Remote sensing datasets are commonly used in the earth sciences to interpret the morphology, location, timing and orientation of geological features. These data types, now routinely delivered by satellite, aerial and UAV platforms, have advanced to the point at which they are widely available at high resolution and in some instances are frequently updated. This proliferation of data has led to a situation in which it is now no longer practical to use manual methods to extract geological information, meaning that the full geological value of high-quality datasets is often not extracted. For example, high- and ultra-high-resolution (centimetre to millimetre) photorealistic reconstructions of geological outcrops (“digital outcrop models”) are becoming widely available (McCaffrey et al., 2005; Pringle et al., 2006; Vollgger and Cruden, 2016), typically acquired using either laser scanning technology (Buckley et al., 2008) or photogrammetric workflows (Bemis et al., 2014; Smith et al., 2015). These techniques, combined with inexpensive and easy-to-use UAV technology, now make it feasible to acquire topographic data at millimetre to centimetre resolution over areas of several square kilometres (e.g. Vollgger and Cruden, 2016; Cruden et al., 2016), providing for the first time an objective method for rapidly collecting detailed 3-D information on geological structures.

There has recently been significant effort to develop automatic or computer-assisted methods for digitising structural data, in particular from orthorectified photographs or image sequences (Seers and Hodgetts, 2016; Vasuki et al., 2014; Jones et al., 2009). Achieving satisfactory automated digitisation is challenging for the mapping of geological structures due to intrinsic variables such as geometry, soft linkage and segmentation over multiple scales and extrinsic variables such as natural variations in colour, shadows, glare and/or incomplete geological exposure. Due to this complexity, fully automatic methods often require significant manual adjustment and vetting to remove false positives while retaining real geological features (Vasuki et al., 2014; Seers and Hodgetts, 2016).

In this paper, we first review existing approaches to the mapping of geological structures from digital data and then describe a novel least-cost-path method that can “follow” structure traces between user-defined control points in both 2-D gridded datasets (imagery, geophysical rasters, etc.) and dense 3-D point clouds (digital outcrop models). We then describe its implementation in two widely used software packages (QGIS and CloudCompare) and introduce four applications demonstrating the efficacy of the method for mapping outcrop structures, earthquake surface ruptures and oceanic fracture zones.

Many automated methods have been developed to extract linear features in the geosciences (e.g. Jinfei and Howarth, 1990; Tzong-Dar and Lee, 2007; Holden et al., 2012; Masoud and Koike, 2017). These use computer vision algorithms for edge and lineament detection and, while often successful in ideal situations, require substantial fine tuning to achieve optimal performance on real-world data. They also have a tendency to detect many false positives related to non-geological features such as shadows, roads or vegetation (Vasuki et al., 2014). Hence, even fully automated methods currently require significant manual effort to remove non-geological features while ensuring that features of interest are correctly detected.

To circumvent these difficulties, several methods have been developed which remain user driven but also use computational power to optimise the interpretation process and improve objectivity and consistency. Vasuki et al. (2014), for example, use an edge detection algorithm (phase congruency; Kovesi, 1999) on orthophotographs to optimise manually defined fracture traces and contacts. This allows the user to quickly define the approximate locations of interesting features and then automatically refine them, speeding up the digitisation process significantly while avoiding problems associated with false positives. Similar computer-assisted approaches have also been applied to improve the interpretation of faults in regional magnetic surveys (Holden et al., 2016) and oceanic fracture zones in global gravity datasets (Wessel et al., 2015).

In many situations, the 3-D orientations of detected features are of interest. This is typically calculated using a digital elevation model to add height information to each trace and then computing a best-fit plane (e.g. Dering et al., 2016; Jaboyedoff et al., 2009; Banerjee and Mitra, 2005). While this method works well in topography containing slopes < 45∘, it is inherently limited to 2.5 dimensions (2.5-D) as elevation data is gridded in the X–Y plane, causing problems when features cross-cut steep or overhanging topography (Pavlis and Mason, 2017). For this reason, direct analysis of 3-D point cloud data is preferable over methods that are limited to 2.5-D. Unfortunately, the unstructured nature of 3-D point data means that methods for trace detection in raster data, including those previously described, cannot be easily applied.

A number of automatic methods for analysing point cloud data have been proposed. These use clustering or plane-fitting algorithms to automatically segment and extract fracture or bedding faces (facets) exposed on the surface of the outcrop, with reasonable success (e.g. Dewez et al., 2016; Lato and Vöge, 2012; García-Sellés et al., 2011). However, structural surfaces are not always directly exposed and instead intersect the outcrop surface to form linear features referred to as structural traces. These traces cannot be detected using facet-based techniques and require a different approach.

Seers and Hodgetts (2016) demonstrate one such approach and automatically extract 3-D structural traces by applying image-based edge detection techniques (phase congruency) to a set of images and then projecting the identified traces into 3-D using depth information derived from photogrammetric reconstructions or associated laser scan data. This approach uses multiple images to overcome issues associated with out-of-plane geometry; however, as with other fully automated methods, a variety of parameters and thresholds require careful calibration and the results must be manually vetted to remove false positives.

To demonstrate the capability of our computer-assisted trace detection approach, we present the results of four case studies. These studies highlight the versatility of our method and its increased efficiency compared to established manual methods.

The first case study involves the interpretation of joint sets in two 10 × 10 m areas from a ∼ 1 cm resolution orthophotograph of a wave-cut platform at Bingie Bingie Point, New South Wales, Australia. The outcrop contains several Cretaceous to Paleogene dykes intruding Devonian diorites and tonalities cross-cut by a series of complex joint sets. The orthophotograph was generated by applying a structure-from-motion multi-view stereo (SfM-MVS) workflow, as implemented in Agisoft Photoscan v1.2.6, to 297 digital photographs captured from a DJI S800 EVO multi-rotor UAV fitted with a 24.3 megapixel Sony Nex-7 camera and a 16 mm F2.8 prime lens (Cruden et al., 2016). The two 10 × 10 m areas (Fig. 2a, b) were selected from the survey as they contain well-exposed dykes and joint sets as well as common confounding effects such as shadows and puddles. For demonstration purposes the selected areas are relatively small, but the workflow is equally applicable to much larger outcrops.

Our second case study focuses on the extraction of 3-D joint traces and orientations, which are interpreted directly on a dense 3-D point cloud. The Cape Woolamai sea stacks, located approximately 115 km southeast of Melbourne (Australia) on Phillip Island, have formed through the erosion of the coarse- to medium-grained Cape Woolamai granite, which intruded Silurian to lower Devonian meta-turbidites during or slightly after the mid-Devonian Tabberabberan Orogeny (a widespread episode of deformation and plutonism across Victoria; Gray, 1997; Richards and Singleton, 1981). Several generations of systematic and non-systematic joints cross-cut this granite, likely related to the cooling of the intrusion and the subsequent deformation and unloading.

For this study, a DJI Inspire 1 multi-rotor UAV and a 20 mm fixed focal Zenmuse X3 camera were used to capture 274 aerial photographs, which were subsequently processed using Agisoft Photoscan v1.2.6. A 45 × 40 × 25 m section of the resulting model containing a single sea stack was then extracted, containing ∼ 2 million points and representing an average ground sampling distance of ∼ 2.5 cm pixel-1. The topographic complexity of this sea stack allows for accurate joint-orientation measurement, but makes interpretation from 2.5-D datasets (orthophotograph + DEM) impractical (Fig. 3a, b).

For the third case study, surface ruptures that formed along the Greendale Fault after the 2010 Mw7.1 Darfield earthquake are extracted from a 1 m resolution lidar-derived DEM. The data were collected a few days after the earthquake and were used along with a variety of other data to measure the surface displacement resulting from the earthquake and to interpret the kinematics of the Greendale Fault (Duffy et al., 2012).

Finally, for our last case study we interpret oceanic fracture zones in the North Atlantic from 30 arcsec bathymetry (Weatherall et al., 2015) and vertical gravity gradient data (Sandwell et al., 2014). From their inception at mid-ocean ridges, fracture zones can be used to constrain plate motion vectors and are widely used in tectonic reconstruction (Williams et al., 2016; Sandwell et al., 2014). Both these datasets provide an opportunity to test our method on global-scale geophysical data.

For each of these case studies, we also assess the similarity of our assisted results to manually derived interpretations. The aim of these comparisons is not to rigorously validate the computer-assisted method because (1) given enough control points our method could match any interpretation and (2) manual interpretation of structures from remotely sensed data is notoriously subjective, so differing results can be expected from different operators and scales (Bond et al., 2007; Scheiber et al., 2015). Instead, we seek to demonstrate the applicability and versatility of the assisted approach and to show that similar results to a manual interpretation can be produced for less time and effort.

For each interpretation, the operator was instructed to digitise every structural feature within the dataset. To ensure that this was an achievable task, the extent of the dataset used in each case study is small compared to its resolution. No attempt was made to ensure that the same number of features was extracted from each dataset, as this would affect timing measurements. Digitisation was performed at or close to the dataset resolution so that the manual interpretations contained similar detail to the assisted approach (which always follows traces at the resolution of the dataset), although the operator was able to freely zoom in and out to inspect the data over multiple scales. The same operator performed the manual and assisted interpretations for case studies 1 and 2, while different operators generated each of the assisted and manual interpretations for case studies 3 and 4. Manual interpretation in QGIS (case studies 1, 3 and 4) was performed by digitising polyline features to a shapefile using the “Add feature” tool, while in CloudCompare (case study 2) the “Trace polyline by point picking” tool was used.

The time required to extract comparable amounts of structural data was significantly reduced using the computer-assisted method (Table 1). This efficiency increase was especially pronounced (61 %) for the point cloud example, as manual methods for digitising linear features on 3-D point clouds are particularly time consuming.

The following four subsections compare and contrast the results of both manual and assisted interpretations in more detail.

The four case studies presented above highlight applications of the least-cost-path method to the interpretation of high-resolution aerial orthophotographs, 3-D point clouds, lidar DEMs and bathymetric data – all datasets commonly used in the earth sciences to interpret and characterise geological features. We discuss here three aspects of our least-cost-path approach: it improves objectivity and reproducibility, allows for automatic refinement if better data become available and, unlike fully automated methods, works in cooperation with expert guidance.

Firstly, the approach is more objective than manual digitisation. Although not as objective as fully automated methods (the location of the trace start and end are interpreted), most of the length of each trace is determined algorithmically and hence will consistently locate in the same spot. Indeed, as demonstrated in Fig. 7, the calculated shortest path varies only slightly when control points are interpreted at different locations. The results from study 2 indicate that this improved consistency might increase the precision of the derived orientation estimates (Fig. 3e, f). Furthermore, each control point can easily be stored, providing a record of the locations at which interpretive decisions were made.

Secondly, similarly to the method outlined by Wessel et al. (2015) for extracting oceanic fracture zones, these control points can also be reused to generate an updated interpretation if higher-resolution or more accurate information becomes available. This possibility is demonstrated in study 4, in which published oceanic fracture zones were reconstructed automatically using an updated underlying dataset and the start and end points of a previous interpretation (Matthews et al., 2011). Although some quality control is required after such an operation, the digitisation process no longer needs to be completely repeated, and interpretations can be rapidly updated as datasets evolve.

When multiple datasets are available, the similarity and total cost of paths reconstructed using different datasets can be used to quantitatively assess the degree to which different datasets support an interpretation. It is common in the geosciences to bring interpretations from multiple types of data into a single synthesis (e.g. Seton et al., 2016; Blaikie et al., 2017), especially when using geophysical datasets such as gravity and magnetics. Limiting factors during such data synthesis include both the time required and the highly subjective nature of multi-data-type interpretations, so a method for rapidly quantifying the extent to which different datasets support an interpretation serves as an important addition. Similarly, sensitivity analyses could be performed by randomly moving control points and measuring the response of the traces to quantify the robustness of the interpretation to uncertainty (similar to Fig. 7).

The time it takes for users to interpret datasets using GeoTrace or Compass will vary significantly between users, and the purpose of this study was not to comprehensively measure the efficiency of our approach. Nevertheless, in each of the case studies, our initial assessment indicates that computer-assisted interpretation required ∼ 35–66 % less user effort, as measured by both average time and mouse clicks per structure trace, when compared to manual methods (Table 1). The resulting traces also appear to be comparable to manual traces in each case (∼±2 pixels), demonstrating that our method can be used to achieve equivalent results.

The Compass implementation of the technique produces especially impressive results, reducing interpretation time in the Cape Woolamai example by 61 %. This is pertinent given the rapid growth in both size and availability of high-resolution point cloud data and the limited range of available tools for extracting structural data from them. Significantly, the implementation of our least-cost-path method in Compass requires only local information such that the calculation time scales with trace length and not dataset size. This means the tool can be used to interpret arbitrarily large point clouds.

Finally, the computer-assisted philosophy behind our method keeps the expert in control of the entire digitisation process, allowing for data vetting and correction during digitisation. The approach ensures that the expert becomes familiar with the particular intricacies of each dataset, a key part of further data analysis and something that is not possible using fully automated methods but is essential for the creative process of understanding and interpreting spatial information.

We have described a least-cost-path-based method for the computer-assisted digitisation of structural traces in point cloud, image and raster datasets. The method enhances an expert's ability to extract geological information from the wide range of high-resolution data available to geoscientists while reducing the required time and effort. The advantages of the method can be summarised as follows:

allows for expert-guided interpretation in a way that seamlessly utilises computing power to significantly optimise the interpretation process and improve objectivity and consistency;