Geological field mapping is a vital first step in understanding geological processes. During the 20th century, mapping was revolutionized through advances in remote sensing technology. With the recent availability of low-cost remotely piloted aircraft (RPA), field geologists now routinely carry out aerial imaging without the need to use satellite, helicopter, or airplane systems. RPA photographs are processed by photo-based three-dimensional (3-D) reconstruction software, which uses structure-from-motion and multi-view stereo algorithms to create an ultra-high-resolution, 3-D point cloud of a region or target outcrop. These point clouds are analyzed to extract the orientation of geological structures and strata, and are also used to create digital elevation models and photorealistic 3-D models. However, this technique has only recently been used for structural mapping. Here, we outline a workflow starting with RPA data acquisition, followed by photo-based 3-D reconstruction, and ending with a 3-D geological model. The Jabal Hafit anticline in the United Arab Emirates was selected to demonstrate this workflow. At this anticline, outcrop exposure is excellent and the terrain is challenging to navigate due to areas of high relief. This makes for an ideal RPA mapping site and provides a good indication of how practical this method may be for the field geologist. Results confirm that RPA photo-based 3-D reconstruction mapping is an accurate and cost-efficient remote sensing method for geological mapping.
Mapping the Earth’s surface is a fundamental step in understanding geological processes. Traditionally, this involved a geologist, on foot, walking through an area, making observations and measurements using a hand compass and notebook. Early mapping expeditions commonly involved surveyors and geologists simultaneously undertaking topographic work and mapping rock outcrop. This limited the accuracy of geological maps, as the locations of lithological units, their contacts, and structural traces across the topography were not well constrained. The accuracy of measuring the location of geological contacts is important, because where geology intersects the topographic surface, it forms distinct map patterns. These patterns are used to measure dip and dip direction of units and structures as well as the plunge and direction of folds. The measurements are used to interpret the subsurface, three-dimensional (3-D) geometry of an area. Therefore, advancements in mapping the location of the surface geology as well as increasing topographic accuracy significantly improve the potential accuracy of geological interpretations.
Several major technological advancements for geologists began with the use of aerial photography and photogrammetry (Ray, 1960). Following this, the first image of Earth from orbit was captured by the Explorer 6 satellite in 1959 (Gupta, 2018), which improved the photographed coverage of Earth. Additionally, the introduction of Global Navigation Satellite Systems (GNSS) has allowed for geospatial positioning. This, combined with the availability of portable computers, means that geological outcrops can be measured more accurately (Pavlis et al., 2010). However, acquisition of these types of remote sensing data sets carries a high financial cost—that is, until recently, due to the availability of commercial remotely piloted aircraft (RPA), commonly referred to as unmanned aerial vehicles (UAVs).
Progress in software and computer technologies also allows even more geological data to be extracted from the images captured by remote sensing methods. An example of this is photo-based 3-D reconstruction, where a 3-D surface is created from a series of overlapping photographs taken from a wide array of angles (Parra et al., 2017; Pavlis and Mason, 2017). Photo-based 3-D reconstruction uses a structure-from-motion (SfM) algorithm to identify the same features (keypoints) from multiple photographs, from which it is then able to calculate an x-y-z location of each keypoint. These locations are internal coordinates, and the reconstruction will have an unknown scale factor. To align the model with real-world coordinates, it must be georeferenced by performing a rigid transformation (scale, rotation, and translation). Thousands of keypoints are calculated per photograph and form a sparse point cloud, which is used to calculate the cameras’ relative position and alignment (Remondino and El-Hakim, 2006; Remondino et al., 2011; Westoby et al., 2012; Bemis et al., 2014; Nex and Remondino, 2014). Additionally, new approaches are being developed to analyze RPA photosets and filter unnecessary photographs to help speed up the SfM process (Dhanda et al., 2018). Once a SfM sparse point cloud is created and cameras are correctly aligned, a multi-view stereo (MVS) algorithm creates a dense point cloud. The dense point cloud is then used to extract a mesh or digital elevation model (DEM).
Such an elevation model can also be referred to as a digital surface model (DSM), a digital terrain model (DTM), or a digital outcrop model (DOM). These terms are often used interchangeably, but some operators may use them with minor differences. In general, DEM is the generic term that includes DSMs, DTMs, and DOMs. A DSM refers to the whole surface with vegetation and buildings, whereas a DTM is only of the Earth’s bare surface (cover removed). DOM is a term used for surface models of geological outcrops, which usually are of small areas.
Textures are then applied to the mesh, creating a photorealistic 3-D model. While SfM and MVS were developed in the 1990s, they have only become widely accessible since ca. 2012. This is due to the availability of user-friendly photo-based 3-D reconstruction software and major advancements in computing power. With the recent addition of low-cost RPA, the ability to operate a camera over any terrain (as opposed to only ground-based photography), combined with photo-based 3-D reconstruction software, provides geologists with a practical remote sensing tool.
Light detection and ranging (lidar) was, until ca. 2012, the preferred topographic surveying method for large spatial coverage in geological applications (Bellian et al., 2005; Buckley et al., 2010; Hartzell et al., 2014). Lidar is a technique where a pulsed laser emits light at a surface which is then reflected back to a sensor. This can be either a ground-based or an airborne setup. The measured time it takes for the light to return to the sensor is used to create a 3-D model of that surface. Lidar is generally more capable of higher resolution (Adams and Chandler, 2002; Nouwakpo et al., 2016) than photo-based 3-D reconstruction can provide. However, photo-based 3-D reconstruction can be used to achieve better resolution than lidar if correctly carried out (Favalli et al., 2012; Fonstad et al., 2013; Javernick et al., 2014; Johnson et al., 2014; Cawood et al., 2017; Corradetti et al., 2017), and is several orders of magnitude more cost efficient. The primary advantage lidar has for geological applications is its capability of penetrating foliage, allowing topographic mapping even in forested areas (Lebow et al., 2017).
In geology, photo-based 3-D reconstruction of RPA data is mostly used for small (<2 km2) areas, such as for monitoring coastal topographic changes (Mancini et al., 2013; Gonçalves and Henriques, 2015), studying landslides (Niethammer et al., 2012; Lucieer et al., 2014) and lava flows (Carr et al., 2019), and mapping geological outcrops and structures (Favalli et al., 2012; Bemis et al., 2014; Vasuki et al., 2014; Johnson et al., 2014; Bistacchi et al., 2015; Chesley et al., 2017; Piras et al., 2017; Saputra et al., 2017, 2018; Corradetti et al., 2018; Fleming and Pavlis, 2018; Madjid et al., 2018). This method has huge potential for efficient large-scale (>2 km2) geological mapping, where there is substantial geologic exposure, when compared to ground-based techniques (Brush et al., 2018). This is because a RPA can quickly cover a large area, and the photo-based 3-D reconstruction is used to measure the orientation of rock layers and faults (dip and dip direction). This means that measurements can be taken anywhere on the 3-D model, which is faster than fieldwork. In addition to increased time efficiency, there are algorithms being developed that automatically digitize and map out features such as rock fractures and faults in photo-based 3D reconstructions (Vasuki et al., 2014). Furthermore, measurements can be taken in locations on the model that are inaccessible in the field. Also, a 3-D model can be analyzed at different scales, from spot measurements to larger areas. Measuring larger areas removes the scatter associated with compass measurements, which are commonly taken on an irregular surface. Finally, a photorealistic 3-D model can easily be revisited to test repeatability and reproducibility of measurements without additional fieldwork.
There are many software options (free and commercial) that provide photo-based 3-D reconstruction with various options (e.g., 3DF Zephyr [https://www.3dflow.net/], Agisoft Metashape (previously named Photoscan) [https://www.agisoft.com/], AutoDesk ReCap [https://www.autodesk.com/products/recap/overview], DroneDeploy [https://www.dronedeploy.com/], Pix4D [https://www.pix4d.com/], VisualSFM (http://ccwu.me/vsfm/), and RealityCapture [https://www.capturingreality.com/]). In this study, 3DF Zephyr Aerial was used.
This survey was carried out as economically as possible over a large area of 3.32 km2, minimizing high-cost specialist tools. Here, a workflow is presented (Fig. 1) that follows four main steps: (1) RPA data acquisition and fieldwork, (2) photo-based 3-D reconstruction, (3) extraction of geological information from the reconstruction, and (4) development of a 3-D geological model. Many tools exist to carry out the 3-D geological modeling (Aswar and Ullagaddi, 2017); here, MoveTM (https://www.mve.com/) was used. Other programs include GeoModeller (https://www.intrepid-geophysics.com/ig/index.php?page=geomodeller), Leapfrog Geo (http://www.leapfrog3d.com/products/leapfrog-geo), SKUA-GOCAD (http://www.pdgm.com/products/skua-gocad/), and Virtual Reality Geological Studio (VRGS; http://www.vrgeoscience.com/). Other useful tools for editing point clouds and meshes generated from photo-based 3D reconstructions include CloudCompare (https://www.danielgm.net/cc/), LIME (http://virtualoutcrop.com/lime), and MeshLab (http://www.meshlab.net/). This four-step workflow results in the generation of a photorealistic 3-D model that can be structurally analyzed, and demonstrates the logistics of relatively large-scale mapping using RPA photo-based 3-D reconstruction.
A transect across the Jabal Hafit anticline, in the United Arab Emirates (UAE), was chosen for the RPA photo-based 3-D reconstruction (Fig. 2). The outcrop exposure of this structure is excellent and vegetation is sparse. The Jabal Hafit anticline is located in the foreland of the Al Hajar Mountains in Oman and the UAE. It has a surface expression of 25 km × 5 km, and protrudes 900 m above the surrounding plains. The area mapped was 3.32 km2 (∼4 km length by 0.8 km width), and includes a 470 m change in relief with some areas being flat and others having cliff faces of up to 170 m in vertical height.
The structure consists of carbonate rocks of early Eocene to middle Miocene age, which have been folded, uplifted, and exposed at the surface through erosion (Warrak, 1987). The subsurface geometry of the Jabal Hafit anticline was previously described with a 3-D geological model that was constrained through detailed mapping and seismic data (Hansman and Ring, 2018). For these reasons (excellent geology exposure and rugged terrain), the Jabal Hafit anticline is an ideal locality for RPA photo-based 3-D reconstruction mapping.
Step 1: Remotely Piloted Aircraft System
To carry out RPA data acquisition, a DJI Phantom 4 quadcopter with three 5350 mAh batteries was used. An Apple iPad Mini4 32 GB was connected to the remote controller, and two iOS applications were used to control the RPA. (1) The DJI GO 4 app (https://www.dji.com/downloads/djiapp/dji-go-4) was used for manual flight around cliff faces with the camera tilted toward the horizon to maintain a perpendicular view of the rock surface. (2) The Drones Made Easy Map Pilot app (https://www.dronesmadeeasy.com/Articles.asp?ID=254) was used with the terrain aware function for a fully automated flight mission. Before an automated flight, the operator defines an area in the app to be mapped. The overlap between photographs and the resolution, which is related to the RPA altitude above ground level, is also set. After the flight path is calculated, the app uses the elevation provided by the Shuttle Radar Topography Mission (SRTM) 1 Arc-Second Global data set to adjust the altitude of the flight path to maintain a constant height above ground level. This is important because if the RPA height increases, the photographs will have a lower ground resolution, and if the height decreases, the overlap between photographs may not be sufficient. A constant height above the terrain provides the best results from the SfM and MVS algorithms. The app then loads the flight path into the RPA, which will then take off, fly the mission, and land autonomously. The Map Pilot app was used for terrain that had low relief variation. At the time of mapping in November 2017, a limitation of this app was that the camera could not be tilted and could only point down vertically. Therefore, the RPA was manually flown around cliff faces to maintain a perpendicular camera orientation to rock surfaces. Also, the maximum number of waypoints that can be loaded into the RPA per mission is limited to 99. However, an area with high relief requires many waypoints to continually adjust the height above ground level. Therefore, the mapping area may need to be reduced, or the altitude increased per flight mission. The SRTM data set has a 30 m resolution, which means that some cliff faces may not be navigated around correctly; this is another reason why the operator should manually fly the drone around vertical cliffs with the DJI GO app. An operator should ensure that their workflow is suitable to carry out the data collection. This is dependent on what terrain types are present in a mapping area, as automated mapping routines may not be suitable for high-relief terrain.
The Phantom 4 has a visible-spectrum (RGB) wide-angle camera, with a focal length of 3.6 mm (35 mm camera equivalent: 20 mm) and a 1/2.3-inch (6.17 mm × 4.55 mm) complementary metal-oxide semiconductor (CMOS) sensor. A total of 5728 photographs were taken at 4000 × 3000 pixels. This data set with image metadata is available online (https://doi.org/10.6084/m9.figshare.7830512), following the guidelines proposed by O’Connor et al. (2017). The height above ground level was between 50 and 80 m, resulting in a resolution of 2.2 cm/pixel to 3.5 cm/pixel. The Map Pilot app calculates the centimeters per pixel for each flight. However, for manual flight, this needs to be calculated and is called the ground sample distance (GSD). The GSD is dependent on the image size, focal length, sensor size, and the height above ground. This needs to be decided on before fieldwork, depending on the project requirements, as it directly impacts the resolution and accuracy of the final photo-based 3-D reconstruction. With the autonomous Map Pilot app, the along-track and across-track overlap of the photographs can be preprogrammed into the flight plan, and both were set to 80%. However, when these were combined with the photographs from manual flying, the overlap of the photograph data set was 60%–96%. The photographs were acquired over 40 flights (9 h 30 min total flight time), where one battery per flight was used. Two to three flights were carried out in the morning and the same number in the afternoon, with batteries being charged at midday (∼1 h per battery). Nine days were required for data acquisition, given the requirement for access to electricity in the vicinity of the outcrop. Ambient temperature ranged between 28 and 34 °C with wind speeds of 0–8 m/s and occasional gusts of up to 15 m/s. The RPA had no issues flying in this environment.
Step 2: Photo-Based 3-D Reconstruction
For the photo-based 3-D reconstruction, a high-performance cloud computer was used to run the reconstruction software. Initially Paperspace (https://www.paperspace.com/) was used, however the system memory was not sufficient. The project was then completed on an Amazon Web Services (AWS; https://aws.amazon.com/) instance (Table 1).
Software: 3DF Zephyr Aerial
The photographs were imported into 3DF Zephyr Aerial Education (ver. 3.503), a photo-based 3-D reconstruction software package. RPA camera autocalibration (FC330–4 mm 4000 × 3000 pixels) was used to fix radial and tangential distortion. The following five processes were carried out. First, the SfM algorithm (Samantha) was run to orient the cameras and generate a sparse point cloud using the Aerial and Default presets (Table 2). A photogrammetry model with >1000 photographs is considered large, therefore this project was split into four areas to create the sparse point clouds, which were then merged. The split areas did not contain any duplicate photographs along the boundaries. This is because the across-track overlap was >80% and is sufficient to not require a larger buffer. When merging two areas, the same three points were selected in each one to assist 3DF Zephyr with the merge. The sparse point cloud was manually edited to remove noise (random points above ground level) and points outside the mapping area. The second process was to carry out rigid scaling using ground control points (GCPs) with GNSS measured locations. Third, the MVS algorithm (Stasia) was used to create a dense point cloud using the Aerial and Default presets. The dense point cloud also contains noise, and the Confidence Tool was used to remove points that 3DF Zephyr had a low confidence in. Fourth, a mesh (using the Sasha algorithm) was created with the following settings: Aerial and Default–Sharp Features presets; Photoconsistency Based Optimization: On; Number of Neighbor Cameras: 3; and Image Resolution: 50%. Lastly, a textured mesh was created on default settings with Image Resolution: 100%. This creates a single texture with a resolution of 8192 × 8192 pixels, which is a size limit of most graphics processing units. Refer to Table 3 for processing times of each stage.
Ground Control Points
Thirteen markers were placed throughout the mapping area and their GNSS coordinates measured to use as ground control points (GCP). Each marker was a black-and-white checker box printed onto A3 paper, taped to the ground (Fig. 3). These markers were identified in the photographs within 3DF Zephyr Aerial and were used to scale and orient the model to improve accuracy and georeference the model into real-world coordinates. The horizontal datum for the model is World Geodetic System 1984 (WGS84), and the vertical datum is Earth Gravitational Model 1996 (EGM96). The quality of the GCP GNSS measurements (Table 4) is indicated by the position dilution of precision (PDOP). This is a measure of the satellite geometry to the GNSS receiver (Langley, 1999). The greater the number of satellites and the greater their dispersion, the better the accuracy (and precision) the receiver can achieve. A PDOP value <4.00 provides excellent positions. The estimated horizontal accuracy of the GCPs in the field (Table 4), with uncorrected (autonomous) GNSS, was 3.2–6.9 m, RMS (root mean square) 63% confidence. This was considered acceptable due to the size of the mapping area (i.e., 4.0 km × 0.8 km). If possible, collecting GNSS GCP positions with decimeter precision is ideal. The accuracy of autonomous GNSS is generally >3 m. With good PDOP values, GNSS corrections can be carried out to achieve decimeter to centimeter precision. This may involve receiving corrections from satellites (and ground stations) using a satellite-based augmentation system (SBAS), such as the Wide Area Augmentation System (WAAS, https://www.faa.gov/about/office_org/headquarters_offices/ato/service_units/techops/navservices/gnss/waas/) in North America. Otherwise, various commercial SBASs are available such as OmniSTAR (http://www.omnistar.com/) and Fugro (https://www.fugro.com/). Additionally, corrections can come from a ground-based augmentation system (GBAS), such as differential-GNSS (DGNSS), real-time kinematic (RTK) positioning, and precise point positioning (PPP), which use ground stations. For smaller areas, the GNSS error will have a larger impact on the accuracy of the model, and may still not correctly orient the model. It is possible to measure a plane in the field with a compass and perform a rigid-body rotation to correct this (Fleming and Pavlis, 2018). The accuracy requirement of a project as well as the available GNSS corrections in the mapping area need to be considered prior to fieldwork.
In addition to the GNSS accuracy of the GCPs, the GCP distribution throughout the field area can contribute to significant spatial errors (Brush et al., 2018; Oniga et al., 2018). GCP density and coverage across the mapping area is critical, as is placement of GCPs at a range of vertical elevations. To assess the spatial error in a photo-based 3-D reconstruction model, a subset of GCPs can be created in the software without adding the corresponding coordinate locations. Then the locations of these GCPs in the output model can be compared with measured GNSS coordinate values. These non-constraint (or control) GCPs can then indicate the spatial error within the model by allowing comparison of real-world coordinates with those of the model (Table 4).
Step 3: Extracting Geological Data
The dense point cloud file (.ptx) was exported from 3DF Zephyr Aerial and imported into Maptek PointStudio (ver. 8; https://www.maptek.com/products/pointstudio/). Note that PointStudio replaced I-Site Studio in 2018. Bedding, contacts, and structures were digitized using polylines in PointStudio. To extract geological data, it is generally better to work on the dense point cloud, as this can be more accurate than the textured triangular mesh. This is because the mesh is extracted from the dense point cloud. Depending on the options used, the meshing algorithm will interpolate a smoother or sharper surface, which may change the actual shape of the terrain. However, in 3DF Zephyr, there is an option to use photoconsistency mesh optimization when creating a mesh, which will recover finer details than the original point cloud. If this is used, the computational time will significantly increase, which may not be practical for large data sets.
There are two ways to extract dip data from the point cloud. The first is used when the planar surface is not exposed and only a perpendicular view of the plane is observable (Figs. 4A and 4B). The planar surface is traced with a polyline and then converted into a surface (Figs. 4C and 4D). The dip and strike query geotechnical tool in PointStudio can then measure the orientation. The second method is used when a planar surface is exposed (Fig. 5A). Points on the surface are selected, and then the geotechnical tool can measure the orientation of that selection. Different-sized areas can be measured (Figs. 5B and 5C), depending on the amount of noise in the point cloud or how rough the surface is in the field. A larger area of the same surface will provide a better average because the point selection can be carried out at various scales on the point cloud. This provides an advantage over single compass hand measurements, which are more subject to error. This is because the azimuth and inclinometer scales can be difficult to read on handheld compasses, and surfaces being measured are commonly rough, which creates scatter between individual measurements. These polylines and dip data can then be imported into MoveTM.
Step 4: Developing the 3-D Geological Model
Finally, to create a 3-D geological model, the textured mesh from 3DF Zephyr Aerial was imported into MoveTM (previously by Midland Valley Exploration Ltd., but during 2018 transitioned to Petroleum Experts Ltd.) as an object file (.obj). This was combined with the geological data extracted from the point cloud, giving a true 3-D geological terrain model. Any additional data, such as well logs and seismic surveys can be added at this stage. From this, a complete 3-D geological model can be created, which involves projecting the geology above and below the topographic surface (mesh). This is accomplished by creating vertical two-dimensional (2-D) cross-sections throughout the model and projecting onto the sections the intersecting topography, stratigraphic contacts, and dip data. Deformation within the 2-D sections is then restored, which includes faulting and folding. The areas of the stratigraphic layers and the line lengths of the layer contacts should remain constant between the final deformed section and the initial, undeformed section to create a balanced 2-D cross-section (Dahlstrom, 1969; Lingrey and Vidal-Royo, 2015). If the section cannot be balanced and there is no known geological reason to explain why (e.g., non-plane strain, erosion, or variation in bed thickness), the subsurface geology needs to be reinterpreted until it is balanced.
Once the cross-sections are balanced, the polylines from PointStudio along the terrain and the corresponding polylines interpreted in the 2-D cross-sections are converted into surfaces to complete the 3-D geological model. This is achieved using a Delaunay triangulation method, which honors the point data of the polylines.
RESULTS AND DISCUSSION
Following the workflow from steps 1 and 2, a RPA photo-based 3-D reconstruction of the Jabal Hafit anticline was carried out using 3DF Zephyr Aerial (Fig. 6). Outputs from this model include an orthophoto (vertical view), a DEM, and a textured mesh (Fig. 7). The quality of these outputs cannot solely be assessed by the DEM resolution, number of points in the cloud, or number of polygons in the mesh. This is because more points do not necessarily mean more accuracy (rather, they could be more noise). A visual comparison between the textured mesh (Figs. 8A and 8B) and real-world photographs (Figs. 8C and 8D) provides a qualitative analysis. However, the precision and accuracy must be tested by comparing real-world measurements with those of the 3-D model. Additionally, during steps 1 and 2, two issues were identified. First, the processing times for the SfM and MVS algorithms were higher than expected. Second, the MVS point cloud density was low. These issues are discussed in the sections below.
Carrying out steps 3 and 4 of the workflow, a balanced cross-section (Fig. 9A) was created in MoveTM constrained by the surface data extracted from the photo-based 3-D reconstruction. This was then used to develop the 3-D geological model (Figs. 9B and 10). This model supplements a larger-scale 3-D geological model of the entire Jabal Hafit anticline (Hansman and Ring, 2018), which was created from ground-based field mapping to quantify the amount of horizontal shortening and the deformation style. A significant advantage that the photo-based 3-D reconstruction provided was the high-resolution DEM. From this, a topographic profile was projected to the cross-section (Fig. 9) and, when combined with lithology contact and dip data, allowed for accurate measurement of the stratigraphic thickness. This was especially useful for estimating the thickness of layer D2, which is constrained by the limbs of the anticline, as well as defining a minimum thickness for layer R (Figs. 9 and 10). Due to the large scale of these stratigraphic layers, it is difficult to directly measure their thicknesses in the field. Therefore, a high-resolution DEM is essential in extracting this type of measurement.
Another advantage of the photo-based 3-D reconstruction over traditional field-based mapping was the ability to extract dip data anywhere in the model, providing good coverage without having spatial data gaps in the mapping area. This can be an issue when carrying out field mapping by foot because geographic barriers, such as rivers and cliffs, can make areas unreachable. In addition, the ability to project these dips onto the photo-based 3-D reconstruction allows for accurate vertical placement of the data. This was especially important in interpreting the geometry of the overturned eastern limb (Fig. 9). In contrast, if a geologist were only to collect individual hand-compass measurements on a 2-D map that were then later combined with a lower-resolution DEM (e.g., SRTM) in MoveTM, the correct geometry of the overturned limb would be difficult or impossible to replicate accurately. Photo-based 3-D reconstruction really outperforms in areas of steep terrain. However, this method may not provide any significant benefit over flat relief.
Photo-Based 3-D Reconstruction Accuracy
The control (or the non-constraint) GCPs indicate the error within the photo-based 3-D reconstruction (Table 4). This error is the distance between the GNSS measured coordinate of the GCP and the coordinates computed for the model. The error calculated by 3DF Zephyr Aerial does not take into account the GNSS errors. Combining the GNSS error with 3DF Zephyr’s computer error will provide the real-world error of the reconstruction, which in this model could be as much as ∼ 15 m. This shows that it is essential to collect accurate GNSS coordinates to provide a better assessment of the error within a reconstruction. Therefore, only a real-time kinematic (RTK) or post-processing kinematic (PPK) GNSS receiver should be used.
When transforming (scaling, rotating, and translating) a reconstruction using GCPs, an additional bundle adjustment can be carried out. Bundle adjustment is a nonlinear optimization method that minimizes the reprojection error of points within a sparse point cloud. The error is measured by reprojecting the predicted 3-D location of a point back to the images it was calculated from. The distance between the reprojected point and the actual point on the image is the reprojection error. For a large data set, scaling with bundle adjustment will considerably increase the processing time (Table 3). Performing a rigid transformation without bundle adjustment will only take minutes. Therefore, it is recommended not to use bundle adjustment for the first transformation. However, if the errors of the GCPs are not acceptable, then running the transformation again with bundle adjustment may resolve this.
Density of the MVS Point Cloud
The density of the MVS point cloud generated was extremely low, with only 4 × 106 points (Table 3). Ideally, 5 × 107 to 1 × 109 points would be generated from 5728 images at this scale. In 3DF Zephyr, the default settings were applied for the MVS step (Table 2). While this dense point cloud can still be used to measure structures at larger scales, it is not dense enough to observe finer geological details.
A second MVS dense point cloud was generated, using the same settings on the same SfM sparse point cloud, but only changing the bounding box dimension (Fig. 11A, bounding box 2). The new MVS dense point cloud (Fig. 11B; also used in Figs. 4 and 5) contained 6 × 106 points. Processing time on the AWS computer was 21 h 40 min. The density of this point cloud is excellent at this scale. Finer features can be observed, and digitizing is easier in PointStudio (Fig. 11B).
The MVS point cloud generated in version 3.5 of 3DF Zephyr is strongly influenced by the size of the bounding box. The default Aerial settings are suitable for areas smaller than 500 × 500 m and should generate ∼5 × 106 points. For larger areas, the main settings to increase are the image input resolution and the discretization values (Table 2). These values directly control the number of points generated. With the new version of 3DF Zephyr (ver. 4.0), the SfM and MVS algorithms have been improved, and the bounding box does not have the same affect on the amount of points generated compared to version 3.5.
When processing a large mapping area, it may be beneficial to carry out the photo-based 3-D reconstruction in “chunks”. Dividing the total area into smaller, more manageable zones that have similar properties (flat versus steep terrain) has two benefits. First, it allows the SfM and MVS algorithms to be processed faster, which increases efficiency when running multiple models to optimize the settings to use within 3DF Zephyr Aerial (or comparable software). Secondly, the software can be optimized for smooth or sharp surfaces. This means that more accurate meshes can be created, because flat terrains may require increased smoothness to remove noise, whereas rugged terrain will need more sharpness to maintain surface edges.
Photo-Based 3-D Reconstruction Processing Time
The computer processing time for running the SfM and MVS steps was high, considering the hardware used. One reason for this is the bundle adjustment, which was carried out after merging (Table 3), which added significant time, but this is not always necessary. Generating the second MVS point cloud, using bounding box 2 (Fig. 11), shows that there is too much overlap in this data set. This is demonstrated by the fact that with bounding box 2, the MVS algorithm processed 1462 images, which is about one-quarter of the entire data set, even though bounding box 2 is significantly smaller than one-quarter of the study area. The reason for this is that bounding box 2 is located on the western margin of the mapping area, where there are many cliff faces. Images acquired here are directed toward the cliff, and are subparallel to the horizon. This results in many images having expansive backgrounds and high overlap from photographs at long distances, which do not add to the accuracy of the model (high GSD). To solve this, masking should be carried out on these photographs. The Masquerade tool in 3DF Zephyr Aerial can be used to define the foreground and background of each image. The background pixels will not be used in the reconstruction (Fig. 12). This will significantly speed up the computer processing time for the SfM and MVS algorithms. This step has been added to the workflow diagram (Fig. 1), and is crucial to improving the performance of 3DF Zephyr Aerial before starting the SfM and MVS processes.
A workflow from RPA fieldwork to a photo-based 3-D reconstruction to a 3-D geological model has been outlined. This paper helps to clarify the logistics for this workflow and provides examples of what hardware and software to use. However, this workflow requires moving data through multiple software packages. A critical issue with this is that the outputs (point clouds and meshes) from photo-based 3-D reconstruction software are not always easily imported into 3-D geological modeling software. Additionally, visualizing and working on these data in 3-D geological modeling programs is commonly associated with poor software performance. Currently, for earth science research, there is a significant need for improved compatibility between these applications, or a software suite that can perform all tasks efficiently.
Software issues aside, the only major limitation when applying this method is when the target area has obstructions to the view of the geological exposure, for example, terrains with dense vegetation cover. In these cases, this method is not suitable. In addition, two pitfalls have been identified: (1) the default MVS settings in 3DF Zephyr Aerial do not scale to large areas adequately, and (2) failing to mask the subhorizontal images significantly increases processing times of the photo-based 3-D reconstruction. The benefits of this workflow becoming a standard for geological mapping, as opposed to pen-and-pad ground-based fieldwork, are numerous. The photo-based 3-D reconstruction allows for considerable analysis to be carried out without the need for additional fieldwork. It is faster to take digital measurements anywhere in a reconstruction without the need to physically be in the field, and locations can be measured even when they are inaccessible by foot. Measurements can also be taken at any scale, from spot measurements (comparable to a hand compass) to larger areas. This provides the ability to quickly and accurately measure irregular or weathered surfaces that would otherwise require multiple hand-compass measurements. A photo-based 3-D reconstruction can be revisited as many times as needed and shared without additional, and potentially expensive, fieldwork. The repeatability and reproducibility of field observation can be carried out effortlessly, and new observations can be made or previous ones tested. A reconstruction also provides an historical snapshot in time of the geology in an area. This is invaluable, as in some localities, geological outcrops are not permanent. For example, geology may be covered or removed by civil engineering projects or excavated through mining, and recent fault ruptures can be weathered or erased by human activity. A snapshot also allows for the analysis of geological processes through comparison of photo-based 3-D reconstructions created before and after an event. This means that erosion, land mass movements, lava flows, and volcanic craters can be accurately studied in a way that was never possible before photo-based 3-D reconstruction became available. Once again, technology is transforming geological mapping, and photo-based 3-D reconstructions are becoming standard practice for field geologists.
This work was funded by the Royal Swedish Academy of Sciences (GS2017-0028) and the Bolin Centre for Climate Research Area 6. Uwe Ring acknowledges funding from Stiftelsen Anna-Greta och Holger Crafoords fond and Stockholm University. Much thanks to Dr. Terry Pavlis for passing on his knowledge regarding the photo-based 3-D reconstruction methodology to us. We are grateful to 3DFlow for providing us with 3DF Zephyr Aerial Education, and especially would like to thank Roberto Toldo and Andrea Alessi for their rapid assistance in helping us with this software. We also acknowledge the use of the MoveTM software suite granted by Midland Valley’s Academic Software Initiative, as well as thank Maptek for giving us an academic license for PointStudio. Finally, we are extremely thankful for the constructive reviews from Stefano Tavani and an anonymous reviewer, whose comments have greatly improved this paper. The RPA photographs used in this study are available online and are open access (https://doi.org/10.6084/m9.figshare.7830512).