A single multi-scale and multi-sourced semi-automated lineament detection technique for detailed structural mapping with applications to geothermal energy exploration

Semi-automated lineament detection methods provide a rapid and robust means of mapping structural features at a multitude of scales. A geological lineament, defined as a mappable rectilinear or curvilinear feature of a surface and distinct from adjacent patterns (O'Leary et al. 1976), can be mapped to infer faults or fractures within the subsurface. The semi-automated approach to lineament detection generally includes five steps: pre-processing; feature extraction; detection; linking; vectorization (Middleton et al. 2015). The increasing resolution of remotely sensed data allows more detailed lineament studies over larger areas, making a completely manual analysis more time-consuming. Therefore, semi-automated methods are becoming increasingly popular for practitioners.

There are a variety of published methods for semi-automated lineament detection available on a range of platforms, including tools within mainstream software packages such as PCI Geomatica and Seequent Oasis Montaj or bespoke algorithms (e.g. Rahnama and Gloaguen 2014a, b; Middleton et al. 2015; Šilhavý et al. 2016; Masoud and Koike 2017; Yeomans et al. 2019). Many of these are able to analyse multi-source data inputs; however, as yet no algorithm has attempted to combine multi-source and multi-scale input data. This would represent an important development in mapping in greater detail over larger areas.

Herein, we use an adaptation of the semi-automated bottom-up Object-Based Image Analysis (OBIA) method of Yeomans et al. (2019). We combine multi-scale and multi-source data from an onshore light detection and ranging (LiDAR) elevation model and offshore bathymetry at three different pixel resolutions (5, 10 and 20 m pixels) to evaluate lineament characteristics over an area of 700 km². This is complemented by two localized manual studies, which validate the semi-automated method and demonstrate the level of structural detail. The study area is the Land's End peninsula and adjacent offshore areas in SW England; the bedrock geology comprises the Land's End Granite and its Devonian host rocks. It has been selected because of its importance for understanding NW–SE fault zones that are currently being targeted farther east in Cornwall as fault-controlled deep geothermal reservoirs (United Downs Deep Geothermal Power Project near Redruth and the Eden Geothermal Project, St Austell). The Land's End area is an ideal locality to study these NW–SE fault systems owing to the accessibility of granite coastal exposures and the quality of bathymetric data. Exposed bedrock in offshore areas reveals a detailed fault network and these areas can be mapped at high resolution to give a representative model of the underlying fault network that may otherwise be obscured in onshore areas.

Lineament networks have played an essential role in defining geothermal systems in other parts of the world such as the Rhine Graben (Bertrand et al. 2017) and in Scotland for coal mine geothermal resources (Andrews 2020). In SW England, the viability of major NW–SE fault systems as reservoirs for deep geothermal energy is being explored. Previous work in SW England has investigated NW–SE structures but has highlighted complexity, with other orientations such as NNW–SSE and NNE–SSW features (Nixon et al. 2012) and ENE–WSW fault-controlled vein (lode) systems further east (Alexander and Shail 1995, 1996; Shail and Alexander 1997). This work details initial attempts to understand the nature of these fault systems at a regional scale. These structures and their interactions with NW–SE systems are investigated based on their host rock and their distance from manually digitized fault traces. We assess the appropriateness of semi-automated and manual approaches for lineament mapping as a precursor to future reservoir modelling and highlight the potential for algorithmic selection bias in semi-automated methods. Orientation data are used to identify target structures and derive an estimate of ‘damage zone width’; however, further characterization of the reservoir (e.g. connectivity and flow modelling) is beyond the scope of this work.

The Upper Paleozoic geology of SW England (Fig. 1) comprises low-grade regionally metamorphosed Devonian–Carboniferous greywacke–mudstone sedimentary successions, with minor intrusive mafic igneous rocks, that were deformed during the Variscan Orogeny (Leveridge and Hartley 2006). These were later intruded by the Cornubian Granite Batholith in the Early Permian (Scrivener 2006). Three regional deformation events (D₁–D₃) are recognized (Alexander and Shail 1995, 1996). D₁ and D₂ structures developed in an NNW-directed thrust–fold belt during Variscan continental collision following the closure of the Rheic–Rhenohercynian Ocean. D₃ structures formed during latest Carboniferous to Early Permian post-Variscan regional extension during which thrust faults were reactivated as top-to-the-SSE extensional faults and new higher-angle ENE–WSW-striking extensional faults formed (Alexander and Shail 1995; Shail and Alexander 1997; Shail and Leveridge 2009; Alexander et al. 2019).

Early Permian magmatism was, in part, synchronous with regional D₃ extension and is largely represented by the Cornubian Batholith, which was emplaced between 293 and 275 Ma (Chen et al. 1993; Chesley et al. 1993; Scrivener 2006; Simons et al. 2016). The Land's End Granite study area is located at the westernmost end of the mainland, although it is worth noting that the batholith continues some 100 km offshore for a similar distance westward across the Cornubian Ridge (Evans 1990). A magmatic–hydrothermal tungsten–tin–copper–zinc orefield was developed contemporaneously with batholith construction and was overwhelmingly fault- and joint- controlled (Chen et al. 1993; Chesley et al. 1993); present in the study area as the St Just Mining District. Extensional fault-controlled vein systems (lodes) are typically ENE–WSW- to east–west-oriented, reflecting NNW–SSE to north–south extension, and formed synchronously with steeply-dipping NNW–SSE strike-slip transfer faults. The latter stages of mineralization, presumed to be associated with the youngest magmatic episodes, are commonly oriented NW–SE to north–south and may reflect a change in regional stress (Shail and Alexander 1997). The development of Early Permian ENE–WNW- to east–west-oriented extensional fault systems in the granites and their host rocks was contemporaneous with the formation of extensional sedimentary basins that host Permian ‘red bed’ successions (Evans 1990; Alexander et al. 2019).

The subsequent structural evolution of fault networks during the Mid-Permian to Mid-Triassic is poorly constrained but two minor episodes of intraplate shortening are identified (Shail and Alexander 1997). Regional ENE–WNW extension during the Triassic brought about the extensional reactivation of Early Permian NNW–SSE transfer faults and development of new faults (Shail and Alexander 1997) and a regional Middle Triassic episode of basinal brine migration through the NW–SE to NNE–SSW extensional fault systems. The resultant basement-hosted ‘cross-course’ veins offset earlier Permian magmatic–hydrothermal lodes (Scrivener et al. 1994; Gleeson et al. 2000, 2001).

Following the Triassic cross-course event, there is little constraint on the onshore structural evolution until the Cenozoic. An Oligocene intraplate strike-slip tectonic regime resulted in both dextral and sinistral reactivation of NW–SE faults, with displacements of up to several kilometres, along the Sticklepath–Lustleigh Fault Zone in the east of the region (Holloway and Chadwick 1986).

The Land's End Granite is the youngest of the granite plutons, with an age of c. 274–279 Ma (Chen et al. 1993; Chesley et al. 1993), having been intruded into the Upper Devonian Mylor Slate Formation of the Gramscatho Group (Fig. 2a). It forms the most westerly mainland exposure of the Cornubian Batholith and provides consistent exposure of the granite and its margins in coastal outcrop. The present-day shape of the pluton is unusual compared with the other plutons in SW England, characterized by a distinct geomorphology controlled by regularly spaced NW–SE-oriented valleys, and is well represented in Figure 2b. These features extend offshore, and are observable in the seafloor, where the bedrock is the Gramscatho Group. The submerged outcrop provides a highly detailed surface upon which to study fracture networks and trace these back to onshore areas where outcrop is more limited. Offshore areas are susceptible to sediment cover, which obscures the desired bedrock exposure, and the occurrence of sand waves upon these sediments can cause false positive results in semi-automated lineament studies. However, these are not extensive in the area selected and have been mitigated during post-processing.

As datasets increase in coverage and resolution, semi-automated lineament detection becomes a more efficient choice to the practitioner for rapid, objective lineament mapping. Built-in tools to mainstream software are commonly applied but there is an increasing prevalence of bespoke algorithms designed for use within different programming languages such as MATLAB (Rahnama and Gloaguen 2014a, b), Python (Šilhavý et al. 2016; Karimi and Karimi 2017) and eCognition's Cognitive Network Language (Middleton et al. 2015; Yeomans et al. 2019). Others can operate as plug-ins to existing geographical information system (GIS) software such as the GeoTrace toolbox for QGIS (Thiele et al. 2017).

Many of these semi-automated methods achieve their results through very different approaches, be it through targeting edges, or minima in the data, or through different methods such as pixel-based compared with object-based (e.g. Rahnama and Gloaguen 2014a, b; Sukumar et al. 2014; Middleton et al. 2015). Regardless, the key to a successful semi-automated algorithm is effective feature extraction to best enhance desirable structures and minimize the inclusion of spurious lineaments. Various feature extraction methods exist, and it is beyond the scope of this study to discuss them all; however, the application of the tilt derivative to LiDAR data (Middleton et al. 2015) and to bathymetry data (Yeomans et al. 2021) has proven highly effective. A comparison of the tilt derivative with more classical enhancement techniques, such as the Gradient, Sobel and Laplacian filters as well as the hillshade transform, found that the tilt derivative was more successful at creating continuous lineaments that were consistently sensed across an entire region of interest (Yeomans et al. 2021).

Despite the focus on semi-automated methods, manual analyses are not without merit. Smaller manual studies, over representative subsets of a much larger study area, can help validate semi-automated lineament sets. Alternatively, it may be necessary to fill in data gaps using another dataset that may not be available for the whole area or that may be impractical as input to a semi-automated algorithm. It should be noted that subjective bias is easily introduced and, over large areas, becomes time-consuming and lacks reproducibility (Masoud and Koike 2006; Scheiber et al. 2015). However, it was suggested by Andrews et al. (2019) that reference to field observations can reduce the subjective bias. Furthermore, manual analyses are widely considered to be better for topological analysis of fracture connectivity. Semi-automated networks are often highly segmented, resulting in short trace lengths and fewer connected branches, and therefore are inferior for connectivity studies.

To date, lineament detection studies, semi-automated or manual, have largely focused on augmenting their results by incorporating multi-sourced datasets. The approach has found success where lineaments that may have different signatures can be detected across different datasets and be incorporated into a final analysis. Some semi-automated methods do this within a single analysis (e.g. Masoud and Koike 2011, 2017; Yeomans et al. 2019). Some studies have looked at different resolution datasets (e.g. Meixner et al. 2017) but not within a single analysis. Combining different resolution data to map larger areas in greater detail is at present the frontier of lineament detection methods.

Three lineament sets are generated within this study area. A semi-automated approach using an adaptation of the bottom-up OBIA method by Yeomans et al. (2019) is conducted to detect lineaments across the whole region of interest and a workflow is presented in Figure 3a. Methods for data processing and lineament detection are presented as well as a detailed account of the required post-processing. A smaller manually digitized lineament set is generated to validate the semi-automated analysis. Both of these lineament sets are generated from a combination of onshore LiDAR and offshore bathymetric data. A third set is generated to fill in the data gap between the onshore and offshore datasets using aerial photography.

The LiDAR data were collected as part of the collaborative Tellus South West project, and the LiDAR survey was conducted by the British Antarctic Survey between July and August 2013. The LiDAR dataset has a spatial resolution of one point per metre and the data are accurate to 10 cm (both horizontal and vertical accuracy) (Gerard 2014). The digital terrain nodel (DTM) was downloaded from the Centre of Ecology and Hydrology repository in ascii grid format at 1 m pixel resolution. The onshore part of the study area covers c. 227 km².

Bathymetric data were downloaded from the Admiralty Data Portal under an Open Government Licence and included five blocks of multi-beam bathymetric data collected between 2008 and 2016. These were downloaded in raster format at 2 m pixel resolution. The multi-beam bathymetry data in the study area revealed an expansive area of submerged bedrock offshore. The data extend from the nearshore environment some 10 km from the shoreline and have an approximate coverage of 423 km².

Proximal to onshore areas, a roughly NE–SW-trending area of sediment-covered seafloor is present, resulting in no bedrock for lineament mapping. More localized patches of seafloor cover are present in other areas but are often small and not detrimental to the overall dataset. In rare, but spectacular cases, sand waves have formed on the seafloor and have the potential to cause artefacts in the data. These potentially problematic areas are included in subsequent analysis and dealt with in the post-processing.

The immediate nearshore areas can lack data coverage, probably owing to tides, poor sea conditions during acquisition or treacherous waters making acquisition too dangerous. This can lead to a gap when combined with the onshore LiDAR and result in the phenomenon referred to as the ‘white ribbon’ (Mason et al. 2008). To mitigate missing data in the onshore–offshore elevation model, optical aerial photography of the coastal zone and immediate nearshore was downloaded from the EDINA Digimap repository under an educational licence. The initial data were supplied in three-band raster format at 25 cm pixel resolution. This dataset was used to supplement lineament mapping in the area and attempt to bridge data gaps where they exist.

Object-Based Image Analysis (OBIA) tools have been increasingly applied in recent years. The approach makes use of raster input datasets to identify groups of pixels that are defined as ‘image objects’ through a process of image segmentation. The approach can use a variety of segmentation methods including top-down (thresholding) and bottom-up (merging) to identify image objects (Diamant 2004; Dragut et al. 2010; Eisank et al. 2014). These image objects are linked through a topology that describes their spatial relationship to one another and allows the calculation of geometric properties and internal statistics based on the subset pixels. The approach provides a profusion of metrics to compare, merge and/or classify image objects.

OBIA has been increasingly used in lineament detection studies such as those by Mavrantza and Argialas (2006), Rutzinger et al. (2007) and Marpu et al. (2008) but most recently through the workflows developed by Middleton et al. (2015) and Yeomans et al. (2019). A key step in these studies is the use of the tilt derivative transform for initial feature extraction prior to applying an OBIA workflow. An initial top-down OBIA method by Middleton et al. (2015) made use of airborne magnetic and LiDAR data separately to generate lineament networks. This approach was developed by Yeomans et al. (2019) to integrate multiple datasets (airborne magnetic, LiDAR and radiometric data) into a single workflow and produce a composite lineament network. A complementary bottom-up method was also produced, which sacrificed some detail in metadata and lineament length but was computationally more efficient and is therefore considered more desirable for larger datasets (Yeomans et al. 2019). Other feature extraction methods have been tested on bathymetric data by Yeomans et al. (2021), exploring the use of gradient and Laplacian filters and the hillshade transform in comparison with the tilt derivative, but were found to underperform where steep gradients (e.g. palaeocoastlines) in the seafloor were present in the data. It is assumed that this extends to subaerial steep gradients such as present-day coastlines.

The five bathymetric data blocks were initially converted from Bathymetric Attribute Grid (.bag) files to a Geotiff format and merged into a single dataset. A visual inspection revealed that, despite the use of nearshore Channel Coastal Observatory (CCO) data, some missing data were still present in the final product. Further, it was noted that the join between nearshore Channel Coastal Observatory (CCO) data and UK Hydrographic Office (UKHO) bathymetric data had a minor step. This is probably due to the higher resolution acquisition of the CCO data and minor differences between the Admiralty Chart Datum and Ordnance Datum to which these datasets are reduced for UKHO and CCO, respectively. The step was noted and revisited during post-processing.

The merged data were resampled to 5 m pixels prior to clipping to the study area and form the first input layer to the semi-automated lineament detection. To generate the two other input layers, the LiDAR data were integrated with the bathymetric data to form a combined single elevation model, which was subsequently resampled to 10 and 20 m pixel resolution.

The tilt derivative is a useful tool for lineament detection methods because it normalizes the magnitude of features preserving minor lineaments in the presence of larger features (Miller and Singh 1994; Verduzco et al. 2004; Fairhead and Williams 2006). It also produces more continuous features where the feature may show minor variations along-strike (Verduzco et al. 2004) and normalizes the data using the arctangent where the zero contour passes over or near the edge of a feature (Miller and Singh 1994).

The use of three input datasets processed at different resolutions (5, 10 and 20 m pixels) allows the capture of a range of lineaments that may display different characteristics. This is particularly effective for identifying fault traces that have a different geomorphological expression in onshore areas, which may be heavily incised, compared with offshore areas. Additionally, it allows the capture of more detailed lineament networks observable in the seafloor, which are masked onshore by soil cover.

The lineament detection workflow applied in this study develops the bottom-up approach first outlined by Yeomans et al. (2019). The workflow is outlined in Figure 3, which adapts the line extraction steps to include different resolution datasets that are tuned to extract lineaments based on the observable geomorphological features. The workflow is conducted in the eCognition software package using the Cognitive Network Language.

First, lineaments are extracted using a rectangular kernel composed of three stripes oriented in the long axis of the kernel. The kernel can be rotated and iterated through 360° and for this study an interval of 5° was selected. A lineament is identified using the central stripe of pixels and is given a weight based on the similarity on either side of the central stripe using the two border stripes. The majority of lineaments in the study area are assumed to be represented by minima in the data where they have been preferentially eroded. The output of the line extraction is a ‘lineness’ raster for each input dataset, all of which are subsequently merged (giving equal weight) into a single raster.

Following line extraction, bottom-up image segmentation is employed using the multi-resolution segmentation tool. The image is divided into many differently sized image objects, which are subsequently merged based on their spectral, statistical, textural, geometric or topological properties. The process also incorporates cleaning steps that remove spurious image objects. Furthermore, this analysis allows the designation of major and minor lineaments in the metadata. The threshold for this is user defined and is based on the relative similarity of features (as defined by the kernel during lineament extraction) rather than a geological measure of importance. Further notes can be found in the Supplementary material.

The final step in the lineament detection processes is to convert image objects to vector format. Given the polygonal nature of an image object, these are simplified to vector lines to produce a skeleton of the image object and a main line (principal axis) of the image object. The two forms allow the main lineament to be identified but also preserve branches should significant lineaments be conjoined. Given that only NW–SE features have been targeted in this instance, the main line vector file was taken forward.

The output vector lines have been post-processed to include segment length and orientation. These were calculated based on the polyline geometry within a GIS where the orientation of polylines was calculated in the range 0–179°. Furthermore, a spatial join was used to create two fields, one for bedrock type and another for location in the onshore or offshore environment. These were appended to the attribute table for the data. Full details of these methods are given in the Supplementary material.

Because of the semi-automated nature of the lineament detection algorithm, due diligence was conducted to ensure lineament quality for both onshore and offshore lineaments. Upon visual inspection it was apparent that areas of sediment cover and sand waves in the bathymetric data had generated artefacts during the transformation using the tilt derivative. Therefore, post hoc removal of potential spurious lineaments was conducted using the approach developed by Yeomans et al. (2021), which implements the Terrain Ruggedness Index (TRI) to map areas of sediment cover. The TRI is used to identify smooth areas that are assumed to represent sediment cover where the submerged outcrop on the seafloor is rough. These areas can be preferentially selected by using a threshold. In this study, the 5 m resolution offshore data were used to calculate the TRI layer, which was normalized to 0–1, and a threshold of 0.0025 was selected using a heuristic approach. This threshold was used to generate a mask (Fig. 3b) that selected all lineaments wholly within the mask and removed them.

Sediment cover in the bathymetric data cannot be fully addressed through a TRI mask. Owing to the presence of sand waves in some areas causing a ripple effect on the surface, the ‘smoothness’ criterion was not a panacea. Therefore, a manual mask was created that identified 11 areas of sand waves and these were removed where lineaments fell wholly within the mask. Additionally, a step in the bathymetry data was noticed around the southern extent of the study area, probably pertaining to a significant time gap between acquisitions. The lineaments generated immediately over the join between the two bathymetric datasets were manually removed by directly editing the shapefile.

Further post-processing of the onshore areas was conducted to remove field boundaries and roads. In Cornwall, these can be particularly problematic to semi-automated lineament detection owing to the presence of ‘Cornish hedges’, tall granite walls covered with earth, which result in a similar feature to desirable lineaments. It is possible that hedges and field boundaries removed in this step follow subtle geological features and result in a loss of data; however, owing to their problematic response and small scale, the accurate mapping of these features is unlikely to be reliable. In this case, it was noticed that most of these spurious lineaments are generated from the 10 m resolution layer whereas the 20 m resolution layer had few errors. On this basis, the 20 m resolution layer was smoother where target values in the tilt derivative would be smeared out and less susceptible to the misidentification. As a consequence, to identify these artefacts and remove them, post-processing began by selecting all onshore lineaments and filtering to reduce the population based on lineaments with a length <300 m and with a TDR value greater than −0.5 in the 20 m resolution layer (i.e. lineaments with TDR values (t) in the range −0.5 > t ≥ 1.57 that are <300 m in length were removed). As an additional step, all lineaments with a length <50 m in onshore areas were also removed.

The extensive post-processing steps described here demonstrate the importance of due diligence when processing large datasets from multiple sources. Careful examination of the lineament set over the region of interest identified probably spurious features caused by a variety of artefacts, each of which required a different approach to remove and ensure quality. Of the original 28 350 lineaments derived from the OBIA algorithm, a total of 10 009 were removed, leaving a final lineament population of 18 341 to be taken forward for analysis; a full breakdown is given in Table 1. This is a high proportion of false positive lineaments and a potential drawback of the semi-automated method when applied to high-resolution data. However, careful post-processing to identify false positives and their defining spatial or geometric characteristics can objectively remove spurious lineaments.

Manual lineament mapping has been conducted twice in this study to complement the semi-automated methods. An area of 7 km² was selected that demonstrates the detail within the offshore data that is beyond the scope of being captured by the semi-automated method used by this study. The study also manually digitized lineaments that were present within the white ribbon using aerial photography. This lineament set attempts to bridge the data gap between onshore LiDAR and offshore bathymetry and provide insight into lineament populations at even higher resolution.

The sub-area of interest is a 7 km² region straddling the west coast of the Land's End peninsula between Botallack in the SW and Morvah in the NE. The fault network was mapped from high-resolution multi-beam bathymetry of the offshore region and airborne LiDAR data into the onshore portion of the area at a pixel resolution of 2 m. The majority of the submerged bedrock is inferred to be Mylor Slate Formation with the exception of bedrock immediately offshore of the Land's End Granite coastal exposure (Goode and Taylor 1988; BGS 2000).

The multi-beam bathymetry and LiDAR data were imported into a GIS for interpretation where a hillshade transformation was applied to accentuate fault traces. It is common practice to generate two orthogonal hillshades and map lineaments in both illuminations to minimize bias (Scheiber et al. 2015). For this study, illumination source azimuths of 315° and 225° with an altitude of 45° were used for the transformation. Analysis of the structures within this sub-area was conducted manually, by hand-digitizing lineaments at a consistent scale of 1:5000. The scale was chosen as a reflection of Tobler's rule where a minimum map scale is determined by multiplying the pixel resolution by 2000. The 1:5000 scale was therefore chosen as close to this minimum map scale but also to reflect common mapping scales.

The nearshore environment is often a problematic area when linking between onshore and offshore datasets. The process of merging a digital elevation model with a bathymetric dataset often results in a gap in the data: the so-called white ribbon (Mason et al. 2008). The missing data in this area can vary depending on the data source, and a workflow by Leon et al. (2013) attempted to create a seamless elevation model over areas that have multiple spatial and temporally separate elevation datasets. Other studies have used field observations and geological mapping to supplement the data gap (Sanderson et al. 2017; Westhead et al. 2018). Although data acquisition in this zone is possible, it is often costly and requires careful planning. Neither geological mapping nor field observations have been permissible to date as a continuous study around the west Cornwall peninsula, therefore in this study, aerial photography was used to map the nearshore, wave-cut platform and immediately onshore areas.

In this study area, the white ribbon is not pervasive around the whole coastline. It is largely constrained to the west and north coasts, which have more inclement weather and have the least protection in periods of high swell compared with the south coast. Mapping of the nearshore environment was conducted around the entire coast in the study area. Aerial photography at 25 cm pixel resolution, available from EDINA Digimap resources on an Education and Research licence, was downloaded and a 250 m buffer around the coast was used to extract and mosaic the relevant image tiles. Lineaments were then manually digitized at a fixed scale of 1:500. Manual mapping was necessary owing to the complexity of the image, with highly varied outcrop shapes including steep slopes and wave-cut platform; the changing environment between shallow water and vegetated areas affecting the image texture; and the difficulty of removing the effects of shadows. However, these aspects must also be considered during interpretation of the derived lineaments as they cause selection biases (Shipton et al. 2019).

Herein, the three different lineament networks generated in this study are presented. This study has examined the influence of bedrock geology and discussed the geological interpretations that can be made using all three lineament networks. Additionally, orientation analysis was used to estimate damage zone widths relating to major NNW to NW structures that may be potential targets for fault-controlled geothermal reservoirs. Finally, we compare and contrast the lineament networks and discuss the benefits and limitations to semi-automated and manual analyses.

Rose diagrams presented herein have been created using the guidelines laid out by Sanderson and Peacock (2020) for equal-area wedge rose diagrams. These diagrams are superior to the conventional equal-radius approach because they better represent more subtle trends and allow for a more robust comparison between networks with different populations of lineaments. For comparison, equal-radius rose diagrams are included within the Supplementary material and illustrate the overemphasized principal orientations.

The semi-automated lineament network (Fig. 4a) shows the whole study area, Figure 4b shows an equal-area rose diagram for the lineament population, and Figure 4c and d highlights the greater density of lineaments mapped in the offshore bathymetry. Results from the offshore manual analysis (Fig. 5a) show the extent of the network and area of interest, Figure 5b shows an equal-area rose diagram and the complexity of the network, and Figure 5c–e shows the structural evolution. Lineaments from the white ribbon lineament set are presented in Figure 6a, with a corresponding equal-area rose diagram (Fig. 6b) and sub-panels highlighting nuances in the lineament network (Fig. 6c–f). Comments on the network and orientations for all three analyses are given in Table 2.

It is worth noting that in Figure 4 onshore areas are dominated by terrestrial drainage compared with offshore areas where the exposed bedrock was probably formerly a subaerial platform (Healy 1996; Waller and Long 2003). This has been submerged and stripped of vegetation and superficial sediment and may have subsequently been modified by wave-dominated processes in the marine environment. The result means that in the offshore areas it is possible to sense a much higher density of lineaments, which undoubtedly exist onshore but are not exposed and do not create a geomorphological signature that can be sensed in the data. Furthermore, there are some cases that can be seen in Figure 4c and d where some structures may have not been detected, and this is probably due to a lack of consistent signal causing small segments that were subsequently removed during the cleaning stages.

In Figure 5, there is a predominant orientation of NNW-trending lineaments, but NE-trending features are also prominent (Fig. 5d). The relationship between these systems is difficult to unpick from lineament analysis alone but both main sets appear to mutually cross-cut each other, suggesting multiple reactivation episodes, as highlighted in Figure 5e.

Figure 6c–f illustrates the variation that exists in the lineament population at this small scale. Figure 6c shows the prevalence of lineaments that can be detected here from aerial photography. The digitized lineaments in this area do not appear to reflect the orientations of those detected in the semi-automated lineament set. This may indicate that at more local scales, the lineament network is more complicated and may represent small relays or transfers or a broader damage zone. Figure 6d shows regular sets of NW–SE-trending lineaments that reflect those detected in the semi-automated set. The manual analysis also contains NNW and NE to ENE orientations that intersperse the NW sets but are not apparent in offshore or onshore areas within the semi-automated method. Figure 6e highlights where the manual lineaments populate the white ribbon in the data and show a complex mix of WNW–ESE, NE–SW and north–south features. This area demonstrates a broad agreement with the general trends observed in the semi-automated lineament set. Figure 6f highlights that major NW–SE lineaments extend from onshore into offshore areas. Between these features, few lineaments are detected in the bathymetry or LiDAR, but manual coastal interpretations show that a complex network exists between large-scale features, reflecting the complexity highlighted in Figure 6c.

Lineaments in all three sets show multimodal populations, some more subtle than others. These modal groupings may be explained by different bedrock types and reflect the protracted structural evolution of the area. For simplicity, the rocks within the study area have been divided into ‘granite’ and ‘slate’ subdivisions as these are the dominant rock types.

In Figure 7, equal-area rose diagrams are presented that depict the granite and slate subdivisions for the semi-automated lineament set. There is a clear change in modal trends between granite and slate subdivisions. The granite subdivision (Fig. 7a) displays a strong NW–SE trend that is much more diffuse in the slate subdivision (Fig. 7b) and may reflect a mechanical control on fault propagation through the slate compared with the granite. Both subdivisions express a strong lineament grouping that trends approximately ESE–WNW. These observations are explored further through the different environments within which one can sense these lineaments.

Figure 8 highlights the differences between lineaments detected in an onshore v. an offshore environment. Figure 8a and b presents lineaments detected over onshore areas for the granite and slate subdivisions, respectively. It can be seen here that granite lineaments have an intense modal population of NW–SE-trending features with a more subdued ESE–WNW trend. In comparison, the slates show a dominant ESE–WNW trend but, perhaps surprisingly, mimic the NW–SE trend observed in the granite subdivision. The slates also demonstrate other orientations of lineaments such as NNW–SSE and NE–SW features that are less prevalent in the onshore granite set. When compared with offshore areas in Figure 8c and d, the lineaments in the offshore granite subdivision have a noticeably subdued NW–SE trend and are dominated by ESE–WNW-trending features. Again, the ESE–WNW group is observed in the offshore slate subdivision and the broad grouping in the NW–SE quadrants that was noted in Figure 9b is also apparent but with a stronger skew towards a NNW–SSE trend.

The lineament data in both the semi-automated and manual offshore networks capture an ESE–WNW set that represents either bedding-parallel faults or recessive features related to the erodibility of different sedimentary packages. The east to ESE orientation contrasts with the dominant ENE to east trend of bedding within the Devonian sedimentary successions to the east of the Land's End Granite (Leveridge 2011; Leveridge and Shail 2011). It is possible that bedding is rotated as a result of granite emplacement or later extensional faulting and varies at different points around the pluton (Hughes et al. 2009). Therefore, the area is considered to highlight an anomalous scenario of potentially important lineaments in the region.

Based on the bedrock analysis of the semi-automated lineaments, both slate and granite subdivisions show a consistent ESE–WNW grouping of features. It is likely that in the slate subdivision this is the result of either lineaments being detected along Variscan (late Devonian–Carboniferous) bedding-parallel faulting or the detection of recessive features in the Devonian sedimentary succession owing to the interbedding of mudstone and sandstone horizons. However, sedimentary origins cannot explain the same set observed in the granite because of its magmatic nature and the much younger Permian age. This lineament grouping is therefore considered to be caused by bedding-parallel faulting. Structures in slate have subsequently been reactivated during Permian (D₃) extension, thus causing faulting of the Permian granite. This interpretation agrees with the model of Shail and Alexander (1997) where extension resulting in reactivation of earlier Variscan thrusts caused zones of distributed shear, detachments and high-angle faults. The only aspect of this theory that is difficult to reconcile is that the ESE–WNW trend observed in this study is at odds with the ENE–WSW observations further east by Shail and Alexander (1997) and the prior works of Alexander and Shail (1995, 1996) but may reflect the rotation of sedimentary rocks through faulting and granite emplacement recorded along the northern margin of the Land's End pluton (Hughes et al. 2009). It is beyond the scope of this study to investigate further but this discrepancy may be due to a number of factors such as smaller scale ENE–WSW structures being observable in the field, a sampling bias either in the field owing to available outcrop or from the semi-automated lineament detection, or owing to the different localities where data for their studies have been collected immediately to the east of this study.

There is a divergence in the orientation of major fault zones between the granite and slate subdivision when analysing lineaments in the NW–SE quadrants. The granite subdivision shows a much more distinct NW–SE trend compared with the diffuse grouping observed in slate (Fig. 8a–d). Additionally, the onshore slate subdivision shows a similarly distinct NW–SE trend that is observed in the granite subdivisions. The strong NW–SE trend in granite is likely to reflect later Permian faulting and the formation of mineralized lodes in the St Just Mining District (oriented NW–SE; Dines 1956) and related to a later ‘reactivation’ episode according to Shail and Alexander (1997). In the case of the granite, the generation of NW–SE faults at this time probably created new features in the rock mass whereas, in the slate, pre-existing features such as Variscan NNW–SSE structures are likely to have accommodated any strain, resulting in reactivated fault zones. Structural inheritance influencing fault systems in this manner is not new and has been attributed to deflections in lineament orientations by a number of studies (e.g. Meixner et al. 2017; Samsu et al. 2020). The trend in onshore slate lineaments is considered to be a local effect of nearby granite, most probably at depth, influencing the fracture pattern observed in what would be the roof zone of the covered pluton.

The lineament networks detected and analysed in this study are critical to understanding the deep geothermal fluid flow pathways in SW England. These are aligned subparallel to the contemporary maximum horizontal stress, which has an approximate NW–SE orientation (Heidbach et al. 2018).

To investigate the width of the damage zones, the regional semi-automated lineament set was first converted to a density map of structures within the orientation (a) range 120° > a > 175° to reflect the maximum horizontal stress and most likely orientation for open structures. This map was used in conjunction with the existing lineament sets and the raw data to identify and manually digitize structures with long strike-lengths at a fixed scale of 1:80 000. This approach was necessary because of the short segments identified in the semi-automated network. The output of this manual analysis resulted in 64 major structures being identified across the study area; these are illustrated in Figure 9. These were used to extract lineaments from both the semi-automated and white ribbon sets within 1000 m each side of a structure and split them into subsets on the basis of their distance from a structure (s): 0 < s ≤ 10 m; 10 < s ≤ 50 m; 50 < s ≤ 100 m; 100 < s ≤ 200 m; 200 < s ≤ 400 m; 400 < s ≤ 1000 m. The offshore manual lineament set was not included owing to its significant overlap with the semi-automated method, which would bias the analysis.

The orientations of derived subsets for these major NW–SE structures are presented in Figure 10. The lineaments at ≤10 m from a structure in Figure 10a show a clear NW–SE to NNW–SSE trend probably representing the main fault system as it rotates owing to changes in bedrock from granite to slate. An ESE–WNW group is also prevalent, suggesting that this trend, originally observed at a regional level, may be pervasive and exist within these fault systems. Subordinate ENE–WSW and NNE–SSW lineaments are apparent. The ENE–WSW set is derived specifically from the white ribbon lineament set (see Supplementary material), and is observable only at small scales of 1:500 or less. These sets, observable at such local scales, are an important observation and may enhance cross flow between the NNE–SSW features or act as potential barriers to flow. The main trends in Figure 10b and c mimic those in Figure 10a; however, the subordinate ENE–WSW trend becomes less distinct. At distances of more than 200 m from the faults (Fig. 10d–f), the NW–SE to NNW–SSE trend is slightly diminished and the large-scale regional ESE–WNW trend becomes dominant, reflecting the global trend captured in Figure 4b. This suggests that damage zones around the main NW–SE faults have a c. 100–200 m width (Fig. 10a–c) and contain more frequent cross-cutting features that may enhance connectivity of the fracture network.

The three lineament networks presented in this study provide a useful comparison not only for quantifying the effectiveness of the multi-scale semi-automated method, but also for identifying sampling bias between different methods and scales. The comparison statistics for lineament numbers, area and trace length are included in Table 3.

The number of lineaments for each network and the area of the mapped region of interest for each are reported with an area-normalized count to give an idea of the density of lineaments detected (Table 3). These networks are of course mapped at different scales where the highest resolution (1:500) identifies the greatest density of lineaments. This decreases significantly compared with the offshore manual lineament network mapped at 1:5000 whereas the semi-automated network shows an increase in density for lower equivalent resolutions of 1:10 000 to 1:40 000 (for pixel resolutions of 5–20 m using Tobler's Rule). The increase in the semi-automated network is likely to be a function of segmented lineaments with multiple segments mapped along the trace of a single structure.

The statistics regarding lineament length in Table 3 show marked variation between the sets. The longest lineament lengths are achieved in the offshore manual lineament set where mean lengths are approximately three times the length of those in the semi-automated lineament set. This is also reflected in the median and standard deviations for the two sets, indicating that it is reasonable to suggest that the semi-automated method underestimates lineament lengths by a factor of three. Practically, several segments along a structure are often sensed so the majority of a lineament may still be captured, so perhaps the more difficult question is how much goes undetected. An empirical approach could estimate this given the difference in lineament density and compare this with trace lengths between the semi-automated and manual offshore networks, but this is considered unwise because there is no geometric or topological point of reference.

For comparison, the orientation data for each lineament set have been reproduced in Figure 11. When examining the two manual sets (Fig. 11b and c), a clear NW to NNW grouping can be seen. The main modal trend for the manual offshore set is c. 325° compared with the white ribbon set, which is aligned at 340°. A broad NW–SE grouping can be seen in the semi-automated lineament set (Fig. 11a); however, the dominant orientation is the ESE–WNW group with a modal trend of 095°.

The rose diagram of the semi-automated lineament network is considered to be a robust example of the true population owing to the objective approach to selecting orientations during the detection phase of the algorithm. The dominance of off-trend ESE–WNW lineaments is considered genuine, although the geological source is perhaps ambiguous. The absence of this trend in both manual analyses may be due to cognitive selection bias towards major NW–SE features where ESE–WNW trends may be more subtle or not considered to represent a fault. Concerning the white ribbon lineament network, there may be a substantial physical selection bias owing to the orientation of the coastline, which is predominantly subparallel to the ESE–WNW trend.

Minor ENE- and NNE-trending structures have been identified in manual studies, with ENE trends particularly evident in the white ribbon lineament network mapped at 1:500 scale. These have previously been noted to be missing from regional onshore studies (Yeomans et al. 2019) and it may be that these structures, well known in the Cornish mining districts to the east, are present but not detectable in the other networks. Therefore, there may be an underlying complexity to these lineament networks that requires further investigation in future reservoir characterization studies.

This study presents a new multi-scale and multi-source semi-automated lineament detection method that allows greater detail to be captured using a semi-automated approach. This is complemented by manual analyses to compare networks, and Table 4 summarizes the respective advantages and limitations of the methods.

Manual approaches offer a more connected network, whereby lineaments are not only longer but contain more detailed branching structures with much less post-processing required. However, the synthesis of multiple large datasets is more time-consuming, requiring many individual studies that are later combined. The incorporation of user knowledge is both an advantage for linking structures and a limitation owing to potential biases, but good practices can mitigate these issues (e.g. Scheiber et al. 2015; Andrews et al. 2019).

Semi-automated methods provide a means of rapid mapping across a breadth of input datasets that can be at different scales. Thus, the resultant lineament network can capture a range of detail and potentially be based on a number of source data (e.g. Yeomans et al. 2019). The trade-off with this approach is a segmented network and the requirement for significant post-processing to identify and remove false positive lineaments.

Additionally, the segmented nature of the semi-automated network therefore means that it is not optimal for detailed investigation of connectivity and reservoir modelling parameters using topological techniques (e.g. Andrews et al. 2020). At present, the connected and more detailed network of manual analyses is better suited for these purposes. However, this study has demonstrated that target structures for fault-controlled geothermal reservoirs can be identified through orientation analysis and an estimate for damage zone widths can be interpreted.

One can identify apparent selection biases for semi-automated methods and how relatable these are to established biases for manual methods (e.g. Scheiber et al. 2015; Andrews et al. 2019; Shipton et al. 2020). Both semi-automated and manual approaches are subject to physical bias in the data such as coverage and whether a lineament is observable in the data (Shipton et al. 2020). This was especially the case for lineaments manually digitized within the white ribbon owing to the shape of the coastline and image shading but was problematic for the input data in semi-automated methods where structures may be discontinuous.

The OBIA method, where pixels are clustered into image objects, is for all intents and purposes an unsupervised machine learning approach and has its own algorithmic bias. The data-driven image segmentation method will successfully identify pixels representing lineaments and group these together but struggles to generalize where the signal is weak or non-existent. This results in the generation of segments that remain unlinked and is a reflection of high bias and low variance, and the bias–variance trade-off from data science transcends the analysis (Friedman 1997). To mitigate high bias, the accuracy can be reduced, allowing more freedom to include weak signal; however, the variance will increase, resulting in more spurious lineaments being detected, which may in turn degrade the lineament network. This is an algorithmic selection bias and is the semi-automated equivalent to the cognitive selection biases that occur in manual analyses.

It has been demonstrated in this study that lineament detection can be conducted across adjacent marine and terrestrial environments, where lineaments are represented by markedly different signatures, in a single analysis. However, it is noted that the nature of onshore and offshore data can yield different subpopulations of lineaments and the implications need to be considered. By using high-resolution bathymetry that contains areas of submerged outcrop the semi-automated lineament detection method is able to map structures that are unobservable in onshore areas. Given the large areas of submerged outcrop in this study, it is clear that simply detecting lineaments in onshore areas would give a considerably biased representation of lineaments in the region. Therefore, where applicable, particularly in coastal regions, it is recommended that bathymetric data should be included as part of an analysis where the data are available.

The Land's End peninsula and the surrounding offshore platforms demonstrate the complexity of the local fracture network. By using a novel multi-scale and multi-sourced semi-automated OBIA method to map the study area, a detailed lineament network has been established. This is complemented by manual analyses that demonstrate the limitations of the semi-automated method. We therefore reach the following conclusions.

1. The incorporation of multi-scale input layers from onshore and offshore datasets allows a single, detailed, composite lineament network to be mapped rapidly over a large area (700 km²). Despite this success, the study has highlighted the need for careful due diligence during post-processing to remove false positives that occur owing to sand waves and artefacts in the data.

2. Comparison of the manual and semi-automated networks demonstrates the discrepancies in trace length of semi-automated lineaments. The segmented nature of the network means that it is inappropriate to use this directly to determine connectivity of the network for reservoir characterization. However, orientation data are considered valuable.

3. Orientation analysis of lineaments proximal to fault zones estimates damage zone widths of 100–200 m that may be potential targets for fault-controlled geothermal reservoirs. High-resolution manual mapping indicates that greater complexity may exist at this scale.

4. The manual analyses demonstrate the detail available in offshore datasets whilst also filling in data gaps in the ‘white ribbon’ between the onshore LiDAR and bathymetry datasets.

5. Major fault zones are demonstrated to change orientation from NW–SE when hosted in granite to NNW–SSE when hosted in slate and are not proximal to the granite margin. The change in orientation is interpreted to be due to reactivation of Variscan NNW–SSE faults in the slate but the propagation of new NW–SE faults in the Permian granite.

6. Although semi-automated methods remain objective, they are not without bias. The study has identified algorithmic selection bias that affects the OBIA method used here. The image segmentation technique inherently works in a data-driven manner that struggles to generalize the data in areas of poor signal. This can be mitigated by reducing accuracy (bias) but will increase variance and may degrade the lineament network.

7. Finally, the study advocates the use of bathymetry to map offshore submerged bedrock to better understand the lineament network that may be obscured in onshore areas owing to cover, which may cause a physical bias towards major structures.

B. Andrews and an anonymous reviewer are thanked for constructive reviews that have greatly improved the structure and clarity of the paper. C. Rochelle (British Geological Survey) is thanked for comments on an early draft of this paper. C.M.Y. is funded by a NERC Highlights grant (NE/S003886/1) on the GWatt project. A. Hart is thanked for helpful suggestions and comments on the final version of the paper. A.H. is funded by a GW4+ NERC DTP grant (NE/L002434/1). C.W. is funded by a NERC Highlights grant (NE/S004769/1) on the GWatt project. The authors would also like to thank A. Matthews and H. Scott of Cornish Lithium Ltd for their support in accessing the bathymetry data. The bathymetry data used in this study have been sourced from the UK Hydrographic Office and accessed via the Admiralty Marine Data Portal. The LiDAR data used in this study have been sourced from the Centre for Ecology and Hydrology. The British Geological Survey is thanked for making the BGS 625k (DiGMapGB-625), BGS Geology 250k (DiGMap250k) and BGS 50k (DiGMapGB-50) data available on an Open Government Licence.

CMY: conceptualization (lead), data curation (lead), formal analysis (lead), investigation (equal), methodology (lead), resources (lead), software (lead), validation (lead), visualization (lead), writing – original draft (lead), writing – review & editing (lead); HC: data curation (supporting), investigation (supporting), resources (supporting), writing – original draft (supporting); AJLH: data curation (supporting), formal analysis (supporting), writing – original draft (supporting); RKS: conceptualization (supporting), investigation (equal), supervision (lead), writing – review & editing (supporting); CW: conceptualization (supporting), methodology (supporting), visualization (supporting), writing – original draft (supporting); ME: investigation (supporting), writing – review & editing (supporting); CH: supervision (supporting), writing – review & editing (supporting)

This work was funded by the Natural Environment Research Council (GB) (NE/S003886/1, NE/L002434/1 and NE/S004769/1).

C.Y. is now an employee of Cornish Lithium.

The data from this study are available on request from the author.