Comparative Assessment of Vertical Precision of Unmanned Aerial Vehicle-Based Geodetic Survey for Road Construction: A Multi-Platform and Multi-Software Approach

Abstract

Accurate geodetic surveys are essential for road design, with altimetric accuracy being particularly critical. UAV photogrammetry offers faster and safer data acquisition than conventional methods, but its applicability depends on whether it can meet engineering accuracy standards. This study investigates the altimetric accuracy of UAV photogrammetry through a comparative assessment of surveys conducted on the same urban roundabout in Osijek, Croatia, in 2016 and 2024. By conducting the surveys eight years apart at the same location, the study allows for an assessment of how technological and methodological developments affect survey outcomes. The research evaluates different UAVs and multiple SfM software packages in a comparative framework, highlighting how UAV–software combinations affect results, rather than attributing accuracy solely to hardware or processing. The results of the conducted research indicate a significant increase in the accuracy of the UAV photogrammetric survey method. Through a proper combination of UAVs and SfM processing software, it is possible to achieve an accuracy within 2 cm and an RMSE of 1.2 cm, which is in line with the accuracy of a standard survey method like GNSS CROPOS. The results underline that UAV photogrammetry, when properly planned and executed, can now deliver altimetric accuracy sufficient for most road construction tasks, providing a reliable and cost-effective alternative to conventional geodetic surveys.

1. Introduction

Road design is a complex process that requires analysing field conditions and weighing various options in order to produce the best possible final product. A high-quality project depends on the accuracy of the data used to assess prospective solutions, and the conditions surrounding road construction will reflect this as well. The geodetic survey, which provides the basic input data for the development of a road project, is extremely important in terms of its precision and detail. Within the survey, it is important to know the horizontal and vertical accuracy of the provided data, particularly if they are to be used for structure stakeout. The data obtained via the geodetic survey are used for the creation of 3D digital terrestrial models (DTMs), which represent the main input data for the design of linear objects such as roads and railways. A few geodetic survey methods are in use today, including the total station, global navigation satellite system (GNSS) and photogrammetry methods. Conventional terrestrial survey methods are often inconvenient for larger-area surveys in terms of the required time, labour and cost. Roads and other long, narrow objects can be surveyed at a reasonable price with less time spent using the photogrammetric method. However, a meticulously conducted survey serves as the foundation for reaching the required degree of accuracy. Setting ground control marks (points), adjusting survey measures to the necessary accuracy and performing aerial surveillance during the low vegetation season, when the ground and all objects are clearly visible, are all important steps in preparing the survey. A photogrammetric survey executed by an unmanned aerial vehicle (UAV) is a highly utilised survey method due to its low cost, high speed of spatial data collection and high safety level for the operator. In the past, UAVs were primarily used for military purposes, and this purpose is increasing again due to active worldwide war zones. However, the global market for transport and logistics utilising UAVs is projected to expand from 11 billion USD in 2022 to 29 billion USD by 2027, while UAV production is anticipated to increase from 2 million units in 2021 to 6.5 million units by 2030 [1]. A significant portion of this market is covered by the civil engineering sector, where UAVs are utilised for diverse purposes. UAV applications within the transportation domain may be divided into three areas: road safety (accident investigation, risk assessment and road network surveillance), traffic monitoring (improved traffic flow analysis methods based on data collected from UAVs) and highway infrastructure management (bridge inspection and monitoring and pavement distress recognition) [2]. Additionally, UAVs are used in all road construction phases, from the monitoring of road construction earthworks [3] and vehicle trajectories [4] to pavement damage monitoring, characterisation and quantification [5]. As suggested in [6], the utilisation of UAVs is expected to minimise or even replace conventional road condition surveys in the field. The application of UAVs for designing purposes increasingly involves using them for spatial data collection through the photogrammetric method, with constant advancement in hardware (vehicle and camera) characteristics and accuracy improvement.

Table 1. Recent studies on UAV photogrammetric surveys.

UAV	SfM	Coordinate	RMSE	Reference Method	GCP	Year	Ref.
Phantom 3 Pro	Agisoft PhotoScan	x, y, z	±16 cm, ±23 cm, ±48 cm	GPS	6	2018	[7]
Phantom 4 Pro	ContextCapture	horizontal, vertical	0.88 cm, 0.38 cm	RTK GNSS	3, 4	2021	[8]
DJI Mavic Pro Platinum	Agisoft Metashape and Pix4dmapper	horizontal, vertical	4–6 cm, 5–6 cm	RTK	21	2021	[9]
DJI M 300	Agisoft Metashape	vertical	21 cm	RTK GNSS	16	2025	[10]
DJI M 300	Pix4dmapper	absolute	1.1 cm	Laser RST	-	2024	[11]
Phantom 4 Pro	Agisoft Metashape	3D	1.72 cm to 7.61 cm	Total station	10	2025	[12]
DJI M 300	Pix4D Mapper	x, y, z	±2.66 cm, ±2.41 cm, ±3.47 cm	RTK GNSS	14	2024	[13]

summarises some of the recent studies on UAV application for road construction survey purposes.

The accuracy of UAV photogrammetric survey for road construction purposes has significantly increased during the last few years thanks to technical improvements (in hardware and software) but also improvements in survey methodology. Many studies applied UAV photogrammetry for traffic monitoring [14], pavement condition [15], or road mapping [16], reporting mixed accuracy levels (sub-cm to several decimetres). Increasing the number of ground control points (GCPs) and checkpoints can improve and enhance the accuracy of the UAV orthophoto [7], and using the right number and distribution of GCPs helps to improve the vertical accuracy in particular [17]. Some recent studies aimed to achieve a high UAV survey accuracy through the elimination of GCPs and the use of direct georeferencing via onboard GNSS RTK equipment. This technique, while it has lower accuracy compared to indirect georeferencing, could be very useful for surveying hazardous, hard-to-access, restricted or prohibited areas or areas in which it would be difficult to establish GCPs; it could also be useful in areas for which the spatial data do not need high accuracy, such as agriculture or infrastructure inspection [9]. However, using a GNSS RTK receiver in a multicopter UAV without external verification may result in a systematic shift in the entire point cloud in the vertical components, while the horizontal components are not affected by this error [18]. Other studies showed UAVs can achieve 2–3 cm accuracy in building monitoring [19,20], or cultural heritage documentation [21], but these are different contexts, not directly road-focused.

Some works compared UAVs and SfM software (e.g., Agisoft vs. Pix4D [9]), showing that results vary depending on sensor size, GCPs, or processing algorithms. Precision variation between the SfM software packages may be due to the particular methods of photogrammetric processing being used. In [9] it was pointed out that Agisoft Metashape had better horizontal and vertical accuracy compared to the Pix4Dmapper SfM software. A comparison of various software products in [22] has also shown that the point clouds differ in density and completeness.

In general, the accuracy can be presented in three ways: the planimetric error (X and Y coordinates), altimetric error (Z coordinate) and total (3D) error. For the planimetric error, generally the accuracy can be increased by increasing the number of GCPs or decreasing the UAV flight height, while there is no clear method for increasing the accuracy of the altimetric error [23]. Additionally, the altimetric error seems to be higher than the planimetric error [7,24]. Also, it is evident that different accuracy levels are achieved by different UAV aircraft and different SfM software. When designing or reconstructing roads, the accuracy of the geodetic survey, especially the altimetric accuracy, is particularly important.

Most prior research examines single UAV setups or isolated case studies, making it hard to separate technology improvements from survey design factors. In order to explore the possibility of applying different geodetic methods in the geodetic survey for road design purposes, various geodetic measurements were carried out and results were compared in this research. This paper is novel because it provides a time-based comparative study (2016 vs. 2024) at the same location, evaluates multiple UAV–SfM combinations, and focuses on vertical accuracy benchmarks relevant for road construction.

The scope of this research is to ascertain the altimetric accuracy attained by various geodetic survey techniques (aerial photogrammetry, GNSS and geometric levelling), with a particular focus on identifying the impact of various UAVs and SfM software on the final altimetric accuracy results. The scope is also a comparative analysis of surveys performed on the same urban roundabout in Osijek, Croatia, first in 2016 and again in 2024. The comparison is based on the mean absolute deviation (MAD) and root mean square error (RMSE) of elevation values, benchmarked against conventional survey methods (GNSS CROPOS and geometric levelling). The goal of this study is to examine technological advancements in this particular engineering field and to compare and assess the suitability of various survey techniques for road design. For this purpose, an urban roundabout was the subject of an initial, preliminary investigation in 2016, and the same location was evaluated once more in 2024 using multiple UAV platforms and SfM software. The aim is to determine whether UAV photogrammetry can now achieve accuracy levels required for road construction and whether it can serve as a reliable alternative to conventional survey methods. The research evaluates different UAVs (DJI Phantom 2, DJI Mavic 3, Parrot Anafi) and multiple SfM software packages (Agisoft Metashape, Pix4D, Reality Capture). This comparative framework highlights how UAV–software combinations affect results, rather than attributing accuracy solely to hardware or processing.

2. Methodology

The preliminary and main surveys were conducted at a small urban roundabout in Osijek, Croatia, located in a transitional traffic zone at the city entrance (Figure 1). Its proximity to an overpass, with notable elevation differences and moderate traffic volumes, made it suitable for obtaining higher-quality results.

2.1. Data Collection in 2016

The aim of the preliminary survey was to define the accuracy of up-to-date technology and to monitor the progress of the accuracy of the UAV photogrammetric survey method with technological progress and methodological refinements. The 2016 survey provides a reference point for UAV photogrammetry accuracy at that time, using the technology and workflows available (DJI Phantom 2 + Pix4D, 4 GCPs). Without this baseline, it would be impossible to quantify how much progress was achieved by 2024. This allows the study to separate progress due to technological and methodological advancements from location-specific factors (terrain, geometry, site conditions).

During the preliminary survey in April 2016, an altimetric geodetic survey was carried out using various measurement methods. The following geodetic measurement methods were applied: the total station, GPS RTK method using two satellite receivers of the TOPCON HIPER V type, the GPS CROPOS method using one satellite receiver of the TOPCON HIPER V type and UAV photogrammetry. CROPOS (Croatian Positioning System) was introduced in Croatia in 2008 as a network of reference stations (‘CORS—continuously operating reference station’) and enables continuous GNSS measurements in real time and the transfer of these data to the control centre. Authorised users of the CROPOS system connect to the control centre via the mobile internet GPRS/GSM. The system enables one to obtain the coordinates of measured points in real time, transformed into a defined official reference system called HTRS96 (Croatian Terrestrial Reference System for the epoch 1995.55). Since the introduction of the system, the number of reference stations has been continuously increasing; there were 33 reference stations available in 2016, and currently there are 37 reference stations in the Republic of Croatia at distances of 70 km, covering most of the territory of the Republic of Croatia.

For the preliminary survey, a DJI Phantom 2 Vision+ was used to obtain photogrammetric images and to evaluate the accuracy of different survey methods. In

Table 2. Characteristics of UAVs used in preliminary and main surveys.

Characteristic/UAV	DJI Phantom 2 Vision+	DJI Mavic 3 Enterprise	Parrot Anafi
CMOS * (in)	1/2.3	4/3	1/2.4
Pixels (MP)	14	20	21
Aperture	f/2.8	f/2.8	f/2.4
Max payload (g)	1350	1050	No payload, weight 320 g

* CMOS—Complementary Metal Oxide Semiconductor.

, the characteristics of the UAVs used for the preliminary and main surveys are presented, while the UAVs themselves are presented in Figure 2.

For the photogrammetric survey, 4 GCPs were set for georeferencing purposes (Figure 3), while 50 detail points, equally distributed on the curbs on all four roundabout legs and on the central island were marked and surveyed using four survey methods. Detail points served as independent validation points for accuracy assessment. All points were stabilised with metal wedges on the asphalt path, with a red rectangle drawn around them for subsequent processing in SfM software. Pix4D software was utilised for aerial image processing. Images were acquired at a flight altitude of 60 m at horizontal speed of approximately 3 m/s. A total of 97 photos were taken with a resolution of 14 MP (4384 × 3288 pixels). Pictures were collected with a front overlap ratio of 90%.

2.2. Data Collection in 2024

For the main survey, three methods were used: geometric levelling, GNSS CROPOS and a UAV photogrammetric survey. For georeferencing purposes, 1 reference point and 8 GCPs (numbered 2–9 and marked red in Figure 4) were set, while the survey was conducted on 20 detail points (numbered 10 to 29 and marked blue in Figure 4) equally distributed on the curb of each roundabout approach (Figure 4). GCPs were used for georeferencing and aligning the UAV photogrammetric models, detailed points served as independent validation points for accuracy assessment while the reference point was used to calculate the altitudes of all other points within geometric levelling used as a reference method. Both GCPs and detailed points were surveyed using the same geodetic instruments and techniques, ensuring that differences in accuracy result from the UAV data and processing workflow rather than from differences in measurement technology. The reference point was stabilised with a metal wedge on the asphalt path. All points were stabilised with metal wedges on the asphalt path, with a red rectangle drawn around them for subsequent processing in SfM software.

The elevations of the GCPs and detail points were calculated using geometric levelling. Measurements were made using the Leica NAK2 level device. The elevations of the detail points obtained from geometric levelling were used as reference values for additional comparative analysis.

The GNSS short-term static approach was used to measure the reference point’s position and altitude over a 30 min period. The coordinates of the point were determined in the Topcon Tools programme using the measurement record (.tps) and data on the virtual base downloaded from the CROPOS virtual reference station (VRS) service. The CROPOS system was used to determine the heights of all orientation and detail points using the Stop&Go method, also known as the pseudo-kinematic method, with a brief observation period of a few seconds.

For a close-up aero-photogrammetric survey, two UAVs were used: a DJI Mavic 3 Enterprise UAV equipped with a wide-angle camera and an integrated GNSS RTK module and a Parrot Anafi UAV. Table 2 displays the aircraft and camera characteristics. Both surveys used the same GCPs and detail points during data processing. Both flights were performed at horizontal speed of approximately 5.3 m/s, with an altitude of 60 m. Pictures were collected with a front overlap ratio of 90%. A total of 175 images were taken with a resolution of 14 MP (5280 × 3956 pixels) and 89 images with a resolution of 14 MP (4608 × 3456 pixels) by DJI Mavic 3 Enterprise and Parrot Anafi, respectively. Data processing was performed by three SfM software programmes: Agisoft Metashape, Reality Capture and Pix4D. To ensure consistency and reproducibility, comparable parameter settings were applied across all three SfM software packages. The same image sets and GCPs were used in all workflows whit GCP coordinates imported in the same coordinate system (HTRS96/TM). In Agisoft Metashape, key point and tie point limits were set to 40,000 and 10,000, respectively. Dense point clouds were generated in high quality mode with mild depth filtering. A normal reconstruction setting was applied for dense cloud generation using the high detail depth map preset in Reality capture software. In Pix4D, the dense point cloud was computed with optimal point density and sharp filtering. Following data processing, an orthophoto plan, a point cloud (Figure 5) and a DTM were produced.

3. Results and Discussion

3.1. Preliminary Survey Results and Discussion

In order to compare the accuracies of different survey methods and the potential use of UAVs for altimetric surveys, the results of the UAV, GPS RTK and GPS CROPOS methods are analysed against those obtained from the total station. With an accuracy of roughly 0.5 cm, the trigonometric approach is used to determine the elevation of the detail points gathered using the total station method. The results of other survey methods are compared in terms of the total station method because this method is the most accurate. It measures angles and distances with millimetre precision while elevation determination (via trigonometric heighting) typically achieves accuracy within ±0.2–0.5 cm.

Elevation deviations were calculated for all detail points by comparing the elevation measured using one of the test measuring methods with the reference elevation values. The resulting elevation deviations were processed using the mean value of the absolute elevation deviations (MAD) and the RMSE. The accuracy analysed based on MAD is calculated according to Equation (1):

$$MAD={\displaystyle \frac{{\sum }_{i=1}^n\left|x_i\right|}{n}}$$

(1)

where x_i represents the deviation values for each detail point and n is the total number of detail points.

RMSE analysis is the most used statistical technique for assessing accuracy. This technique assesses the discrepancy between Z_R—a detail point elevation derived from a given survey method (UAV, GPS RTK, GPS CROPOS), and Z_M—a detail point elevation obtained via the survey method taken as the reference one (total station). For the RMSE calculation, Equation (2) is used:

$$RMSE=\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{(Z_R-Z_M)}^2}{n}}}$$

(2)

The deviation in the altimetric result compared to the total station results for all 50 detail points are presented in Figure 6. The calculated MAD and RMSEs are presented in

Table 3. Results of calculated MAD and RMSEs for 2016 survey results.

Calculated Parameter	UAV	GNSS CROPOS	GPS RTK
MAD (cm)	5.72	0.79	1.64
RMSE (cm)	6.78	1.04	1.70

.

For all 50 detail sites, Figure 6 clearly shows that the GPS RTK method provides higher point elevation values than the total station method. The average difference of 1.6 ± 0.5 cm indicates that this survey method has a systematic measurement inaccuracy. Typical vertical accuracy of GPS RTK method is ±1–2 cm under good satellite geometry and stable connection, but systematic errors (e.g., atmospheric delay, multipath) can shift all elevations by a similar bias. That is why this survey method was not included in the main research conducted in 2024. The calculated MAD value of 0.79 cm for the GNSS CROPOS method presents very high compatibility with total station measurements. The UAV method results have the highest deviations compared to the total station results, making this method the least accurate one. The calculated standard deviation of the mean value of the absolute elevation deviations (MAD) is 3.7 cm, pointing to a high variation in the measured elevation (Figure 7). Nevertheless, the calculated MAD value of 5.72 cm indicates a relatively reliable result considering the technological possibilities of 2016. It is worth noting (Figure 6) that more than 85% of the measured detail point elevations were over-measured compared to the reference elevation points. This could also potentially suggest a certain systematic error. Even though this is a significant loss of accuracy compared to the GNSS CROPOS method, it indicates the potential of using UAVs as an alternative, relatively cheap and fast survey method, with applicability in projects that do not require highly accurate measurements.

The highest accuracy in GNSS point determination is achieved through relative positioning, where the coordinates of an unknown point are computed in relation to a known base point using two receivers. In the classic RTK method, corrections are transmitted from a single local base station to a rover over short baselines (typically less than 5 km), which minimises residual satellite orbit, clock, and atmospheric errors because both receivers observe similar satellite constellations. The CROPOS method, as a network RTK system, uses corrections derived from a continuously operating network of reference stations spaced 40–60 km apart. This approach enables broader spatial coverage but can introduce slightly larger residuals in the vertical component due to differences in satellite geometry and atmospheric conditions between the rover and reference stations. As a result, small systematic differences between RTK and CROPOS height determinations may appear, especially when measurements are taken at different times of day or under varying atmospheric conditions. It should also be noted that the 2016 measurements were performed with earlier-generation instruments and software, whose performance was more limited compared to current technology.

3.2. Main Survey Results and Discussion

The elevation of the detail points was first determined by geometric levelling in relation to the reference point. These heights are reference values in a comparative analysis with other measurement methods applied in this study (GNSS CROPOS and two UAV photogrammetric methods processed by three SfM software programmes). Recordings from two UAV photogrammetric surveys were processed by three SfM software programmes, Agisoft Metashape, Pix4D and Reality Capture, with the aim of assessing their differences and applicability for specific, road design and construction purposes. Elevation deviations were calculated for all detail points by comparing the elevation measured using one of the test measuring methods with the reference elevation values. The resulting elevation deviations were processed based on MAD, the standard deviation (σ) of these deviation values and RMSE using Equations (1) and (2), as in the preliminary research. Here, σ indicates the variability in the deviation values around their mean for each survey method, not the population standard deviation of the entire dataset. The measured detail point elevation results for different survey methods relative to the geometric levelling reference survey method results are presented in Figure 8 and

Table 4. Statistical data of the results from conducted survey.

	GNSS CROPOS	DJI Mavic 3			Parrot Anafi
		Agisoft Metashape	Reality Capture	Pix4D	Agisoft Metashape	Reality Capture	Pix4D
MAD (cm)	1.21	1.07	1.62	2.93	8.36	5.56	4.26
RMSE	1.61	1.20	1.98	4.31	9.74	6.41	5.70
σ (cm)	1.08	0.57	1.17	4.25	5.13	3.27	5.85

.

The improved accuracy observed in the 2024 survey can be attributed to an optimised workflow that integrates several methodological refinements. These include: (i) careful pre-survey planning with optimised flight parameters (altitude of 60 m, 90% overlap, 5.3 m/s flight speed); (ii) the use of a greater number and improved spatial distribution of ground control points (eight GCPs compared to four in 2016); (iii) consistent application of reference coordinate systems and geometric levelling as control; and (iv) standardised data processing using identical parameter settings across all SfM software. Together, these workflow enhancements minimised systematic deviations and increased the reliability and vertical precision of the UAV-based results.

A direct comparison between the DJI Phantom 2 Vision+ with Pix4D (2016) and the DJI Mavic 3 Enterprise with Pix4D (2024) clearly demonstrates the technological advancement in UAV photogrammetric surveying over the eight-year period. In 2016, the DJI Phantom 2 combined with Pix4D achieved a mean absolute vertical deviation (MAD) of 5.72 cm and an RMSE of 6.78 cm relative to total-station reference data, while in 2024 the DJI Mavic 3 Enterprise processed in Pix4D achieved a MAD of 1.62 cm and an RMSE of 1.98 cm compared to geometric levelling results. This represents roughly a 70% improvement in altimetric accuracy and precision, moving from engineering-grade to survey-grade performance. The improvement stems primarily from advances in UAV hardware—larger and higher-quality sensors, mechanical shutter and integrated RTK in the Mavic 3—as well as algorithmic enhancements in Pix4D processing. However, accuracy in photogrammetry is influenced by many other factors beyond just hardware and software progress. The number and distribution of GCPs, the choice of reference methods, flight planning parameters (altitude, overlap, speed), environmental conditions, and operator experience all affect the final results. The higher accuracy achieved in 2024 is therefore a result of both technological progress and refinements in survey methodology. Consequently, UAV-based photogrammetry that once yielded decimetre-level accuracy now routinely attains centimetre-level results suitable for road design and construction applications.

From the results presented in Figure 8, it is evident that the highest deviation from the reference elevation results occurs for the Parrot Anafi UAV, regardless of the SfM software used. All other survey methods display similar results. Less accurate results were recorded by using the UAV photogrammetry method conducted with the Parrot Anafi UAV and processed by Agisoft Metashape software. The lowest MAD and RMSE values, i.e., the most accurate results, are recorded by the UAV photogrammetric survey method using the DJI Mavic 3 UAV and Agisoft Metashape software. It is important to note that this method yielded even more precise results than the GNSS CROPOS method. An interesting trend may also be observed when the deviations from the reference results are analysed from the perspective of over-measured and under-measured recorded values. Namely, all the survey methods used presented a similar trend, with 55–65% over-measured values, except for the DJI Mavic 3 and Agisoft Metashape combination. This method recorded only 25% over-measured values.

A consistent positive bias was observed in both surveys, where more than 85% of the 2016 elevations and 55–65% of the 2024 elevations were slightly overestimated relative to the reference values. Several factors may explain this systematic deviation. First, UAV photogrammetry tends to produce slightly higher elevation estimates when the ground surface includes bright or reflective materials (such as water [25], asphalt or concrete) that cause local image saturation, affecting tie-point matchings. Second, the vertical component of UAV photogrammetric models is particularly sensitive to small errors in camera calibration, which can propagate as a uniform shift in the model [26,27]. Third, insufficient or uneven GCP distribution can introduce a vertical tilt or bias in the photogrammetric block adjustment, especially when GCPs are located mainly along the periphery of the survey area [28,29]. Although these effects were reduced in 2024 through a larger number and better distribution of GCPs and improved RTK georeferencing, minor positive bias remained, likely due to inherent systematic offsets in UAV direct georeferencing and the influence of image geometry under near-nadir flight conditions. Future studies should include an explicit bias correction step to further minimise this effect.

As presented in Figure 9, with a proper combination of UAVs and SfM software, an accuracy within 2 cm may be achieved; this is suitable for most road construction projects, particularly for designing purposes. It can also be concluded that different SfM software are more or less suitable for particular image processing purposes, depending on the UAV used. For DJI Mavic 3 UAVs, image processing using Agisoft Metashape gave more accurate results compared to the processing of the same recording by Reality Capture and Pix4D software. On the contrary, image processing using Pix4D software seems to be better for processing recordings obtained by Parrot Anafi UAVs. Comparing the three used software programmes, all of them offer the ability to create 3D spatial data, such as point clouds and georeferenced orthophoto maps. They are used in GIS applications, architecture, geodesy, cultural heritage documentation, visual effects production and the indirect, so-called non-contact measurement of objects of different sizes (e.g., when calculating the cubic capacity of masses).

The Reality Capture programme is currently supported only on computers running the Windows operating system with NVIDIA graphics cards, which are designed to perform demanding computer operations—processing large amounts of data in a shorter amount of time. Designed for speed and efficiency, it uses GPU-accelerated algorithms and is highly optimised for handling very large datasets quickly. It relies heavily on dense image matching and fast bundle adjustment but has fewer options for user control (e.g., you cannot directly build surfaces from point clouds). Reality Capture tends to produce models faster, but accuracy may be slightly worse in datasets with less stable geometry or small sensor UAVs. This aligns with findings that Reality Capture showed larger deviations, particularly with Parrot Anafi.

The Agisoft Metashape programme is compatible with the Windows, macOS and Linux operating systems. It uses a feature-based SfM pipeline. Key steps include SIFT-like keypoint detection, bundle adjustment (based on least-squares minimization), and dense image matching (often semi-global matching). It allows for chunk-based processing, meaning the dataset can be split and recombined, improving stability and alignment for complex or large datasets. It is strong in robust tie-point detection and small-scale digitisation, but processing is slower compared to other packages. This explains why Metashape often delivers slightly better accuracy (lower RMSE) but with longer processing time.

Pix4D is a software package for photogrammetric data processing. The Pix4D Capture module allows one to plan how to capture images with compact cameras, DSLR cameras, drone cameras or large-format cameras. It allows one to plan how to capture images for aerial photogrammetric tasks and create 3D models. It is especially adapted for capturing images with Anafi and DJI drones with GSD 1–2 cm. Pix4D Mapper is software designed to process images collected by RGB, multispectral and LiDAR sensors. The end result consists of a point cloud, a mesh, an orthomosaic, a high-accuracy DTM (1–2 cm) and a report on the accuracy and quality of the image capture. It follows an SfM workflow like Agisoft Metashape, but optimised with its own proprietary multi-ray matching algorithm, which can be more efficient for large UAV datasets. It is known for fast processing and flexibility with many UAV models (including integration with capture apps). However, Pix4D often balances between speed and density, which can sometimes result in lower vertical accuracy compared to Metashape, especially if camera calibration or GCP distribution is not optimal. Pix4D performed better with the Parrot Anafi dataset, likely due to its tuning for smaller, consumer-grade sensors.

The Agisoft Metashape software provides fast and easy operation due to its more accessible, so-called ‘user-friendly’ interface, and it requires a low level of manual work. However, the data processing time is longer compared to the Reality Capture software, as is pointed out in [21,30]. The difference in accuracy may also be the result of methodology differences in the used software since different pieces of SfM software use different algorithms and processing options. While in Agisoft Metashape, it is possible to directly specify the data source when generating a point cloud, while in Reality Capture, it is not possible to build a surface model based on a point cloud from dense image matching [30]. Also, in Agisoft Metashape, there is a significant increase in time for tie point detection and image block alignment, but it has the ability to create chunks and combine them for a complete model, contrary to Reality Capture [21,30]. Finally, Agisoft Metashape has more reliable photo alignment for small-scale digitisation [21]. As for Pix4D software, it presented the most accurate results for the Parrot Anafi UAV records, but as presented in [31], it may be less suitable for surveying and reconstructing underwater features. Finally, choosing the right software is dependent on the survey type and purpose and may affect the final results; using the same software will give users greater confidence if the same object is to be monitored and observed over a longer period of time. The accuracy of the conducted surveys is also affected by the UAVs used and their on-board cameras. The Parrot drone’s camera has a smaller CMOS 1/2.4-inch sensor. The DJI Mavic’s CMOS sensor is superior due to its larger size, variable aperture, wider ISO range and ability to use a mechanical shutter. These features allow for better image quality, especially in low-light conditions and when shooting moving subjects. The DJI Mavic also offers advanced features such as obstacle avoidance, a longer flight time and greater wind resistance, making it ideal for professional tasks in demanding conditions. The Parrot Anafi drone is lighter and more compact, making it suitable for users who value portability and ease of use. Thus, as in the choice of appropriate software, camera capabilities should also be considered depending on the specific application of the methodology and the expected level of detail.

The results are also influenced by the flight characteristics. In order to increase the accuracy of any photogrammetric survey, it is necessary to select the desired level of accuracy. Based on this, parameters like the flight level, image overlapping and the number of GCPs, which determines the recording resolution, can be set. Adjusting the flying height in particular could be utilised to increase the accuracy of UAV surveys, but the safety issue needs to be addressed—operating UAVs in densely populated regions can be dangerous for both the UAV equipment and any bystanders, such as people or other traffic. Additionally, a flight level that is too low, with the camera being too close to the ground, results in insufficient distinguishing features to stitch the images together, thus warping the models created [8].

Finally, it is also important to note that, despite the significant improvements in UAV technology and accuracy, the application of UAV photogrammetry is still subject to regulatory restrictions that define where and how UAVs may be operated. Most national aviation authorities, including those in the European Union under EASA regulations, prohibit or strictly limit flights over densely populated areas or near airports without special authorization. Consequently, when planning surveys in urban or infrastructure zones—such as the roundabout in this study—it is essential to ensure that flight operations comply with local airspace regulations and safety requirements. These regulatory boundaries directly influence flight planning, allowable altitudes, and data acquisition strategies in professional UAV-based surveying.

4. Conclusions

The aim of this study was to assess the accuracy of different survey methods for road construction purposes. The scope was the comparison of two UAV photogrammetric surveys over the same area and an eight-year period (2016–2024), with the specific objective of determining the applicability of a specific hardware/software combination based on the obtained accuracy. The added value of the research is the analysis of the accuracy of the UAV photogrammetric survey method over a period of 9 years through analyses of survey results at the same test site. Based on the conducted research, the following conclusions can be drawn:

• The altimetric accuracy improved from 6 cm to better than 2 cm. This demonstrates that UAV photogrammetry can now provide results comparable to the GNSS CROPOS method, making it suitable for road design and construction tasks where vertical accuracy is critical. Hardware and software developments—such as larger CMOS sensors, integrated GNSS RTK modules, and more robust SfM algorithms—contributed significantly to improved precision. However, accuracy is not determined by technology alone. The number and distribution of GCPs, the choice of reference methods, flight planning parameters (altitude, overlap, speed), environmental conditions, and operator experience all affect the final results. The higher accuracy achieved in 2024 is therefore a result of both technological progress and refinements in survey methodology and workflows.
• Through a proper combination of UAVs and SfM processing software, it is possible to achieve an accuracy within 2 cm, which is in line with the accuracy of a standard survey method like the GNSS CROPOS method.
• A minor positive vertical bias, consistent across survey years, suggests that UAV photogrammetry may systematically overestimate elevations unless vertical correction or calibration is applied; this should be considered in future workflows.

This study makes a distinct contribution by providing a long-term comparative assessment of UAV photogrammetry accuracy over an eight-year interval, combining comparative analyses of multiple UAV platforms and SfM software packages. By systematically examining how hardware and software developments, together with methodological refinements, influence vertical precision, the research demonstrates that contemporary UAV photogrammetry systems can now reliably achieve the centimetre-level accuracy demanded for engineering and road construction applications. These findings confirm that UAV-based photogrammetric surveys have evolved from an experimental tool into a mature, engineering-grade method capable of complementing or even replacing conventional geodetic techniques in many practical contexts. Consequently, this study provides a clear benchmark for future UAV-based survey methodologies in road construction practice. Nevertheless, careful planning of survey parameters and processing strategies remains essential to achieve the highest possible accuracy.

However, several limitations should be acknowledged. The analysis was based on a limited number of UAV platforms and therefore does not represent the full range of available UAV technologies or sensor configurations. The survey area was a single, small urban roundabout and different terrain types such as mountainous roads, dense vegetation, or large linear alignments may produce different accuracy outcomes. The software parameters and flight settings were optimised for this site and may need adjustment for other environments or UAV platforms. Future research should include a broader range of UAVs, cameras, tests under varied environmental and terrain conditions, and larger-scale linear infrastructure projects to confirm scalability.