Effect of photogrammetric RPAS flight parameters on plani-altimetric accuracy of DTM

2.1 Study area

The study area is located in the city of Culiacan, Sinaloa, Mexico. Specifically, in the lower part of Culiacan river basin, in coordinates 24°48′40.3″ – 24°48′69.2″ N, 107°24′ 53.1″ – 1 07°24′24.5′ W according to World Geodetic System 84 (WGS84) reference system. High-intensity rainfalls and tropical storms occur in the study area during rainy season [24]. In particular, the lower part of Culiacan river basin was selected as the study area (Figure 1). This zone is a sloping terrain characterized by constant flooding caused by the rise of the Tamazula River. These river floods have caused severe damage to the urban area. Therefore, high detail DEM is needed for future hydrological modeling in this area.

Figure 1

Study area. Source: Own elaboration using QGIS 3.10 La coruña (Open Source Geospatial Foundation, Beaverton, USA, 2020).

2.2 DEM construction process

The methodology used in this study is presented in Figure 2. DEM construction process was characterized by three main phases: the first phase consisted of photogrammetric flight planning project, where the flight parameters of RPAS were established and the photogrammetric surveys were carried out. Data processing was the second phase and consisted of defining a photogrammetric block and some information about a scene during image capturing process was obtained (i.e., camera orientation and image point correlation). At this stage, the 3D reconstruction was also carried out through the extraction process (point cloud extraction). Then, DEM generation was carried out by using different mathematical algorithms with SfM technology (SIFT, Hamming Distance, Poisson reconstruction, SemiGlobal Matching (SMG) algorithms and Delaunay triangulation). Finally, the third phase involved the DEM accuracy assessment using root mean square error (RMSE). Then, statistical analysis (analysis of variance [ANOVA]) was carried out to evaluate the effect of photogrammetric flight parameters on DEM accuracy.

Figure 2

DEM process generation diagram. Source: Own elaboration.

2.3 First phase: photogrammetric project definition

Step 1.1 Project Definition. A Quad Copter DJI Phantom 4 Pro was used for image acquisition. This RPAS has a global positioning system (GPS)/GLONASS and a sensor mounted on a gimbal stabilizer. A Sony RX100 camera with 35 mm focal length CMOS sensor of 20.2 effective megapixels was used. The strategies used in a conventional photogrammetric flight planning were implemented in the RPAS flight. Photogrammetric project began defining the area of interest: a polygon was drawn on a geo-referenced base map, camera and sensors onboard were calibrated (tested).

Step 1.2 Flight Planning. Free areas were located as landing points of the RPAS in case of an emergency. Then, the photogrammetric parameters were defined. The RPAS parameters considered in this study were flight mode, flight altitude, flight speed, camera tilt and the longitudinal and transversal overlaps of the images. An orthogonal array L18 was used to determine the different flight parameters for image acquisition (Table 1). This orthogonal array was used to measure the effect of these six photogrammetric flight parameters on the accuracy of DEMs generated.

According to Zhao et al. [20], a double grid flight is recommended to obtain a more accurate point cloud. In this sense, the flights were made in single and double grid. Flight altitudes were carried out at 60, 70 y 80 m as indicated by Hussain and Bethel [25]. In this study, longitudinal overlaps between 40% and 60% were defined as suggested by Reshetyuk and Mårtensson [26], while lateral overlaps between 60% and 80% were established as suggested by Harwin and Lucieer [16] and Dall’Asta et al. [27]. Under these conditions, high quality topographic maps have been reported. Regarding the camera tilt, positions close to nadir (90°) were chosen as suggested by Wasklewicz et al. [28] and Bemis et al. [29]. Oblique angular positions between 65° and 85° were also chosen. Concerning the RPAS, speed values of 3, 6 and 9 m/s were chosen to analyze distortions generated due to camera instability at high velocities.

According to Mora-Felix [30], image resolution is related to GSD. In this study, images resolution values were considered between 2 and 3 cm/pix. These values correspond to the resolutions of the most widely lightweight cameras installed on RPAS [31]. Once the photogrammetric parameters were stablished, mathematical models proposed by Vollgger and Cruden [32] and Mora-Felix et al. [30] were implemented to compute RPAS flight trajectories and programming camera shots, which are the parameters required to execute the real-time flight.

Step 1.3 Control points establishment. GNSS Trimble R10 was used for setting control points on Earth surface. This is high precision equipment that works with satellite system GLONASS, SBAS, Galileo and GPS. Each checkpoint was marked with physical signals on the ground. These marks were established before the surveys to be identified by the RPAS. Crosses in orange color were marked on the ground because they contrast with the background and are observed perfectly in the images. Control points were distributed throughout the study area as 20 × 20 cm marks to improve the triangulation and the DEM accuracy [33]. The marks were located on physical objects, such as drains, crosswalks and wood fixed to the surface by stakes were used in zones where no asphalt or concrete was present.

Step 1.4 Flight execution. During the processing of photogrammetric parameters, a file was generated containing the flight instructions that the RPAS must execute (waypoint file). This file contained GPS coordinates of a group of points that represent the photogrammetric flight to be followed sequentially by RPAS. This file also contained flight speed, flight altitude and shooting frequency configuration. Waypoint file was loaded and then the photogrammetric flight started. Through the navigation system, RPAS autonomously took position at the starting point and then automatically captured the images according to the sequence previously established. RPAS flight was monitored from a central station and furthermore an image dataset was generated.

2.4 Second phase: dataset processing

Image dataset processing was performed by SfM [34]. Image processing was characterized by two main sub-phases: first, a photogrammetric block configuration was performed. At this sub-phase, image dataset is calibrated. Then, detection and feature correlation processes were carried out. Atypical data was eliminated and image orientation was carried out through aerial triangulation. Points relating images to each other were identified and finally the block adjustment was done.

The second sub-phase consisted of the 3D reconstruction. At this stage, disperse point cloud was generated and a densification process was performed on the point cloud. Subsequently a georeferencing process was performed and finally a triangular mesh (triangulation Delaunay) was used for the DEM generation. The dataset processing phase is fully described below:

2.5 Subphase one: photogrammetric block configuration

Step 2.1 Camera calibration. Before the flight execution, a camera calibration process was carried out. This calibration process is a fundamental requirement for the metrical reconstruction of images in photogrammetry [35]. In addition, this process reduces any alteration or noise in the cartographic product. In this study, automatic calibration was implemented as suggested by Fraser [36]. During camera calibration process, extrinsic and intrinsic parameters were obtained, which are related to the external and internal orientation processes of DEM. Extrinsic parameters refer to the coordinates in a reference system and orientation. Intrinsic parameters refer to the focal length, the central point of the image and distortions of the camera lenses. Noise was reduced by adjusting the tangential (p₁, p₂) and radial (k₁, k₂, k₃,) distortion coefficients based on Brown’s distortion model [37].

Step 2.2 Feature extraction and correlation. In a photogrammetric survey with RPAS, the dataset of a photogrammetric block corresponds to the images captured during flight survey and the parameters that correlate them (overlap) with the recorded scene (position and orientation). The photogrammetric block configuration process was carried out by implementing detection and feature correlation algorithms. The most used detector algorithm in SfM artificial vision systems is the scale invariant feature transform (SIFT) algorithm [38]. This algorithm was used to extract image features. SIFT algorithm generated a set of key points between stereoscopic pairs of images that were invariant (robust) to image scale, translation, rotation and partially invariant to changes in camera illumination. The set of key points generated from images were then processed through Hamming Distance correlation algorithm [34].

Step 2.3 Image orientation. An image orientation process was applied to determine external orientation parameters (position coordinates and tilt angles) for each image. The aerial triangulation was implemented for this process. The exterior orientation elements include camera position on the platform in the coordinates x, y, z and the three angles of camera tilt (ω, φ, κ). These coordinates and angles are relative to the ground coordinate system. The positions of coordinates x, y, z were obtained by a GPS system on board. Camera angles were measured by using the INS (Inertial Navigation System) and IMU (Inertial Measurement Unit) devices. For absolute orientation, tie points (homologous points) between images were used. In addition, an image alignment process was done by using four ground control points (GCPs) defined at the corners of the model to ensure a more accurate result and eliminate the bowling effects of the RPAS data [39]. Extrinsic and intrinsic camera parameters were adjusted to perform the block adjustment procedure. Finally, Levenberg–Marquardt (LM) algorithm was used to minimize the re-projection error. This is a robust algorithm based on the non-linear least squares method. LM minimizes the re-projection error between the observed and predicted image points [40].

2.6 Subphase two: 3D reconstruction

The 3D reconstruction consisted of four steps clearly identified: scattered point cloud generation, point cloud densification, georeferencing and DEM generation. These stages are described as following.

Step 2.4 3D point cloud generation. The coordinates of the extracted points were determined and a disparity map was computed by calculating the stereoscopic correspondence of each object. Based on this map, the distance between ground objects and the camera was obtained. The depth z of a point P(x, y, z) in a 3D reconstruction was obtained with equations (1)–(3), considering a triangular geometrical relationship.

where b is the value of baseline, f is the focal length, I_l, l_r are the left and right images respectively, d is the distance from the camera to a point on the ground. Finally, the resulting data is a dispersed point cloud referred to a local coordinate system Universal Transverse Mercator, WGS84 zone 13. For this procedure, Helmert transformation was implemented [41].

Step 2.5 Point cloud densification. Consisted of extracting a larger amount of 3D points from the scene that complements the disperse point cloud generated in the previous stage. SMG was implemented due to its high precision and computational cost reduction [42]. Once the dense map was available, atypical points were eliminated (error points). Then, the Delaunay triangulation algorithm was implemented for the generation of the triangular irregular network. In this algorithm, the elevation values were interpolated (triangulated) for the generation of a regular matrix in raster format (triangular mesh) [43]. The created mesh is textured by re-projecting the 3D points with the original images to obtain the corresponding color.

Step 2.6 Georeferencing process. Before the DTM is generated, a georeferencing process is suggested. This process consisted of linking identifiable locations in the images with the locations of GCPs that were spatially related. Identifiable locations, such as sidewalks, corners, culverts, were used for georeferencing. Then, polynomial transformation was generated to adjust the locations to its correct spatial location [44].

Step 2.7 DEM generation. After constructing the triangular mesh, a smoothing model was constructed to fill the gaps in the surface and repair the errors generated. Poisson algorithm was implemented for this surface reconstruction [45]. This surface reconstruction is represented in a DSM. Finally, a Digital Terrain Model (DTM) was generated with the algorithm proposed by Unger et al. [46]. This algorithm removed man-made objects (i.e., buildings) and vegetation. Afterwards, a DTM was regularized through Huber norm.

2.7 Third phase: DEM validation

Step 3.1 RMSE-validation. DEM accuracy refers to the proximity of the observation of a coordinate point to a true value. Plani-altimetric error must be minimized to obtain a more accurate model [47]. According to National Standard for Spatial Data Accuracy, DEM errors meet a normal distribution and this is a fair assumption for open terrains [48]. In this study, DEM validation consisted of a comparison between plani-altimetric features with the GNSS points acquired in the study area. Planimetric and altimetric errors were calculated for the 18 DTMs generated. The most common statistic used to describe planimetric and altimetric errors in DEMs is the RMSE [49]. The RSME measures the difference between the values predicted by the DEM and the values observed (GNSS points).

Altimetric accuracy (RMSE_z) of DTM was carried out by using equation (4). This statistical parameter consists of a comparison between the elevation points obtained with GNSS and those marked on ground.

In this equation, n refers to the total points measured, Z_DTM corresponds to the elevation of the control points marked on the ground visible in the image, Z_CP corresponds to the elevation of the GCP using GNSS. Planimetric validation (RMSE_x,y) was determined by the distance between the GCPs and the marked points (CPs) on the ground for both x and y coordinates. This validation was carried out using equation (5).

where n refers to the total points to be measured, X_DTM, Y_DTM are the coordinates of the points in the model and X_CPY_CP are the coordinates of the validations GNSS points.

Step 3.2 Statistical analysis. An ANOVA was carried out to measure the effect of RPAS flight parameters on DTM accuracy (α = 0.05). The planimetric (RMSE_xy) and the altimetric (RMSE_z) errors were used as response variables for this statistical analysis. ANOVA identified the flight parameters that showed a significant influence on the response variables. An optimization strategy was then carried out to find out the optimum flight configuration and to obtain the DTM with the best plani-altimetric accuracy.

3.1 Camera calibration

Servomotors service and camera calibration were carried out to achieve the best survey performance and to avoid mechanical failures. A self-calibration process was done according to the suggested by Fraser [36]. The camera calibration parameters used in this study are presented in Table 2. The main objective of the calibration process is to solve the radial (k₁, k₂, k₃) and tangential (p₁, p₂) distortion parameters to reduce uncertainties or errors of the principal points x and y. This calibration is needed to obtain the most accurate results during the key point detection and correlation processes in stereoscopic pairs. Based on the errors obtained in this study, the calibration process could be considered as satisfactory for the 3D reconstruction process.

3.2 Ground control points and checkpoints

In this study, 51 points were established, from which 27 points (GCPs) were used for georeferencing and 24 points (checkpoints, CPs) were used for validation. Figure 3 shows the GCPs and CPs distributed in the study area.

Figure 3

Location of GCPs in the study area. Source: Own elaboration using QGIS 3.10 La coruña (Open Source Geospatial Foundation, Beaverton, USA, 2020).

3.3 Photogrammetric project definition (execution)

Table 3 shows the GSD, frames, areas covered and flight lines obtained in the photogrammetric surveys (treatments) that were conducted for image acquisition. Each treatment generated an average of 328 images, with a mean GSD of 2.53 cm/pix. Approximately 6,000 photographs were taken and 6.3 km were covered. It is noteworthy that a greater number of flight lines were registered in the last nine treatments. These surveys were carried out under double grid flight mode. Likewise, it was demonstrated that the flight lines depended on the overlap of images and flight mode. It is also noteworthy that the quantity of frames acquired under normal grid flight mode is significantly smaller than the ones acquired when the double grid flights were performed.

3.4 Image processing

Figure 4 shows the results of keypoint extraction process of treatment 13 by overlapping a pair of images. In this figure, it is evidenced that most of the keypoints were correctly matched, especially in areas of dense vegetation where erroneous coincidences are possible due to the presence of similar local texture.

Figure 4

Keypoint detection, feature correlations and stereo pair assembling during image processing with SIFT algorithm. Source: Own elaboration using SIFT algorithm in Python 3.6.5 (Python Software Foundation, Wilmington, Delaware, USA, 2011).

During the 3D reconstruction, the algorithm identified a mean of 37,336 keypoints per image (Table 4). Despite the keypoints per image were similar in all treatments, the number of match points found in double grid treatments was greater than the observed during normal grid. This situation could explain the fact that the number of keypoints obtained in the 3D densification process was greater when double grid flight mode was carried out.

3.5 Planimetric (RMSE_xy) and altimetric (RMSE_z) accuracies

The real coordinates (x, y, z) were obtained with GNSS system. These points were compared with the CP marked on the ground and observed in the digital model.

Figure 5 shows a map where planimetric and altimetric differences are observed in the generated DTM. Figure 5a shows a cross-section with a distance of 117 m, where a slope is observed. In this figure, three validation points are present: CP17, CP16 and CP14. The elevation differences observed between the DTM and GNSS measurements were less than 0.15 m. Figure 5b shows the horizontal errors in CP16. In this figure, the horizontal difference between the DTM and GNSS measurements was less than 0.02 m. Due to the differences found, equations (4) and (5) were used to obtain the planimetric (RMSE_xy) and altimetric (RMSE_z) accuracies for each photogrammetric survey.

Figure 5

Planimetric and altimetric differences in the generated DTM. (a) Elevation differences observed in a cross-section. (b) Planimetric difference identified in GCP 16. Source: Own elaboration using QGIS 3.10 La coruña (Open Source Geospatial Foundation, Beaverton, USA, 2020).

3.6 Statistical analysis

A georeferencing process was implemented on the DTMs obtained. 27 GCP were used as georeferencing points and the accuracy of georeferenced DTMs was then evaluated using the 24 points that were established as CP. Table 5 shows the results of the plani-altimetric accuracy obtained in the eighteen surveys before and after the georeferencing process. The planimetric accuracy (RMSE_xy) before georeferencing process ranged from 9.01 to 1.46 m. This high variation was reduced after implementing the georeferencing process. RMSE_xy after georeferencing ranged from 1.05 to 0.11 m. The same situation occurred for altimetric accuracy (RMSE_z). Lower values of RMSE_z (higher altimetric accuracies) were registered after georeferencing. According to statistical analysis, when the georeferencing process was performed, a significant reduction in error was observed in both planimetric and altimetric accuracies (p < 0.05).

ANOVA was carried out to assess the effect of photogrammetric flight parameters on plani-altimetric accuracy of DTMs (Table 6). Based on statistical analysis performed, longitudinal overlap and flight mode were flight parameters that showed a significant influence on planimetric accuracy. According to main effect analysis, lower RMSE_xy values were registered when higher percentages of longitudinal overlap are used. Likewise, higher planimetric accuracies were observed when double grid flight mode was carried out.

Statistical analysis for altimetric accuracy showed that only the flight mode had a statistical significance (p < 0.05). Lower RMSE_z values were observed in DTMs generated when double grid surveys were performed.

3.7 Optimization (the most accurate DTM generated)

According to statistical analysis, treatment 13 showed the highest plani-altimetric accuracy. Under these flight operational conditions, values of 0.38 and 0.11 m were achieved for RMSE_xy and RMSE_z. In both cases, the digital models were adjusted by the georeferencing process using the GCP. In contrast, treatment 4 presented the lowest accuracy (RMSE_xy = 1.05 m and RMSE_z = 1.33 m). The resulting DTM shows neither artifacts nor unnatural terrain features, and it represents the real morphology of the terrain.

4.1 Effect of the image overlapping

Figure 6 shows the amount of images that were overlapped in treatments 13 and 4. Treatment 4 generated with normal grid flight mode (Figure 6b) has a greater amount of red and yellow areas than the one generated with double grid (Figure 6a). The red and yellow zones are located mostly on the margins and these zones are characterized by the overlapping of one to three images. The green zones reflect that five or more images were overlapped. Therefore, the DEM accuracy could be highly related to the amount of images overlapped, as it is suggested by Dandois et al. [50].

Figure 6

Images overlapped in DEMs generated in treatments 13 and 4. Source: Own elaboration using Pix4D (Pix4D SA, Laussane, Switzerland, 2011). (a) Overlapped images in treatment 13, (b) overlapped images in treatment 4.

According to Miřijovský and Langhammer [33], the greater number of intersected photographs, the greater the precision in the 3D reconstruction process is. This situation is demonstrated in Figure 7, where lower RMSE_z values were obtained when higher key points were identified (Figure 7a). Besides, higher altimetric accuracy was obtained when higher match points were correlated (Figure 7b). Altimetric accuracy was also correlated with the amount of frames. Treatments with a greater number of photographs (double grid) observed higher altimetric accuracies (Figure 7c). Therefore, the altimetric accuracy of a DEM depended on the densification points obtained during 3D reconstruction process, as it is also demonstrated in Zhao et al. [20].

Figure 7

Relationship between altimetric accuracy and 3D reconstruction process. (a) Correlation between altimetric accuracy and keypoints identified. (b) Correlation between altimetric accuracy and matched points. (c) Correlation between altimetric accuracy and frames acquired. Source: Own elaboration using Statgraphics Centurion 18 (Statgraphics Inc., The Plains, USA, 2020).

In this study, high accuracy was achieved because images were acquired from different perspectives. However, it is important to determine the optimum overlap to achieve the desired accuracy without falling into long processing computation times. Altimetric precision must be optimized by balancing the number of intersected photographs. The stereoscopic multivision effect must be achieved taking into account the computational processing times. The greater the points correlate in a scattered point cloud, the denser the point cloud is generated with better detail and therefore greater accuracy.

4.2 Effect of the georeferencing process

The effect of the georeferencing process on the plani-altimetric accuracy of the DTM is shown in Figure 8. Plani-altimetric accuracy is strongly dependent on this georeferencing process. The Box and Whisker plot shows the planimetric (Figure 8a) and altimetric (Figure 8b) accuracies before and after georeferecing process. In general, a high error is observed in the DTMs generated before georeferencing. According to the results obtained in this study, georeferencing is a critical process for improving the accuracy of DTMs. Therefore, technological alternatives must be developed for the georeferencing process. The complexity of terrestrial areas and the extensive time consuming for the establishment of GCPs on the ground represent a major opportunity area in RPAS photogrammetry. The georeferencing process is being integrated as an indirect system on the RPAS [51]. In the future, this situation will replace the use of GCPs.

Figure 8

Box and Whisker plots showing the importance of georeferencing process on the DTM accuracy. (a) The effect of georeferencing onplanimetric accuracy. (b) The effect of georeferencing on altimetric accuracy. Source: Own elaboration using Statgraphics Centurion 18 (Statgraphics Inc., The Plains, USA, 2020).

Luhmann et al. [52] suggest the use of at least 20 GCPs to achieve high precision in DEMs. In this study, 27 GCPs were considered for georeferencing. Few studies are given about the effect of the georeferencing process in DEM accuracy. More in-depth studies are needed to determine the optimal amount of GCPs that must be used for the generation of DEMs with high accuracy.

4.3 Effect of the photogrammetric flight parameters

Planimetric accuracy results varied from 0.11 to 1.05 m. The planimetric accuracies obtained are similar to the results obtained in other photogrammetric studies. Seifert et al. [53] reported a planimetric accuracy range of 0.10 to 0.15 m, while Matthews [54] reported a planimetric accuracy of 0.08 to 0.40 m. These high planimetric accuracy results were obtained using double grid flight. This coincides with the result reported by Zhao et al. [20], who found that double and triple overlapping concentrate a large number of intersected pixels, which increased the planimetric accuracy of DEM. The stereoscopic multivision and 360° perspective given by the multiple intersections of photographs increased DEM accuracy. SfM algorithms showed a better result when high redundancy block configuration was performed. A higher correlation of overlapping characteristics in multiple photographs was observed under these conditions.

Rabah et al. [51] suggest implementing a high longitudinal overlap, but also a high vertical overlap to obtain high DEM accuracy. In particular, Westoby et al. [55] and Eisenbeiss [56] report that high percentages of longitudinal and transversal overlaps result in low RMSE. However, the results obtained in this work indicated that the transversal overlap did not have a significant effect on planimetric accuracy (p > 0.05).

Flight altitudes of 30, 40, 50, 60, 70 and 80 m were set for the production of orthomosaics for archaeological applications in the study by Mesas-Carrascosa et al. [39] and they concluded that flight altitude is the main flight parameter because of its influence on RMSE_xy. However, in this study, low flight altitudes (<60 m) were discarded because this implied the processing of a high amount of images and consequently, high computation times. Flight altitudes of 60, 70 and 80 m were established for photogrammetric surveys and no significant effect was found on RMSE_xy (p > 0.05). Mesas-Carrascosa et al. [39] also suggested that camera angle close to nadir (90°) is required to obtain high planimetric accuracy. This study demonstrated that the camera tilt was not a significant flight parameter for planimetric accuracy (p > 0.05).

Some studies have suggested that certain configurations are determinant for altimetric accuracy. Henriques et al. [57] suggested that transversal overlap has a significant influence on the density of the point cloud created for DEM generation. However, the results obtained in this research indicate that transversal overlap was not statistically significant on altimetric accuracy (p > 0.05).

Rabah et al. [51] report that the nadir view (90°) generates many occlusions in the image data set. According to Dandois et al. [50], steep hills, natural or artificial constructions, such as walls or buildings, generate hidden areas in the images. Both studies suggest that this effect could be reduced with the camera tilt or with a greater GSD. In this study, the statistical analysis (ANOVA) demonstrated that the camera tilt did not have any significant effect on the DTM accuracy (p > 0.05). However, in our study, a greater DTM accuracy was observed when surveys were carried out with great obliquity (65°).

Likewise, Torres-Sánchez et al. [58] indicate that high flight altitudes resulted in deterioration of DEM accuracy. In this study, altimetric accuracy was not influenced by the flight altitude (p > 0.05). According to statistical analysis, double grid flight was the unique determinant flight parameter for altimetric accuracy. The altimetric accuracy obtained in this study could be considered as high, according to the Udin and Ahmad’s report [59], who reported altimetric accuracies ranging from 0.045 to 0.52 m for DEM generation.

4.4 Field validation

The performance of the optimum flight parameters was evaluated in different topographical conditions. Culiacan Institute of Technology was chosen since it is a terrain with steeper slopes (more than 50%). Figure 9 shows a distributed regular network of 20 GCPs, of which 12 GCPs were designated for georeferencing (indicated with asterisk in Table 7) and 8 GCPs designated for DTM validation. These points were marked on the ground to be easily distinguished from the background during the processing of the aerial images. The entire methodology proposed in this study was used for image acquisition, data processing, and DTM validation.

Figure 9

GCPs and CPs established for DEM validation at Culiacan Institute of Technology. Source: Own elaboration using QGIS 3.10 La coruña (Open Source Geospatial Foundation, Beaverton, USA, 2020).

According to Table 7, the RMSE_z and RMSE_xy obtained were 0.37 and 0.05 m, respectively. These results were very similar to those obtained for DTM generated for Tamazula river. Figure 10 shows the DTM generated for Culiacan Institute of Technology. In analytical and modern photogrammetry performed with manned aircraft, the elevations in a study area should not vary 5% with respect to the flight altitude [60]. According to this criterion, the scale differences (GSD) in the images must not affect the accuracy of the final digital model. In the field validation, DTM elevation ranged from 54 to –102 m. In this sense, the methodology proposed in our study was validated and can be used in different zones with different topography, in particular in areas with steeper slopes. However, the methodology proposed must also be evaluated in areas with high vegetation density [61]. Knowing the importance of the vegetation density on the methodology proposed and the digital model generated is an opportunity area identified in our study.

Figure 10

Contour lines and DTM generated during field validation. Source: Own elaboration using QGIS 3.10 La coruña (Open Source Geospatial Foundation, Beaverton, USA, 2020).