1 Introduction

Accurate extraction of fault patterns is crucial for both structural interpretation and modeling, thereby enhancing the understanding of different geological structures and structural and combined hydrocarbon traps (Zeng 2010; Ashraf et al. 2020; Zhou et al. 2022; El Dally et al. 2023). In general, small-scale faults (SSFs) are accountable for a minimal percentage of seismic activities. These have an offset of around 25–50 m (Clausen and Korstgåd, 1993), and may reach approximately 2–4 km; however, they only have minor throws in the range of forty meters (Jia 2013). The throw is mainly influenced by tiny variations in fault trace geometry because of the irregularity of the exposure (Camanni et al. 2023). Moreover, as the throw on a fault surface increases, the relevance of throw segmentation caused by a bend diminishes (Roche et al. 2021). The exploration through evaluation of the SSFs in varied reservoir rocks is a challenging step that affects the interpretation accuracy in basin analysis, tectonic analysis, and reservoir exploration (Yin et al. 2018). Furthermore, since SSFs include all types of faults and may have various complex characteristics, it can be challenging to classify them effectively (Boro et al 2014; Wennberg et al. 2016). SSFs give critical information relating to tectonics, the flow, and the trapping of hydrocarbon (Zeng 2010).

SSFs have a tendency to cause highly complex hydrocarbon reservoirs, mainly in those fields that demonstrate a high degree of variability and poor uniformity of lateral facies distributions (Zhang et al. 2011). In addition, SSFs frequently make reservoirs extremely challenging to extract. Limits and directions may be obtained by the analysis of SSFs, which is essential for the prospect characterization and development of reservoirs (Ameen and Hailwood 2008). It is standard procedure for interpreters to manually identify faults on seismic data to complete thorough fault interpretation, a process that may be challenging and time-consuming. Whenever faults and fractures are manually interpreted, there is a very significant level of uncertainty (Peace et al. 2018), and this is certainly relevant in basins where there is a lack of seismic data. Several publications recently presented examples that highlight the significance of fault classification using a wide range of methods in CCS, EGS, and hydrocarbon exploration. Because of this, an approach that is both efficient and accurate is required to identify these SSFs (Basir et al. 2013; Riaz et al. 2019).

Generally, various ML methods have been used for automatic horizon picking (Yang and Sun 2020), facies interpretation (Bressan et al. 2020; Alzubaidi et al. 2021; Hou et al. 2023), automatic channel detection (Pham et al. 2019; Ismail et al. 2022), salt dome detection (Waldeland et al. 2018), signal enhancement and seismic image denoising (Liu et al. 2022), and gas chimney detection (Ramya et al. 2020; Ramu et al. 2021; Ismail et al. 2022; Eid et al. 2023). To improve fault interpretation, many approaches for calculating fault parameters have been developed. One of the most often utilized methodologies is coherence attributes, which are described by fault likelihood (FL) and thinned fault likelihood (TFL) characteristics and evaluate 3D structural and stratigraphic features and geobodies (Hale 2013; Ashraf et al. 2020; Zhou et al. 2022).

In addition, there is a significant amount of scientific interest in employing artificial intelligence approaches for automatic fault detection, such as the Convolutional NNs (CNNs) (Lou et al. 2022a), supervised with a synthetic dataset (Wu et al. 2019; Li et al. 2022; Lou et al. 2022b), 3D-FaultSeg-UNet (Van-Ha and Thanh-An 2022), seismic fault attribute estimation using a local fault model (Lou et al. 2019), unsupervised learning method (Ashraf et al. 2020; Liu et al. 2022), and deep neural networks are presented for pixel-by-pixel segmentation (Kadlec et al. 2008; Schultz et al. 2009; Di et al. 2018; Guo et al. 2018; Hale 2013; Zhou et al. 2022). Furthermore, deep learning is used for fault interpretation to minimize the gap between synthetic and real data (Zhou et al. 2021).

2 Geologic setting

The Taranaki Basin in the western part of the New Zealand subcontinent represents an enormous Late Cretaceous-Cenozoic depositional system (Fig. 1). The basin is subdivided into two main structural provinces: a stable platform in the west and the eastern mobile belt (King 2000; Uruski et al. 2003; Stagpoole and Nicol 2008; Radwan et al. 2021; Radwan and Nabawy 2022). The Western province constitutes a 150 km wide platform stretching below the continental shelf. This region comprises a broad basement province covered with a thick sedimentary pile (~ 6000 m). The eastern belt is 80 km wide mobile province consisting of a series of rift basins. The Taranaki Basin was developed related to a rifting phase in the Late Cretaceous. During this period, the separation between the New Zealand plate and the Australian plate developed. Therefore, a series of half-grabens were formed and the syn-rift deposition took place until the latest Paleocene. This sequence was dominated by clastic facies that rapidly-filled the half-grabens (Forder and Sissons 1992; King 2000). In the Early Eocene, the basin became a passive margin, and widespread shallow marine sand, and coal were precipitated due to a marine transgression period. Thermal subsidence prevailed throughout the Eocene due to the effects of sag-related thermal decay.

Fig. 1
figure 1

a Regional geological and structural map of New Zealand, b base map of the Kerry 3D seismic cube showing inline and crossline ranges

Full size image

The change from extensional to compressional tectonism occurred after the end of the Eocene. Oligocene compressive phase resulted in development of a rapid subsiding fore-deep in the eastern part of Taranaki Basin (Abdelmaksoud and Radwan 2022; Leila et al. 2022). Moreover, a foreland basin was formed at that time, several pronounced compressional structures were developed during the early Miocene. An uplift occurred in the southern part of New Zealand offering an enormous source of terrigenous materials that entered the deep water settings as turbidities (King 2000; Ainsworth and van der Pal 2004).

An extensional-wrenching phase occurred during the latest Miocene, this event was associated with a back-arc spreading. It is formed by a voluminous subsidence in the mobile eastern belt and a rapid accumulation of thick, deep marine clastics. The western boundary fault- the Cape Egmont fault system- of the eastern mobile belt was created at this time. In the west, the western platform remained unaffected by these structural regimes where a passive sag continued virtually uninterrupted throughout the Paleogene-Neogene.

The Cape Egmont Fault, bordering the Maui field from the east, was a major reverse fault in the Miocene (Thrasher 1990, 1991), but has been downthrown since that time. The hanging wall basement block of Taranaki Fault overthrust Taranaki Basin sedimentary sequence, with up to 6 km of vertical basement offset. The Tarata Fault Zone consists of a set of echelon thrusts, mainly east-dipping, and associated anticlines, as well as backthrusts at some localities. The vertical offset is generally in the order of 700–1000 m, reaching around 1500 m in the Waihapa structure to the south (Bulte 1989, Palmer and Bulte 1991). The Waimea-Flaxmore Fault System (WFFS) represents one of the largest, active reverse fault system situated along the western side of the plate boundary in the northern South Island. WFFS dissects the Alpine Fault at its southern termination. To the northeast, this system joins the Manaia Fault in the offshore areas (Bull et al. 2019; Ghisetti et al. 2020; King and Thrasher 1996; Langridge et al. 2016; Rattenbury et al. 2016). The Manaia Fault is the most important reverse fault in the area south of the Kupe South field, as the reverse offset of the nearby Taranaki Fault decreases rapidly to the south. Near the southern end of the Taranaki Fault, basement is offset by up to 3.5 km on the Manaia Fault (McDougall and Thrasher 1991).

Taranaki Basin hosts a complex and heterogeneous sedimentary succession (Fig. 2). From Cretaceous to Eocene, a thick sequence of continental sandstones and swamp coal interbeds was accumulated in the southern sector of the mobile belt (North Cape Formation, Farewell Formation) (Pilaar and Wakefield 1978; Palmer and Bulte 1991). This sequence changes laterally northwestward todeltaic/parallic sediments, barrier bars and offshore turbidites and mudstones (Turi Formation). In the Late Eocene, the eastern part of the basin was dominated by fine-grained clastic sediments, due to the rapid subsidence of a fore-deep (Higgs et al. 2012; Strogen et al. 2017; Leila et al. 2022; Radwan et al. 2022). The Turi Formation is followed by the Tikorangi carbonates that were accumulated in both the western and eastern parts of the basin (King and Thrasher 1996; Elmahdy et al. 2023). During the Miocene, a rapid subsidence of the basinal fore-deep prevailed resulting in the deposition of the deep-water bathyal mudstones of the Manganui Formation. Moreover, synchronous deposition of the Moki and Mount Messenger formations took place in palaeobathymetric lows of the eastern belt (Radwan et al. 2021, 2022). Additionally, the Giant Foresets represents a southeast prograding wedge of deposits that were accumulated during a period of increased detritus supply. Another phase of rapid subsidence took place during the Early Pliocene which produced a local deep-water conditions, however, in the Late Pliocene the basin became fully-filled with prograding shelf/slope siliciclastics (Knox 1982; Kamp et al. 2004). The study area (Kerry Field) locates in southern edge of Manaia Anticline. The region is highly-fractured with a NW–SE trending normal fault pattern (Schmidt and Robinson 1990) (Fig. 1). The reservoir section in the Kupe–Momoho area is a +400 m thick sequence of Paleocene lithic sandstones and shales overlain by Oligocene mudstones of the Otaraoa Formation. The Paleocene interval has been divided into two units: the Puponga and Farewell formations. On seismic, the boundary between these units is represented by a prominent seismic marker (“Near Top Puponga” seismic marker) that can be mapped over a large part of the Kerry 3D seismic survey area (Alotaby 2015; Kozak 2018).

Fig. 2
figure 2

A generalized lithostratigraphic column of the Taranaki Basin showing the main lithostratigraphic units (King and Thrasher 1996; Strogen et al. 2014)

Full size image

3 Material and method

The Kerry 3D marine dataset was acquired by GECO-PRAKLA on board the MV Geco Resolution vessel. Reprocessing of the KS96 Kerry 3D marine seismic volume performed in 1996. The approximate survey and processing volume is 410 km2. The processing is for the period March 2004 to November 2004 and is registered under CGG’s (Compagnie Générale de Géophysique). Processing was carried out according to a sequenced framework agreed upon between Origin Energy and CGG. Extensive testing was performed for each processing step, with results documented and presented in the form of technical notes. Origin Energy reviewed and approved all test parameters before production. The processing was done with a processing record length of 5000 ms, a processing sample rate of 2 ms that was resampled to 4 ms, and a grid size of 12.5 m × 25 m.

3.1 Workflow

In this study’s workflow, data conditioning, estimation, and collection of seismic attributes, as well as UVQ calculations, were used (Fig. 3). The study’s methodology included the examination of data conditioning, analysis, and selection of appropriate seismic attributes, automatic fault extraction, and the UVQ-NN process. The data were prepared in the first phase, which included data conditioning. The data conditioning involved geostatistical and geometrical filters to improve and polish the seismic data's structural features. In the second stage of the selection process, several geometrical attributes were evaluated and tested. The seismic attributes that extracted the best results from discontinuities were prioritized. After that, the TFL is inserted as input to compute the VCF and generate the fault patches automatically. The final phase involved testing and training the prioritized seismic attributes for neural network analysis.

Fig. 3
figure 3

The detailed workflow chart shows the building of the SSF model through three main steps: data conditioning, optimum seismic attribute selection and extraction, VCF, automatic fault extraction, and finally, UVQ analysis

Full size image

3.2 Seismic conditioning

Noise disturbances can distort the visibility of geologic features and faults in the seismic data. Additionally, this causes the quality of the data to decrease, making it extremely challenging to precisely map features (Alves et al. 2015). Hence, to address these challenges, improve signal quality, and decrease undesirable (noisy) signals, the data should be effectively conditioned. To effectively eliminate undesired noise from the 3D seismic volume, we have employed median and diffusion filters in a systematic manner (Höcker and Fehmers 2002). First, the 3D seismic data was examined for noise and distortions that might not have been eliminated (or may have been added) during the processing process, such as acquisition footprints and processing artifacts. In order to exclude these later distortions from the 3D seismic dataset, a semi-automated geostatistical filtration was used (Magneron and Petit 2008; Piazza et al. 2015).

A dip-steering (DS) volume was then calculated as the next stage of the data conditioning process. First, we build the DS seismic volume by getting the dip and azimuth parameters along the events we want to find (Tingdahl and de Rooij 2005). This procedure produced two separate steering cubes: (1) the detailed steering cube, which stores small seismic signal variations in the 3D volume (filter step-out in the present work: crossline/inline/sample: 1/1/3); and (2) the extracted 3D background steering volume, which stores the overall pattern of all these variations (filter step-out in the present work: crossline/inline/sample: 5/5/5) in order to follow the target of data conditioning (Kumar 2016). Moreover, the background steering (BS) volume is provided through the conditioning of seismic data.

The FFT of the BS volume was used to create a steering cube after a geostatistical filter was applied initially. Utilizing the dip and azimuth measurements, the DS volume was constructed (Tingdahl and de Rooij 2005). Multiple filters were included in the BS volume. After that, the BS volume was used to examine faults that are linked with the dip. The seismic volume’s stratigraphic and structural data were therefore extracted using FFT. Each seismic window sample was bordered by a cube of 3 × 3 × 3 samples, within which the dip was calculated.

The DSMF was employed in order to reduce noise and enhance the azimuth and dip data of the DS volume that was created. The flattening effect on the faults was then eliminated by applying a diffusion filter to the DSMF. The surrounding faults in the noisy seismic data were improved by the DSDF. To create an FEF, both DSMF and DSDF were combined. The results were adjusted to the DSMF apart from irregularities and sharpened DSDF at SSFs locations, and a predefined level of similarity was used to compute the FEF. The FEF improved the visualization of the SSFs, minimized the noise, and improved reflector smoothness.

At inline 653, the output gives a comparison between the original seismic data (Fig. 4a), DSMF output (Fig. 4b), DSDF output (Fig. 4c), and FEF output (Fig. 4d). Based on the displayed data conditioning and quality enhancement output, the FEF achieved a remarkable performance in identifying the SSF faults that were denoted by the black arrows.

Fig. 4
figure 4

Seismic sections at inline 653 from the 3D seismic volume: a the original seismic section; b DSMF output; c DSDF output; and d FEF output. The black arrows mark the major faults by showing the effect of each filter in the discontinuity detection. The FEF shows signal-to-noise enhancement, improvement of the reflectors’ lateral continuities, and clear visualization of discontinuities

Full size image

3.3 Seismic attributes

The subsurface’s heterogeneity frequently makes it challenging to represent a desired geological feature, such as SSF and fractures, most practically. Whenever seismic discontinuities are detected using several attributes and the appropriate filters, they may indeed be described more precisely (De Groot et al. 2001; Kuznetsov et al. 2016, 2017; Naseer 2020, 2021a, 2021b; Tayyab et al. 2017; Ismail et al. 2020, 2021; Gammaldi et al. 2022; Nasser 2023). Throughout this study, various complex and geometrical attributes were tested after extraction. The similarity, polar dip, and maximum curvature attributes combined with the FEF, DSMF, DSDF, VCF, and TFL were the most helpful in defining the structural configuration of the 3D seismic volume.

Evaluating various attribute values and employing multiple attributes may yield the best results. The optimal attributes gather valuable information and are combined into one optimized fault volume. Coherent noise was attenuated using the DSMF, and discontinuities were then detected and improved via seismic TFL attribute extraction.

To determine the greatest variety of structural changes, geometrical attributes were used. Initially, we choose the attributes that provided the most detail on structural discontinuities. This part includes a summary of the geometric attributes that were selected. A significant amount of effort has been invested in developing geometrical attributes for seismic object visualization and identification (Gersztenkorn et al. 1999; Tingdahl et al. 2001; Tingdahl 2003). Geometrical attributes are best determined by finding the direction of maximum similarity at each point. Dip can find the best similarities between every two traces. Similarity may be calculated using hyperspace distance and the trace samples as vectors. By integrating similarities and other attributes, users may potentially get better results in a short time. Consider trace segments as hyperspace vectors.

3.3.1 Similarity attribute

The coherence attribute is an advanced method that is widely utilized in the process of defining seismic discontinuities. It might be challenging to extract the coherence characteristic in an efficient manner from field data (Lou et al. 2021). Different coherence algorithms can distinguish between all different types of discontinuities and have found widespread use in the process of extracting geological faults (Lou et al. 2019; Li et al. 2021). We utilized the similarity attribute, a class of coherence attributes that represents the degree to which two or more trace segment appearances are similar (De Rooij and Tingdahl 2002; Tingdahl and De Rooij 2005; Ismail et al. 2021). The concept of similarity originated in elementary mathematical concepts: the points on the trace are interpreted as the distinct elements of a vector, and the similarity itself is considered in terms of the distance between points in hyperspace. The similarity attribute describes how closely two or even more trace segments match each other where it belongs to a sort of “coherency” attribute. A similarity ranged from 0 to 1, where a similarity of 1 represents perfect waveform matching between the trace segments, whereas a similarity of 0 represents total dissimilarity. According to this attribute, small faults may be more clearly detected. The major faults and SSFs that were not visible in the seismic volume (Fig. 5a) are clearly identifiable in Fig. 5b, which displays the similarity attribute at time slice 1500 ms. Moreover, the similarity attribute was extracted from the FEF seismic cube to evaluate the effect of the FEF on the SSFs and discontinuities in the seismic volume in Fig. 6.

Fig. 5
figure 5

a The original amplitude seismic time slice at 1500 ms; b after extracting the similarity attribute, which shows that low similarity values near zero represent faults, discontinuities, and fracture zones (marked by white arrows)

Full size image
Fig. 6
figure 6

The extracted similarity attribute from the FEF seismic volume: a the original 3D seismic volume; b seismic time slice at 1500 ms. a and b show enhancement of the SSFs' detection, where low similarity values near zero represent faults, discontinuities, and fracture zones (marked by white arrows)

Full size image

3.3.2 Curvature attribute

A surface’s curvature in a particular position is described as having curvature. It may be described as the reversal radius of the osculating circle at a certain position on a curvature in two dimensions (Roberts 2001). Using characteristics of curvature to detect the extent of faults and fractures (Roberts 2001; Chopra and Marfurt 2007). Generally, curvature absolute values show the faults’ and fractures’ intensity (Samson and Mallet 1997; Sigismondi and Soldo 2003). The extensions of the faults that were not identified in the original seismic volume (Fig. 7a) are readily identifiable in Fig. 7b, which displays the maximum curvature attribute at a time slice of 1500 ms.

Fig. 7
figure 7

a The original amplitude seismic time slice of 1500 ms; and b after extraction of the maximum curvature attribute, which shows the high curvature values represent faults, discontinuities, and fracture zones

Full size image

3.3.3 Polar dip attribute

The horizontally calculated polar dip (true dip) was calculated, with a range that was always positive. The faults of the study area, indicated by a white arrow, were distinguished by the polar-dip attribute on the time slice at 1500 ms (Fig. 8b). In terms of detecting faults and discontinuities, the polar dip outperformed most other dip attributes (Ashraf et al. 2020).

Fig. 8
figure 8

a The original amplitude seismic time slice of 1500 ms; and b after computing the polar-dip attribute, which shows the extensions of the faults and fracture zones (marked by white arrows)

Full size image

3.3.4 Thinned fault likelihood (TFL)

The sample location fault probability is indicated by the FL and TFL attributes. 3D structure-oriented semblance volume scanning is fault probability (FL), where mathematically, FL resembles semblance. To improve fault likelihood, it searches for semblance in numerous fault orientations. The fault plane has the lowest semblance and highest probability (maximum FL). The scanning ends with visualizations of maximal fault probability (FL) and fault strikes and dips (Hale 2012; Ashraf et al. 2020; Imran et al. 2021). The TFL attribute was created with razor-sharp faults for further fine-tuning. This automated technique uses Hale’s fault-oriented semblance algorithms to clearly indicate fault planes. The derived attributes demonstrate a detailed structural framework. It contributed to the seismic fault modeling and recorded various minor features that were challenging to recognize in the original seismic volume. Figure 9 shows the computed TFL attribute from the DSMF seismic volume.

Fig. 9
figure 9

The computed TFL from the FL volume shows all the razor-sharp discontinuities across the seismic volume, which include the major faults, SSFs, and fractures. White pointers indicate 6 faults of large-scale faults

Full size image

3.3.5 Fracture density

This attribute refers to the volume of faults and fractures that might indeed be detected within a user-defined radius for scanning and calculating the number of fractures within the given radius. In this instance, this attribute demonstrated an improvement in the visibility of possible fracture zones. The FEF was utilized to generate the fracture density volume. The fracture network zones are shown in Fig. 10b, where the highest fracture zones appear in black.

Fig. 10
figure 10

a The original amplitude seismic data; and b after computing the fracture density attribute, which shows where the highest fracture zones appear in black

Full size image

3.4 Artificial neural networks

ANN methodologies have been effectively used throughout recent years in seismic data analysis and petroleum exploration (Rebai et al. 2019; Shalaby et al. 2019; Farfour and Mesbah 2020; Yang and Sun 2020; Ismail et al. 2022). NN processing (Sheng et al. 2019; Ashraf et al. 2020) can be used to get detailed results by using several seismic attributes to analyze a huge dataset efficiently and precisely. This improves the extraction and understanding of various features obtained from seismic data. In general, ANN algorithms operate as a multi-channel computational unit, with a number of processors with defined inputs and entirely optional defined outputs for SSFs and fracture detections.

3.4.1 Unsupervised neural network

The UVQ works in the same way as the K-means clustering algorithms by organizing data in groups with similar structures in different clusters and then defining all targets (clusters). UVQ-NN is a computer-based neural network model that works in a sequence of parallel layers where each one receives a number of nodes interconnected with different weights based on the contribution of each node to the network. The input weight values are critical parameters in the error and misclassification calculations through iteration and output extraction. The UVQ-NN trains through a non-linear filtering process. The filtering process includes the initialization and evaluation of the input nodes (6 nodes) which consist of meta-seismic attributes tested and chosen from tens of attributes, these attributes classes are associated with the abrupt change in the amplitude of the faults, discontinuities, reflection terminations, and the dip variation where each of them works to improve the faults visualization and auto-detection. Subsequently, each neuron inside the hidden layer, consisting of three nodes, establishes connections with the neurons residing in the following layer. The synaptic connections between neurons are referred to as weights, which undergo updates during the learning phase. After that, the iteration process is performed to compute the UVQ-NN average matching. To follow the UVQ-NN performance through the iteration process, the average match of the input nodes (clusters) is tracked along the process which increases in a step-function by finding a new cluster in each step. The size of the dataset used, the software, and the parameter settings for the UVQ-NN model are shown in Table 1.

Table 1 The size of the 3D dataset and parameter settings for the UVQ-NN model
Full size table

The UVQ-NN is trained using a back-propagation NN by the following equations:

$$x = \Sigma^{n} w_{i} \times A_{{\dot{l}}}$$
(1)
$$P = \frac{2}{{1 + e^{ - x} }} - 1$$
(2)

where A is the input sample vector, \(x\): weighting function, w: weighting vector, n: number of input vectors, P: output.

Unsupervised NN classifies the input information into various classes depending on their attributes. It operates as a self-organized network. In 3D space, the UVQ strategy is typically focused on identifying and categorizing the data from inserted input samples into clusters. UVQ analysis is crucial in petroleum exploration and structural analysis in developing concessions because it allows the seismic volume to be restructured and divided into various classes, each with unique features or parameters, entirely depending on the details obtained from the 3D seismic volume and without the need for supplementary information collected from logs or cores (Saggaf et al. 2003; Ismail et al. 2022). Meta-attributes, such as the TFL, FEF on similarity, fracture density, steered similarity, polar dip, and maximum curvature attributes, are considered to visualize and allow discrimination of the faults and fracture zones that are distinguished by extremely low amplitudes from seismic volume. Before selecting the optimal attributes package to serve as input for the UVQ network procedure to image SSFs and fractured zones, each of these attributes is put through a set of tests.

The UVQ analysis organizes the input vectors defined by the tested and chosen attributes package and identifies cluster centers in the seismic volume that provides various colors for various classes (faults and fractures in class 1 and seismic reflectors without discontinuities in class 2) (Fig. 11). The average matching between both the selected and tested meta-seismic attributes package and the number of training vectors demonstrates how the output vector has a high matching percentage of around 89% (Fig. 12). Table 2 shows training attribute weights.

Fig. 11
figure 11

The UVQ-NN with the input layer contains six nodes, the hidden layer has three nodes, and the output layer has one node. The colored legend shows the weight and contribution of every single attribute

Full size image
Fig. 12
figure 12

The average matching percentage shows increasing match rates, which reached 89% after 410,000 training vectors

Full size image
Table 2 During UVQ training, the contributions (weights) of each input node were shown by the weight of each seismic attribute, with red showing the highest weight and light-yellow showing the lowest
Full size table

4 Results

At inline 653, data conditioning and quality enhancement generate three outputs represented by the (Fig. 4b), DSDF output (Fig. 4c), and FEF output (Fig. 4d) where the FEF contributed effectively to identify the SSF highlighted by the black arrows. The similarity attribute is generated to extract small faults that may be more clearly identifiable in Fig. 5b, which displays the similarity attribute at a time slice of 1500 ms. In addition, the similarity attribute was derived from the FEF seismic cube to assess the effect of the FEF on the SSFs and discontinuities in Fig. 6. Another geometrical attribute that assists to detect the extensions of the faults that were not identified in the original seismic volume is the curvature attribute (Fig. 7b), which displays the maximum curvature attribute at a time slice of 1500 ms. Moreover, the polar dip, TFL, and fracture density attribute results allow discrimination of the faults and fracture zones that are distinguished by extremely low amplitudes from seismic volume (Figs. 8, 9, and 10).

3D matching and segmentation volumes yield the best UVQ-NN results (Fig. 13). The segment and corresponding 3D seismic volumes (Fig. 13a) match major faults, SSFs, and fracture zones at an average of 89% matching. The 3D segment volume uses two-color classes to segment clearly (Fig. 13b). Class 1 shows major faults, SSFs, and fracture zones in black, whereas class 2 shows no faulted or fractured zones. 3D segment and match volumes produce effective segmentation and matching images. A new volume was computed and ranked based on discontinuity body size, and the VCF was utilized to provide the outputs of the connected fracture bodies, as illustrated in Fig. 14. After that, a completely automated fault extraction approach was applied to the TFL volume for fault auto-tracking and fault patching. Figure 15a demonstrates that both major and SSF patches were detected, whereas Fig. 15b shows just the filtering and extraction of the SSF patches, which are more obvious in the deeper part of the model.

Fig. 13
figure 13

a The matching 3D volume distinguishes the background (zones that are neither faulted nor fractured) from the main faults, SSFs, and fracture zones. b The 3D segment volume shows a clear and exact segmentation, where class 1 shows major faults, SSFs, and fracture zones in black and class 2 shows areas that are not faulted or fractured

Full size image
Fig. 14
figure 14

The VCF derived from the TFL depicts all major and SSF faults as well as minor discontinuities. The white elliptical shape shows a high fracture zone, while the major faults are marked with white arrows

Full size image
Fig. 15
figure 15

a A 3D visualization of all major faults and SSFs’ surfaces. b The 3D view shows the surfaces of filtered SSFs

Full size image

The fracture bodies were primarily oriented northeast-southwest, and major fault patches were generated in the seismic volume’s shallower zone, while SSF bodies were discovered at greater depth. The Rose diagrams (Fig. 16a, b) depict all TFL samples over a threshold. The Stereonet showed around 350 fault patches in various orientations from the calculated faults (Fig. 16c). The results demonstrated that the majority of patches were concentrated in the northeast-southwest direction.

Fig. 16
figure 16

Rose diagrams (dip) in (a) and (strike) in (b), respectively. c Stereonet automatically displays the orientation, dip, and azimuth of all major faults and SSFs surfaces

Full size image

5 Discussion

The recording and processing of seismic data could create noise in the data. Extracting structural-based attributes before data conditioning may generate geometrical errors linked to disconformities and faults. Conditioning seismic data is a solution for reducing noise and enhancing the smoothness of reflectors, geobodies, features, and fault sharpness. The data conditioning procedure is able to enhance the quality of the real signal generated by the presence of stratigraphic and structural feature variations. Therefore, the primary purpose is to increase the signal-to-noise ratio while keeping important features such as main faults, SSFs, and amplitude variation linked to different other geological features.

Therefore, significant seismic volume enhancements can be achieved by designing effective filters such as DSMF, DSDF, and FEF (Fig. 4) that enhance different types of faults before using the generated output cube as input to extract various geometrical attributes.

Understanding the structural framework of the subsurface is essential for constructing an accurate and reliable geological model. This has typically been performed by manually identifying the faults detected on the 3D seismic cube. This is extremely significant for low-quality seismic volume. Using advanced methods helps the interpreter identify all discontinuities in the 3D seismic cube. The interpreter can construct a preliminary fault network from a 3D seismic volume using automatic fault segmentation with very minimal operator interaction. The detected faults can be efficiently adjusted afterward in a considerably shorter period.

Automated fault recognition and extraction developments have increased dramatically during the previous two decades. Several seismic attributes, like similarity, TFL, polar dip, fracture density, and curvature attributes, have been developed and used to find fault positions and extensions (Figs. 5, 6, 7, 8, 9, and 10).

Unsupervised learning includes several major learning algorithms such as clustering algorithms (K-means, DBBSCAN, agglomerative, mean shift, and fuzzy c-means), associate rule learning algorithms (Fb growth, euclat, apriori), and dimensionality reduction algorithms (t-SNE, PCA, LSA, SVD, and LDA). All these algorithms have the advantage of working without direct supervision, but the disadvantage is that the accuracy of the final output model depends on the quality of the original dataset and the weight values of the attributes inserted into the input layer. Therefore, if the dataset contains a high noise ratio or the inserted attributes are not tested, the final output model will have a high error percentage. Therefore, the UVQ-NN is chosen because it is considered one of the simplest algorithms for computing the optimal centroids of all clusters using optimal and tested attributes, which helps discriminate between fault and non-fault zones in the output model.

The present method does not include displacement analysis and maybe this is one of the limitations of the present work but the selected meta-attributes, such as the TFL, FEF on similarity, fracture density, steered similarity, polar dip, and maximum curvature attributes gather valuable information and are combined into one optimized fault volume and considered to visualize and allow discrimination of the faults, cross-faults, and fracture zones.

5.1 UVQ models

Prototype vectors are updated until they fall in cluster centers. Cluster centers are sorted in the training and iteration processes of the input nodes (seismic attributes). The output layer of the network extracts two 3D volumes (outputs), the first shows the winning clusters of the network target (segmentation) and the second output is the match (confidence), which is a matching value between 0 and 1 computing and displaying how the input vector is close to the winning cluster vectors. Generally, the step function boosts the average match. Every step represents the identification of an additional cluster by the network. Once the average match (confidence) has attained an acceptable level of stability, training can end.

The 3D matching volume indicates the confidence level between the input data (multi-attributes) on a scale ranging from 0 to 1, where 0 means vectors that become dissimilar and 1 means similar (winning) vectors. The optimal and most acceptable scale range is 0.5–1 (Fig. 13a). The matching process effectively produces a new 3D volume that distinguishes the background (zones that are neither faulted nor fractured) from the main faults, SSFs, and fracture zones. The 3D segment volume categorizes seismic events linked to derived attributes into two classes with different colors (Fig. 13b).

The highest quality result of the UVQ-NN is given by 3D matching and segmentation volumes (Fig. 13). These two new 3D seismic volumes are the segment and matching 3D seismic volumes, whereas the 3D matching volume (Fig. 13a) identifies the major faults, SSFs, and fracture zones with an average of 89% matching. The 3D segment volume displays a clear and exact segmentation utilizing basically two-color classes (Fig. 13b). Class 1 depicts major faults, SSFs, and fracture zones in black, whereas class 2 represents zones that are not faulted or fractured. In the presentation of the main faults, SSFs, and fracture zones, both 3D segment and match volumes produce segmentation and matching images.

5.2 Automatic fault and fracture network extraction

However, the attribute values can be affected by seismic signal inhomogeneities near the faults (Fehmers and Hocker 2003; Philit et al. 2019; Ashraf et al. 2020). This makes it hard to see where the faults are. For the aforementioned reasons, the VCF, supported by geometrical attributes and filters, is utilized to generate a 3D structural model that displays SSFs and key faults with a color bar that ranks these faults depending on size. The resulting TFL volume (Fig. 9) was used as input for the VCF to estimate the fracture zones.

The VCF was performed employing the envelope of amplitudes greater than 0.6 which is redefined to generate the fracture bodies. Propagation was accomplished in all directions (26) by making use of all volume sides and corners. The extracted volume was computed and ranked based on discontinuity body size, and the VCF was used to present the outputs of the linked fracture bodies that were created, as shown in Fig. 14.

After choosing the appropriate parameters, a fully automated fault extraction technique was performed on the TFL volume for fault auto-tracking to create fault patches. Figure 15a shows that both major and SSF patches were found, while Fig. 15b shows only the filtration and extraction of the SSF patches, which are more visible in the deeper part of the model.

The SSFs and fracture patch orientations were shown in the results. The fracture bodies were primarily oriented northeast-southwest, and major fault patches were generated in the seismic volume's shallower zone, while SSF bodies were discovered at greater depth.

After the fault surfaces have been automatically produced, we extracted both stereonet and Rose diagrams (dip and strike) (Fig. 16); the Rose diagrams (Fig. 16a, b) gather and display all TFL samples over a predetermined threshold. On the Stereonet, the outputs of the derived faults were demonstrated (Fig. 16c), and they revealed more than 350 fault patches dispersed in a variety of orientations. The results demonstrated that the majority of patches were concentrated in the northeast-southwest direction.

The rose diagram (Figs. 16a, b) and final UVQ-N, VCF, and automated fault extraction from the TFL models (Figs. 13, 14, 15, and 16) show high matching for most of the fault trends and orientation with recent studies derived by using conventional interpretation and modeling in the Taranaki basin (Haque et al. 2016; Kenyon 2016; Qadri et al. 2017; Thota et al. 2021) and compatible with the tectonic stress pattern (Rajabi et al. 2016).

In summary, this work developed an innovative methodology that detected and automatically extracted both SSFs and larger faults. It is fast and may be accomplished in a relatively short amount of time compared to manual interpretation.

6 Conclusion

This study introduces an improved workflow for automating the recognition and extraction of large-scale faults, FFSs, and fracture network zones in a 3D seismic volume. It is based on both UVQ-NN and automatic fault extraction combined with geometrical and unconventional attributes that deal with 3D fault-oriented semblance and therefore able to recognize SSFs and fracture networks even when reflectors are relatively uninterrupted through fracture networks and faults. Its application to reservoirs with various faults extracted the major and minor faults. Providing precise structural modeling, the outputs are encouraging and could play a significant role in prospect discovery. Combined with both the dip and azimuth considerations, this new workflow innovation shows high accuracy and effectiveness in evaluating and developing mature reservoirs. The ANN–UVQ and automated fault extraction analyses indicated a fault with a minimal heave and throw orientation (NNE–SSW). The strategies utilized in the work for the modeling of major faults, SSFs, and fracture networks are automated, adaptable, time and cost-efficient, accurate, and innovative. The implemented approach is innovative and can be employed in the future for the identification, extraction, and classification of different-scale discontinuities in any region of the world.