Potential of Low-Cost UAV Photogrammetry for Documenting Hard-to-Access Interior Spaces Through Building Openings

Abstract

Unmanned aerial vehicles (UAVs) are primarily used in the field of cultural heritage for mapping the exteriors of larger objects and documenting the roofs and façades of tall structures that cannot be efficiently or feasibly measured using conventional terrestrial technologies and methods. However, due to the considerable diversity of cultural heritage, there are practical demands for the measurement of complex and inaccessible objects in interior spaces. This article focuses on the use of two different off-the-shelf UAVs for partial photogrammetric reconstruction of the attic of a mining house, which was only visible through a window in the gable wall. Data from both UAVs were compared with each other and with terrestrial laser scanning. Despite the lower quality of the results from the DJI Mini 4 Pro compared to the DJI Mavic 3 Enterprise, the results from both UAVs would still be suitable for documenting the interior attic spaces. However, a detailed analysis of the photogrammetric data indicates that, when selecting a UAV for this purpose, it is necessary to consider the limitations of the camera system, which may lead to a reduction in the geometric accuracy and completeness of the point clouds.

1. Introduction

Photogrammetric methods for documenting heritage objects have been employed since the second half of the 19th century, when architect Albrecht Meydenbauer almost fell from the side aisle of the cathedral in Wetzlar, which led him to the idea of replacing direct façade measurements with safer indirect measurements on photographic images [1]. Today, for this purpose, we not only use close-range terrestrial photogrammetry but also UAV photogrammetry, particularly for taller objects, as it significantly reduces the financial costs of documentation due to the easy availability of low-cost UAV systems equipped with relatively high-quality cameras [2,3]. Another advantage is the automation of photogrammetric processing thanks to software solutions based on computer vision techniques. Using algorithms such as Structure from Motion (SfM) [4,5], Scale-Invariant Feature Transform (SIFT) [6], and Random Sample Consensus (RANSAC) [7], it is possible to perform not only fully automated image orientation but also to refine the interior camera orientation elements within the bundle adjustment process, thus performing camera calibration [8]. Self-calibration, in particular, allows for the use of non-metric cameras that are often equipped on commercially available low-cost UAVs for photogrammetric measurements [9].

An important asset of UAV photogrammetry is its applicability in areas where direct access would be risky. This includes primarily post-catastrophic sites where it is necessary, for instance, to map the extent of damage to cultural heritage objects [10,11,12] or to analyze the safety and stability of the environment where the heritage site is located [13]. UAV photogrammetry is very often combined with terrestrial laser scanning (TLS), which provides higher robustness of the measured geometry and, compared to photogrammetry, enables reliable data collection even in poor lighting conditions in interiors or on surfaces with unsuitable textures [14,15,16]. The use of photogrammetry supported by SfM in indoor environments is generally uncommon, especially when the interior walls are treated with uniform smooth coatings. However, when utilizing specialized systems such as Matterport, photogrammetry also has the potential for interior documentation, although with lower accuracy [17].

Virtually every commercially available multi-rotor UAV equipped with a camera also contains a GNSS (Global Navigation Satellite System) sensor, which is used not only to enhance the UAV’s stability in the air, for instance, in situations where wind may start to drift the drone but also for approximate georeferencing of the images. While most UAVs are equipped with a sensor that provides global coordinates’ accuracy only at the meter level, which is sufficient for stabilization purposes, professional solutions equipped with real-time kinematic (RTK) systems or enabling a post-processing kinematic (PPK) approach [18,19,20] also exist. These solutions, when coupled with an inertial measurement unit (IMU), improve the accuracy of determining the reference coordinates of the camera’s projection center to a centimeter level, which in turn enhances the accuracy of photogrammetric processing. In indoor spaces where GNSS signals are not available, obstacle detection systems are used to enhance flight stability. These systems can be based on various principles and are generally categorized into passive (image-based), active (sensor-based), or hybrid systems [21,22]. However, such systems do not record reliable global coordinates of the projection center in the image file metadata, so their contribution to photogrammetric measurement lies primarily in minimizing motion blur and increasing flight safety in confined spaces. Safety can also be enhanced with the use of additional protective elements, such as propeller guards or specialized cage constructions like those used on the UAV Elios, whose application was described in [23]. From a photogrammetric perspective, an important advantage of the available localization system is the ability to automate navigation along a predefined trajectory, which facilitates data collection during mapping. Over the past decade, numerous articles have been published focusing on automated navigation even in indoor spaces [24,25,26], but low-cost off-the-shelf UAVs generally require manual navigation.

1.1. Related Research

Due to the predominant use of low-cost UAVs for exterior documentation, publications focusing on the use of UAVs in indoor environments are generally less common [27,28]. Nevertheless, the advantages of this technology have led several researchers to use it, for instance, for the documentation of an altar using the ultra-light DJI Spark [29], for documenting the interiors of a church using the ultra-light Parrot Anafi [30], or for a crime scene using the DJI Mini 2 [31]. Photogrammetric documentation of indoor spaces using UAVs brings numerous challenges, which are associated not only with confined spaces and potential collision hazards but also with unsuitable lighting conditions [28]. These conditions can be improved with additional lighting, but it may be difficult to position it at a sufficient height where the UAV will operate. A solution may involve equipping the UAV with additional spotlights or manually adjusting the UAV and camera settings. Some UAVs allow modifications to parameters such as aperture and exposure time, which result in sufficiently bright images with lower noise levels. However, longer exposure times may cause motion blur if the UAV is unstable. This can be partially mitigated by capturing images only from static UAV positions, i.e., avoiding image acquisition during flight, as is common in exterior applications. To facilitate this, some UAV models offer a “Tripod Mode”, which enhances stability by reducing speed and sensitivity to control inputs, making the drone more stable in confined spaces.

1.2. Case Study—Mining House in Banská Štiavnica, Slovakia

The aim of this article is to explore the challenges associated with the use of low-cost UAVs for photogrammetric documentation of cultural heritage objects in interior spaces by analyzing data collected within a case study focused on surveying a mining house in the town of Banská Štiavnica, Slovakia, whose attic could only be measured through a window opening in the gable wall.

Banská Štiavnica (Figure 1) is the oldest mining town in the central Slovak mining region. It has developed in connection with mining activities aimed at extracting gold and silver since at least the 11th century. Due to its intactly preserved urban and architectural structure and the numerous objects documenting its mining activities, the town has been listed as a UNESCO World Heritage Site.

The town has preserved a variety of typological building types, including burgher houses, craftsmen’s houses, sacral buildings, schools, monasteries, and others. A unique feature of the town is the scattered development of architecturally simple mining houses along the slopes surrounding the town center. The object selected for documentation is situated on a northeast-facing slope above the valley of the town center (Figure 2). The area is called “Resla” (Figure 2, right) and is named after the famous miner Erasmus Rössel (1375/83?–1402), who granted miners the freedom to build residences near the mining operations.

The subject of verifying the possibilities of various methods of surveying was a mining house with an ambiguous designation of its original function. It is a single-wing structure built on a rectangular floor plan, consisting of two stories that are not connected by a staircase. The entrances to each floor are located at ground level, as the building is embedded in a relatively steep slope. It was constructed from quarry stone. The interior was later divided by a partition into two rooms, but, originally, both floors consisted of a single space. The two-story house is topped with a gable roof, with the ridge perpendicular to the street line. The roof is finished with wooden gable ends on both sides. Originally, two smaller skylights with intricately carved frames were placed on the street-facing side of the house. Currently, the original openings in the gable have been destroyed, and there is now one large opening.

Given its atypical layout and unusual two-story design for mining residential houses, it is presumed that its original function was related to mining processing facilities or technical support for mining activities. It was only when this original function became obsolete that the building was used for residential purposes. Therefore, it represents an important and unique typological type of structure whose function has yet to be interpreted. Currently, the building is in a state of emergency due to long-term neglect and poor construction and technical condition. Without detailed surveying and documentation using all possible contemporary geomatic methods, any trace of its existence and material substance would be erased, resulting in the loss of study material for an important typological type of building in the town.

For this reason, we decided to digitize at least the accessible parts of the building and create a virtual digital copy in its current state. The measurements were carried out as part of an interdisciplinary student workshop organized under the auspices of a cultural-educational project focused on the digitization of critically endangered heritage objects in Banská Štiavnica, supported by the Ministry of Education, Science, Research, and Sport of the Slovak Republic (see Funding section). Students of the Geodesy and Cartography study program at the Faculty of Civil Engineering of the Slovak University of Technology in Bratislava acquire practical skills in working with modern technologies such as TLS, as well as close-range and UAV photogrammetry during these regular workshops, collecting data that genuinely serve to preserve national cultural heritage. The selection of objects for digitization is consulted with the Monuments Board of the Slovak Republic to avoid duplicate measurements, prioritizing objects that the board has not yet managed to digitize within its own campaigns.

2. Materials and Methods

A combination of several technologies and methods was used for the digitalization of the object. Each has its advantages and disadvantages, and, in practice, it has been shown that for comprehensive documentation of heritage objects, it is advisable not to rely solely on a single preferred technology. Just as terrestrial laser scanning (TLS) does not allow for the capture of roofs, photogrammetry generally does not provide reliable data in areas with poor lighting conditions. The main advantage of TLS is the high accuracy and robustness of the captured geometry, while photogrammetry excels in the high variability in camera network configuration around the object and the high resolution of textural information. Moreover, measurements using these two main technologies typically need to be supplemented by geodetic measurements of ground control points (GCPs), which serve not only for the transformation of point clouds and models into a national reference coordinate system but also for verifying the quality of image orientation or registration of laser scans.

The measurement was conducted in two epochs:

• During the original measurement in November 2023, the UAV DJI Mavic 3 Enterprise (M3E) (Figure 3 left,

  Table 1. Parameters of DJI Mavic 3 Enterprise.

  Parameter	Value
  Sensor size	17.3 × 13 mm
  Sensor resolution	5280 × 3956 pixels
  Pixel size	3.3 µm
  Focal length	12.3 mm
  Aperture	f/2.8–f/11
  Minimum Focus Distance	1 m
  ISO Sensitivity	100–6400
  Shutter Speed	Mechanical Shutter 8 to 1/2000 s
  UAV Class by weight	C2
  Price	3500€

  ), DSLR camera Nikon D850 (

  Table 2. Parameters of Nikon D850 DSLR with fisheye lens.

  Parameter	Value
  Sensor size	35.9 × 23.9 mm
  Sensor resolution	8256 × 5504 pixels
  Lens	AF Fisheye-NIKKOR 16 mm f/2.8 D
  Focal length	16 mm
  ISO Sensitivity	64–25,600
  Shutter Speed	30 to 1/8000 s

  ), terrestrial laser scanner Trimble X9 (

  Table 3. Parameters of Trimble X9 laser scanner.

  Parameter	Value
  Range	0.6–150 m
  Range noise	<1.5 mm/30 m
  Range accuracy	2 mm
  Angular accuracy	<16″
  3D point accuracy	2.3 mm/10 m, 4.8 mm/40 m

  ), and GNSS equipment Trimble R6 were used.
• During the supplementary measurement in August 2024, only the UAV DJI Mini 4 Pro (M4P) equipped with a universal attachable spotlight (Figure 3 right,

  Table 4. Parameters of DJI Mini 4 Pro.

  Parameter	Value
  Sensor size	9.7 × 7.3 mm
  Sensor resolution	4032 × 3024 pixels
  Pixel size	2.4 µm
  Focal length	6.72 mm
  Aperture	f/1.7
  Minimum Focus Distance	1 m
  ISO Sensitivity	100–6400
  Shutter Speed	Electronic Shutter 2 to 1/8000 s
  UAV Class by weight	C0 (C1 with spotlight)
  Price	800€

  ) was employed.

The M3E is equipped with an RTK module, which allows for real-time determination of the spatial coordinates of the camera’s projection center during flight, meaning it is generally unnecessary to measure GCPs on the ground. However, our experience indicates that measuring at least some check points using GNSS equipment is justified. The aforementioned UAVs not only belong to different weight classes (C0 and C2) but also fall into significantly different price categories. While the M3E is considered a professional mapping solution, the M4P is primarily designed for video recording, reflected in its price of under EUR 1000.

Since one side of the object was very close to the fence, it was necessary to capture images from very short subject distances, prompting the use of a Nikon D850 digital SLR camera with a fisheye lens for terrestrial photogrammetry during the first measurement (see Table 2).

Photogrammetric processing was conducted using Metashape Professional version 2.1.2 by Agisoft LLC (St. Petersburg, Russia).

In addition to photogrammetric measurements, terrestrial laser scanning was performed using the Trimble X9 device, the parameters of which are listed in Table 3. During the scanning process, “outdoor” settings were used with a scanning density of 6 mm/10 m and point cloud coloring enabled.

Georeferencing of the laser scans was ensured through four GCPs around the object measured using the Trimble R6 GNSS instrument. Both the M3E and the Trimble GNSS apparatus utilized a connection to the state permanent observation service SKPOS, achieving centimeter-level accuracy. The obtained ETRS89 coordinates (EPSG:4258) were subsequently transformed into the national coordinate system S-JTSK (EPSG:5514), and ellipsoidal heights (EPSG:4937) were converted to the Bpv height system (EPSG:8357). The accuracy of the determined coordinates reached levels of 0.02–0.04 m.

The analysis of the obtained data was performed using the freely available software Cloud Compare version 2.13—more details can be found in Section 5.

3. Data Collection on the Attic

The greatest challenge of this case study can be considered the capture of the interior spaces of the building’s attic, which is why we will discuss it in more detail within this article. Although we documented only a small portion of the attic through a window opening using UAV, photogrammetry is generally not employed for this purpose, as it is typically used solely to document exterior parts of buildings, such as façades and roofs. Therefore, the primary contribution of our study is to highlight the potential for extending photogrammetric documentation and to illustrate its limitations in cases where it is impossible to position the camera inside the structure.

The interior of the building was only partially accessible, as seen in the cross-section in Figure 4 on the right. The entrance to the second floor was located at the back of the building and was permanently closed.

Since the building was situated on a slope and surrounded by other structures and vegetation, terrestrial laser scanning of the attic through the openings in the gable wall of the front façade was very limited. Out of a total of 14 positions of TLS, there was only one position (No. 7) from which it was possible to see deeper into the attic; however, this position was also significantly far away and low (see Figure 5).

The only option to partially capture the attic space was to use a UAV and take photos of the interior through the opening in the northern gable wall (Figure 6).

The configuration of the camera network for photogrammetric measurements during both the initial and supplementary data collection is illustrated in Figure 7. Both UAVs were sufficiently small to enter through the window; however, as we did not have propeller guards available, and the obstacle detection systems prevent flying closer than 1 m to obstacles, we decided not to risk damage to the equipment and captured the interior of the attic only from the outside through the window.

During the initial measurement, a total of 380 images were captured, with 159 taken using a UAV and 221 using the DSLR. In the supplementary measurement, 102 images were captured using a UAV.

4. Data Processing

Laser scans were registered automatically in Trimble RealWorks using the surface-based registration method. During the registration of the laser scans, a total Cloud-to-Cloud error of 1.7 mm was achieved. The transformation of the scans into the reference coordinate system was performed using four GCPs with a total error of 9.5 mm. The resulting point cloud of the object, downsampled to a usable resolution of 2 mm, contained 64 million points.

As for the results of photogrammetric processing, they can be influenced not only by the configuration of the camera network and the overall quality of the data but also by the specific software solution used for processing the images. In the case of combining terrestrial and UAV photogrammetry, it is sometimes necessary to increase the overlap between image blocks or reduce vertical disparities between terrestrial and UAV image blocks for the selected software, based on Structure from Motion (SfM) principles, to properly orient the images. Furthermore, not all photogrammetric software supports calibration models for fisheye lenses. Thus, capturing images needs to be adapted to the available photogrammetric software. This is why we chose to use Agisoft Metashape Professional, which not only supports a distortion model for fisheye lenses but also demonstrates the ability to orient images even under less-than-optimal configurations of the camera network [33].

The primary and secondary measurements were processed in separate projects, as different meteorological conditions prevailed during the data collection (overcast/sunny), and approximately nine months elapsed between the measurements, which could have led to physical changes in the object and environment. During the processing of all photogrammetric projects, the same settings for image orientation were used: accuracy (high), key point limit (40,000), and tie point limit (10,000). The only difference was the selection of the calibration model for the DSLR camera (fisheye), which did not play a role in processing the attic, as it was only captured using a UAV.

The same settings for generating detailed point clouds were also applied to both projects: quality (high) and filtering mode (mild). As mentioned earlier, the UAV M3E equipped with an RTK module was used for the initial measurement. Since the accuracy of the selected SKPOS observation service for determining reference coordinates was at the centimeter level, it was not appropriate to merge data from TLS and UAV photogrammetry solely based on their individual transformations into the reference coordinate system (point clouds from TLS through GCPs and from a UAV using RTK). Logical discrepancies would arise between the point clouds obtained using different technologies, and, given the relatively small size of the building object and UAV image block, there could also be a reduction in the accuracy of the photogrammetric point cloud scale. To enhance the reliability of the scale and overall comparability of multiple data sets, we used identical points determined by selecting prominent features from the already registered TLS point clouds (Figure 8 left). The total error after transforming the photogrammetric project from the first measurement (M3E) using six naturally marked points reached 6 mm. The secondary photogrammetric project from the additional imaging (M4P) was then aligned with the original project (M3E) based on automatically detected tie points in the texture. For the reliability analysis of this operation, the point cloud obtained from the M4P was compared with the cloud from the M3E. From the histogram of deviations in Figure 8 on the right, it is evident that most deviations did not exceed 5 mm (RMSE = 2 mm).

Currently, it is possible to effectively and fully automate the orientation of images and laser scans, and this functionality is also supported by Metashape [34]. However, the available workflow is not meaningful from a photogrammetric perspective. After importing laser scans into Metashape, the software does not fix the position of the already registered laser scans; while it can orient images and scans relative to each other, they no longer retain their original position in the reference coordinate system. Even the software developers themselves recommend performing an additional transformation of the data into the reference system using GCPs after this step [35]. The only software we know of that meaningfully merges laser scans and photogrammetric images is RealityCapture by Capturing Reality (Bratislava, Slovakia). However, it does not support a distortion model for fisheye lenses, which is why we ultimately decided to manually connect TLS and photogrammetry using identical points in Metashape. The results from the processing of the photogrammetric projects are presented in

Table 5. Statistics from the photogrammetric processing of the primary and secondary photogrammetric measurements.

Parameter	Basic Project	Additional Project
Aligned images	380	102
Cameras	DJI M3E + Nikon	DJI M4P
Tie points	773,390	115,972
Rms reprojection error	1.08 pixels	1.13 pixels
Matching + Alignment time	9 min	2 min
Point cloud from UAV images	17.4 mil. points	15.2 mil. points
Depth maps + Point cloud generation time	53 min	42 min

.

The resulting point cloud created by combining both ground and UAV images contained up to 68 million points and served for the comprehensive documentation of the heritage site. However, since the subject of this study is solely the interior spaces captured through the window opening using a UAV, we will further analyze only the data from the attic obtained from the UAV and TLS.

5. Analysis of Results

The achieved results can be analyzed from various perspectives. Since different photographic techniques were used for data collection, we will separately address the photogrammetric results and subsequently compare them with TLS, where the advantages of using different technologies for comprehensive documentation of heritage sites will be clearly visible.

5.1. Comparison of M3E and M4P

The quality of results from the photogrammetric measurements using the two different UAVs was primarily influenced by the implemented optics and sensors. From Table 1 and Table 4, it was already clear that the M3E is equipped with a higher-quality camera featuring a sensor that not only has a higher geometric resolution in pixels but also a larger physical size. While the pixel size on the 12 MP sensor in the M4P is only 2.4 μm, the professional M3E with a 20 MP sensor has a size of 3.4 μm. In this case, it is always true that the larger the pixel on the sensor, the more light can reach it during a given exposure time, resulting in a higher quality final image. This disadvantage was attempted to be compensated for by using an attachable spotlight on the M4P, but, as we will see later, other factors had a greater impact.

5.1.1. Camera Calibration

A higher geometric resolution (number of megapixels) allows for a greater level of detail in the image, which translates to a smaller ground sampling distance (GSD). For example, from a distance of 2 m, a GSD of 0.57 mm can be achieved for the M3E and 0.72 mm for the M4P. Therefore, the M3E allows capturing smaller details from the same distance.

Another factor that cannot be overlooked in photogrammetric processing is the reliability of the camera system calibration. Metashape employs the Brown distortion model when calibrating cameras with the “frame” setting. The results of camera calibration can be evaluated in the Metashape environment using various tools, with the most important being the ability to display the residuals after applying the mathematical model. As shown in Figure 9, the residual errors after distortion correction for the M4P are significantly larger than those for the M3E.

The distribution and magnitude of the residuals in Figure 9 indicate the adequacy of the camera system as described by the chosen mathematical model. The residuals are averaged for individual image cells and subsequently across all images captured by the respective camera. In addition to graphical representation, Metashape provides numerical expressions of the residuals in terms of RMS and maximum values—the M3E achieved an RMS residual of 0.17 pixels (maximum 0.45 pixels), while the M4P had an RMS residual of 0.38 pixels (maximum 1.3 pixels).

Noteworthy results were also obtained regarding the elements of the camera’s interior orientation. Image geometry can be affected by various distortions, with two fundamental types typically determined during camera calibration: radial and decentering distortion. Of these, radial distortion usually has the most significant impact, whereas decentering distortion can sometimes be compensated for by adjusting the coordinates of the principal point during camera calibration. In photogrammetric measurements, there is a tendency to utilize lenses with minimal radial distortion; however, it is generally accepted that the size of the distortion is less important than the degree of agreement with the deterministic model. This was confirmed in our case, as illustrated in the graphs in Figure 10. Since pre-calibrated parameters are also provided for the M3E camera, we present them in

Table 6. Statistics from the calibration of the M3E and M4P cameras, including final adjusted interior orientation parameters and their mean errors.

Parameter	M3E			M4P
	Precalibrated	Adjusted	Error	Adjusted	Error
F (pix)	3713.29	3715.55	0.03	2985.08	0.08
Cx (pix)	7.02	23.80	0.04	16.79	0.18
Cy (pix)	−8.72	−28.74	0.04	−4.84	0.17
K1	−0.1125750	−0.1101450	0.0000262	0.0693043	0.0000948
K2	0.0148744	0.0098727	0.0000583	−0.0889360	0.0002993
K3	−0.0270641	−0.0247238	0.0000416	0.0429116	0.0003051
P1	0.0000001	−0.0000541	0.0000014	0.0002508	0.0000164
P2	−0.0000857	−0.0005270	0.0000016	0.0006589	0.0000154

along with the resulting adjusted parameters after calibration in the project. The values clearly indicate that the mean errors of the parameters are several times larger for the M4P camera compared to the M3E.

F stands for focal length; Cx and Cy refer to the image coordinates of the principal point; K1, K2, and K3 represent the coefficients of radial distortion; and P1 and P2 represent the coefficients of tangential (decentering) distortion.

Despite the lens of the M3E camera exhibiting a maximum radial distortion of approximately 300 pixels, previous results indicate that its progression is significantly better modeled mathematically compared to the M4P, where the maximum radial distortion reaches only about 50 pixels. Thus, we hypothesize that the geometric quality of the point clouds, illustrated in the figures in the following sections, is primarily influenced by the quality of the camera system regarding its calibration.

5.1.2. Completeness and Quality of Point Clouds

An initial visual inspection of the results reveals that the point clouds in the attic area obtained using the M4P (Figure 11e,f) are less complete and exhibit more noise compared to those captured by the M3E (Figure 11b,c).

Some beams are completely missing, while others are only partially reconstructed, even though they were undoubtedly within the field of view, with suitable overlap between adjacent images, optimal baseline ratios, and similar object distances achieved, as shown in Figure 7.

The quality of point clouds can also be assessed photogrammetrically based on the confidence parameter, which Metashape calculates for each point in the cloud. Generally, the greater the number of depth maps (stereo pairs of images) used for the reconstruction of a point, the higher its confidence will be. The software also allows for point cloud filtering based on this parameter. As shown in Figure 12, the M4P produced significantly fewer reliable points in indoor spaces compared to the M3E, despite the use of additional lighting and sunny weather conditions.

The low confidence level observed in the M4P (right in Figure 12) is primarily related to the generation of point clouds from multiple depth maps and their subsequent fusion into a single compact point cloud within Metashape. As shown in Figure 13b, while the M4P cloud exhibits a higher point density, the dispersion of deviations from the regression plane aligned with the selected beam is nearly double that of the M3E (see Figure 13a).

In the cloud section depicted in Figure 13, there are approximately 1866 points obtained from the M3E and as many as 13,563 points obtained from the M4P. This significant difference in point count is attributed to the substantial discrepancies in the partial clouds generated from multiple depth maps with the M4P, which the software failed to effectively merge into a single compact point cloud, as was achieved with the M3E. It is assumed that Metashape retained all generated points from multiple depth maps even in overlapping areas, artificially increasing the point cloud density at the expense of an elevated noise level. Thus, the results indicate that a higher density of points in photogrammetric measurements does not necessarily equate to superior quality.

5.2. Comparison with TLS

As illustrated in Figure 14, a primary advantage of capturing the attic space using UAVs was the increased completeness of the scanned surfaces compared to TLS. The limited scanning angle due to the scanner’s position and the small size of the window opening only allowed for the scanning of a small portion of the attic, capturing fragments of beams and lathing.

The overlaps between the photogrammetric point clouds and the TLS cloud are minimal; however, in both cases, most photogrammetric points are located within 10 mm of the TLS points (green color), which provides sufficient accuracy for the documentation of architectural objects. On the other hand, the data from TLS are more accurate and less noisy. The maximum dispersion of points relative to the regression plane overlaid on the identical section of the beam illustrated in Figure 13 reached only ±3 mm for TLS.

6. Discussion and Conclusions

The utilization of various technologies, such as TLS, terrestrial and UAV photogrammetry, and geodetic measurements of control points, proves to be an effective approach for achieving comprehensive and precise documentation of heritage sites. Each of these methods has its advantages as well as limitations, which are important to consider when planning data collection. TLS provides exceptionally accurate data regarding the geometry of objects, which is invaluable for documenting architectural details and complex structures. Conversely, photogrammetry enables efficient capturing of textures and color information of the object at high resolution; however, its accuracy and reliability can be influenced by lighting conditions and the availability of camera positions. In this case, the combination of TLS and photogrammetry has proven to be the most suitable solution for enhancing the completeness of the documentation of the inaccessible attic of the mining house.

Although the DJI Mavic 3 Enterprise offered very good results due to its high-quality optics, the use of the smaller UAV, the DJI Mini 4 Pro, in supplementary measurements proved to be an effective solution. Despite its camera having a smaller sensor and lower resolution, its deployment allowed for the acquisition of data that remains usable for reconstructing the internal spaces of the object, all at a fourfold lower acquisition cost. With both UAVs, it was possible to reconstruct parts of the internal wall surfaces beneath the roof structure, which could not be captured by the terrestrial laser scanner (Figure 15).

Although the reconstruction was only partial, and in some areas limited to the pointing between the bricks, these details were sufficient for vectorization and determining the dimensions of the internal space, as well as for assessing the thickness of the walls on the second floor. From the measurement results, it is possible to confirm some indicators of the roof structure and only hypothesize about others, considering the established principles of construction. It is evident that this is a simple triangular roof construction with a tie beam, which reinforces the roof triangle at the upper third of the structure. The tie beam is connected to the rafter through lapping joints. Additionally, it can be inferred from the measurements that every third rafter is connected to a purlin to form a full truss (complete triangle). The first and last full trusses are placed on stone perimeter walls. In the center, the roof was likely later reinforced with a longitudinal tie beam placed perpendicular to the full trusses, on which vertical or sloping braces for the rafters are located. However, for a complete reconstruction of the structural principles of the attic space, additional information is needed regarding the connections of the rafters to the wall plates and how the rafters are connected to the purlins within the full trusses.

As established in the study, the origin of significant noise in the photogrammetric data should be attributed to a combination of unsuitable lighting conditions and, notably, the lower quality of the camera system and the resulting modulation transfer function of the less expensive UAV. In architectural applications, an accuracy and detail level of 1 cm is typically required, based on the commonly requested scale of drawings (1:100) and line thicknesses in analog form (0.1 mm). Therefore, utilizing a low-cost UAV to supplement data collection in hard-to-reach interior areas of architectural objects can still be justified despite the higher noise levels. Furthermore, smaller UAVs have the potential to penetrate interior spaces with protective rotor covers, allowing them to capture images not only from the outside through openings but also directly from within, providing entirely new perspectives on otherwise inaccessible structural elements. In this case, however, it would be necessary to disable obstacle detection sensors and rely solely on live image data during flight, which could pose risks for the UAV and lead to unwanted collisions with interior objects. The protective covers usually shield the rotors only from the sides and cannot prevent contact with overhanging elements when changing flight heights. We consider the examination of this aspect to be a primary limitation of the presented study, as we did not have protective rotor covers available, and we did not want to risk entering an otherwise inaccessible space without them.

The results of this study confirm that no single technology is capable of independently providing all the necessary data for comprehensive documentation of historical objects. Therefore, it is essential to combine various technologies and methodologies to achieve the best possible results. Even with the emergence of UAVs equipped with RTK systems, geodetic measurements of control points remain crucial, whether through GNSS or total stations, particularly for quality control and data unification from different technologies, allowing for precise transformation of models into national reference coordinate systems.

Future developments of these technologies may lead to even greater automation and integration between TLS and photogrammetric systems, potentially enhancing data processing and analysis efficiency. Overall, the use of TLS and photogrammetry for the documentation of heritage sites represents a robust and flexible approach that enables detailed and accurate reconstruction of historical objects.