*2.5. Methods*

As stated, the aim of this research is to obtain digital surface models (DSM) on the terrain corresponding to the natural cover of the underground cellars. A methodology was defined in several phases for this purpose, as shown in the diagram in Figure 5. In the first phase, the 45 GEP points were analysed after obtaining the data with methods based on backpack MMS and UAV technologies. These points were identified using Agisoft Photoscan® Professional software and included in the aerotriangulation process, and then identified and extracted from the DSM generated from the dense point cloud in the UAV photogrammetric project. Each GEP point was located in the data from the photographs registered by backpack MMS and measured by stereoscopy with ArcExplorer (Esri, USA) software. The GEP were also identified in the LiDAR point cloud from the backpack MMS. These four measures were compared with the GEP coordinates obtained by the RTK-GPS method.

The results of the comparisons led to the selection of the point cloud-based methods and the rejection of the photogrammetric methods, since they were insufficiently dense to generate DSMs that could be guaranteed to detect possible collapses. However, the photogrammetric backpack MMS was the most accurate method (see results section).

The data obtained with the backpack MMS define the tracks from which to obtain the point cloud covering the terrain. These tracks were used as a common element to establish the zones that cover the total study area and allow the evaluation of the DSM.

In a second phase, the results from the two massive data recording techniques are statistically analysed to compare the numerical and graphic products obtained and to define the DSM more clearly.

As has been mentioned, a supported network of control points was defined with RTK-GPS to serve as the basis for the UAV flight and the MMS backpack. This network established the real precision of both models and allowed the study of parameters such as the distance to the track in the case of backpack MMS, the point density according to the method used, and the veracity of the model with regard to walls, roofs, steeply sloping areas, and other elements. These points acted as a geometric control of the parameters to be assessed. The methodology applied in each system is shown in the diagram in Figure 5 and described below in more detail.

#### *2.6. Processing the UAV Point Cloud*

The previously mentioned software was used to generate the point cloud with the images recorded with UAV, and also included their corresponding metadata and the RTK-GPS coordinates of the photocentres of the images referenced on the ground. This type of software uses sfM-MVS algorithms for the orientation and calculation of point clouds and is widely used in UAV work processes due to the large number of images, type of cameras, and auxiliary data taken into account [57–60].

The sfM-MVS processing workflow used was as follows: initially a feature detection is performed by identifying a large number of key points in each image. With these points, an image matching process is carried out to identify and match these features in all the images in which they are registered. Subsequently, a blunder adjustment is performed with the camera's self-calibration, photocenter coordinates, and other parameters. Thus, the photogrammetric solution of the external orientation parameters of the images is obtained together with the 3D sparse model composed of the feature points detected in these images, as described in Reference [61].

This study has specifically included the use of GCPs as field control, GEPs as manual photogrammetric tie points, and feature points as automatic tie points or sparse points (see Figure 5, right) in the same photogrammetric adjustment process. Lastly, the digital terrain model (DTM) was created to produce the orthophotograph on which to identify the assessment points and verify the quality of the LiDAR DSM.

#### *2.7. Processing the Backpack MMS Point Cloud*

The point clouds from the LiDAR data registered with the backpack MMS platform were computed using the Leica Pegasus Manager software, which also allows the management of the data captured by means of MMS. It is composed of several modules in a workflow that ranges from the prior planning of the work to be done, the acquisition of the data, the subsequent processing and refinement with other sensors and algorithms, and the automatic and manual extraction of the characteristics of interest within the point clouds.

A total of six tracks were obtained (labelled A to F, Figure 4). The points were measured by photogrammetry with the Leica ArcExplorer application. The software includes a tool that allows the stereoscopic measurement of any point in the images from the perspective centres recorded and their rotations, and, thereby, obtains the coordinates of the GEPs. The last step was to identify the GEPs on the orthophotograph obtained by UAV and to measure them in the LiDAR MMS point cloud (see Figure 5, left).

#### *2.8. Comparison of the UAV-Backpack MMS Point Clouds*

After comparing the coordinates for each GEP with the DSMs obtained from the UAV and backpack MMS data, the information from each registration system was analysed. This comparison uses the tracks defined with backpack MMS to identify the different sections for analysis, while taking into account the following factors.


The backpack MMS tracks A-F were used as identifiers for the assessment, and grouped according to the similarity of their features. Tracks A and B are perimetral, defined in areas of broad corridors with little influence of vegetation at a distance of 2.5 m each side of the axis. Tracks C, D, E, and F are located on the interior and, therefore, have a presence of vegetation and irregularities in the terrain, low walls, etc. These six tracks were divided into sections based on the similarity of the type of terrain in order to enable a better comparison between UAV and MMS data. Hence, Track A was divided into five sections, Track B was divided into four sections, Tracks C and E were divided into two sections, and, lastly, Tracks D and F each have only one section. Obstacles caused by constructions were eliminated during the digital processing of the point cloud to avoid errors due to vertical and/or horizontal measuring in the two systems.

For the assessment, a DSM was generated with a resolution of 0.05 m from the point clouds obtained with backpack MMS, and a grid with a point-to-point resolution of 0.10 m was projected on the DSM. The same procedure was followed for the UAV point cloud. Therefore, the resolution and geolocation of the DSMs coincide and allow their subsequent comparison.

Lastly, for the assessment of the DSM comparison at 7.5 m from the axis of the track taken by backpack MMS, the sections with no errors due to obstacles were maintained, and sections A4, A5, B1, B2, and C2 were removed.

#### **3. Results**

Before comparing the results of the different methods, it should be noted that a mean square error of 1.9 cm in height was obtained in the calculation of the GEP points.

Ten of the 59 initial GEPs were eliminated since they could not be measured in all four methods including some that were impossible to identify and others had insufficient resolution for providing real coordinates with regard to other methods, such as the eaves of a warehouse in the testing area. Four more points were also eliminated due to problems in identifying them, such as the corner of a bench which—although it could be adequately measured in photogrammetry—was difficult to discern from the LiDAR point clouds.

The following results were obtained from the various methods, according to the distance criteria 0 to 2 m, 2 to 3.5 m, and 3.5 to 10 m.

#### *3.1. Point Cloud Processed with UAV*

• The assessment of coordinates was obtained by triangulation with the UAV images. The GEP points included as linkage points in the aerotriangulation process gave the following results. The mean square error for the three distances is 4.7 cm in distance and 8.9 cm in height. The points were measured on an average of 16 photographs and their internal precision was 1.2 pixels. The standard

deviation was up to 6.2 cm in height. This is a "low density" point cloud (2 points/square metre) (Table 2 and Figure 6). Five minutes were required to measure each point in photogrammetry, since the points must be measured in all the photographs. Sixty points were recorded. parallax must be manually cancelled in each point. Table 2 shows the precision and densities obtained with each method. They are then compared with the RTK-GPS points network measured in the area, and the data are grouped by distance to the

process, with one point measured every 10 min. The points are identified manually, and the

*Sensors* **2019**, *19*, x FOR PEER REVIEW 9 of 20

• The assessment of coordinates was obtained by photogrammetry with ArcExplorer software. The mean square error for the three distances was 10.3 cm in distance and 3.6 cm in height. The

• This section discusses the assessment of points in the UAV dense point cloud. The GEPs identified and measured in the dense point cloud gave the following results. The mean square error for the three distances was 5.0 cm in distance and 10.7 cm in height. The standard deviation of the points measured in the dense point cloud was 7.5 cm (Table 2). The points in the dense clouds were measured. Ten hours were required to make the dense model of 2,400,000 points extracted from this work. track with the laser scanner located in the backpack: 0 to 2 m (measurement angle of the laser scanner on the terrain from 90° to 5°), 2 to 3.5 m (measurement angle of 45° to 22°) and over 3.5 m (angle of less than 22°). The greatest precisions, particularly in terms of height, are obtained in the photogrammetric measurement on the backpack images, even though this is the least efficient method in terms of density and production. When the point is clearly defined, no significant loss of precision was observed with regard to distance.



**Figure 6.** Comparison of precision based on the method used and the distance from the axis of backpack MMS. **Figure 6.** Comparison of precision based on the method used and the distance from the axis of backpack MMS.


The standard deviation in height was up to 3.2 cm (Table 2). This is the least productive measuring process, with one point measured every 10 min. The points are identified manually, and the parallax must be manually cancelled in each point.

Table 2 shows the precision and densities obtained with each method. They are then compared with the RTK-GPS points network measured in the area, and the data are grouped by distance to the track with the laser scanner located in the backpack: 0 to 2 m (measurement angle of the laser scanner on the terrain from 90◦ to 5◦ ), 2 to 3.5 m (measurement angle of 45◦ to 22◦ ) and over 3.5 m (angle of less than 22◦ ). The greatest precisions, particularly in terms of height, are obtained in the photogrammetric measurement on the backpack images, even though this is the least efficient method in terms of density and production. When the point is clearly defined, no significant loss of precision was observed with regard to distance.

#### *3.3. Comparison of Backpack MMS vs. UAV DSMs*

The result of the precision analysis for the DSM at a distance of 2.5 m from the axis of the track is shown in Table 3 and Figure 7. The two DSMs are evaluated with a square grid with a resolution of 0.1 m based on the point clouds obtained with UAV and backpack MMS. The height precision for the tracks on the exterior is 3.8 cm, whereas the precision in interior areas is 5.7 cm in height. Tracks A4, A5, B1, B4, D1, and E2 have a vertical displacement between 0.10 and 0.15 m. All these areas have an influence due to construction walls or other obstacles.

**Table 3.** Results of the comparison between UAV and backpack MMS DSMs with points measured at a distance of at least 2.5 m from the backpack MMS track.


If the distance from the backpack MMS track increases from 2.5 m to 7.5 m on each side of the axis from which the point clouds defining the DSM are obtained, the maximum angle of measurement of backpack MMS on the terrain would decline from 45◦ to 20◦ , which increases the error due to the noise produced in the measurement. There are also a higher number of errors due to obstacles. The number of zones in the study was, therefore, reduced in order to avoid errors caused by built elements or obstacles. The study for these distances at 7.5 m is limited to zones A1, A2, A3, B3, and B4 in the exterior part and to zones C1, D1, E1, E2, and F1 in the interior. In overall terms, the mean square error in height ranges from 0.04 m to 0.21 m (Table 3 and Figure 8).

Table 4 shows a comparison between the data obtained with points located at a distance of up to 2.5 and 7.5 m from the backpack MMS track. The correlation coefficient R<sup>2</sup> allows the incidence of isolated errors to be analysed when comparing both DSMs.

**4. Discussion** 

**Track Perimeter (m)** 

A4, A5, B1, B4, D1, and E2 have a vertical displacement between 0.10 and 0.15 m. All these areas have

**Table 3.** Results of the comparison between UAV and backpack MMS DSMs with points measured at

A1 370 564 184 357,165 13,694,022 55,578 −0.02 0.04 A2 189 420 90 255,611 11,128,095 41,606 −0.1 0.04 A3 135 196 63 116,012 4,222,443 19,171 −0.1 0.01 A4 95 117 44 69,940 5,290,734 11,562 −0.09 0.03 A5 59 90 26 52,414 7,885,791 8869 −0.16 0.02 B1 98 138 45 80,114 2,921,953 13,628 −0.16 0.02 B2 157 182 77 111,226 5,390,511 17,678 −0.1 0.05 B3 106 159 50 94,737 4,219,886 15,508 −0.08 0.06 B4 236 286 116 180,494 13,471,066 27,982 −0.12 0.07 C1 339 348 168 226,162 9,852,318 33,758 −0.04 0.04 C2 262 339 126 243,346 16,488,277 33,363 −0.09 0.08 D1 69 83 31 52,109 1,691,763 8062 −0.15 0.02 E1 169 232 83 167,189 10,230,845 22,552 −0.07 0.07

F1 236 464 116 302,089 14,207,240 45,617 −0.06 0.06

**Backpack MMS Resolution** 

**Grid Resolution**  **Precision** 

**(m) RMS (m)** 

**UAV Resolution** 

an influence due to construction walls or other obstacles.

**Area (m2)** 

a distance of at least 2.5 m from the backpack MMS track.

**Distance (m)** 

**Figure 7.** Distribution of the precise comparison of the point clouds based on the backpack MMS track. Letters denote the different tracks while numbers depict the different sections in the tracks. **Figure 7.** Distribution of the precise comparison of the point clouds based on the backpack MMS track. Letters denote the different tracks while numbers depict the different sections in the tracks. E2 14,961 72,501 −0.16 −0.2 0.07 0.17 0.9967 0.9567 F1 45,617 158,948 −0.06 −0.09 0.06 0.11 0.9975 0.9908

**Figure 8.** Precision distribution with the influence of 7.5 m with regard to the backpack MMS track. Letters denote the different tracks while numbers depict the different sections in the tracks. **Figure 8.** Precision distribution with the influence of 7.5 m with regard to the backpack MMS track. Letters denote the different tracks while numbers depict the different sections in the tracks.

eliminated due to of the fact that identifying these points in all the methods is not possible.

The methods for acquiring and processing the data provided by UAV and backpack MMS are

The first part of the work develops the methods of data capture using photogrammetry and a laser scanner with UAV and backpack MMS. The most challenging task was to select the points that could be identified using the different methods. Additionally, 20% of the control points were


**Table 4.** Comparison of precision when the width is increased from 2.5 to 7.5 m.

#### **4. Discussion**

The methods for acquiring and processing the data provided by UAV and backpack MMS are of sufficient quality to generate valid DSMs in the study area or similar.

The first part of the work develops the methods of data capture using photogrammetry and a laser scanner with UAV and backpack MMS. The most challenging task was to select the points that could be identified using the different methods. Additionally, 20% of the control points were eliminated due to of the fact that identifying these points in all the methods is not possible.

Except in the case of measurement by photogrammetry, where backpack MMS provides the best precision results with a little under 5 cm in height, the rest of the methods had a precision of around 5 and 10 cm. Better overall results are obtained with UAV methods than with backpack MMS, but this depends on the distance of the points from the axis of the backpack MMS track. Backpack MMS is also better over a short distance under 3 m, with 6.8 cm. At higher distances, the UAV methods obtain a lower precision of over 8 or 9 cm.

In addition to precision, performance and production were also analysed. Manual photogrammetric methods were discarded since they required longer execution times, which represent a higher production cost. The registration of the dense cloud points with UAV proved to be the fastest and most economical method, whereas the registration via MMS gave the best performance, but had a higher production cost.

Given these results, the comprehensive comparison of UAV and backpack MMS point clouds in different zones that are more or less devoid of vegetation and small obstacles allows the analysis of their precision and the influence of each method on the definition of the DSM to be characterised. The precision in clear zones or zones near the track trajectory is 5 cm (Figure 9).

More substantial differences are found at greater distances, and measurement problems also arise due to the lack of information or measurements in an orography such as the one in the study area, with holes/gaps or even obstacles, as can be seen at the most extreme points of the model (Figure 10). Although they are few and more distant from the backpack trajectory, they present a proportionally higher error, which can be seen in Table 3 and Figures 7 and 8.

The clearest example of these differences can be seen in the correlation graphs in Figure 11 (external tracks) and Figure 12 (internal tracks), where the different backpack MMS trajectories and their comparison with the UAV points not only show the R<sup>2</sup> coefficients and the correlation equations indicated above, but also highlight the difference in the number and height of the points outside the trend line for the analysis of each track.

*Sensors* **2019**

obtain a lower precision of over 8 or 9 cm.

performance, but had a higher production cost.

Except in the case of measurement by photogrammetry, where backpack MMS provides the best precision results with a little under 5 cm in height, the rest of the methods had a precision of around 5 and 10 cm. Better overall results are obtained with UAV methods than with backpack MMS, but this depends on the distance of the points from the axis of the backpack MMS track. Backpack MMS is also better over a short distance under 3 m, with 6.8 cm. At higher distances, the UAV methods

In addition to precision, performance and production were also analysed. Manual photogrammetric methods were discarded since they required longer execution times, which represent a higher production cost. The registration of the dense cloud points with UAV proved to be the fastest and most economical method, whereas the registration via MMS gave the best

Given these results, the comprehensive comparison of UAV and backpack MMS point clouds in different zones that are more or less devoid of vegetation and small obstacles allows the analysis of

The precision in clear zones or zones near the track trajectory is 5 cm (Figure 9).

**Figure 9.** Identification of zones with significant differences between UAV and MMS DSMs. **Figure 9.** Identification of zones with significant differences between UAV and MMS DSMs. , *19*, x FOR PEER REVIEW 13 of 20

**Figure 10.** Differences in height according to the type of system, the width of the scan, and the obstacles or concealed elements. (**Left**): example of a zone with a 2.5 m track width. (**Right**): same area with a 15 m track width. Black areas depict concealed zones not registered in the backpack MMS. **Figure 10.** Differences in height according to the type of system, the width of the scan, and the obstacles or concealed elements. (**Left**): example of a zone with a 2.5 m track width. (**Right**): same area with a 15 m track width. Black areas depict concealed zones not registered in the backpack MMS.

The clearest example of these differences can be seen in the correlation graphs in Figure 11 (external tracks) and Figure 12 (internal tracks), where the different backpack MMS trajectories and

trend line for the analysis of each track.

*Sensors* **2019**, *19*

**5. Conclusions** 

precision and a greater level of detail.

*Sensors* **2019**, *19*, x FOR PEER REVIEW 14 of 20

**Figure 11.** Comparison of UAV vs. backpack MMS by external tracks (**A**,**B**). **Figure 11.** , x FOR PEER REVIEW Comparison of UAV vs. backpack MMS by external tracks (**A**,**B**).

15 of 20

**Figure 12.** *Cont.*

**Figure 12.** Comparison of UAV vs. backpack MMS by internal tracks (**C**–**F**).

When comparing the data extracted individually, a similar precision is obtained with an average of around 5 cm in height with both methods, and an average of 10 cm with distance or identification problems. This precision is perfectly acceptable for defining DSMs in this work environment.

Since the precision of both systems is known to be similar in small environments or at short distances from backpack MMS, denser trajectories must be used to reduce errors due to obstacles. A mixed method such as the proposed MMS–UAV technique represents a useful tool for identifying areas that are difficult to determine in the DSM and which lead to the most significant differences in height in the comparison. Given the very high number of points obtained with these procedures, the detection and elimination of these points would make it possible to obtain a DSM with greater

The assessment of the DSMs reveals that the tracks followed by backpack mobile mapping present a scarcity of information in some spaces. There are various obstacles and hidden areas in this terrain. In summary, the UAV provides a more homogeneous or stable DSM even though the DSM obtained with backpack mobile mapping is more accurate. This is due to the optimum distance range and the spaces where there are no obstacles or strong rupture lines. Both techniques can, therefore, be considered complementary and reliable for obtaining DSMs for the area where these underground cellars are located. This type of application can help detect deformations in the ground a posteriori.

**Figure 12.** Comparison of UAV vs. backpack MMS by internal tracks (**C**–**F**).

#### **5. Conclusions**

**Figure 12.** Comparison of UAV vs. backpack MMS by internal tracks (**C**–**F**). When comparing the data extracted individually, a similar precision is obtained with an average of around 5 cm in height with both methods, and an average of 10 cm with distance or identification problems. This precision is perfectly acceptable for defining DSMs in this work environment.

**5. Conclusions**  When comparing the data extracted individually, a similar precision is obtained with an average of around 5 cm in height with both methods, and an average of 10 cm with distance or identification problems. This precision is perfectly acceptable for defining DSMs in this work environment. Since the precision of both systems is known to be similar in small environments or at short distances from backpack MMS, denser trajectories must be used to reduce errors due to obstacles. A mixed method such as the proposed MMS–UAV technique represents a useful tool for identifying Since the precision of both systems is known to be similar in small environments or at short distances from backpack MMS, denser trajectories must be used to reduce errors due to obstacles. A mixed method such as the proposed MMS–UAV technique represents a useful tool for identifying areas that are difficult to determine in the DSM and which lead to the most significant differences in height in the comparison. Given the very high number of points obtained with these procedures, the detection and elimination of these points would make it possible to obtain a DSM with greater precision and a greater level of detail.

areas that are difficult to determine in the DSM and which lead to the most significant differences in height in the comparison. Given the very high number of points obtained with these procedures, the detection and elimination of these points would make it possible to obtain a DSM with greater precision and a greater level of detail. The assessment of the DSMs reveals that the tracks followed by backpack mobile mapping present a scarcity of information in some spaces. There are various obstacles and hidden areas in this terrain. In summary, the UAV provides a more homogeneous or stable DSM even though the DSM obtained with backpack mobile mapping is more accurate. This is due to the optimum distance range and the spaces where there are no obstacles or strong rupture lines. Both techniques can, therefore, be considered complementary and reliable for obtaining DSMs for the area where these underground cellars are located. This type of application can help detect deformations in the ground a posteriori.

The use of methods of mass capture offers an excellent opportunity in such a complex area as the exterior surface of the underground cellars of El Plantío in Atauta. These methods have different limitations, such as the irregularity of the terrain, difficult-to-isolate low-growing vegetation, and mobile or fixed obstacles. However, the general precision is high and in line with the data results necessary for their study and preservation.

A novel development is the inclusion of parameters such as the distance to the scanning point, the angles of incidence with regard to the ground, and the study of irregularities in the terrain. A clear comparison of both technologies that conclusively reveals the pros and cons of their use would be impossible without considering these aspects. Their combined use has also been proposed, which may be a source of further improvements in future studies.

The results obtained point to the conclusion that both techniques—albeit not without difficulty—provide DSMs that are capable of defining terrain stability. Due to the registration speed and the precision achieved, both systems allow the assessment of the underground wine cellars. Their use over time will make it possible to establish the necessary priorities to guarantee the conservation of such unique and important spaces such as this site of El Plantío in Atauta.

**Author Contributions:** Conceptualisation, S.L.-C.M. and E.P.-M., Methodology, S.L.-C.M., E.P.-M., J.F.P., J.V., and T.R.H.T. Investigation, J.F.P., J.V., and T.R.H.T. Data backpack MMS registration M.Á.C.M., A.E.-C., and E.P., Data UAV registration S.L.-C.M. and J.A.d.M. GPS field observation, J.F.P. and J.V. Backpack MMS data processing E.P.-M., M.Á.C.M., and A.E.-C. UAV data processing, S.L.-C.M. and J.A.d.M. GPS data processing, J.F.P., J.V., and T.R.H.T. Visualisation, E.P.-M., M.Á.C.M., and J.A.d.M. Supervision, T.R.H.T. and A.E.-C. All the authors contributed to writing and original draft preparation.

**Funding:** This work is part of the results of the "Red de Investigación en Paisajes Culturales" research project in the national R+D plan with reference number RED2018-102558-T. It is also supported by the programme for collaboration among the GIAPSI (ETSIINF), GESYP (ETSIAAB) and GIPC (ETSAM) research groups promoted by the Universidad Politécnica de Madrid (UPM) with reference VJOVUPM17IMS.

**Acknowledgments:** We would like to thank the Atauta Town Hall (San Esteban de Gormaz) and the companies TOPCON and LEICA for their help with the cultural landscape, including administrative and technical support (materials used for experiments). We also want to acknowledge Pru Brooke-Turner (M.A. Cantab.) for her English language and style review.

**Conflicts of Interest:** The authors declare that there are no conflicts of interest.

#### **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
