Equipment and Software
The authors of the research decided to create a virtual 3D model on the basis of low-level aerial imagery. The crucial task was to create a cartometric 3D model that would reflect the geometry and actual size of the object studied and could thus be used for all the necessary surveys. The unmanned aerial vehicle DJI Phantom Advanced + , equipped with the camera with 1″ CMOS matrix and the 20 Mpx lens. The platform used is equipped with Global Navigation Satellite Systems (GNSS), however, it is used only to secure the correct flight of the platform, which means that the platform fails to use corrections collected from reference stations to determine the position more accurately or save precise coordinates for particular images. Moreover, the authors assumed that the model would be defined in the selected coordinate system so that it would be then possible to integrate it with other cartographic analyses. Such assumptions determined the necessity to carry out aerotriangulation of the images obtained on the basis of ground control points (Nex and Remondino 2014). The GNSS Trimble R4 model 3 of the receiver was used for surveying. The images were then converted to the 3D form in the dedicated Agisoft Metashape Professional software.
The first activity was to plan the flight path. In the research, it was conducted on the basis of orthophotomap obtained from a national geoportal (Fig. 1). The aim of preparing the flight path was to determine the procedure of obtaining images and locations of ground control points. Digital photogrammetry is a technique that by the use of algorithms allows one to reconstruct land’s topography as a 3D model providing spatial information from visible elements on two or more images obtained from different perspectives (Westoby 2012). On the basis of the images obtained it is possible to receive point clouds of a very high definition along with Digital Terrain Model (DTM), orthophotomaps and accurate 3D models of the land’s surface (Colomina and Molina 2014; Clapuyt et al. 2017). The process of creating photogrammetric visualisations is usually conducted with the use of software with the Structure-from-Motion (SfM) algorithm that can calculate a 3D model on the basis of the sequence of overlapping images that record the modelled object from different perspectives (Westoby 2012). To be able to use the SfM algorithm in calculations, the authors of the research decided to obtain the images in such a way as to record the modelled object from all sides. Moreover, as Rossi et al. (2018) point out, taking into consideration the time necessary to prepare the model and its low costs, compared to classic photogrammetric techniques, the SfM model can be repeated on a regular basis and thus effectively used for monitoring and studying changes that occur on the particular area in a given time. The choice of 3D modelling with the use of the SfM algorithm, along with the designed flight path, imposed the necessity to obtain oblique images. The use of oblique images obtained using UAV technology is particularly effective for registering cultural heritage objects, due to the possibility of reproducing their details during 3D modeling (Aicardi et al. 2016a, b).
During the preparation process, except for the suitable planning of the flight path, it is of great importance to plan the network of ground control points carefully (Eugster and Nebiker 2008; Wang et al. 2008; Barazzetti et al. 2010; Anai et al. 2012; Nex and Remondino 2014). According to Gonçalves and Henriques (2015), the GNSS receiver and the IMU system installed on the UAV’s platform are used mainly for navigation, stabilising the platform’s flight or the external orientation of the images obtained. Furthermore, Gonçalves and Henriques (Gonçalves and Henriques 2015) indicate that for photogrammetric use it is best to employ the GCP network. The process of aerotriangulation, based on GCPs assumed, is time-consuming but allows one to adjust the imagery to the particular coordinate system (Aicardi et al. 2016a, b; Gerke and Przybilla 2016; Padró et al. 2019). Salvini and others (2015) indicate that surveying GCPs is possible through tacheometric surveying, which provides the best angular and linear accuracy between GCPs. Real-Time Kinematic (RTK) measurements ensure obtaining satisfactory measurement results, hence they are often used to measure photogrammetric control points (Clapuyt et al. 2016; de Kock and Gallacher 2016). It should be emphasized, that measurements of the GCPs with the use of RTK technology are possible only when we have access to differential corrections. It is mainly related to the area of measurements. They differ in terms of access to the network of permanent reference stations and with access to the cellular network, through which corrections are sent with appropriate protocols. When access to differential corrections is impossible, it becomes necessary to perform classic tacheometric measurements. It should also be noted that while it is possible to perform satellite measurements, currently the measurement of the GCPs is possible through hybrid measurements using two techniques: tacheometric and satellite. Such a solution allows to obtain overtime observations, increasing the certainty of the measurement. Having considered the necessity to obtain oblique images to record the windmill from each angle as accurately as possible, it was established that the points should be located around the object, which would secure visibility of GCPs on all the images obtained. This approach resulted in establishing four GCPs on the opposite sides of the windmill along its diagonals, which allowed one to receive a particular “frame” in the centre of which the windmill was situated. The location of GCPs in the research area was demonstrated by white points in Fig. 4 (1–4 points).
Surveying each point included 30 periods with the recording interval of 1 s and surface corrections from the nearest reference station, located by the vector of 23 km from the research area, were used during the process. Table 1 presents the levelled coordinates of GCPs. The coordinates of GCPs, obtained through RTK, were defined in the rectangular PL-2000 coordinate system (zone 6) that was effective in Poland at the time.
Conducting the Flight
Weather conditions during the flight path were favourable. There was no rain, the wind speed was app. 4 m/s and the sky was clear. The flight took place at the average altitude of 35 m AGL, along the flight path on different ceilings and lasted app. 20 min. As a result of the flight path, 51 images of the 5472 × 3078 definition recording the windmill from all sides were obtained (Fig. 5).
Aerotriangulation, whose aim was to define the images obtained in the selected coordinate system, was another stage of the research. The first step was to reconstruct their mutual internal orientation (Siebert and Teizer 2014). In the research, the images were adjusted to the same coordinate system in which GCPs were defined. Reconstructing mutual internal orientation is possible thanks to EXIF files of the images obtained that include metadata describing the images. It allows one to determine their approximate location in space. Then, the mutual orientation of images was reconstructed again, however, this time on the basis of GCPs assumed that had been previously surveyed by means of GNSS. The Agisoft Metashape Professional software was used for this purpose. According to Uysal et al. (2015), the software mentioned is used particularly frequently for working out images obtained from UAVs and allows one to generate Digital Terrain Model and orthophotomaps in the coordinate system defined by the user. As a result of aerotriangulation based on GCPs, the value of RMSE, describing deviations between tie points and the points calculated from the photogrammetric model generated, was calculated. On the basis of the results from Table 2, one can conclude that point no 2 had the largest error (1.5 cm), and the smallest error was determined for point no 4 (0.8 cm). The average value of RMSE calculated for all GCPs was 1.2 cm.
Creating a 3D Model
The next stage of creating a 3D model of the windmill studied was to generate 3D point clouds based on low-level aerial imagery. In the research, the original size of the imagery obtained (i.e., 5472 × 3078 pixels) was used for the purpose, which allowed to get 1 cm of ground sampling distance (GSD). The point cloud hereby obtained consisted of 17,084,208 points and its density was 2310 point/m2. Then, using the point cloud, the 3D model of the land’s surface, consisting of 1,138,925 planes connected by 576,902 vertices, was generated. Figure 6 demonstrates the model with the texture applied.
The 3D model, calculated in Agisoft Metashape Professional, was exported to OBJ format in the WGS 84 (EPSG:4326) coordinate system.
3D Model of Archival Map
Next, the 3D model of the archival map was created. The terrain included a small area around the windmill in the village of Kamionka. To create the model, the authors used a part of Messtischblätter from 1911 at the scale of 1:25,000. The map included a contour line drawing of the 1.25 interval (Lorek et al. 2018; Lorek and Medyńska-Gulij 2019; Cybulski et al. 2020). Each contour line was vectorised as a polygon (Fig. 7a) in QGis 3.12 (tracing) and assigned height in tabular data (Horbiński and Medyńska-Gulij 2017).
The vectors obtained were used for generating a digital elevation model (DEM). The authors employed the Rasterize (Vector to Raster) tool (Fig. 7b).
Then, the authors moved on to creating the 3D model, using the qgis2threejs plugin to QGIS 3.12. The plugin allows one to create a 3D model for the map in Raster or Vector format (Horbiński and Medyńska-Gulij 2017). It also makes it possible to insert other 3D elements (e.g., simple solid figures), however, for model files there are still problems with loading and, particularly, exporting the finished project. Therefore, the authors decided to create the 3D model of the archival map in qgis2threejs without inserting the previously created 3D model of the existing historical topographic object, namely the windmill. During the creation of the 3D model of the archival map the authors used the generated DEM onto which the archival map was applied. Resampling level was set at the value of 1 (Grid Size: 138 × 73 px, Grid Spacing: 0.00016 × 0.00016 px) to tone down the transition between particular heights assigned during vectorisation. Hence, the 3D model presented in Fig. 7c, which was then exported to the gITF format and will be used in the next stages, was produced.