1 Introduction

Field observation on surface orientation is time-and-cost intensive and limited by factors such as the accessibility of the outcrops, the ruggedness of the terrains, and the pandemic-hit periods, such as the global pandemic – COVID-19. As a potential alternative, remote sensing image acquisition techniques, such as Laser imaging Detection And Ranging (LiDAR), Terrestrial structure from motion (Tsfm), and Aerial structure from motion (Asfm) by Unmanned Aerial Vehicle (UAV) could be used for extraction of 3-D data from outcrops (Passchier and Exner 2010; Westoby et al. 2012; Riquelme et al. 2014; Dewez et al. 2016; Wang et al. 2019; Giordan et al. 2020; Herrero et al. 2022; Ghosh et al. 2023). These semi-automatic methods have been extensively tested for fracture orientation in rock mass characterization studies (Lapponi et al. 2011; Vöge et al. 2013; Vasuki et al. 2014; Assali et al. 2016; Casini et al. 2016; Chen et al. 2016, 2017; Dewez et al. 2016; Viana et al. 2016; Cao et al. 2017; Guo et al. 2017; Wang et al. 2017; Akara et al. 2020; Kong et al. 2021; Wu et al. 2021; Daghigh et al. 2022; Herrero et al. 2022). In a comparative study, Cawood et al. (2017) demonstrate that the UAV models provide more accurate results than those of the LiDAR or the Tsfm models.

Several recent software packages, such as DSE (Discontinuity Set Extractor, Riquelme et al. 2014), FACETS in cloudcompare (Dewez et al. 2016), and RANSAC (Vasuki et al. 2014), are available for semi-automatic extraction of orientations of geological surfaces from the UAV models (Bieniawski 1989; Hudson 2001; Riquelme et al. 2014; Dewez et al. 2016; Wang et al. 2019). The success of these methods has been tested mostly on a single fracture, slope, or in a relatively small area of a few hundred square meters (Riquelme et al. 2014; Dewez et al. 2016; Herrero et al. 2022). Here, we use UAV images for demonstration of the DSE and the FACETS software in an 840 m2 area of folded and fractured rocks in the Himalayan terrain and compare the results with the direct field observations.

2 Geological setting of the study area

The basic framework of geology of the Garhwal–Kumaun Lesser Himalaya was laid down by Auden (1934), Bhargava (1972), Valdiya (1980), Ravi Shanker Kumar et al. (1993) and Kumar (2005). The Main Boundary Thrust, an orogen-scale lineament, is the most conspicuous tectonic structure in the outer parts of the Himalaya. The thrust dips to the north and emplaces the Lesser Himalaya over the Sub-Himalaya. Several isolated Krol-synclines, each comprising the Proterozoic sedimentary rocks predominantly, occur on the hanging wall of the Main Boundary Thrust (figure 1).

Figure 1
figure 1

Krol synclines on hanging wall of the Main Boundary Thrust and the study area (after Auden 1934; Ravi Shanker Kumar et al. 1993; Rai et al. 2021; Xiao et al. 2022). The Shell limestone and Subathu Formations are not shown due to the limitation of scale. DD: Dehradun, RK: Rishikesh, NT: Nainital.

Full size image

The study area lies on the southwestern limb of the Garhwal syncline, one of the several Krol-synclines (figure 1). The syncline contains the Baliana Group, the Krol Group, and the Tal Group rocks in younging order. The Baliana Group consists of Blaini diamictite, thin bands of sandstone and shale, cap dolomite, and the shale/phyllite of the Infra-Krol Formation. The overlying Krol Group consists of three formations; Mahi Formation or Lower Krol (Krol-A), Jarashi Formation or Middle Krol (Krol-B), and Kaudiyala Formation or Upper Krol (Krol-C, D and E). The core of Garhwal syncline is occupied by the Tal Group, which is overlain by shell limestone and Subathu Formation.

Situated on the southwestern bank of the Ganges River near Tapovan, N 30°08.038′, E 78°19.979′, the study area exposes a carbonate sequence that belongs to the Krol-A or Mahi Formation (figure 2). The rock types vary from shaly-limestone to slaty-limestone, with cm-thin intercalations of clay-rich bands. The bedding surface, marked by colour contrast and/or compositional layering, is folded into cm- to m-scale folds (figure 3a). At a few outcrops where the carbonate rocks assume the slaty character, a well-developed axial plane slaty cleavage cuts the bedding surface, and the bedding/cleavage intersection lineation parallels fold hinge lines (figure 3b and c). Fold geometry, in general, varies from isoclinal in the southwestern part to open to gentle box folds in the north-eastern part of the study area (figure 2). The poles to the bedding surface lie on a broad great circle girdle due to the non-cylindrical geometry of map scale folds (figure 3d). The mesoscopic fold hinge lines, paralleling intersection lineation, plunge at varying angles towards SE on NW-striking upright axial surfaces (figure 3e, f).

Figure 2
figure 2

The study area. Map pattern of folds traced by bedding surfaces in the study area (mapped by Syed Shahid Akhtar). The folds are isoclinal in the southwestern part and open to gentle in the north-eastern part of the area.

Full size image
Figure 3
figure 3

Mesoscopic structures. (a) Open folds in shaly-limestone interbanded with thin clay layers. Camera facing SE. (b) Bedding (S0) and cleavage (S1) in slaty-limestone. Camera facing SE. (c) Camera facing NW striking upright cleavage surface (S1) in slaty-limestone. Dashed white runs parallel to S0/S1 intersection lineation. The bottom surface of a bedding plane is annotated as S0. The fracture runs perpendicular to the bedding surface. (d) Poles to bedding surfaces (S0). (e) Poles to intersection lineations and fold hinge lines. (f) Poles to axial surface and slaty cleavage (S1). Great circle-Modal axial plane dips 87°/204°. Lower hemisphere equal-area projections (contours 2% per 1% area).

Full size image

Among diversely oriented multiple fracture sets, the bedding-perpendicular fractures are the most dominant and ubiquitous. The extensional nature of these fractures is evident from the termination of fractures against lithological boundaries, lack of shear offset across the fractures, and consistent spacing between successive fractures (figure 4a, b). In general, the extension fractures strike NE and dip at sub-vertical angles (figure 4c). In this study, we focus on bedding surfaces and extension fractures.

Figure 4
figure 4

Extension fractures. (a) Termination of fractures at lithological boundaries. The fractures are at a high angle to the bedding surface. Camera facing NW. (b) Lack of shear offset in the bedding-parallel carbonate vein running across extension fractures. The carbonate vein is older than the fractures. Camera facing N. (c) Poles to extension fractures. Lower hemisphere equal-area projection (contours 2% per 1% area).

Full size image

As described later, the UAV image, combined with the field evidence, shows two distinct types of geological surfaces: the bedding surfaces and the extension fractures running at high angles to the bedding surface. By processing the UAV images, we extract orientations of the bedding surfaces and the extension fractures and compare the results with direct field observations.

3 Material and methods

Our approach includes the UAV image acquisition and photogrammetric processing for the generation of sparse point clouds, dense point cloud, and 3-D virtual models (figure 5). Both the FACETS and the DSE software extract the modal orientations of the planar surfaces by processing the dense point cloud. Using this approach, we obtained modal orientations of bedding surfaces and extension fractures. In addition, we also extracted the orientation of a few individual bedding surfaces and fracture surfaces using the compass tool available in the cloudcompare (FACETS).

Figure 5
figure 5

Workflow of the methodology.

Full size image

A brief on the major steps of the methodology is as follows:

3.1 UAV image acquisition

The study area was photographed using a DJI Phantom 4 Pro (Pix4D Ltd) UAV with a 12-megapixel camera (figure 6). The parameters, such as the area of interest, overlap percentage, and flight height or desired GSD (ground sampling distance), were specified within the application. We set a moderate speed, 20 km/h, and 80% overlap for the automatic mapping by the DJI Phantom 4 Pro. The camera angle was 90° from the horizontal, and the camera face was downward. The Pix4DCapture, an Android smartphone application, was used to plan and program the auto-flight of the UAV. We scheduled 38 flights, each lasting less than 10 minutes (figure 7). A total of 611 overlapping UAV images were captured with a GSD of 0.82 cm/pixel from an elevation of 20 m. These images were automatically geotagged during the image acquisition by GPS-mounted UAV.

Figure 6
figure 6

UAV model shows the carbonate sequence of the Krol-A Formation. The area in red rectangle is used for semi-automatic extraction of orientation of geological surfaces.

Full size image
Figure 7
figure 7

Camera positions on 3-D virtual outcrop model. Each black vertical line indicates the camera position and each array of black lines is a flight. Red rectangle: Area used for automatic extraction of data from software.

Full size image

3.2 Data processing and dense point cloud

All the 611 images were processed using the structure-from-motion technique in the AGISOFT Photoscan Professional Version 1.1 software (AGISOFT PhotoScan L.L.C. 2014). Using the processed images, we obtained the sparse and dense point clouds that contain 3, 01, 931; and 56, 585, 676 points, respectively (figure 8a–c). The direct georeferencing of point clouds was enabled using the global coordinate system WGS 1984. The entire point cloud creation took 11 hours of processing in Lenovo core i7 11th generation system. This time may vary with the processing speed of the system. The processing of dense point clouds gave the 3-D virtual model showing the ground features (figure 6).

Figure 8
figure 8

Point clouds for a part of the study area (red rectangle in figure 7). (a) Sparse point cloud. (b) Dense point cloud. (c) Points in zoomed image of area marked by white rectangle in (b).

Full size image

3.3 The DSE and the FACETS software

The DSE is an open-source MATLAB-based software developed by the University of Alicante, Spain (Riquelme et al. 2014). The software searches such a set of points that correspond to a consistent planar orientation. Once the sets of points have been detected, the DSE estimates the best-fit plane using the co-planarity test (Riquelme et al. 2014). Finally, it gives the modal orientation of several sets of surfaces, including those representing noise, from the 3-D point clouds in a tabular form. With the help of field evidence and a virtual model, the user selects the desired surfaces from the table displaying the DSE results.

The FACETS is a plugin of cloudcompare software. It extracts a planar surface by segmenting the 3-D point clouds with a user-defined degree of co-planarity using the Kd-tree or fast marching method (Dewez et al. 2016). The Kd-tree method, used in this study, subdivides the 3-D point clouds into quarter cells down, in recursion, until the points contained in the cell define the best-fit plane. Cells containing less than six points are discarded. The output is a virtual model that contains several differently coloured domains corresponding to different sets of surfaces. A user can directly delete the domains of noise on the virtual model. The step produces an output that contains modal orientation of the desired surface only.

In practice, the output from both the DSE and the FACETS, display many sets of surfaces that are marked by different colours on virtual model (figures 9, 10). A substantial number of such surfaces are noise corresponding to the eroded or broken outcrops. For a meaningful interpretation, it is imperative to filter out the noise with the help of field observations. It is obvious that the time taken to filter the noise varies with the number of undesired surfaces in different orientations.

Figure 9
figure 9

Result of the DSE. Each colour denotes a particular set of planes having similar orientations. Multiple colours are due to large noise. The green and the blue are the fracture surface (FS) and the bedding surface (BS), respectively.

Full size image
Figure 10
figure 10

Result of the FACETS after filtering out the noise. Green: fracture surface, yellow: bedding surface.

Full size image

In addition to the above, we used Breithaupt Kassel Clar Compass, with accuracy ±1–2°, for field observations.

4 Errors in modal surface orientation

Both the FACETS and the DSE software gave modal orientations of all the bedding surfaces and the extension fractures in the study area. For error estimation, we compared the software-derived modal orientations with the modal orientations corresponding to the maximum of contoured stereographic plots (figures 3d, 4c, and 11a, b).

Figure 11
figure 11

Comparison of modal orientations. Modes of bedding and fractures (black great circle), observed directly in field, correspond to the maxima of contoured poles. These modes are compared with modal orientations, red and blue great circles, given by DSE and FACETS software, respectively. (a, b). Filled coloured circles: pole to modal surface. Black: field observations, blue: DSE, and red: FACETS.

Full size image

In the DSE, the absolute errors in dip angle and dip direction of bedding are of the order of 3° and 5°, respectively. In the FACETS, these errors are 2° and 0°, respectively (table 1). For extension fractures, the errors in dip angle and dip direction are 7° and 14° in the DSE and 3° and 1° in the FACETS, respectively (table 2). Our results show that the FACETS gives more accurate results than the DSE. This inference is consistent with a recent study by Herrero et al. (2022). The higher accuracy of FACETS is largely due to the option for filtering out the noise.

Table 1 Comparison of the modal bedding surface orientations obtained from field observations with those given by the DSE and FACETS.
Full size table
Table 2 Comparison of the modal fracture orientations obtained from field observations with those given by the DSE and FACETS software.
Full size table

5 Errors in individual surface orientation

For comparing the individual surface orientations, we selected such bedding and fracture surfaces that are distinct on the virtual model and also in the field. We extracted individual surface orientations by placing the virtual cloudcompare compass directly on these surfaces. The same surfaces were then spotted in the field for direct observations by Breithaupt Kassel Clar compass.

The comparison of individual surface orientations reveals that the dip angle obtained by using the compass tool in cloudcompare differs from the direct field observation by 10–20° (tables 3, 4). However, the difference in the dip direction is lower, 2–6° (tables 3, 4). A much larger difference, of the order of 23° in dip angle and 45° in dip direction, is reported by Herrero et al. (2022). The dip angle obtained by direct field observations is generally steeper than that obtained by the compass tool. The reason for these variations is, as yet, unknown.

Table 3 Comparison of the individual bedding surface orientations obtained from field observations and by using the compass tool in the Cloudcompare software. n = number of observations. Dip angle/dip direction.
Full size table
Table 4 Comparison of individual extension-fracture orientations obtained from field observations and by using compass tool in the Cloudcompare software. Dip angle/dip direction.
Full size table

6 Discussion and conclusions

The UAV-based techniques are particularly useful for the determination of the orientations of planar surfaces in inaccessible outcrops. Steep cliffs in the gorge and canyon walls, the glacial valleys, and steep road or rail cuts are examples of such situations.

Cawood et al. (2017) is the only study that uses semi-automatic methods to estimate the orientations of the fold axis of the Stackpole Quay syncline, West Wales. This is achieved by intersecting the modal bedding orientations of opposite limbs of the syncline. The Stackpole Quay syncline is an ideally exposed simple syncline that folds bedding surfaces and allows UAV imaging and direct measurements. In geologically complex areas where different sets of axial plane cleavages, multiple sets of fractures, faults and erosional surfaces cut the rocks, using semi-automatic methods may be time-intensive. Similarly, in the areas of isoclinal folds with relatively small hinge zones and long limbs, the software-derived modal orientations correspond to limbs. A careful field check is, therefore, essential for correct structural interpretation. To quote Herrero et al. (2022), ‘For the moment, 3D models will not replace field surveys because the information obtained by an experienced geologist’s eye and by the rest of the senses cannot be replaced’.

One of our main objectives was to compare the results given by two alternative software, DSE and FACETS. The results show that the FACETS gives relatively more accurate results than the DSE. In our experience, the semi-automatic methods give relatively more accurate results, provided that 3-D surfaces are distinct on the virtual model. The accuracy of semi-automatic methods primarily depends on the noise level, the resolution of the photographs, and the quality of the 3-D exposure on the virtual model. In addition, the height of flight and the angle between the camera and exposed surfaces also control the accuracy. The higher the resolution and the denser the point cloud, the more accurate the results. However, a very dense point cloud may be too heavy to process by the semi-automatic software.