3D Drone Inspection: Comparing DJI Flight Methods on a Real-World Case

Inspecting bronze statues perched more than 50 metres above ground on a palace facade raises a simple question: which automated flight method produces the most accurate 3D model without physical access to the structure. On this heritage project carried out in Brussels in December 2025, four DJI flight methods were tested and compared on the same target: Area Route, Slope Route, Geometry Route and Smart3D Route. Geometry Route came out on top with a georeferencing RMSE (Root Mean Square Error) of 0.012 metres and a 100% RTK Fix, while using 3.7 times fewer images than Smart3D for an equivalent result. Here is the complete method, the parameters of each flight and the measured results.

The context: a heritage building inaccessible without 3D documentation

The building inspected is Palais 5 at Brussels Expo, erected for the 1935 World’s Fair on the Heysel plateau north of Brussels, and designed by architect Joseph Van Neck in a monumental Art Deco style.

Palais 5 is a feat of engineering: 51 metres high, 165 metres long, more than 90 metres wide, the largest reinforced concrete frame in the world in 1935, with parabolic arches spanning 86 metres without intermediate support. Four bronze statues adorn the facade, symbolising Navigation, Steam, Transport and Aviation. They were made under the direction of Egide Rombaux, with a contribution from Zhang Chongren, a Chinese artist and close friend of Hergé, to the design of the facade. The palace served as the main entrance hall for Expo 58 and is today part of a complex of 12 halls totalling 115,000 m², the largest exhibition site in the Benelux.

These 90-year-old structures require regular heritage monitoring. The technical department of Brussels Expo is responsible for maintaining these 12 Art Deco halls. Its starting constraint is common in the heritage sector: no 3D documentation of the current state of the structures existed. The statues, more than 50 metres above the ground, are inaccessible without an aerial platform or scaffolding, with a risk to operators working at height.

The objective of the mission, carried out within the CTR of Brussels Airport: produce an accurate 3D model of the facade and the statues, detect anomalies invisible from the ground, establish a baseline for tracking degradation over time, and above all, comparatively validate the automated flight methods against safety and accuracy criteria.

The equipment: a fully integrated DJI Enterprise ecosystem

The operation was carried out with an entirely DJI ecosystem, without any third-party software. The DJI Matrice 4E was chosen for its integrated RTK module with antenna (no additional positioning equipment required), its 20 MP wide-angle camera optimised for photogrammetry, and its compact form factor suited to constrained urban environments. Classified C2 under the European regulation, the Matrice 4E can be operated in subcategory A2 of the Open category, an advantage for urban missions near people. DJI FlightHub 2 was used for mission planning based on a real 3D model. DJI Terra handled the photogrammetric processing. An RTK network provided real-time centimetre-level positioning.

The choice of an RTK network rather than ground control points (GCPs) is decisive. The DJI Matrice 4E integrates an RTK antenna that connects to the national GNSS network and provides real-time centimetre-level positioning corrections. The drone thereby achieves sub-centimetre absolute georeferencing in RTK Fix mode, without deploying physical targets on site: faster deployment, with equivalent absolute accuracy during reconstruction.

The palace sits in controlled airspace. Prior coordination with air traffic control (ATC) was carried out before take-off, as required by EU Regulation 2019/947, with a restricted flight window. In Belgium, the ministerial decree of 21 December 2020 establishing UAS geographical zones also requires checking airspace access via the DAA Planner tool before each flight. The take-off point was chosen on the public square facing the facade.

The method: first generate a reference 3D model

The whole approach rests on one principle: a close-range inspection is not planned on a generic map background, but on a real, georeferenced 3D model of the building. The first step is therefore to generate this reference model.

Flight 0, an oblique Area Route, was used to generate this reference model. Acquisition parameters:

• Coverage: 12,170.09 m²
• 236 images, 4 min 53 s
• GSD: 1.62 cm/px (oblique GSD 2.28 cm/px)
• Altitude: 60 m AGL, with real-time terrain following

The 3D reconstruction in FlightHub 2 gives:

• Aerotriangulation GSD: 1.859 cm/px
• Coordinate system: WGS 84 / UTM zone 31N (EPSG:32631)
• 3D reconstruction time: 28 min 14 s
• Reprojection RMSE: 1.092 px
• Georeferencing RMSE: 0.018 m

This 3D model serves as the reference and planning basis for all subsequent inspection flights, replacing Google Maps.

The flight methods compared

Three inspection methods were then carried out on this reference model, to determine which gives the best accuracy depending on the target.

The Slope Route (Flight 1) covers the entire facade. Objective: assessing the general condition of the facade. Acquisition parameters:

• Coverage: 725.77 m²
• 237 images, 4 min 33 s
• GSD: 0.17 cm/px
• Distance to slope: 6 m
• Altitude: 81 to 107.2 m ASL, following the geometry of the facade

Reconstruction in FlightHub 2:

• Aerotriangulation GSD: 0.24 cm/px
• Duration: 1 h 16 min 47 s
• Reprojection RMS: 0.995 px
• Georeferencing RMSE: 0.919 m

RTK Fix reached 72% on this flight (284 out of 393 images): the vertical scanning geometry orients the drone away from signal reception on certain passes. A second overlapping pass, or the use of a D-RTK 3 base station, would correct this. The result remains valid for a qualitative assessment of the facade.

The Geometry Route (Flight 2) targets the statues only. Acquisition parameters:

• Coverage: 639.64 m²
• 483 images, 7 min 53 s
• GSD: 0.20 cm/px
• Overlap: 80% frontal, 70% lateral
• RTK Fix: 100% (483 of 483 images)

This is the best-performing method of the test for a complex, constrained geometry.

The Smart3D Route was flown in addition to compare reconstruction accuracy with that of Geometry Route. Acquisition parameters:

• Type: DJI Pilot 2 Smart 3D
• Coverage: 540.07 m²
• 1,065 images, 10 min 26 s
• GSD: 0.14 cm/px
• Overlap: 70% frontal, 70% lateral
• Distance to slope: 5 m

Smart3D offers wider coverage (107,796 m²) but requires 3.7 times more images and 2.9 times more processing time than Geometry Route for the same area of interest.

The accuracy results

The table below compares the three inspection methods on aerotriangulation and georeferencing criteria, based on the quality reports generated by DJI Terra.

Parameter	Geometry Route	Smart3D Route	Slope Route
Images	483	1,065	393
Calibrated images	483 (100%)	1,065 (100%)	392 (99.7%)
RTK Fix	483/483 (100%)	1,065/1,065 (100%)	284/393 (72%)
GSD	1.151 cm/px	1.334 cm/px	0.17 cm/px
Georeferencing RMSE	0.012 m (best)	0.060 m	0.919 m
Reprojection RMS	1.096 px	1.110 px	0.955 px
Tie points	65,573	323,001	128,393
3D processing time	33 min	1 h 38 min	1 h 17 min
Coverage area	639.64 m²	107,796 m²	725.77 m²

RTK Fix per mission confirms the ranking: Geometry Route 100% (mean standard deviation H 0.483 cm, V 0.706 cm), Smart3D 100% (H 0.307 cm, V 0.478 cm), Slope Route 72% (H 5.604 cm, V 6.978 cm). The GSD of the Slope Route is measured relative to the slope surface at a distance of 6 m; its RMSE of 0.919 m reflects the inherent difficulty of aerotriangulation on a vertical surface without GCPs, and remains valid for a qualitative inspection.

The technical conclusion is clear: Geometry Route is the optimal method for a complex, constrained geometry, with a georeferencing RMSE of 0.012 m, a 100% RTK Fix, and 3.7 times fewer images than Smart3D for an equivalent result on the area of interest. Smart3D retains its value for wide-area coverage, at the cost of a significantly higher image volume and processing time.

A word of caution: plan on a 3D model, never on a generic map background

A check was carried out to measure the gap between planning on a real 3D model and planning on a standard map background. A mission was recreated in DJI Pilot 2, reproducing the geometry of the Geometry Route, but on a Google Maps background. When compared with the georeferenced 3D model from FlightHub 2, a position offset of about 2 metres appeared. At a flight distance of 6 metres from the statues, this offset would have caused a collision had the mission been executed automatically without obstacle-avoidance sensors. The mission was cancelled before execution.

The lesson is a point of method: an automated close-range inspection must be planned on a real, georeferenced 3D model, not on a generic map background. The margin of error of such a background, harmless for large-scale mapping, becomes critical a few metres from a structure.

The actionable result: a defect invisible from the ground

The textured 3D mesh of the statues from Geometry Route was generated at a resolution of 1.15 cm/px in DJI Terra, using local standalone processing on a workstation (AMD Ryzen 7 5800X3D, RTX 4070 Super, 64 GB RAM, high point density, high quality, RTK POS data enabled). This resolution is sufficient to identify surface defects on the order of a millimetre on the bronze. The model can be directly viewed and measured in FlightHub 2: distances, angles and surface conditions can be assessed remotely.

Inspection of the 3D model revealed an anomaly: a broken mounting bracket on one of the statues, invisible from the ground and identified solely thanks to the model. The defect was reported to Brussels Expo’s technical department with a precise, annotated 3D location. DJI Terra produced a textured 3D mesh, a point cloud and a 2D orthophoto (TDOM at 1.01 cm/px). Aerotriangulation: 2 min 29 s; 3D reconstruction: 33 min; georeferencing RMSE: 0.012 m.

Integration into Brussels Expo’s workflow

Brussels Expo’s technical department is responsible for the heritage maintenance and structural monitoring of the 12 halls, including Palais 5. Its initial constraint, the lack of 3D documentation and the reliance on physical access for every inspection, is resolved by the deliverables. The textured 3D model of the facade and statues is loaded directly into FlightHub 2, accessible remotely, without any third-party software (B3DM format). Dimensions, crack widths and heights of inaccessible elements are measurable without physical intervention. The point cloud, in LAS format, can be imported into AutoCAD for structural analysis. Reproducing identical flight plans at regular intervals, using the same Geometry Route plan, produces comparable models, enabling automated degradation monitoring. The mission routes saved in FlightHub 2 are reproducible exactly for future monitoring cycles.

Summary

On the technical side, Geometry Route is optimal for a complex, constrained geometry (RMSE 0.012 m, 100% RTK Fix, 3.7 times fewer images than Smart3D); the RTK network replaces ground control points (GCPs) for equivalent absolute accuracy during reconstruction and faster deployment; and the deliverable is directly accessible in FlightHub 2 for the end user. The complete DJI ecosystem (Matrice 4E, FlightHub 2, DJI Terra) required no third-party software. The operation took place in controlled airspace with ATC coordination, in compliance with EU Regulation 2019/947.

On the applied side, the method removes the need for heavy, costly equipment, cuts operational costs by half to a third, eliminates the safety risk for workers at height, and produces accurate 3D documentation of 90-year-old heritage structures with no pre-existing baseline. The outlook is direct: extending the digitisation to all 12 halls for a complete heritage record, annual cycles using Geometry Route with automated model comparison to detect structural changes, and application to other listed buildings in Belgium and the Benelux where access to facades is restricted.

What this mission demonstrates for your inspection project

The choice of flight method determines the quality of the deliverable. For a complex, constrained geometry, FlightHub 2’s Geometry Route, combined with an RTK network and a reference 3D model generated using Area Route, achieves a georeferencing RMSE of 0.012 m and detects millimetre-scale defects on inaccessible elements, without ground control points and without physical access.

Drone Parts Center, authorized DJI Enterprise dealer for the Benelux, supports facility managers, engineering firms and operators on this type of drone inspection mission: equipment selection, acquisition methodology, photogrammetric processing and integration of deliverables into your workflow. To assess the feasibility of a 3D inspection on your site, contact our technical team in Nivelles.

Frequently asked questions

Which DJI flight method is the most accurate for statues or complex geometry?

DJI FlightHub 2’s Geometry Route. On this mission it achieved a georeferencing RMSE of 0.012 m with a 100% RTK Fix, while using 3.7 times fewer images than the Smart3D method for an equivalent result on the area of interest.

What is the difference between Geometry Route and Smart3D Route?

Geometry Route targets a precise area of interest with a reduced number of images and a short processing time (483 images, 33 min in this case). Smart3D Route covers a much wider area (107,796 m²) but requires 3.7 times more images and 2.9 times more processing time. For a complex, well-defined geometry such as statues, Geometry Route is more efficient; for wide-area coverage, Smart3D retains its value.

Do ground control points (GCPs) need to be deployed?

Not with a drone fitted with an integrated RTK receiver such as the DJI Matrice 4E connected to an RTK network. The drone achieves sub-centimetre absolute georeferencing in RTK Fix mode, without physical targets on site, which speeds up mobilisation for equivalent absolute accuracy.

Can a drone detect defects invisible from the ground?

Yes. On this mission, the textured 3D mesh at 1.15 cm/px made it possible to identify a broken mounting bracket on a statue located more than 50 metres above the ground, a defect invisible from the ground and precisely located in the 3D model, without any physical access.

Why plan an inspection on a 3D model rather than on Google Maps?

A generic map background can show a position offset of up to about 2 metres compared with the real geometry. Harmless for large-scale mapping, this offset becomes critical for a close-range inspection a few metres from a structure, where it can cause a collision during automated flight. A georeferenced 3D model generated in FlightHub 2 removes this risk.