Evaluation of the Accuracy of Direct Georeferencing of Photogrammetric Products in a Large Area with Steep Topography

Highlights

What are the main findings?

• The comparison between the coordinates measured with GNSS and those of the orthomosaic shows bias of −0.045 m on the X-axis, 0.001 m on the Y-axis, and 0.599 m on the Z-axis.
• Direct image georeferencing provides better results in horizontal coordinates, primarily along the X-axis. The standard deviation indicates significant anisotropy between the components. The X-axis shows a dispersion of 0.221 m, while the Y-axis exhibits a variability of 0.791 m, and the Z-axis shows the worst performance with a dispersion of 1.130 m.

What are the implications of the main findings?

• A significant positive bias indicates that the orthomosaic is consistently inferior to the GNSS by 0.60 m. The bias in the X and Y-axes shows little systematic trend. This suggests greater accuracy in both axes.
• UAV surveys in areas with complex topography require improved acquisition strategies, such as greater overlap, additional control points, and possibly multi-angle imagery. The distance from the takeoff point is not a reliable predictor of the accuracy of direct georeferencing of images using NTRIP/CORS.

Abstract

Technological advancements have revolutionized photogrammetry, with the implementation of unmanned aerial vehicles for capturing images from different angles and the ease of obtaining sensor position information at the time of capture. This study evaluates the accuracy of direct georeferencing via Networked Transport of Radio Technical Commission for Maritime Services Via Internet Protocol, in the orthomosaic as a photogrammetric product in a large urban area with steep and highly variable topography, comparing it with the coordinates of nine checkpoints obtained with GNSS equipment connected to the National Active Geodetic Network, managed by the National Institute of Statistics and Geography of Mexico. An orthomosaic of the historic center of Zacatecas was obtained with a resolution of 2.70 cm/pixel. The orthomosaic coordinates, compared to those of the GNSS equipment, show a root mean square error (RMSE) of 0.78 m in the horizontal coordinates and an RMSE of 1.22 m in the vertical coordinates. Previous studies prove the efficiency of the Continuously Operating Reference Station module and network with other aircraft; this study determines that this is true for large areas with high coverage and quality in the internet network, but with rugged topography, the results are not accurate.

1. Introduction

The use of unmanned aerial vehicles (UAVs) in aerial photogrammetry has surged. In contrast to manned aircraft, drones are more affordable [1], making them cost-effective, especially for large projects. Given their ability to provide higher-resolution images and more precise positioning than satellite imagery [2], data collection in topographic surveys is streamlined. Their ease of operation, clarity of detail, high resolution, low production cost, and use have made them a more widely adopted method for topographic mapping in engineering surveys [3,4]. Although remote sensing platforms such as satellites offer cost-effective and wide coverage, their positional altitude and spatial resolutions are not suitable for generating the high-spatial-resolution terrain data required for detailed topographic mapping [5]. On the other hand, although laser scanning and terrestrial LiDAR are capable of providing terrain information at high spatial resolution, their limitation lies in the cost of these techniques and the experience required in data collection and processing [6]. This is why unmanned aerial vehicles (UAVs) are increasingly appearing on the market to obtain data with greater spatial, spectral, temporal, angular, and radiometric resolution, thus reducing costs, time, and human resources [7]. The direct georeferencing of these photogrammetric models or products has constantly evolved thanks to the technology integrated into UAVs. Direct georeferencing significantly reduces fieldwork requirements while enabling the generation of high-precision geospatial products, such as Digital Surface Models (DSMs), orthomosaics, and topographic maps. Unlike traditional GNSS surveys, direct georeferencing integrates sensor data collected aboard an unmanned aerial vehicle (UAV) with geometric correction of the image position [8].

Real-Time Kinematic (RTK) and Post-Processed Kinematic (PPK) technologies represent significant advancements in the evolution of UAV-based digital aerial photogrammetry. RTK is a key technique for real-time positioning in surveying and other applications requiring high accuracy [9]. It is based on the transmission of carrier-phase differential GNSS corrections from a reference receiver located at a known position to one or more mobile receivers aboard the UAV. These corrections, typically transmitted via ultra-high-frequency (UHF) radio links or internet-based communication channels, allow the compensation of common GNSS error sources and enable centimeter-level positioning accuracy during image acquisition [10]. When RTK corrections are not available, UAV surveys can rely on the Post-Processing Kinematic (PPK) approach. Unlike RTK, which applies positioning corrections during the flight, PPK computes differential GNSS corrections after data acquisition [11], enabling accurate camera position estimation even in remote or poorly connected areas. Although this approach requires additional post-processing time, PPK can provide positioning accuracy comparable to RTK under appropriate operational conditions. PPK-based UAV photogrammetry has been successfully applied in several geoscientific contexts, including geological and geomorphological mapping [12], and archeological applications [13]. One of its main advantages is the ability to generate accurate photogrammetric products—such as orthomosaics, Digital Elevation Models (DEMs), and dense point clouds—while reducing or eliminating the need for Ground Control Points (GCPs). This characteristic makes PPK particularly suitable for large or inaccessible areas, where traditional ground-based control is difficult to deploy. UAVs are not only capable of performing RTK surveys with a fixed ground station, but they can also carry out field measurements. Currently, there are drones on the market with the ability to receive corrections via the internet, known as Networked Transport of RTCM via Internet Protocol (NTRIP) to Continuously Operating Reference Stations (CORS). This means that each image the drone acquires carries the phase center coordinate of the RTK antenna already corrected and with high accuracy, eliminating the need to position ground control points or use a GNSS base station, thus reducing time, costs, and fieldwork. The combination of NTRIP and CORS improves the effectiveness of RTK correction by enabling wider coverage, reducing dependence on local base stations, and improving the availability and accuracy of correction data [14]. Evaluating the accuracy of UAV photogrammetry is essential because it determines the quality of large-scale measurements. Early consumer UAVs provided great convenience for surveying and mapping tasks [15].

According to [16], the accuracy and precision of photogrammetrically generated DSMs depend on geometric and physical parameters, such as: image scale, ground sampling density (GSD), the ratio between stereo base length and object distance, camera grid geometry (nadir, transverse, and oblique strips), strip overlap percentages (front and side), accuracy and distribution of ground control points (GCPs), camera calibration, surface texture and albedo, lighting conditions, air refractive index, sensor quality (signal-to-noise ratio and dynamic range), image sharpness (no blur), as well as processing methods, including SfM, image matching, point cloud noise, and outlier removal algorithms. Ground control points (GCPs) are typically measured using real-time kinematic (RTK) GNSS surveys, either relative to a nearby base station or with differential corrections sent by a Continuously Operating Reference Station Network (CORS). Both techniques have an accuracy of 1 to 2 cm in horizontal coordinates and 2 to 3 cm in elevation (i.e., to a level that often coincides with the ground-state distance).

Ref. Costantino et al. [17] describes current prospecting methodologies based on photogrammetric techniques to produce 3D models, orthophotos with high geometric resolution, and Computer-Aided Design (CAD) representations useful for restoration or maintenance activities of cultural heritage. They used a Mavic 2 Pro, Agisoft photoscan, Pix4D Capture, and GNSS. The work demonstrates the enormous potential offered by photogrammetric techniques and those based on the use of active sensors for the documentation and representation of sites; this approach is particularly useful in the digitization of cultural heritage environments.

Muhammad et al. [3] determined that producing an orthophoto requires at least five ground control points (GCPs) in the field. The more control points an image has, the greater its accuracy. GCPs must be permanent structures, and their positioning must cover all study areas. The GCP should not only be placed near the boundary of the study area but should also cover its center. In the same project, it was established that the optimal overlap and lateral overlap levels for producing topographic maps are 80% to 50% and 70% to 40%, respectively. With optimal flight altitude and overlap, the best results can be obtained for digital elevation models (DEMs), orthophotos, and topographic maps.

The research by Teppati et al. [18] analyzes the possibility of eliminating the use of traditional ground control points (GCPs) in the photogrammetric process in order to reduce the costs and time of survey operations. DJI Matrice 300 UAV was used to document built heritage and Agisoft Metashape processed the information. The results obtained were considered satisfactory for meeting the documentation process needs in the field of Cultural Heritage.

Hugenholtz et al. Ref. [19] compared the accuracy of orthophotos and Digital Surface Models (DSMs) produced from UAV imagery using georeferencing (GCP) and direct georeferencing (i.e., without GCP). The results showed that the horizontal accuracy of orthophotos produced by direct georeferencing of RTK UAV imagery was very similar to the accuracy obtained with non-RTK UAV imagery using GCP; however, the vertical accuracy of the DSMs differed by a factor of 1.9, with the latter yielding a lower RMSE. Direct georeferencing with RTK UAV data is suitable for projects requiring the highest possible horizontal accuracy thresholds, but for the highest possible vertical accuracy thresholds, practitioners should use GCP.

Syetiawan et al. Ref. [20], centimeter-level position accuracy per photograph was achieved. The authors conclude that the use of NTRIP-RTK UAVs shows great potential for generating high-precision maps. According to the accuracy test results for the horizontal and vertical orthomosaic in DSM, the NTRIP-RTK accuracy on the UAV is 0.775 m and 0.215 m, respectively. However, in their study, both the GNSS receiver and the aircraft obtain the correction using the same network.

Table 1. Comparative table of flight parameters used by other authors.

Investigation	UAV	Height (m)	Forward Overlap	Side Overlap	Area/Length
[20]	DJI Mavic 3 enterprises RTK	120 m	80%	70%	Not specified
[21]	Autel Explorer EVO II pro	65 m	80%	70%	1.38 ha
[22]	DJI Phantom 4 RTK	35 m	80%	80%	3.24 ha
[23]	DJI Matrice 300 RTK	140 m	10, 30, 60 y 80%	10, 30, 60 y 80%	3 ha
[24]	DJI Phantom 4RTK	100 m	70%	50%	229.8 ha
[25]	DJI Phantom 4 RTK	25 m	80%	80%	2 ha
[26]	DJI Phantom 4 Pro V2	150 m	60–90%	60–90%	0.25 ha
[27]	Not specified	20 0m	80%	70%	16 ha
[28]	DJI Phantom 4 PRO V2	60 m	70–80%	70–80%	1.5 has
[29]	DJI Phantom 4 RTK	80 y 120 m	80%	70%	9.43 ha

shows the flight parameters used by other authors, with different aircraft and study areas with specific characteristics. The altitude is defined based on the sensor characteristics and the expected resolution of the photogrammetric products. Forward overlap must ensure that the center of one image is contained within another and vice versa. In areas with rugged topography, the minimum overlap is not recommended to guarantee stereoscopic coverage of the entire study area.

In topographic map production, the optimal flight altitude is between 20 m and 120 m [30]. If the altitude is increased to 120 m, image features become difficult to identify and interpret, even though a larger area is covered in each photograph. For clear visualization and easy interpretation, the altitude should be below 120 m. The use of GCP for RTK UAV work is still used by most professionals to ensure the accuracy of georeferencing images. However, UAV companies are constantly improving the characteristics of the antennas so that they can do without them, especially when placement is an almost impossible task, depending on the conditions of the site to be studied. The objective of this study is to evaluate the accuracy of direct georeferencing via NTRIP/CORS, in the orthomosaic as a photogrammetric product in a large urban area with steep and highly variable topography, comparing it with coordinate checkpoints obtained with GNSS equipment connected to an RGNA station.

2. Materials and Methods

2.1. Equipment

For the development of this study, a Tersus GNSS system Oscar Ultimate, manufactured by Tersus GNSS Inc., Shanghai, China. (base) and Basic (rover), a Sokkia GRX2 GNSS system manufactured by Sokkia Co., Ltd., a company based in Atsugi, Japan. (Base and Rover) and a DJI Mavic 3T UAV created by Da-Jiang Innovations, located in Shenzhen, China were used. In addition to the aforementioned equipment, the following software was used: Trimble Business Center 5.52, Agisoft Metashape 2.2.0, and ArcMap 10.8.

2.1.1. Tersus GNSS Systems

The Oscar GNSS receiver is a next-generation RTK GNSS system. It supports tilt compensation without calibration, is immune to magnetic disturbances, and requires no leveling rod. With a high-performance, multi-constellation (GPS, GLONASS, BeiDou, Galileo, QZSS, SBAS) and multi-frequency (supports multiple frequencies for each constellation) internal GNSS board, the Oscar GNSS receiver provides high accuracy and stable signal detection. The high-performance antenna accelerates time to first fix (TTFF) and improves interference protection. The integrated UHF radio module enables long-range communication. The Oscar GNSS receiver is available in three versions: Ultimate, Advanced, and Basic. RTK Accuracy: horizontal, 8 mm + 1 ppm; vertical, 15 mm + 1 ppm [31].

2.1.2. Sokkia GRX2 GNSS System

The Sokkia GRX2 instrument features a fully integrated dual constellation, bringing a new level of versatility and flexibility to precision applications. Key features of the GRX2 include: RTK and static surveying operations, 226 channels, multi-constellation (GPS, GLONASS, SBAS), and RTK accuracy of 10 mm + 1 ppm horizontally and 15 mm + 1 ppm vertically [32].

2.1.3. DJI Mavic 3T UAV

The DJI Mavic 3T UAV features a 1/2-inch wide-angle CMOS camera; effective pixels, 48 MP; an RTK module compatible with GPS + Galileo + BeiDou + GLONASS constellations, with horizontal positioning accuracy of 1 cm + 1 ppm; vertical positioning accuracy of 1.5 cm + 1 ppm [33].

2.2. Metodology

The methodology of this study comprises three phases. These are divided into corresponding activities, as shown in Figure 1.

2.2.1. Study Area

The polygon’s boundaries were defined according to the area known as the Historic Center of Zacatecas [34]. The study area covers approximately 109.15 hectares [35], located in the state of Zacatecas, municipality of Zacatecas, with coordinates of 22°46′31.742″ north latitude and 102°34′24.712″ west longitude, approximately, as shown in Figure 2.

2.2.2. Checkpoints

The positioning of checkpoints consisted of defining twelve points to verify the accuracy of the results. These points were distributed across the length and width of the polygon to be surveyed. Once the locations of the checkpoints were defined, they were marked on the ground with red paint in the shape of a cross (approximately 1 × 1 m) with a white center and a screw and washer anchored to the ground. The checkpoints were distributed considering the elevation variations according to the DEM, and spaces free of obstructions at ground level on a flat surface were prioritized so that the mark would not be distorted (Figure 3). Once the location of the 12 points was defined, the coordinates were obtained using the RTK method. The points were surveyed with the Tersus GNSS receiver for 3.4 h, receiving data to obtain the coordinates. The same survey was then performed with the Sokkia receiver for 3.5 h. Data processing was carried out using the IZAC and INEG stations of the National Active Geodetic Network (RGNA), managed by the National Institute of Statistics, Geography and Informatics (INEGI). The post-processing of each base point was performed separately, obtaining the coordinates shown in

Table A1. Vector components with GNSS Tersus.

	From: INEG	To: Tersus Base	Vector
X	780,685.1403	749,893.0988	−30,792.041
Y	2,419,383.0498	2,520,961.9352	101,578.885
Z	1901.7065	2583.8049	682.0984

and

Table A2. Vector components with GNSS Sokkia.

	From: INEG	To: Sokkia Base	Vector
X	780,685.1403	749,893.0986	−30,792.042
Y	2,419,383.0498	2,520,961.9376	101,578.888
Z	1901.7065	2583.8147	682.1082

. In relation to the INEG station using Trimble Business Center 5.52 software, Table A1 shows the post-processing report of the base point with the Tersus equipment, to which all RTK point surveys will be linked. Table A2 shows the same post-processing report of the vector formed by the base point and the RGNA antenna, but with the Sokkia equipment. According to Table A1 and Table A2 the base for checkpoints surveyed with both instruments vary by 0.2 mm, 2.4 mm, and 9.8 mm in the X, Y, and Z coordinates, respectively.

Once the point was processed, the baseline was corrected to the RTK survey.

Table A3. Coordinates of the checkpoints obtained with both GNSS equipment.

Point	${\mathit{X}}_{\mathit{Sokkia}}$	${\mathit{Y}}_{\mathit{Sokkia}}$	${\mathit{Z}}_{\mathit{Sokkia}}$	${\mathit{X}}_{\mathit{Tersus}}$	${\mathit{Y}}_{\mathit{Tersus}}$	${\mathit{Z}}_{\mathit{Tersus}}$	HD	VD
1	749,445.927	2,521,736.707	2470.270	749,445.918	2,521,735.788	2471.984	0.919	1.714
2	–	–	–	749,368.068	2,521,475.573	2425.678	–	–
3	749,352.486	2,521,214.893	2415.005	749,352.509	2,521,214.889	2415.108	0.023	0.104
4	749,138.824	2,520,786.051	2417.841	–	–	–	–	–
5	749,207.535	2,520,468.386	2395.818	749,207.516	2,520,468.354	2395.882	0.037	0.064
6	748,831.691	2,520,435.222	2408.812	748,831.695	2,520,435.252	2408.849	0.031	0.037
7	–	–	–	748,802.048	2,520,127.543	2424.722	–	–
8	749,320.037	2,520,038.221	2390.355	749,320.052	2,520,038.188	2390.344	0.035	−0.011
9	748,698.499	2,519,699.923	2430.648	748,698.436	2,519,699.915	2430.731	0.065	0.083
10	–	–	–	749,258.288	2,520,782.997	2420.197	–	–
11	749,150.157	2,520,249.379	2395.004	749,150.153	2,520,249.363	2395.023	0.016	0.019
12	749,351.743	2,521,217.205	2415.201	749,351.740	2,521,217.223	2415.308	0.018	0.107

compares the coordinates obtained with Sokkia and Tersus and the separation of one coordinate from the other on the same point. The data obtained from both GNSS systems underwent a three-dimensional translation transformation and were applied to the coordinates obtained with TERSUS, using SOKKIA homologous points. One outlier (point 1) was identified, and was not excluded from the model calculation and is retained in the analysis. The results show adequate agreement after the correction, with no evidence of significant rotation or scaling. Points with no correspondence between the two systems were not used in the transformation model calculation; however, the resulting correction was applied to all points in the TERSUS system, ensuring the spatial consistency of the dataset. Once the coordinates from both sources of information were compared, the orthomosaic and the measurements with the GNSS, the difference between them (dX, dY, dZ) was obtained to calculate the root mean square error (RMSE).

The combined horizontal error (dX,dY) is obtained with Equation (1), which calculates the horizontal root mean square error $RMS{E}_H$.

$$RMS{E}_H=\sqrt{RMS{E}_X^2+RMS{E}_Y^2}$$

(1)

The $RMS{E}_{3D}$, which summarizes the total spatial error between both datasets, is obtained with Equation (2).

$$RMS{E}_{3D}=\sqrt{RMS{E}_X^2+RMS{E}_Y^2+RMS{E}_Z^2}$$

(2)

The DEM was used to classify the slopes and analyze the representation of the nine checkpoints in the model. According to Figure 4 the moderate slope class (7–15%) appears to be overrepresented in the final dataset. This is explained by the fact that, although a total of 12 checkpoints were originally planned and positioned, several points could not be used during the flight campaign due to the presence of metallic structures that obstructed their visibility. As a result, the usable checkpoints were reduced, leading to a relative concentration of points within the moderate slope range. Nevertheless, the Chi-square goodness-of-fit test applied to the three slope classes (0–7%, 7–15%, and >15%) yielded a value of $X^2$ = 2.00, which is lower than the critical value at the 0.05 significance level $(X_{0.05,2}^2$ = 5.991). This indicates that there are no statistically significant differences between the observed and expected distributions. Therefore, despite the apparent overrepresentation of the moderate slope class, the checkpoints can still be considered to provide an adequate and statistically consistent coverage of the topographic conditions within the study area.

2.2.3. Flight Plan

The flight plan was designed in an east–west direction, which allowed the entire area to be covered using the least number of images and in the shortest time. The perimeter was increased, so in total more than 150 hectares were surveyed, as shown in Figure 5. The photogrammetric flight was conducted at an altitude of 85 m above ground level, as this is the optimal altitude for obtaining fewer, high-resolution images. Above this altitude, the number of photos decreases, as does the size of the Ground Sample Distance (GSD). With a speed of 9 m/s, a gimbal angle of 90°, 80% lateral overlap, and 80% frontal overlap, the overlap was defined equally in both directions due to the variability of the terrain topography, to ensure stereoscopic coverage of the entire mosaic, given that it is an urban area and was only flown over in one direction with images of the vertical axis. The aircraft was connected to the CORS network, specifically to the antenna located approximately 7 km from the study area. Before each flight, the aircraft obtained a fixed coordinate, which, in all flights, showed an accuracy of approximately 3 cm in the coordinates corrected using NTRIP on the three axes (X, Y, Z). The wind speed recorded during the flight was 18 km/h in the east direction. A total of 5076 images were captured and processed using Metashape/Agisoft software with high quality and aggressive filtering. The dense point cloud consists of 1,392,820,566 points, the DEM is 26,897 × 49,581 pixels with a size of 5.39 cm/pixel, and the orthomosaic is 53,794 × 99,162 pixels with 2.70 cm/pixel. The overlaps established during the flight yielded good results, as there were no gaps in stereoscopic coverage, and the orthomosaic has no gaps in information.

3. Results

The correlation coefficient of the RTK coordinates of the other points obtained with the Sokkia instrument is 0.716, while with the Tersus instrument it is 0.760. Therefore, the coordinates obtained with the Tersus instrument will be used for verification. Checkpoint number 12 corresponds to a plate from the National Passive Geodetic Network located near point number 3. Since it was not marked, it is not visible on the orthomosaic, so a comparison cannot be made. It was included only for control purposes, in case considerable variations are obtained with both instruments. Points 10 and 11 were obstructed by metal structures on the day the images were captured; therefore, they will also not be taken into account, as they are not visible in the orthomosaic.

Figure 6 shows the checkpoints identified in the orthomosaic obtained from the photogrammetric process to extract the coordinates of each one, which are shown in

Table A4. Coordinates of the checkpoints obtained in the orthomosaic compared with those recorded in the GNSS.

	XOrtomosaico	YOrtomosaico	ZOrtomosaico	XGNSS	YGNSS	ZGNSS
1	749,445.983	2,521,734.850	2470.813	749,445.925	2,521,735.907	2471.723
2	749,367.968	2,521,474.679	2426.113	749,368.075	2,521,475.692	2425.417
3	749,352.508	2,521,214.251	2416.125	749,352.516	2,521,215.008	2414.847
4	749,138.681	2,520,786.128	2419.667	749,138.824	2,520,786.051	2417.841
5	749,207.524	2,520,468.886	2397.579	749,207.522	2,520,468.474	2395.621
6	748,831.340	2,520,435.397	2409.752	748,831.702	2,520,435.372	2408.588
7	748,801.907	2,520,128.046	2425.240	748,802.054	2,520,127.663	2424.461
8	749,320.505	2,520,039.483	2389.825	749,320.058	2,520,038.308	2390.083
9	748,698.295	2,519,700.800	2429.327	748,698.442	2,519,700.034	2430.470

.

It should be noted that the heights from both the orthomosaic and the GNSS are ellipsoidal heights. The study area shows an average undulation between the ellipsoidal and orthogonal heights of 14.52 m. Although the heights measured with the GNSS were converted, given that the orthomosaic provides ellipsoidal heights and that the variation between the orthogonal and ellipsoidal heights ranges from 14.51 to 14.53 m in the study area, it was decided to use ellipsoidal heights. Using the data from

Table 2. Difference between the coordinates obtained in the orthomosaic and the measurements with the GNSS.

Point	dX	dY	dZ
1	0.058	−1.057	−0.910
2	−0.106	−1.013	0.696
3	−0.007	−0.757	1.278
4	−0.143	0.076	1.826
5	0.001	0.412	1.958
6	−0.362	0.026	1.164
7	−0.148	0.383	0.779
8	0.446	1.175	−0.258
9	−0.148	0.766	−1.143

, the RMSE between the orthomosaic and the GNSSS equipment was calculated. The values measured in the orthomosaic differ considerably, showing an RMSE of 0.213 m on the X-axis, 0.746 m on the Y-axis, and 1.222 m on the Z-axis. The horizontal RMSE (dX, dY) is 0.776 m, while the 3D RMSE which summarizes the total spatial error between the two datasets is 1.448 m.

Table 3. Statistical analysis.

Variable	Bias (m)	SD (m)	IC 95% − (m)	IC 95% + (m)	Correlation
dX	−0.045	0.221	−0.190	0.099	0.594
dY	0.001	0.791	−0.516	0.518	−0.943
dZ	0.599	1.130	−0.139	1.337	−0.537

shows the statistical analysis of the differences between the orthomosaic coordinates and those obtained with the Tersus GNSS. Position error analysis indicates marked anisotropy between the components. The dX-axis shows low bias (−0.045 m) and low dispersion (SD = 0.221 m), with a 95% confidence interval including zero (−0.190–0.099 m), indicating a stable and statistically insignificant systematic error, although a moderate correlation (r = 0.594) suggests some structured residual behavior. In contrast, dY exhibits insignificant bias (0.001 m) but substantially greater variability (SD = 0.791 m), with a wide confidence interval (−0.516–0.518 m) and a strong negative correlation (r = −0.943), indicating dominant systematic effects likely related to acquisition geometry or GNSS distortions. The dZ component exhibits the worst performance, with the greatest bias (0.599 m) and dispersion (SD = 1.130 m), and a wide confidence interval (−0.139–1.337 m), confirming low vertical reliability; its moderate negative correlation (r = −0.537) further supports the structured error patterns.

Figure 7 shows the horizontal error (eXY) at each point. Green circles indicate points with an error less than 0.5 m, corresponding to checkpoints 4, 5, 6, and 7. Yellow circles indicate points between 0.5 and 1 m, specifically checkpoints 3 and 9, while checkpoints with an error greater than 1 m are 1, 2, and 8, marked with red circles. The checkpoints with the best horizontal accuracy are located in the center of the model. The horizontal error increases towards the edges of the photogrammetric model, on the surface with high elevation where overlap is less effective, resulting in less ray redundancy and angular intersection due to the direction in which the flight was conducted.

Figure 8 shows the vertical error (eZ) at each checkpoint. Green circles indicate points with an error less than 0.5 m, corresponding to checkpoint 8. Yellow circles indicate points between 0.5 and 1 m, specifically checkpoints 1, 2 and 7, while checkpoints with an error greater than 1 m are 3, 4, 5, 6 and 9, marked with red circles. The vertical error (eZ) is dominated by the complexity of the terrain, does not show a defined radial or geometric pattern, depends on the slope and variability is greater in transition zones. The most accurate value is found in the lowest elevation zone.

Figure 9 shows the distance from takeoff from each checkpoint. The near checkpoints are 4 and 5, while control points 2, 3, 6, 7, and 8 are far checkpoints. The farthest checkpoints are 1 and 9. The takeoff point is positioned in the center of the study area to provide coverage of the entire zone.

4. Discussion

The coordinate comparison with two GNSS devices validates the data at eight of the nine checkpoints. At checkpoint number 1, the horizontal distance between the point measured by the Tersus and Sokkia devices is 0.9193 m, and the vertical distance is 1.7143 m. Correcting the image position by connecting to the Cord network provides better results in horizontal coordinates, primarily along the X-axis. The checkpoints with the smallest residuals between the coordinates obtained with the GNSS and those on the orthomosaic correspond to the central checkpoints, close to the aircraft’s takeoff point but also to the area where the topography is least variable. The checkpoints with the greatest difference correspond to the highest checkpoint (1) and the lowest checkpoint (8), with a difference in elevation between them of 81 m. Checkpoint 2 also has large residuals and is located within the slopes exceeding 15%, according to Figure 4. The error distribution shows a clear dominance of vertical discrepancies, with RMSE values exceeding 1 m, while horizontal errors remain relatively lower but exhibit spatial variability. The Z-axis has a significant positive bias, indicating that the orthomosaic is systematically lower than the GNSS by 0.60 m. The bias in the X and Y-axes is small (little systematic trend), indicating greater accuracy. The confidence interval is wider in the Z-axis, showing greater variability. The coordinate values obtained from the orthomosaic, compared to those obtained with the GNSS, show good accuracy (bias) in the X and Y-axes, while in the Z-axis, the orthomosaic exhibits a systematic vertical bias of +0.599 m that must be corrected. Regarding accuracy (RMSE), the X-axis has better accuracy (0.21 m) and the Z-axis is less accurate (1.22 m). In the 95% confidence intervals (CIs) for the three axes, all ranges from negative to positive values, therefore, all include zero. It cannot be statistically concluded that there is a systematic bias between the compared methods. Although the differences show some variability, the true bias could be zero, positive, or negative. On the X-axis, the confidence interval (IC) (−0.190 to 0.099) is the narrowest, therefore more reliable. On the dZ-axis, it is very wide, exhibiting greater variability and lower precision. The correlation on the dX-axis (0.594) is moderately positive, indicating a size-dependent bias (proportional error), but not severe. On the dY-axis (−0.943), the correlation is very strong and negative, which is interpreted as exhibiting very clear systematic behavior. On the dZ-axis (−0.537), there is a moderate negative correlation, indicating a dependence on the value, but not as pronounced.

Overall, the results demonstrate direction-dependent horizontal accuracy and significantly degraded vertical performance, with predominantly systematic rather than random error behavior, highlighting the limitations of positioning based solely on CORS/NTRIP corrections in UAV photogrammetry and the need for GCP to improve vertical accuracy. The results obtained differ from those obtained by [20], who obtained accurate results using the CORS network, possibly due to the uniformity of the terrain (around 1 m), the size of the studied area (16 ha), the signal quality, and the fact that both the aircraft and the GNSS received correction from the same network. Our study area is 10 times larger, is completely urbanized, and has complex terrain (the elevation across the entire area differs by almost 100 m). The aircraft received CORS/NTRIP correction, and the GNSS from the RGNA. For large areas, the use of sufficient and well-distributed GCPs, spaced no more than 100 m apart and arranged in a zigzag pattern, is still necessary to guarantee topographic-quality accuracy, as suggested by [36]. Study [37] “Development of GNSS Receiver for Mobile CORS with RTK Correction Services Using Cloud Server,” determined that the ideal range for CORS stations should be less than 50 km and that to obtain accuracies between 2 and 5 cm, the maximum distance between stations is 20 km. Therefore, despite using the antenna located in Guadalupe, which is approximately 7 km from the area under study, this separation is within the recommended range. Studies of [7] indicate that the CORS network transmits correction data in real time via a cellular network. Low signal coverage can compromise the availability and reliability of real-time corrections due to latency or delay in receiving the correction positioning data, thus reducing the accuracy of direct georeferencing. The observed pattern of horizontal error suggests a clear dependence on the geometry of the photogrammetric block. The central region of the study area benefits from higher image overlap and stronger ray intersection geometry, resulting in improved planimetric accuracy. Conversely, points located toward the edges of the block show increased errors, which can be attributed to reduced redundancy and weaker intersection angles. Unlike horizontal error, vertical accuracy is strongly influenced by terrain characteristics. The integration of the Digital Surface Model (DSM) with the spatial distribution of eZ highlights a clear relationship between elevation variability and error magnitude. Areas characterized by abrupt topographic transitions, particularly in the central–southern region, correspond to the highest vertical errors. These zones exhibit complex surface geometry, including rapid elevation changes and heterogeneous textures, which complicate image matching and reduce the reliability of elevation estimates. In contrast, areas with more homogeneous topography, such as the southeastern sector (checkpoint 8), show significantly improved vertical accuracy despite being located at greater distances from the takeoff point. This finding confirms that terrain variability, rather than absolute elevation or distance, is the dominant factor controlling vertical error in UAV photogrammetry. The comparison between error distribution and distance from the takeoff point shows no consistent relationship, particularly for the vertical component. While horizontal error exhibits a mild increase toward the edges of the survey, this trend is not sufficient to establish distance as a primary controlling factor.

5. Conclusions

The coordinates measured directly on the orthomosaic are compared with those obtained using GNSS equipment. The difference between the two measurements is related to variations in the topography. The drone flight maintains a constant altitude of 85 m above the ground, thus generating a variation in scale relative to the variation in ground level. This limitation is evident in the elevated vertical errors observed in the central zone with significant topographic variation. Increasing flight line diversity, such as incorporating cross-flight patterns or oblique imagery, could improve vertical accuracy by enhancing ray intersection geometry. The flight is at a constant altitude above the ground, thus eliminating the direct influence of the data on topographic variation. Point number 1 has the highest elevation (2471.98 m) and the least overlap between the images, which explains the variation with the orthomosaic coordinates. The coordinates obtained with GNSS are considered more accurate. For reconnaissance work and where precision is not the most important factor, working with this technology proves effective due to the reduction in fieldwork time, material and human resources, and image processing. Otherwise, for greater certainty in the accuracy of information obtained with photogrammetric products, the use of GCPs remains a necessary tool. A reduction in accuracy due to delay and latency in receiving positioning correction is not ruled out, but it is considered unlikely given that the study area corresponds to the urban area of the study zone, where there is sufficient cellular network coverage, and also because the analysis does not show a relationship between accuracy and takeoff distance to each of the points. To use the NTRIP correction network, mobile data connectivity (4G/5G) and a subscription are required. Therefore, flight planning must be performed very carefully, considering weather conditions, provider response time, and the subscription period. The study evaluated the positional accuracy of the Mavic 3T aircraft using real-time corrections via the CORS/NTRIP network, comparing it with coordinates obtained using GNSS at nine checkpoints. The results showed that accuracy degrades in large urban area with rugged topography and full 4G coverage, but the results are useful for projects where elevation precision is not required. Further studies are recommended to rule out the influence of atmospheric or climatological conditions on the results. The results demonstrate a clear differentiation between the factors controlling horizontal and vertical accuracy. While horizontal error is primarily influenced by the photogrammetric block geometry, showing increased values toward the edges of the survey, vertical error is strongly controlled by terrain variability, with the highest discrepancies observed in areas of abrupt elevation change. The longitudinal flight configuration, although effective for planimetric reconstruction, introduces limitations in vertical accuracy due to reduced angular diversity. These findings confirm that UAV photogrammetric accuracy is governed by a complex interaction between acquisition geometry and topographic conditions. For future work, cross flights with oblique images are planned to map completely urbanized steep areas, in order to rule out discrepancies generated by the direction of flight, and also to monitor weather conditions to analyze their influence on the accuracy of the results.