorate (ŘSD), have adopted hybrid inspection protocols for motorway network assessment that reflect exactly this logic. [5]

5. Implications for Asset Management and Budget Allocation

The financial implications of confusing the diagnostic domains of the two technologies are significant in both directions.

Relying exclusively on optical inspection for structural assessment leads to under-detection of subsurface failures. Sections that appear to be candidates for surface treatment are in reality structurally compromised — and applying a surface dressing to a structurally failed base is not a maintenance action, it is deferred cost at a multiplied scale. The AASHTO Pavement Design Guide identifies premature surface treatment application over structurally deficient base layers as one of the primary contributors to lifecycle cost overruns on managed road networks. [6]

Conversely, deploying GPR for network-level surface condition monitoring is both technically inappropriate (it does not measure surface defects) and economically inefficient — GPR survey costs per lane-kilometre are typically 4–8 times those of drone-based optical inspection for equivalent network coverage.

An asset manager who understands both technologies will use each for the diagnostic question it is designed to answer, and will recognise the condition signatures — in optical data — that indicate when subsurface investigation is warranted. That recognition is the analytical skill that separates reactive pavement management from genuinely predictive infrastructure stewardship.

Conclusion

Computer vision aerial inspection and ground-penetrating radar are not competing technologies. They are complementary diagnostic tools that operate on different physical principles, detect different failure modes, and answer different questions in the asset management decision chain.

The most consequential misuse of either technology is deploying it outside its diagnostic domain — treating a surface condition score as evidence of structural adequacy, or expecting GPR to replace the spatial defect inventory that optical inspection provides. Both errors lead to the same outcome: maintenance budgets allocated to the wrong interventions, at the wrong locations, at the wrong time.

IDL's platform addresses the surface diagnostic domain at network scale. Where survey outputs indicate conditions warranting structural investigation, we flag this explicitly — because knowing the boundary of your diagnostic capability is as important as the capability itself.

References

[1]  Benedetto, A., Tosti, F., Bianchini Ciampoli, L. & D'Amico, F. (2017). 'An overview of ground-penetrating radar signal processing techniques for road inspections.' Signal Processing, 132, pp. 201–209. DOI: 10.1016/j.sigpro.2016.05.016.

[2]  Maeda, H., Sekimoto, Y., Seto, T., Kashiyama, T. & Omata, H. (2018). 'Road damage detection and classification using deep neural networks with smartphone images.' Computer-Aided Civil and Infrastructure Engineering, 33(12), pp. 1127–1141. DOI: 10.1111/mice.12387.

[3]  Napoli, E., Chirico, G.B., Cafaro, G., Mele, B. & Flora, A. (2021). 'Probabilistic analysis of sinkhole triggering caused by undermining due to lateral erosion.' Acta Geotechnica, 16, pp. 2261–2275. DOI: 10.1007/s11440-020-01097-w.

[4]  Loizos, A. & Plati, C. (2007). 'Accuracy of ground penetrating radar models in pavement layer thickness estimation.' NDT & E International, 40(2), pp. 147–157. DOI: 10.1016/j.ndteint.2006.09.001.

[5]  Sybilski, D. & Mechowski, T. (2019). 'Network-level pavement management in Central European countries: current practices and integration challenges.' Transportation Research Record, 2673(4), pp. 538–547. DOI: 10.1177/0361198119838945.

[6]  American Association of State Highway and Transportation Officials (AASHTO). (2008). Mechanistic-Empirical Pavement Design Guide: A Manual of Practice. Washington, D.C.: AASHTO.