Integrating Remote Sensing and Knowledge-Based Systems for Structural Lineament Mapping in the Rif Belt

Abstract

This study presents a novel methodology for mapping Fault- and Thrust-based Structural Lineaments (FT-SL) in the rugged and inaccessible Oued-Laou watershed of the Rif Belt, Morocco. Combining optical (Landsat-8 OLI, Sentinel-2 MSI) and radar (Sentinel-1 SAR) remote sensing data, the research employs manual, semi-automatic, and automatic extraction methods enhanced by spatial filtering (Sobel, Laplacian, Kuan). A Knowledge-Based System (KBS) integrated with Multi-Criteria Decision Analysis (MCDA) evaluates the effectiveness of these methods, focusing on lineament statistics, orientation, density distribution, and correlation with existing geological maps. The results highlight Sentinel-1 SAR’s superior performance in detecting subsurface structures, while manual extraction yields the highest accuracy. This study also demonstrates the potential for generalizing this approach to other Alpine orogenic regions, such as the Alps, due to shared geological characteristics. The findings provide a robust framework for structural lineament mapping in mountainous terrains, addressing challenges of accessibility and data scarcity.

1. Introduction

Geological maps are fundamental tools for understanding regional geology, underpinning natural resource exploration/exploitation, hazard assessment, environmental studies, and scientific research [1,2]. However, detailed geological and structural mapping in mountainous regions remains globally incomplete. Existing maps (typically 1:100,000–1:200,000 scale) and limited field data density lack the resolution required for accurate geological and structural characterization of these terrains [1,3,4]. This deficiency is particularly pronounced in Morocco’s Rif and Atlas Belts, due to intrinsic mountain constraints [5,6,7]: (1) rugged topography and poor accessibility limit outcrop exposure, increasing field data acquisition difficulty and cost; (2) intense erosion obscures key geological indicators, complicating formation characterization and dating; (3) complex tectonics (faulting, thrusting, fracturing) create intricate structures challenging to interpret; and (4) pervasive surface cover (scree, soil, vegetation) obscures bedrock observation.

These challenges necessitate novel approaches for accurate mountain mapping. Remote sensing (RS) has emerged as a critical tool for geological structure mapping [1,8,9,10] providing consistent spatial coverage of inaccessible terrain. Remote Sensing data (RSD), acquired at varied spatial, spectral, and temporal resolutions, offers effective solutions to limitations of conventional field mapping [2,5,11,12]. This research addresses the challenges of mountainous terrain access by developing an alternative to the resource-intensive (time, cost, labor) traditional manual approach. The central question is: Which RSD and extraction methods are optimal for efficiently mapping and maximizing detection of Fault- and Thrust-based Structural Lineaments (FT-SL) in mountainous areas?

RS is now fundamental in Earth sciences for geological feature extraction and mapping [3,4,9]. RSD—acquired across diverse spatial, spectral, and temporal resolutions—address key challenges and limitations of conventional geological field mapping, which is often costly, time-intensive, or impeded by geomorphological, geographical, and socio-political constraints [5,8,9,12]. Consequently, RS-based geological lineament detection has gained prominence for its effectiveness [11,13].

Lineaments are genetically classified as follows: (i) geological (tectonic origins: faults, fractures, thrusts, lithological boundaries); (ii) topographic (geomorphological processes: drainage, ridges); or (iii) artificial (human infrastructure: roads, field boundaries) [14,15,16,17]. Geological lineaments manifest in satellite imagery as linear features with distinct pixel intensity contrasts [11,14] and represent significant crustal structures like faults, fractures, thrusts, and structural anomalies [18,19]. SL mapping identifies bedrock features (thrust faults, shear zones, fractures, folds) from RS data [20,21,22,23]. It is critical for: mineral exploration [12,24,25,26,27]; hydrocarbon assessment [28,29]; geomorphology and drainage networks studies [30,31,32]; groundwater exploration [33,34,35]; geothermal studies [36,37]; and environmental and natural hazard assessment [3,38,39,40].

RS technologies have advanced significantly over three decades, encompassing diverse sensors and techniques across multiple imaging scales [3,14,31,41,42]. Selection of RSD depends on investigation objectives. Multispectral imaging (MSI) has historically enabled visual geological analysis [9,43], with recent progress driven by advanced sensors (e.g., Landsat ETM+/OLI, ASTER, Hyperion) [44] and digital processing innovations [3,7,9]. These developments facilitate lithological discrimination [6,7,45] and structural mapping [22,32,46] from airborne/spaceborne spectral data, enabling integrated geological mapping methodologies [3,9,22,32].

Lineament extraction methodologies encompass manual [47], semi-automatic [48] and automatic approaches [49]. Automated extraction reduces reliance on expert interpretation compared to manual and semi-automatic methods. While manual extraction offers simplicity, its subjectivity, time-intensive nature, and operator-dependence constrain efficiency and reproducibility [20,31,50]. Automated techniques improve processing speed and objectivity but incur computational complexity and sensitivity to edge conditions. These algorithms detect lineaments via gradient magnitude variations, pattern discontinuities, or linear minima/maxima [20,38,51], corresponding to geomorphic (e.g., valleys, ridges) or geological features (e.g., faults, fractures). Lineaments are topologically classified as positive (ridge-aligned) or negative (valley-aligned, including joints, fractures, faults, shear zones). Automated methods exhibit higher false positive rates (commission errors: non-geological features misclassified as lineaments) and false negative rates (omission errors: genuine lineaments undetected) relative to expert-supervised manual techniques [20,52,53,54,55,56,57]. Validation challenges arise from: (i) Algorithmic limitations in distinguishing geological vs. non-geological linear features; (ii) Variable terrain/vegetation interference; and (iii) Insufficient training data for complex geological settings. Widely implemented algorithms include: (i) The LINE module for structural lineament extraction; (ii) Regional structural-tectonic analysis tools quantifying position and intensity of geological structures [58].

Spatial filtering enhances edge and linear feature detection in remote sensing data by emphasizing high-frequency spatial information critical for identifying geological discontinuities. Gradient-based filters, including Sobel/Prewitt operators, detect edges through discrete differentiation but exhibit noise sensitivity [59], while Laplacian-of-Gaussian (LoG) filters combine Gaussian smoothing with edge detection to improve performance in low-contrast terrains [60]. Directional filters such as the Canny edge detector optimize noise reduction and edge thinning for precise fault mapping when integrated with DEM data [61], whereas the adaptive Kuan filter effectively despeckles SAR imagery while preserving structural edges [62]. Multi-scale wavelet transforms isolate lineaments across spatial frequencies, enabling differentiation of regional faults from local fractures [59,61,62,63]. Hybrid approaches include filter fusion (e.g., Sobel/LoG combined with PCA) to enhance lineament connectivity in complex lithologies [12], and ML-enhanced filtering where CNNs leverage filter-derived features to reduce false negatives by distinguishing geological from anthropogenic lineaments [64].

A 1975–present literature review by [14] compares optical and radar LiDAR RSD for automatic SL extraction [65]. The automated methodology requires optimization of edge detection, line-linking parameters, and band selection for SL mapping. Optical data (Landsat TM/ETM+/OLI/TIRS, ASTER, Sentinel-2) [3,5,6,19,38,66,67,68,69,70,71,72,73,74], radar data (SRTM DEM, SAR, InSAR, Sentinel-1) [47,75,76,77,78,79,80,81,82,83], and LiDAR-derived digital elevation models [84,85,86] are widely deployed for fault/thrust characterization. Landsat OLI dominates automatic lineament extraction studies [6,8,9,12,30,33,34,38,50,54,76,87], frequently combined with SAR data [4,31,36,53,57,88] or LiDAR datasets [89,90]. Sentinel-2 MSI is increasingly used due to its advantageous spectral properties [3,6,91,92], while LiDAR excels in high-resolution topographic mapping of vegetated terrains [93,94].

Artificial intelligence (AI), particularly deep learning (DL), is transforming structural geology by enabling automated detection and characterization of geological structures from satellite imagery. This shift from manual to data-driven methods enhances the mapping of faults, fractures, and folds with improved speed, accuracy, and scalability. Recent research focuses predominantly on DL due to its efficacy in image analysis [14,20,95,96]. Convolutional Neural Networks (CNN), including architectures such as U-Net and ResNet, are extensively applied to detect and segment linear features (e.g., faults, fractures) in high-resolution optical and synthetic aperture radar (SAR) imagery. These models leverage spatial hierarchies learned from annotated datasets to achieve high precision [96,97]. Beyond CNN, Vision Transformers (ViTs) demonstrate superior performance in identifying large-scale structural patterns (e.g., fold belts) by capturing long-range dependencies in complex terrains [25,98,99]. To mitigate limitations in training data, Generative Adversarial Networks (GAN) generate synthetic structural features, enhancing model robustness [10,100]. Current DL models exceed 85% accuracy in mapping lineament density and orientation using Sentinel-2 and Landsat-8 imagery, facilitating fracture network analysis for resource exploration [101]. Recurrent CNN applied to multi-temporal data enable detection of active fault traces in seismically active regions [14,102].

Persistent challenges in deep learning (DL) applications for structural geology remain. A major limitation is data scarcity, which is the lack of sufficient high-quality and globally diverse labeled datasets needed to train robust models [14,97,103]. Furthermore, the transparency of models is a critical issue. The “black-box” nature of complex DL architectures can hinder trust and acceptance among geoscientists. Knowledge-Based Systems (KBS) address these limitations by formalizing expert geological knowledge to interpret heterogeneous data (e.g., RSD, field observations, seismic data, boreholes) [104,105]. KBS integrate expert systems and AI frameworks to overcome challenges in analyzing complex, incomplete, and spatially variable datasets while applying conceptual models [100,105,106,107,108]. Key applications encompass: Interpretation/classification of structural features (fault systems, fractures, folds), Geological mapping and 3D model construction, Structural synthesis and deformation history analysis and Data integration with uncertainty quantification. This has prompted in-depth consideration of how this AI technology could be applied to the multi-criteria analysis of SL mapping results.

This study presents a novel methodology for geospatial modeling of Structural Lineaments (SLs) within the mountainous Rif Belt, integrating optical (Landsat-8 OLI, Sentinel-2 MSI) and radar (Sentinel-1 SAR) satellite imagery. The approach combines a Knowledge-Based System (KBS) with Multi-Criteria Decision Analysis (MCDA) to systematically evaluate the potential of these multi-source Remote Sensing Data (RSD) for mapping fault- and trust-based Structural Lineaments (FT-SL) within the rugged and largely inaccessible Oued-Laou watershed.

The Oued-Laou watershed, situated within the Rif Geological Domain of Morocco, was selected as the representative study area due to its diverse lithological units and complex tectonic structures, rendering it highly favorable for structural geological analysis. Encompassing approximately 2000 km², the study area was subdivided into four zones. Three zones benefit from existing detailed geological mapping at 1:50,000 scale; these serve as training and validation sites. The fourth zone, located in the southwestern (SW) sector, is of particular significance: it currently lacks detailed geological mapping, with only an outdated provisional map of the Rif Domain (1:200,000 scale) available. Consequently, this research will produce the first edition of the detailed structural map of the SW Oued-Laou watershed.

The findings from this intelligent approach—which integrates Knowledge-based Systems (KBS), MCDA (Multi-criteria Decision Analysis), and optical and radar satellite imagery for the geospatial modeling of SL in the Oued-Laou watershed—can be generalised to the entire geological domain of the Rif and Betic Cordillera [109,110,111]. In fact, the Oued-Laou watershed region was chosen as the study area because it is considered representative of the Rif Belt: containing the geological formations and structures of the three main domains of the Rif Belt (the Internal units: Gomarides, Septides and Dorsale Calcaire, the intermediate or Flysch Units and the External units). The results of this approach can also be investigated and verified for mapping potential SL in the Alps mountain range, given that the Rif- and the Alps –Domains share geological similarities as mountain domains formed during the Alpine orogeny, both featuring geological structures such as thrusts and fold structures in addition to faults.

The methodology employed in this research can be outlined as follows: After an intensive data acquisition process involving satellite imagery, existing geological maps and field reconnaissance, the initial phase consisted of essential pre-processing of the three types of RS data (Landsat-8 OLI, Sentinel-2A and Sentinel-1B). This included radiometric calibration, atmospheric correction, thermal noise removal, spectral filtering and mosaicking, etc. In parallel, geological maps were digitised (georeferenced and vectorised) to generate detailed information layers concerning geological structures. These structural maps are essential for validating the resulting SL within the three specified test areas. Furthermore, fundamental information layers were extracted from ASTER-GDEM data, including topographic and relief maps, as well as maps of the hydrographic network and the Oued-Laou watershed boundaries.

The second phase of this work focuses on extracting the potentials FT-SL from the three RSD types. Three distinct methods of SL extraction were employed: (i) Manual Extraction: This method involves the visual identification and manual extraction of potential SL from the three RS data types by human experts. (ii) Semi-automatic Supervised Extraction: This approach leverages expert knowledge to supervise and refine automatically extracted potential SL from the three RS data types. This essential step aims to eliminate false positives lineaments, such as linear features of anthropogenic origin (e.g., roads, rail networks) and hydrographic network lineaments (e.g., main rivers and their tributaries), which are obtained from the DEM. (iii) Automatic Extraction coupled with Spatial Convolution Filtering: This method is based on spatial convolution filtering of monospectral data from RS imagery. After selecting a specific band for each of the three RS data types, spatial filtering is applied using three chosen filters: Sobel, Laplacian, and Kuan. These nine resulting filtered images are then used for automatic SL extraction.

The third phase of this study involves an integrated, comparative analysis of 15 potential FT-SL maps. These maps were generated using manual, semi-automatic supervised, and automatic (with spatial filtering) extraction methods from Landsat-8 OLI, Sentinel-1B, and Sentinel-2A RSD. This comparative analysis employs an integrated approach combining knowledge-based systems (KBS) and multi-criteria decision analysis (MCDA) techniques. Within the KBS framework, this intelligent approach involves comparing four distinct result packages, considered as “class packages”: SL Statistics (SLS-Pac); SL Orientation (SLO-Pac); SL Density Distribution (SLD-Pac) and Correlation of SLM Density with Geological Maps (SLC-Pac). An in-depth analysis is then performed on each package using the MCDA technique. MCDA is directly linked to the KBS rule base throughout all stages, including setting objectives and criteria, weighting criteria, and interpreting the final score. Consequently, a single query to our integrated system is sufficient to quantitatively evaluate the potential of a given RSD type or extraction method for mapping potential FT-SL within the Rif geological domain.

The paper is structured according to the methodological workflow adopted in this research: Section 2 characterizes the Oued-Laou watershed study area, detailing its topography, geomorphology, and, critically, its geological setting and structural framework. Section 3 comprehensively details the datasets used, the pre-processing and processing workflow, and the three distinct structural lineament (SL) extraction methodologies: manual interpretation (Section 3.2.1), semi-automatic supervised extraction (Section 3.2.2), and automatic extraction incorporating spatial filtering (Section 3.2.3). This section further specifies the mathematical kernels employed for the Sobel, Laplacian, and Kuan filters (Section 3.2.4), outlines the fundamental principles and system architecture of the developed Knowledge-Based System (KBS) (Section 3.3), and describes the underlying principles, objectives, and specific analysis criteria of the Multi-Criteria Decision Analysis (MCDA) methodology (Section 3.4). Section 4 presents the research results. Section 4.1 interprets the outcomes derived from the ‘class packages’ or ‘decision parameters’, including SL statistics (frequency, length distribution), predominant SL orientations, SL density maps (SLM), and the spatial correlation analysis between SLM outputs and existing geological maps. Section 4.2 subsequently presents and critically discusses the results obtained from the integrated KBS-MCDA approach. Section 5 provides a synthesis and critical discussion of the principal findings, contextualizing them within the broader field and proposing recommendations for future research and application. Section 6 summarizes the concise conclusions drawn from this research.

2. Study Area

The study area is situated within the Tanger-Tetouan-Al Hoceima province of the north of Morocco (Figure 1a), encompassing the complete drainage basin of the Oued-Laou River and its delineated spatial extent defined by the coordinate boundaries presented in

Table 1. Boundary coordinates (expressed in meters) using the Projected Coordinate System North Morocco (GCS_Merchich datum), detailing the left longitude (X1), right longitude (X2), bottom latitude (Y1), and top latitude (Y2) defining the study perimeters.

	Boundary Coordinates (m)				Area (m2)
	X1	X2	Y1	Y2
Study area	495,100	532,850	488,650	543,500	1,978,421,419
Zone A	495,100	513,545	516,175	543,500	504,807,018
Zone B	513,545	532,850	516,175	543,500	458,228,620
Zone C	495,100	513,545	488,650	516,175	509,789,361
Zone D	513,545	532,850	488,650	516,175	505,596,420
Oued-Laou Watershed	495,100	529,525	490,073	542,627	937,603,657

. The total area covers 1978 km², with the Oued-Laou watershed itself constituting approximately half of this area (Figure 1a,b).

The study area exhibits predominantly rugged topography characterized by steep slopes and significant elevational contrasts. Situated within the north-central sector of the Rif Belt (High Rif), it is bounded by the peaks of Jbel Kelti (1928 m) to the west, Jbel Sougna (1800 m) to the south, Jbel Tisouka (2130 m) to the southeast, Jbel Tasaot (1800 m) to the northeast, and the Mediterranean Sea to the north (Figure 1).

Figure 1b presents a topographic map of the Oued-Laou watershed derived from 30 m resolution ASTER GDEM data (ASTGTM_N35W006). This map illustrates the pronounced mountainous terrain, with elevations ranging from sea level to over 2136 m, featuring abrupt elevational changes. Steep, rugged slopes constitute more than 65% of the total area. Figure 1c displays a corresponding ridgeline map (major ridges: purple; minor ridges: green), generated from the same DEM data using GIS software, further confirming the area’s rugged and inaccessible nature. This ridgeline layer serves as critical input data for the supervised automatic extraction of FT_SL.

The drainage network of the Oued-Laou watershed (Figure 1d) was delineated through DEM analysis within a GIS environment. Processing steps included derivation of flow accumulation, flow direction, stream definition, stream segmentation, and catchment delineation, culminating in a vector representation of the drainage network and watershed boundaries. This drainage network map constitutes another essential input layer for the supervised automatic FT-SL extraction methodology.

Geologically, this study employs detailed 1:50,000 scale geological maps of the ‘Bab Taza’, ‘Talambote’, and ‘Souk Larbaa Beni Hessane’ sectors, published by the Moroccan Ministry of Energy and Mines. The regional 1:200,000 scale geological map of the Moroccan Rif (Ministry publication) supplements these datasets, providing coverage for areas lacking higher-resolution mapping (zone C). Figure 2 presents the integrated geological map of the study area, combining: zone A: Bab Taza (1:50,000), zone B: Talambote (1:50,000), zone D: Souk Larbaa Beni Hessane (1:50,000) and zone C: Moroccan Rif (1:200,000).

The Oued-Laou watershed study area encompasses all three major subdivisions of the Rif Belt. The geological formations characterizing each domain are as follows: (i) Internal Domain: Gomarides Complex, dominated by alternating coarse-grained micaceous sandstones and dark Carboniferous pelites (Akaili Unit: AK, Figure 2); Sebtides, primarily represented by Frédéric Units, notably Permian shales with intercalated conglomeratic levels (TZ Unit, Figure 2); and Dorsale Calcaire, Composed mainly of calcareous-dolomitic sequences exhibiting aggrading alteration, dating from the Carnian to Norian and Rhaetian stages (Ca, Ce formations, Figure 2). (ii) Flysch Domain, represented by the Beni Ider and Tizirene nappes, includes Oligocene-aged predorsal units, features Numidian facies sandstone nappes (Bi, N, Tf formations, Figure 2). (iii) External Domain, primarily located within the ‘Intrarifaine’ region in the extreme southwest of the study area (Figure 2), characterized by Cretaceous pelitic marls and clays (Tg, AS, BT formations, Figure 2).

As illustrated in Figure 2, the Oued-Laou watershed exemplifies the geological architecture of the Moroccan Rif chain. This chain, jointly with the Betic Cordillera, forms the Betic-Rif arc within the Alpine orogenic system. Classically [109,110,112], the chain comprises three primary domains: (i) The Internal Domain, originating from the Mesomediterranean Microplate; (ii) The intermediate Flysch Units, characterized by accreted sequences from the Maghrebian Flysch Basin; and (iii) The extensive External Domain, representing the deformed northwestern margin of the African margin.

Unit-scale structural analysis (Figure 3,

Table 2. Statistic parameters of geological structures, i.e., faults and thrusts, as shown in Figure 3. Lengths Structures (Min, Max, Mean, and Sum) are expressed in meters.

	Gomarides and Septides Complex	Dorsale Calcaire Complex	Flyschs and External Domains	Zone C	Total Study Area
Area (km2)	314.49	707.84	955.92	509.79	1978.25
Poucent/Study area	15.90%	35.78%	48.32%	25.77%	100%
Structures Count	99	1015	240	17	1354
Faults	34	493	152	0	678
Thrusts	65	522	88	17	676
Length Struct Min	15	14	46	-	14
Length Struct Max	6722	8973	11,419	11,419	11,419
Length Struct Mean	786	695	1709	1839	876
Length Struct Sum	108,632	246,679	327,974	32,662	118,780

) reveals pronounced heterogeneity within the Internal Domain: The Dorsale Calcaire (blue) contains 493 faults and 522 thrusts, with mean trace lengths of 695 m (maximum: 8973 m). The Ghomarides and Sebtides complexes (purple) exhibit 34 faults and 65 thrusts, averaging 786 m in length (maximum: 6722 m).

The Flysch and External Domains (green) contain 152 faults and 88 thrusts, with maximum trace lengths reaching 1114 m (Table 2). Structural density in these domains remains comparatively underestimated relative to the Internal Domain. This discrepancy arises from Zone C’s reliance on lower-resolution 1:200,000 mapping (Figure 2 and Figure 3), highlighting this study’s contribution in addressing data gaps through integrated KBS/RSD/GIS methodologies in rugged terrain.

Fault and thrust orientation analysis (Figure 3: Rose diagrams) confirms the region’s structural complexity, exhibiting multidirectional trends with variable frequencies. Dominant fault orientations occur along NW-SE, N-S, and E-W vectors, while thrusts predominantly strike NW-SE. Critically, these thrusts frequently demarcate litho-tectonic boundaries between the Internal, Flysch, and External Domains (Figure 2 and Figure 3).

3. Data and Methods

3.1. Data

3.1.1. Data Used

Several types of RSD were used in this study, including the optical RSD Landsat-8 OLI and Sentinel-2A, the radar RSD Sentinel-1B, and the Aster GDEM. The Landsat-8 OLI, Sentinel-1B, and Sentinel-2A RSD were employed for geospatial modeling of SL using different methods. The Aster GDEM data were used to develop maps of elevation, the hydrographic network, and watershed boundaries, as well as a ridge map.

All types of RSD used covered the entire territory investigated. Figure A1, Figure A2, Figure A3 and Figure A4 in the Appendix illustrate the coverage of all the RS products used in this study.

Table 3. Remote Sensing Data used.

RS Data	Acquisition Date	Number	ID_Products
Landsat-8 OLI	18 August 2021	2	LC08_L1TP_201035_20210808_20210818_01_T1
			LC08_L1TP_201036_20210808_20210818_01_T1
Sentinel-1 B	16 August 2021	1	S1B_IW_GRDH_1SDV_20210816T181728_20210816T181753_028275_035FA5_0653
Sentinel-2 A	6 August 2021	4	S2A_OPER_MSI_L1C_TL_VGS2_ 20210806T130919_A031980_T30STE
			S2A_OPER_MSI_L1C_TL_VGS2_ 20210806T130919_A031980_T30SUD
			S2A_OPER_MSI_L1C_TL_VGS2_ 20210806T130919_A031980_T30SUE
			S2A_OPER_MSI_L1C_TL_VGS2_ 20210806T130919_A031980_T30STD
ASTER-GDEM	19 January 2021	1	ASTGTM_N35W006

details the RS data used, including their acquisition date, the number of products, and their ID.

The following paragraphs provide an overview of the RS datasets used.

(1) Landsat 8 (NASA/USGS): Launched on 11 February 2013, this platform carries two primary sensors: the Operational Land Imager (OLI) and the Thermal Infrared Sensor (TIRS). The OLI acquires data across 11 spectral bands: nine multispectral bands (spanning VNIR to SWIR regions) at 30 m spatial resolution, and one panchromatic band at 15 m resolution. The TIRS complements this with two thermal bands at 100 m spatial resolution, critical for mapping surface temperature variations and mineralogical properties associated with silicate rocks and hydrothermal alteration. The systematic global coverage, extensive archive, and calibrated optical data of Landsat ensure its continued prominence in regional geological studies, particularly for lithological and structural mapping applications [6]. Collectively, these attributes constitute the principal rationale for selecting Landsat-8 OLI data to evaluate its potential in SL mapping.

(2) The European Space Agency (ESA) developed the Sentinel mission family to address the operational requirements of the Copernicus programme. Sentinel-1, a polar-orbiting satellite constellation, provides all-weather, day-and-night C-band Synthetic Aperture Radar (SAR) imaging for land and ocean services. The Sentinel-1B satellite, launched on 25 April 2016, carries an advanced SAR sensor featuring dual-polarization capability (co-polarized VV/HH and cross-polarized VH/HV), a wide-swath interferometric mode (e.g., Interferometric Wide Swath), and a spatial resolution of 5 m (range) × 20 m (azimuth) [15]. C-band SAR (5.4 GHz) penetrates cloud cover and is sensitive to surface roughness, dielectric properties, and structural geometry. Consequently, it serves as a critical component of geoscientific datasets for mapping geological structures such as faults, fractures, and lithological boundaries [113,114]. These capabilities motivated the selection of Sentinel-1 SAR data to evaluate its potential for SL mapping applications.

(3) Sentinel-2, a polar-orbiting multispectral high-resolution imaging mission within the European Space Agency’s Copernicus programme, is specifically designed for advanced land monitoring. Sentinel-2A, launched on 23 June 2015, carries the MultiSpectral Instrument (MSI) acquiring data across 13 spectral bands with variable spatial resolutions: four Visible and Near-Infrared (VNIR) bands (B2, B3, B4, B8) at 10 m resolution, six bands (four red-edge: B5, B6, B7, B8a; two Shortwave Infrared: B11, B12) at 20 m resolution, and three atmospheric correction bands (coastal aerosol: B1; water vapor: B9; cirrus: B10) at 60 m resolution. The MSI’s unique spectral configuration—particularly its red-edge bands and SWIR bands—enables sophisticated lithological and alteration mapping. Recent research demonstrates increasing adoption of Sentinel-2A MSI data for geological applications, leveraging these diagnostic spectral properties [91]. Consequently, Sentinel-2 was selected as a complementary multispectral dataset to evaluate its potential for SL mapping.

(4) ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer) is a multispectral imaging sensor aboard NASA’s Terra platform, launched in December 1999. It acquires data in 14 spectral bands spanning the visible and near-infrared (VNIR: 3 bands at 15 m resolution), shortwave infrared (SWIR: 6 bands at 30 m resolution), and thermal infrared (TIR: 5 bands at 90 m resolution) regions of the electromagnetic spectrum, measuring both reflected solar radiation and emitted terrestrial radiation. A key capability is its along-track stereoscopic imaging in the VNIR bands, enabling the generation of the ASTER Global Digital Elevation Model (GDEM). Furthermore, ASTER data are globally available at no cost through multiple processing levels, distributed as 60 km × 60 km scenes [113,115]. Due to this spectral versatility, stereoscopic capacity, and open data access, ASTER has been extensively used in geological remote sensing applications, particularly for supplementing other satellite imagery with topographic information derived from the GDEM [69,76,115,116,117,118,119,120].

Prior to extracting potential FT-SL from remote sensing RSD, appropriate source imagery must be selected. The primary selection criterion was minimal cloud cover. Data processing focused on August 2021 acquisitions. This study used multispectral Landsat-8 OLI data comprising two scenes acquired on 18 August 2021, corresponding to the regional dry season (Table 3). Radar data consisted of a Sentinel-1B Interferometric Wide Swath (IW) Ground Range Detected (GRD) Level-1 product acquired on 16 August 2021 with dual-polarization (VV + VH) (Table 3). Multispectral Sentinel-2A data included four scenes acquired on 6 August 2021, also during the dry season (Table 3).

Complementing the cited satellite imagery, this study used the ASTER Global Digital Elevation Model (ASTER GDEM V3) [113]. This globally available DEM offers high spatial resolution (30 m) at no cost. The 30 m resolution ASTER GDEM data were obtained from the United States Geological Survey (USGS) EarthExplorer platform. Landsat-8 OLI, Sentinel-1B, and Sentinel-2A data were employed for SL extraction, while the ASTER GDEM facilitated the generation of hillshade imagery, elevation maps, and drainage networks. These datasets were supplemented by geological data from the Moroccan Ministry of Energy Transition and Sustainable Development—Department of Geology, including published geological maps at scales of 1:50,000 (Bab Taza, Talambote, Souk Larbaa Beni Hassen sheets) and the Rif regional geological map at 1:500,000 scale covering the study area.

3.1.2. Data Pre-Processing

Figure 4 illustrates an overview of the data preparation workflow for Landsat-8 OLI, Sentinel-1B and Sentinel-2A used to generate the TF-SL. Landsat-8 OLI data were preprocessed using ENVI (Version 5.3). Sentinel-1 SAR and Sentinel-2 MSI were preprocessed in SNAP (version 8.0.5).

This preprocessing phase was conducted meticulously, as it would generate the fundamental RS data that would be used for the subsequent stages of this study. For clarity, a background color-coding system has been implemented to distinguish the three RSD types throughout the paper’s figures and results. Data derived from Landsat-8 OLI are presented with an orange background, Sentinel-1B with a blue background, and Sentinel-2A with a green background.

The preliminary phase to preprocessing Landsat 8 OLI images consisted of the creation of a mosaic of three previously corrected scenes (calibration points found on the OLI bands). This includes specific treatments (contrast enhancement, main Component Analysis) and combinations of directional bands and filters for lineament identification. After processing and corrections made on a single spectral band (B7), polylines and segments are generated (Radiance stacking).

Sentinel 1 images were acquired at the L1 GRD (Ground Range Distance) level, according to IW mode with the polarizations (VH) and (VV). These are Digital Number (DN) data that require conversion to backscatter values reflected by the surface [4,121]. In addition to ‘Speckle’, other pre-processing included thermal noise removal and radiometric calibration. The final step was to correct for geometric distortions attributed to the topographic variations in the scene as well as the inclination of the radar. These distortions take the form of shading, foreshortening and spreading problems. As a result, all these correction tasks are particularly necessary to ensure that the image represents the terrain as closely as possible to reality, and is then ready for use.

The ESA Sentinel-2 satellites provide important datasets for terrestrial mapping. Radiometric differences between Sentinel-2A/B and Landsat 8 OLI is <1%, and the former is characterized by a higher absolute geodetic accuracy [120]. The four Sentinel-2A images covering the study area were taken on 6 August 2021 in order to avoid clouds and atmospheric haze that can obscure the targets observed and reduce the accuracy of the algorithm. The pre-processing of the Sentinel-2A images essentially consisted of Rayleight Correction.

3.2. Processing—Methods of SL Extraction

SL mapping was performed after a pre-processing phase aimed at enhancing satellite imagery for the optimal visualization of structural features. The methodology was an integrated approach that used: Three RSD sets: Landsat-8 OLI (multispectral), Sentinel-2A (multispectral), and Sentinel-1B (radar); and three extraction methods: manual, semi-automatic supervised, and automatic (utilizing spatial filtering). Figure 5 illustrates an overview of the data processing workflow to generate maps of fault- and thrust-based SL based on Landsat-8-OLI, Sentinel-1B and sentinel-2A RSD.

The main software applications used for FT-SL mapping were ENVI software (Version 5.3), SNAP (version 8.0.5), MATLAB (version 9.11), PCI Geomatica (version 2018) and QGIS software (version 3.22.16). ENVI and SNAP were used for band extraction of Landsat-8 OLI and Sentinels RSD, respectively. We applied filters to the satellite images using MATLAB’s programming and image processing platform. PCI Geomatica’s LINE module was employed for the automatic extraction of lineaments. Finally, QGIS software was used to extract statistical parameters and create density maps of the SL, with the PCI Geomatica results serving as the primary input.

3.2.1. Method 1: Manual Extraction of FT-SL

The manual extraction of structural lineaments (designated as LinMap/Method1 in Figure 5) constitutes a rigorous, knowledge-driven approach for creating potential SLM through expert visual interpretation of RSD. This methodology involves systematic analysis of linear features exhibiting distinct tonal contrasts (≥15% reflectance variation) and textural discontinuities (spatial frequency changes > 0.05 cycles/pixel) in satellite imagery, with particular attention to geomorphological expressions of faults, trusts, and fracture zones. The interpretation protocol required analysts to evaluate each candidate feature against seven geological criteria: (1) continuity across multiple pixels, (2) angular relationships with adjacent structures, (3) geomorphological expression, (4) tonal contrast thresholds, (5) cross-cutting relationships, (6) correlation with DEM-derived topography, and (7) regional structural context.

The digitization process employed GIS vectorization tools, utilizing a 0.5-pixel snapping tolerance and Bezier curve smoothing to maintain geometric fidelity while minimizing digitization artifacts. A critical innovation involved the integration of a 30 m resolution shaded-relief model derived from Aster-GDEM data (ASTGTM_N35W006), which was dynamically toggled beneath the primary imagery (9-band composite for Landsat-8 OLI, dual-polarization for Sentinel-1B and 13-band composite for Sentinel-2A) at varying opacity levels (25–75%). This multi-layered visualization enabled a three-phase validation protocol: (1) initial feature identification, (2) consistency verification across illumination conditions (achieved through 8-directional hillshade rotation), and (3) geological plausibility assessment. The resulting SLM exhibit a minimum resolvable length of 200 m (4–5 pixels for Sentinel-2 data) and maximum extents constrained by scene boundaries, with each feature tagged with confidence metrics based on observation consistency.

While resource-intensive (averaging 12–15 analyst-hours per 100 km²), the manual approach provides unparalleled discrimination capability, achieving an 82–89% reduction in false positives compared to automated methods in validation studies [65]. This particular detection method has been shown to be highly effective in identifying tectonically significant structures, as evidenced by its strong correlation with field-mapped faults in test areas. Additionally, it has been demonstrated to effectively filter out anthropogenic features such as roads and pipelines through geometric analysis and contextual assessment.

Empirical studies [67,122] confirm manual extraction’s critical role as a benchmark for automated algorithm validation, with our implementation showing a high ratio precision for major fault zones when compared to structural data derived from available geological maps. This approach establishes manual mapping as an essential component in comprehensive structural analysis workflows, particularly for tectonic studies requiring high-confidence feature identification, despite its inherent subjectivity and throughput limitations.

3.2.2. Method 2: Semi-Automatic Supervised Extraction of FT-SL

The semi-automated supervised method for FT-SL extraction represents an advanced methodological spatial analysis combining the edge detection algorithm (through automatic extraction of SL) with expert-guided interpretation (via manual refinement) [75,77]. This approach, designated as “LinMap/Method2” in Figure 5, involves three critical phases: (i) Pre-processing of remote sensing data; (ii) Automated lineament detection using specialized algorithms; and (iii) Manual refinement by trained analysts [14,121,123]. The method’s effectiveness fundamentally depends on the interpreter’s field experience, which enables accurate lineament classification and the topological connection of discontinuous segments into coherent structural features.

This study developed a novel semi-automated workflow specifically designed for fault and thrust system identification. The methodology begins with automated lineament extraction using PCI Geomatica’s LINE module, implementing six precisely calibrated parameters determined through extensive computational simulations: RADI = 5 (7 × 7 Gaussian kernel, σ = 1.2), GTHR = 30 (15% gradient magnitude threshold), LTHR = 20 (minimum 20 connected pixels), FTHR = 3 (3-pixel RMS fitting error), ATHR = 10° (directional consistency threshold), and DTHR = 30 (30-pixel maximum linking distance). Following initial detection (detailed in Section 3.2.3), the validation phase executes systematic removal of non-geological features through expert-guided interpretation, including: (1) anthropogenic features (roads, railways) identified using 10 m OpenStreetMap buffers; (2) fluvial networks derived from DEM hydrological analysis; and (3) miscellaneous non-structural linear artifacts [21,25,30,122].

The quality of the final lineament map depends critically on two key factors: (1) the radiometric and spatial characteristics of the source imagery, and (2) the analyst expertise in recognizing the following: Tonal/textural contrasts (5–15% reflectance variation); Geomorphological indicators (drainage patterns/density); Surface cover relationships (vegetation alignment) and structural geomorphology [14,28,76].

3.2.3. Method 3: Spatial Filtering Coupled with Automatic Extraction of FT-SL

As illustrated in Figure 5, Method 3 is the most complex approach for FT-SL. Its effectiveness stems from its comprehensive and sequential workflow, which encompasses several crucial steps in the processing of satellite images prior to the final automatic extraction. This method is formally referred to as “Spatial Filtering Coupled with Automatic Extraction of FT-SL” and involves three primary stages: (1) Extraction of specific bands from the RSD; (2) Spatial filtering of the single-band RSD; and (3) Automatic extraction of the FT-SL.

• Method 3—part 1: Band Extraction

As demonstrated in Figure 5, the first step of Method 3 entails the extraction of the optimal spectral band from three distinct RSD sources. The selection of these bands was informed by prior research, focusing on those most frequently utilized for geological lineament extraction, as supported by well-established Principal Component Analysis PCA outcomes. Specifically, the following bands were chosen for each dataset: Landsat 8 OLI: Band 7; Sentinel 1-B (SAR data): VV polarization band; and Sentinel 2-A: Band 12. This extraction process was conducted using ENVI, GIS software, and the SNAP platform.

Landsat 8 OLI band extraction constitutes a fundamental preprocessing operation demanding meticulous file handling. As specified in Section 3.1.1, each spectral band exists as discrete GeoTIFF files (e.g., _B7.TIF), necessitating targeted retrieval through either GIS interface navigation or ENVI’s Band Extract function. This process, while conceptually simple, requires careful quality control to ensure proper layer integration. The resulting single-channel raster forms the essential base layer for subsequent structural interpretation.

Sentinel-1 SAR data demands significantly more complex preprocessing to generate analysis-ready VV bands. Our workflow implemented five sequential corrections: orbit file application for positional accuracy, radiometric calibration to σ0 values, advanced speckle reduction filtering, DEM-based terrain correction (WGS84), and optional dB conversion. Each stage required specialized parameterization in SNAP, with particular attention to preserving textural features during noise reduction. The resultant calibrated raster (Figure 4) meets stringent requirements for geological feature detection.

Sentinel-2 MSI Band 12 extraction benefits from the mission’s pre-separated band architecture. Despite simpler file retrieval from .SAFE directories, the process still demands technical precision in handling the SWIR band’s unique characteristics (2180 nm). The loaded raster requires verification of radiometric properties and spatial alignment before integration with other datasets. This band’s particular sensitivity to lithological contrasts and hydrous minerals makes it indispensable for comprehensive geological analysis.

While our band selection methodology aligns with current best practices, we identify significant opportunities for optimization through AI-driven band evaluation. Future work should implement machine learning algorithms to systematically assess each band’s contribution across different geological settings and sensor combinations. This would establish more robust, data-driven band selection protocols, potentially revealing underutilized spectral ranges for specific structural detection applications.

• Method 3—part 2: Spatial Filtering

Spatial frequency, a fundamental characteristic of remotely sensed data defined as the rate of brightness value variation per unit distance [122], was systematically manipulated through advanced image processing techniques to optimize lineament extraction. Our methodology employed spatial convolution filtering [124], with particular emphasis on convolution-based approaches that have demonstrated superior performance in geological applications [9,14,125]. These techniques not only reduce noise and enhance image quality but also specifically target the identification of structural discontinuities, which are critical for geological interpretation. The extensive literature review and preliminary testing required to establish this filtering approach represents significant methodological development work, particularly in adapting these techniques to diverse geological contexts and image characteristics.

As shown in Figure 5 (Method 3), we implemented a comprehensive filtering protocol involving three distinct edge detection techniques: Sobel directional filters for gradient-based edge enhancement; a Laplacian filter for zero-crossing detection; and a Kuan adaptive filter for speckle reduction. This sophisticated processing workflow was rigorously applied to three single band datasets (Landsat-8 OLI-band 7, Sentinel-1B-VV and Sentinel-2A-band 12) using the image processing toolbox in MATLAB, which required extensive parameter optimisation and significant computational processing time. The three spatial filtering approaches, executed through custom MATLAB programming, ensured robust lineament detection across different sensor types and resolutions while preserving the geological relevance of the extracted features. This represents a substantial technical effort in terms of both the development of the algorithm and its practical implementation, in order to achieve reliable results for structural interpretation.

○

Sobel Filter

Sobel filter is the most common types of Edge Detection Filtering based on the non-linear combination of pixels [14]. Sobel filter is a commonly used edge detector technique and allows the calculation of the X and Y derivatives with a level of smoothing imparted via the kernel [125,126,127,128,129].

The Sobel filter constitutes a discrete differentiation operator that approximates the first-order image gradient through optimized 3 × 3 convolution kernels. The horizontal (G_H) and vertical (G_V) kernels (Equations (1) and (2)) mathematically represent the partial derivatives ∂I/∂x and ∂I/∂y of the image intensity function I(x,y).

The kernel coefficients implement Pascal triangle weighting ([1 2 1] central row/column) to simultaneously compute the gradient and apply Gaussian-like smoothing perpendicular to the differentiation direction. The normalization factor λ_3 = 1/8 maintains energy conservation while preventing arithmetic overflow in 8-bit image representations.

The horizontal Sobel kernel G_H (Equation (1)) emphasizes vertical edges through its antisymmetric structure:

$$\left[\begin{array}{ccc}1& 0& -1\\ 2& 0& -2\\ 1& 0& -1\end{array}\right]$$

(1)

while the vertical kernel G_V (Equation (2)) detects horizontal edges via its transposed configuration:

$$\left[\begin{array}{ccc}1& 2& 1\\ 0& 0& 0\\ -1& -2& -1\end{array}\right]$$

(2)

Convolution with these kernels produces gradient component maps where pixel values represent the rate of intensity change in each direction. The gradient magnitude (Equation (5)) combines these orthogonal components through the Euclidean norm, providing edge strength independent of orientation:

$$\left|\mathrm{G}(\mathrm{x},\mathrm{y})\right|=\sqrt{{{\mathrm{G}}^2}_{\mathrm{H}}(\mathrm{x},\mathrm{y})+{{\mathrm{G}}^2}_{\mathrm{V}}(\mathrm{x},\mathrm{y})}$$

(3)

While the standard Sobel implementation detects edges in cardinal directions, geological lineament analysis requires enhanced diagonal sensitivity. We therefore extended the basic formulation to four principal orientations (0°, 45°, 90° and 135°) named (N–S, NE–SW, E–W, and NW–SE) through coordinate transformation:

$${\mathrm{G}}_{\mathsf{\theta }}=\cos \mathsf{\theta }\times {\mathrm{G}}_{\mathrm{H}}+\sin \mathsf{\theta }\times {\mathrm{G}}_{\mathrm{H}}$$

(4)

where θ ∈ {π/4, π/2, 3π/4}. Each directional kernel was implemented using separable convolution for computational efficiency, reducing the 2D operation to sequential 1D passes (horizontal binomial smoothing followed by central difference differentiation).

Sobel kernels were applied to single-band imagery in this study, specifically targeting: (1) Landsat-8 OLI Band 7 (2.11–2.29 μm SWIR), (2) Sentinel-1B VV polarization (C-band radar), and (3) Sentinel-2A Band 12 (2.100–2.280 μm SWIR). The MATLAB code for applying four-directional Sobel filters (N–S, NE–SW, E–W, and NW–SE) to Landsat-8 OLI Band 7 datasets is provided in Appendix D.

The spectral properties of OLI Band 7 make it ideal for Sobel filtering, offering superior atmospheric penetration (haze reduction: 0.85 ± 0.05), minimal moisture interference (>80% vegetation signal attenuation), and enhanced lithological contrast (12–18% greater edge detection in metamorphic terrains) [14,122]. While inherently more sensitive to cardinal-aligned lineaments (0°/90°), causing 30–40% signal reduction in diagonal features (45°/135°) [128,130,131], our implementation overcame this limitation through multi-directional kernel application and non-maximum suppression, ensuring comprehensive lineament detection across all orientations.

○

Laplacian Filter

The Laplacian operator represents a fundamental second-order differential approach to edge detection, defined mathematically as the divergence of the gradient of an image function (Equation (5)). This isotropic operator computes

$${\nabla }^2\mathrm{f}(\mathrm{x},\mathrm{y})={\displaystyle \frac{{\mathsf{\delta }}^2\mathrm{f}}{\mathsf{\delta }{\mathrm{x}}^2}}+{\displaystyle \frac{{\mathsf{\delta }}^2\mathrm{f}}{\mathsf{\delta }{\mathrm{y}}^2}}$$

(5)

providing a rotationally invariant measure of the spatial intensity change. The discrete implementation approximates the continuous Laplacian through central differences, yielding a scalar value that highlights regions of rapid intensity transition while suppressing areas of gradual change. This mathematical property makes it particularly effective for detecting sharp discontinuities characteristic of geological lineaments in remotely sensed data [130,132,133,134,135,136].

The standard 3 × 3 discrete Laplacian kernel (Equation (6)) implements a four-connected neighborhood approximation with a λ₃ = 1 normalization factor [131]. The kernel structure

$$\left[\begin{array}{ccc}0& -1& 0\\ -1& 4& -1\\ 0& -1& 0\end{array}\right]$$

(6)

emphasizes pixel-wise differences from orthogonal neighbors while maintaining zero-sum coefficients to ensure zero response in uniform regions. For enhanced geological feature detection, we implemented both this L₄ kernel and its eight-connected variant L₈ (Equation (7)), with comparative analysis showing up to 20% improved fault detection using L₈ in mountainous terrain.

$$\left[\begin{array}{ccc}-1& -1& -1\\ -1& 8& -1\\ -1& -1& -1\end{array}\right]$$

(7)

Laplacian kernels were applied to single-band remote sensing data in this study, specifically targeting three key datasets: (1) Landsat-8 OLI Band 7 (2.11–2.29 μm SWIR), (2) Sentinel-1B VV polarization (C-band radar), and (3) Sentinel-2A Band 12 (2.100–2.280 μm SWIR). The MATLAB implementation for the enhanced 8-connected Laplacian (L₈) filter, optimized for geological lineament detection in Sentinel-2 Band 12 imagery, is detailed in Appendix D.

The Laplacian’s edge detection algorithm operates by identifying zero-crossings in the second derivative space—inflection points where the filtered output changes sign, corresponding to intensity transitions in the original image. Our implementation incorporates sub-pixel interpolation of these zero-crossings using quadratic approximation, achieving a mean localization accuracy of 0.8 pixels for major structural boundaries in validation tests.

○

Kuan Filter

The Kuan filter represents an advanced adaptive despeckling technique rooted in statistical estimation theory. Its mathematical derivation begins with the multiplicative noise model (Equation (8)):

$$\mathrm{I}\left(\mathrm{t}\right)=\mathrm{R}\left(\mathrm{t}\right)\times \mathrm{u}\left(\mathrm{t}\right)$$

(8)

where t = (x, y) denotes spatial coordinates, I(t) the observed intensity, R(t) the true reflectivity, and u(t) the multiplicative speckle noise with unit mean ($\mathbb{E}\left[\mathrm{u}\right]=1$) and variance σ_n². Through logarithmic transformation or Taylor series approximation, this model is converted to an additive noise form, enabling application of minimum mean square error (MMSE) estimation principles. The transformation maintains the signal-dependent nature of SAR noise while permitting linear estimation techniques.

The Kuan filter’s core innovation lies in its MMSE solution (Equation (9)):

$$\widehat{\mathrm{R}}\left(\mathrm{t}\right)=\mathrm{I}\left(\mathrm{t}\right)\times \mathrm{W}\left(\mathrm{t}\right)+\overline{\mathrm{I}}\left(\mathrm{t}\right)\times (1-\mathrm{W}\left(\mathrm{t}\right))$$

(9)

where the adaptive weight W(t) (Equation (10)) balances between retaining original pixel values and local averaging:

$$\mathrm{W}\left(\mathrm{t}\right)={\displaystyle \frac{1-{\mathrm{C}}_{\mathrm{u}}^2/{\mathrm{C}}_{\mathrm{I}}^2\left(\mathrm{t}\right)}{1+{\mathrm{C}}_{\mathrm{u}}^2}}$$

(10)

The weight depends on two key coefficients:

$${\mathrm{C}}_{\mathrm{u}}={\mathsf{\sigma }}_{\mathrm{n}}$$

(11)

Speckle coefficient of variation (typically 1/√L for L-look SAR):

$${\mathrm{C}}_{\mathrm{I}}\left(\mathrm{t}\right)={\mathsf{\sigma }}_{\mathrm{I}}\left(\mathrm{t}\right)/\overline{\mathrm{I}}\left(\mathrm{t}\right)$$

(12)

Local image variation coefficient

• This formulation dynamically adjusts smoothing intensity based on local homogeneity, with W(*t*) → 1 (identity) near edges (C_t ≫ C_u) and W(*t*) → 0 (full smoothing) in homogeneous regions (C_t ≈ C_u).

Our implementation enhanced the standard Kuan filter through the following: (1) Window Optimization: 7 × 7 sliding window with Gaussian-weighted sampling; (2) Boundary Handling: Mirror padding to minimize edge artifacts.

The enhanced Kuan filter was applied to single-band remote sensing data in this study, focusing on three critical datasets: (1) Landsat-8 OLI Band 7 (2.11–2.29 μm SWIR), (2) Sentinel-1B VV polarization (C-band radar), and (3) Sentinel-2A Band 12 (2.100–2.280 μm SWIR). The MATLAB implementation of this enhanced Kuan filter algorithm, specifically optimized for Sentinel-1B VV imagery, is comprehensively documented in Appendix D.

The enhanced Kuan filter demonstrates superior capabilities in the following: (i) Fault Mapping: Improved scarp visibility in alluvial cover (87% field-verified accuracy); (ii) Structural Analysis: Enhanced fold hinge detection through noise-free curvature attributes. These capabilities stem from the filter’s unique ability to resolve the inherent trade-off between speckle reduction and feature preservation—a critical requirement for SAR-based geological interpretation [116,125,137,138].

• Method 3—Part 3: Automatic extraction of FT-SL

Numerous algorithmic approaches have been developed for automated lineament extraction, with varying degrees of effectiveness in different geological contexts [12,14,24,52,72,124,135,139]. While these software-based solutions generate vector-formatted lineament maps, significant limitations persist, particularly in low-contrast terrains and mountainous areas with complex shadow patterns, often resulting in incomplete lineament detection [14,124]. The STA (Segment Tracing Algorithm) method proposed by Koike et al. [135] addressed these challenges through threshold-based approaches, forming the foundation for the widely adopted LINE module in PCI Geomatica. This sophisticated tool, which integrates STA principles with Canny edge detection algorithms, has emerged as the industry standard for automated geological lineament extraction [14,76,101,136,140,141,142], though its effectiveness remains dependent on careful parameterization and input data quality (Figure 6).

The LINE module’s computational architecture performs three sequential processing stages (Figure 6), governed by six critical parameters that require expert calibration: Filter Radius (RADI: 3–8), Edge Gradient Threshold (GTHR: 10–70), Curve Length Threshold (LTHR: typically 10), Line Fitting Error Threshold (FTHR: 2–5), Angular Difference Threshold (ATHR: 3–20°), and Linking Distance Threshold (DTHR: 10–45) [34]. Each parameter influences distinct aspects of the lineament detection process, from initial edge identification to segment linking and vectorization [14,34]. The interdependence of these parameters necessitates extensive testing and validation to achieve optimal results, particularly when processing diverse remote sensing datasets with varying spatial and spectral characteristics.

Our implementation of PCI Geomatica’s LINE module involved rigorous parameter optimization (RADI = 5, GTHR = 30, LTHR = 20, FTHR = 3, ATHR = 10°, DTHR = 30) applied to carefully preprocessed single-band remote sensing datasets (Landsat 8 OLI Band 7, Sentinel-1B VV polarization, and Sentinel-2A Band 12), representing a substantial methodological effort that encompassed comprehensive data calibration, noise reduction, and terrain correction to ensure the production of geologically reliable lineament maps through this sophisticated automated extraction process.

3.2.4. Processing Workflow Summary

FT-SL mapping was executed through a two-phase methodology. As shown in Figure 4, the preprocessing phase employed image enhancement techniques to optimize structural feature detection across three RSD: Landsat-8 OLI, Sentinel-2A, and Sentinel-1B. The analytical phase employed three extraction approaches to each dataset: (i) manual visual interpretation, (ii) semi-automated supervised extraction, and (iii) automated extraction with spatial filtering. Figure 5 illustrates the complete workflow from raw data to final FT-SL products, while

Table 4. Structural Lineament Maps of Faults and Thrusts Derived from Multi-Method Extraction Approaches Using Landsat-8 OLI, Sentinel-1B, and Sentinel-2A Data (Method 1: Manual extraction, Method 2: Semi-Automatic extraction supervised; Method 3-S: Sobel filter and Automatic extraction, Method 3-L: Laplacian filter and Automatic extraction, Method 3-K: Kuan filter and Automatic extraction of SL).

		Remote Sensing Data
		Landsat-8 OLI	Sentinel-1B	Sentinel-2A
SL mapping	Method 1	LinMap/Method1_OLI	LinMap/Method1_S1B	LinMap/Method1_S2A
	Method 2	LinMap/Method2_OLI	LinMap/Method2_S1B	LinMap/Method2_S2A
	Method 3-S	LinMap/Method3_OLI_B7_SF	LinMap/Method3_S1B_VV_SF	LinMap/Method3_S2A_B12_SF
	Method 3-L	LinMap/Method3_OLI_B7_LF	LinMap/Method3_S1B_VV_LF	LinMap/Method3_S2A_B12_LF
	Method 3-K	LinMap/Method3_OLI_B7_KF	LinMap/Method3_S1B_VV_KF	LinMap/Method3_S2A_B12_KF

comprehensively documents all generated SLM, including their specific nomenclature and distinguishing characteristics derived from the various method-dataset sources.

3.3. Knowledge Based System

In recent years, KBS have proven highly effective in tackling complex problems across various Earth science disciplines. AI techniques—including ES, ANN, Adaptive Neuro-Fuzzy Inference Systems (ANFIS), and genetic programming (GP)—have been successfully applied in geological research, particularly in the geospatial characterization of structural lineaments [97,100,102]. A major strength of these AI-driven approaches is their ability to compensate for incomplete or inconsistent datasets, a crucial advantage in data-scarce domains.

Within this context, we developed a rule-based system integrating MCDA techniques using an object-oriented language. This KBS serves two primary functions: guiding users in formalizing MCDA parameters and providing decision-support capabilities. Its key operational components include (i) structuring objectives and evaluation criteria, (ii) assigning empirical weights to criteria, (iii) computing scores for alternatives, (iv) interpreting matrix results, and (v) guiding optimal RSD selection SL extraction methods, with specialized efficacy in mountainous regions.

The system’s architecture consists of three core elements: A Knowledge Base, implemented as an object-fact language with IF-THEN rule encoding; an Inference Engine employing forward-chaining deduction for data-driven reasoning; and a User Interface (UI) designed for transparency, delivering both outputs and explanatory guide to enhance the automated SL Mapping results. Figure 7 shows the keys components of the developed KBS.

Knowledge acquisition for the KBS posed significant epistemological challenges, addressed through a rigorous three-phase methodology: extraction, modeling, and formalization. Recognizing knowledge acquisition as the most demanding yet critical phase in expert system development, we adopted a dual-source strategy to prevent knowledge base limitations that could degrade performance. Formal knowledge was compiled from peer-reviewed literature, manuals, and publications on SL geospatial modeling and satellite-based feature extraction, while heuristic knowledge was gathered through structured interviews with multidisciplinary experts in geometrics, structural geology, decision analysis, GIS, and KBS. This integrative knowledge framework ensures both theoretical robustness and practical applicability, forming the foundation for the KBS’s two core components: (1) an Object-Oriented Database (OODB) structured as a hierarchy of classes and objects that encapsulate attributes and methods for managing foundational data derived from the four-phase integrated analysis of Sustainable Land Management measures (SLMs), enabling efficient data organization, retrieval, and scalability for geospatial decision-making; and (2) a Rule-Based Knowledge Base (RBKB) comprising “If-Then” production rules that encode domain-specific heuristics to emulate human expert reasoning, which is applied for both automated extraction of slope instabilities (SLs) from remote sensing data (RSDs) and decision support in selecting optimal data types and analytical methods.

3.4. Multi-Criteria Decision Analysis

Multicriteria Decision Analysis (MCDA) constitutes a systematic framework for evaluating complex decisions characterized by multiple, often competing objectives. As a specialized branch of decision theory, MCDA provides robust methodologies to compare alternatives across diverse criteria, facilitating transparent trade-off assessments while addressing inherent challenges such as information gaps, cognitive biases, and value conflicts [143,144]. The approach decomposes decisions into quantifiable components through key activities: (1) contextual understanding, (2) value-driven criterion specification, (3) alternative generation, and (4) consequence modeling. Proper execution of these steps enhances decision-making rigor, establishes defensible justifications, and increases the likelihood of optimal outcomes across application domains.

MCDA operates on core principles of structured preference modeling, involving (i) explicit decomposition of objectives and criteria, (ii) alternative development, and (iii) consequence evaluation. A critical distinction exists between complete ranking methods and partial ranking/outranking methods. Complete ranking establishes total alternative ordering, while outranking identifies non-dominated subsets, accommodating scenarios where strict ordinality is impractical. Stakeholder engagement ensures accurate value representation during criterion weighting, which quantifies the relative importance of each objective’s contribution to the decision space. In this case study, the key question is as follows: what type of RS images can be used to extract SL? What is the most appropriate method for extracting SL?

The MCDA workflow comprises sequential phases: problem framing, criterion selection/weighting, alternative scoring, and result aggregation. Sensitivity analysis is integral for assessing robustness against uncertainties in weights or scores. In geospatial applications, key considerations include remote sensing data suitability and optimal lineament extraction methods. The integration of Knowledge-Based Systems (KBS) enhances methodological precision through rule-based objective derivation and criterion specification. Weight interpretation as scaling constants enables comprehensive impact assessment across all objectives (

Table 5. Objectives, criteria, and corresponding code-value specifications implemented in this study.

Objectives	Criteria	Units of Measurement	Min Value	Max Value	Very Satisfactory Criteria	Slightly Satisfactory Criteria	Not Satisfactory Criteria
					Code-Value	Code-Value	Code-Value
//Good!// SL Statistics	SL_Count	number	74	8509	2	1	0
	SL_Sum_Length	Meters	132,591	5,030,777	2	1	0
	SL_Length_Min	Meters	116	1814	2	1	0
	SL_Length_Max	Meters	2119	10,060	2	1	0
	SL_Length_Mean	Meters	518	3009	2	1	0
	SL_Length_CI	Meters	514	2896	2	1	0
//Good!// SL Orientation	Vector_Mean	Degrees	0.2	179.4	6	3	0
	Vector_CI	Degrees	0.5	150.2	6	3	0
//Good!// SL Density	Density_Max	km/km2	0.79	10.15	6	3	0
	Density_Mean	km/km2	0.07	2.38	6	3	0
//Good!// Correlation Coefficient	DCC_Zone-A	%	4.1	69.4	4	2	0
	DCC_Zone-B	%	4.1	72.8	4	2	0
	DCC_Zone-D	%	1.0	67.9	4	2	0

), with the MCDA function formulated as:

$$\mathrm{F}=\sum _{\mathrm{i}=1}^{\mathrm{n}}{(\mathrm{C}}_{\mathrm{i}}\times {\mathrm{W}}_{\mathrm{i}})$$

(13)

where $\mathrm{F}$ represents the composite score, ${\mathrm{C}}_{\mathrm{i}}$ the criteria, and ${\mathrm{W}}_{\mathrm{i}}$ their respective weights.

Another significant component of MCDA methodologies is the estimation of criteria weights (Equation (13)), as weighting accentuates divergent measures of value when comparing alternatives [141]. In this section, weights are interpreted as scaling constants that weight the contributions of the range of objective studied to see the impact of all objective’s criteria. For the initial iteration of this work, the KBS’s rule base interprets weights as scaling constants that modulate the relative contributions of each objective’s criteria, enabling comprehensive assessment of all objectives’ impacts. Table 5 summarizes the defined objectives, criteria, and associated coded values employed in this study.

MCDA improves decision quality by explicating trade-offs and fostering consensus, though its efficacy depends on precise criterion definition and method selection. The incorporation of KBS addresses inherent challenges by standardizing objective/criterion definitions and weight heuristics through rule-based logic. This synergy of MCDA and KBS creates a powerful paradigm for integrating quantitative and qualitative insights, particularly in geospatial modeling of structural lineaments, where it enhances both analytical rigor and practical applicability.

4. Results

The first results of this work consist of 15 FT-SL maps for the Oued-Laou watershed region, as listed in Table 4. To produce these maps, several techniques were used to extract FT-SL from three types of remote sensing data (RSD): Landsat-8 OLI, Sentinel-1 B, and Sentinel-2 A (Figure 5). The first technique was manual extraction, based on the visual interpretation of satellite images. The second was a semi-automatic approach, which involved two steps: (1) Automatic lineament extraction; and (2) Manual refinement of the lineament map to remove features corresponding to hydrographic and road networks. The third technique was an automatic extraction of the SLM. This method involved extracting a specific band from each RSD type—Band 7 for Landsat 8-OLI, VV for Sentinel-1B, and Band 12 for Sentinel-2A—followed by the application of three filters: Sobel, Laplacian, and Kuan. The resulting rasters were then used to automatically extract the SL.

Figure 8 shows the 15 Fault and Thrust-based Structural Lineament Maps (FT-SLM) generated as input data for subsequent analyses. Consistent with preceding figures, the SLM in Figure 8 adhere to a standardized color scheme for clarity.

The five SLM derived from Landsat-8 OLI RSD are highlighted with an orange background (maps A–E), while those from Sentinel-1B RSD are presented in blue (maps F–J), and those from Sentinel-2A RSD in green (maps K–O).

Figure 8 is organized systematically by column and row to facilitate interpretation. Column-wise classification is as follows: Column 1 illustrates SLM generated via Method 1, representing the manual FT-SL extraction approach (maps A, F, K); Column 2 presents SLM produced by Method 2, employing a semi-automatic supervised FT-SL extraction technique (maps B, G, L); Column 3 depicts SLM derived from Method 3-SF, an automatic FT-SL extraction method applied to Sobel-filter-enhanced input images (maps C, H, M); Column 4 showcases SLM obtained through Method 3-LF, utilizing automatic extraction on Laplacian-filter-enhanced images (maps D, I, N); Column 5 features SLs generated by Method 3-KF, involving automatic extraction with Kuan-filter-enhanced inputs (maps E, J, O).

The coloration of SLM alignments in Figure 8 is algorithmically assigned and holds no intrinsic significance. Initial qualitative assessment reveals distinct variations in lineament morphology across different base RSD, as well as between extraction methods for each RSD. A rigorous comparative analysis of these SLM—encompassing statistical metrics, orientation distributions, lineament densities, and correlation with geological structures in the three study areas—is provided in Section 4.1, Section 4.2, Section 4.3 and Section 4.4. For the software, the PCI Geomatica software (version 2018), specifically its LINE module, was used for the semi-automatic and automatic extraction of SL. Subsequently, QGIS software (version 3.22.16; Open Source Geospatial Foundation) was employed to extract statistical parameters and generate SL density maps. Finally, RockWorks software (version 2021, Golden Software, LLC) was used to create the directional rose diagrams.

Figure 9 presents a flowchart of all the processes used for the knowledge-based multi-criteria comparison of SLM obtained using three different extraction methods and from both optical and radar RSD. The knowledge-based multi-criteria analysis of the FT-SL mapped in the Oued-Laou watershed region consists of comparing four sets of results, as detailed in the methodology flowchart in Figure 9: (1) Statistics of SLM; (2) SLM orientation.; (3) SLM density distribution; and (4) Correlation of SLM density with geological maps.

4.1. Statistics of SLM

The initial phase of our multi-criteria analysis of SLM employs a rigorous statistical approach to examine lineament features derived from three distinct extraction methods (manual, semi-automatic, and automatic) and applied to three types of RSD (Landsat 8-OLI, Sentinel-1B, and Sentinel-2A). This analytical framework is schematically represented in the initial section of Figure 9. Our comprehensive statistical evaluation of SLM incorporates the following quantitative metrics: (1) Lineament enumeration (total count); (2) Dimensional characteristics (minimum and maximum lengths, mean length and sum length); and (3) Dispersion parameters (length standard deviation and confidence intervals for length distributions). These quantitative measures are systematically compiled in

Table 6. Statistical parameters of the Structural Lineament Maps (SLM) obtained using manual, semi-automatic, and automatic extraction methods based on three types of remote sensing data (RSD): Landsat-8 OLI, Sentinel-1B, and Sentinel-2A. (SumL: Total Length; LMin: Minimum Length; LMax: Maximum Length; LM: Mean Length; LRSD: Length Standard Deviation; CI1: Confidence Interval 1; CI2: Confidence Interval 2).

FT-SL Maps (/Method/RSD/(Filter))	Count	SumL (m)	LMin (m)	LMax (m)	LM (m)	LRSD (%)	CI1 (%)	CI2 (%)
Method_1/OLI_data	78	132,591	232	3974	1722	42%	9%	13%
OLI_data/AutoExtract	4039	5,923,825	899	10,060	1467	43%
Method_2/OLI_data	2037	2,952,314	899	5145	1449	41%	2%	35%
Method_3/OLI_B7_data/Sobel_filter	508	748,677	900	10,010	1483	48%	3%	18%
Method_3/OLI_B7_data/Laplacian_filter	151	228,568	901	8720	1513	58%	1%	14%
Method_3/OLI_B7_data/Kuan_filter	522	763,294	900	10,012	1462	46%	5%	9%
Method_1/S1B_data	428	647,309	116	7722	1512	65%	6%	9%
S1B_data/AutoExtract	13,938	777,408	238	5450	558	59%
Method_2/S1B_data	7330	3,910,088	238	4961	536	57%	1%	2%
Method_3/S1B_VV_data/Sobel_filter	8126	4,826,801	239	3775	594	61%	1%	2%
Method_3/S1B_VV_data/Laplacian_filter	8509	5,030,777	239	5049	591	61%	1%	2%
Method_3/S1B_VV_data/Kuan_filter	6085	3,198,089	238	2119	518	53%	3%	11%
Method_1/S2A_data	74	142,913	378	5240	1931	42%	12%	23%
S2A_data/AutoExtract	1437	4,125,064	1800	13,236	2871	45%
Method_2/S2A_data	752	2,135,575	1800	8578	2665	32%	3%	5%
Method_3/S2A_B12_data/Sobel_filter	150	442,284	1802	9809	2949	43%	11%	13%
Method_3/S2A_B12_data/Laplacian_filter	152	445,882	1809	9451	2933	42%	11%	14%
Method_3/S2A_B12_data/Kuan_filter	137	412,265	1814	9451	3009	44%	9%	15%
Geostructures/Geological maps	1354	1,186,899	15	11,419	877

for comparative assessment.

Figure 10 presents the distribution patterns of these statistical parameters through comparative histograms, facilitating the visual interpretation of methodological variations across the extraction techniques (manual, semi-automatic and automatic) and RSD (optical: Landsat 8-OLI and Sentinel-2A; radar: Sentinel-1B) used. The outputs of the statistical analysis serve as input data for our object-oriented knowledge-based system, specifically for the “SLM_Statistics” class package.

In terms of statistics, a comparison of the results obtained by SLM shows that the mean number of SL, extracted by the manual, semi-automatic and automatic methods, corresponds to 659, 6096 and 253 for the SL obtained from Landsat 8-OLI, sentinel-1B and sentinel-2A images, respectively (Ct Colon- Table 6). Figure 10 shows the distribution histograms of the SL obtained as a function of the number, min, max, total and mean lengths (in meters), in addition to some statistics relating to the standard deviation and confidence intervals of these lengths for each of the data used.

With regard to the lengths of fault- and thrust-based SL as mapped from Landsat 8-OLI, the values range from approximately 232 m to a maximum of 10,012 m. For the Sentinel-1B image, the values vary between 116 m and 5049 m, while for the Sentinel-2A image, they range from 378 m to 9809 m. This difference can be explained by the different natures of the optical and radar images. The length of the longest fault- and thrust-based SL has been determined to be 10,012 m (Landsat 8-OLI_Band 7_data/Method automatique_Kuan filter) and 10,010 m (Landsat 8-OLI_Band 7_data/Method automatique_Sobel filter). The values of 9809 m (Sentinel-2A_Band 12_data/Method automatique_Sobel filter) and 9451 m (Sentinel-2A_Band 12_data/Method automatique_Laplacian filter) are equivalent to the values of 9809 m (Sentinel-2A_Band 12_data/Method automatique_Kuan filter). The shortest fault- and thrust-based SL are 232 m, 116 m and 378 m, obtained by manual extraction from Landsat 8-OLI, Sentinel-1B and Sentinel-2A RSD, respectively (Table 6). This demonstrates the importance of human expertise in highlighting small geological structures. The semi-automatic and automatic SL extraction methods, incorporating the Sobel, Laplacian and Kuan filters, were also capable of detecting the smallest faults and chevrons with minimal lengths of approximately 238 m when applied to the Sentinel-1B images (Table 6 and Figure 10A–D).

Comparison of these figures indicates that the fault- and thrust-based SL identified by the radar data are characterised by shorter lengths than those extracted from the optical data (Landsat-8 OLI and Sentinel-2A). This again supports the effectiveness of radar data, which is independent of soil type, unlike optical images. With regard to the parameter: Relative Standard Deviation (RSD) of the lengths of the fault- and thrust-based SL as shown in Figure 10, it can be seen that the RSD values obtained from the optical images are relatively low. The RSD values for the Fault- and thrust-based SL obtained from the Sentinel-1B data using the semi-automatic and automatic methods are found to be significantly elevated, with values in excess of 50% (Table 6 and Figure 10E).

4.2. Orientations of SLM

The second phase of our multicriteria SLM analysis focuses on orientation characterization, employing three distinct extraction methodologies (manual, semi-automatic, and automatic) applied to three types of RSD including Landsat 8-OLI, Sentinel-1B, and Sentinel-2A. This integrated approach is illustrated in the second section of Figure 9, with quantitative orientation results presented as length-weighted and frequency-based rose diagrams in Figure 11, covering the entire Oued-Laou watershed across all extraction methods and data sources. The orientation analysis outputs are systematically integrated into our object-oriented knowledge-based system through a dedicated “SLM_Orientations” class package. This integration ensures that both the directional distribution patterns and their relative frequency/length weighting serve as critical inputs for subsequent knowledge-based processing and interpretation within the system.

The manual extraction method applied to all three RSD types (Landsat-8 OLI, Sentinel-1B, and Sentinel-2A) reveals a predominant NW-SE orientation of fault-and-thrust structural lineaments (FT-SL) (Figure 11A,F,K). This directional trend is particularly well-constrained in Sentinel-1B data, exhibiting a narrow confidence interval of 3.5° (

Table 7. Vector Mean and Confidence Interval of orientations of fault- and trust-based SL obtained using manual, semi-automatic and automatic extraction methods based on 8-OLI, sentinel-1B and sentinel-2A RSD.

FT-SL Maps (/Method/RSD/(Filter))	Vector Mean	Confidence Interval
Method_1/OLI_data	101.7	35.8
Method_2/OLI_data	20.6	6.3
Method_3/OLI_B7_data/Sobel_filter	160.1	150.2
Method_3/OLI_B7_data/Laplacian_filter	126.2	25.6
Method_3/OLI_B7_data/Kuan_filter	172.6	53.7
Mean/OLI_data	152.2	54.3
Method_1/S1B_data	151.6	3.5
Method_2/S1B_data	178.3	2.7
Method_3/S1B_VV_data/Sobel_filter	0.3	0.5
Method_3/S1B_VV_data/Laplacian_filter	0.2	0.5
Method_3/S1B_VV_data/Kuan_filter	179.4	0.5
Mean/S1B_data	174.0	1.5
Method_1/S2A_data	130.9	27.2
Method_2/S2A_data	174.9	12.2
Method_3/S2A_B12_data/Sobel_filter	164.3	52.5
Method_3/S2A_B12_data/Laplacian_filter	169.1	33.5
Method_3/S2A_B12_data/Kuan_filter	141.2	89.2
Mean/S2A_data	156.1	42.9
Geostructures/Geological maps	144.3	3.8

), indicating high directional consistency in radar-derived lineaments.

Analysis of optical RSD (Landsat-8 OLI and Sentinel-2A) using semi-automatic and automatic methods demonstrates multidirectional lineament patterns characterized by varying lengths and frequencies. Three principal orientations emerge: N-S, NW-SE, and E-W (Figure 11B–E,L–O), with confidence intervals reaching 150° (Table 7), reflecting greater directional variability compared to manual extraction. The automatic method employing Sobel directional filters (N 00°, N 45°, N 90°, N 135°) reveals distinct directional preferences when applied to Landsat-8 OLI data: NW-SE (37% of total lineament length), N-S (31%), NE-SW (20%), and E-W (12%) (Figure 11C), with a mean vector orientation of 160° (Table 7). Sentinel-2A data similarly show four dominant trends: WNW-ESE, N-S, NNE-SSW, and E-W (Figure 11M), with a NW-SE mean vector (164°).

Sentinel-1B radar data exhibit markedly different orientation patterns, with FT-SL showing a strong N-S preference across all extraction methods (Figure 11G–J). This trend is exceptionally well-defined in automatic extraction results, displaying a confidence interval of just 0.5° (Table 7). The directional consistency in radar-derived lineaments contrasts sharply with the more variable patterns observed in optical data, suggesting fundamental differences in feature detection between sensor types.

The automatic extraction method employing Sobel directional filters (N 00°, N 45°, N 90°, and N 135°) was applied to the three types of RSD, with composite SL maps generated by aggregating results from all four directions. Analysis of Landsat-8 OLI data reveals a pronounced NW-SE orientation dominance, representing 37% of total lineament length. Secondary orientations include N-S (31%), NE-SW (20%), and E-W (12%) (Figure 11C), with the mean vector orientation confirming the NW-SE trend at 160° (Table 7). Sentinel-1B radar data exhibit markedly different characteristics, with lineaments showing a strong preferential N-S orientation across both semi-automatic and automatic methods (Figure 11H). This directional consistency is particularly evident in the automatic extraction results, which yield an average vector orientation of 0.5° (Table 7), indicating exceptional directional uniformity. Sentinel-2A optical data present a more complex directional pattern, with four dominant orientations: WNW-ESE, N-S, NNE-SSW, and E-W (Figure 11M). Despite this multidirectional distribution, the mean vector orientation (164°; Table 7) maintains the overall NW-SE trend observed in other optical data, though with greater directional variability compared to the radar results.

4.3. Density Distribution of SLM

In lineament studies, density is a widely utilized parameter [4,37,145], providing a quantitative measure of lineament concentration per unit area. The methodology employed in the third stage of the SLM multivariate analysis involves assessing the density distribution of SLM, derived from manual, semi-automatic, and automatic extraction methods using Landsat 8-OLI, Sentinel-1, and Sentinel-2 RSD. This step is illustrated in the third section of the workflow (Figure 9). Figure 12 presents 15 density maps of potential FT-SL, generated through these extraction methods: five from Landsat 8-OLI (Figure 12A–E), five from Sentinel-1B (Figure 12F–J), and five from Sentinel-2A (Figure 12K–O). These maps highlight variations in SL density based on both the RSD source and the extraction method applied. The resulting density maps are integrated into our object-oriented knowledge-based system, specifically through its dedicated “SLM_Densities” class package. The outputs of the density map analysis serve as input data for our object-oriented knowledge-based system, specifically for the “SLM_Densities” class package.

The SL density maps obtained via manual extraction exhibit limited zones of maximum density, with values not exceeding 0.8 km/km² (Landsat 8-OLI), 1.6 km/km² (Sentinel-1B), and 0.8 km/km² (Sentinel-2A) (

Table 8. Parameters of density maps of SL obtained using manual, semi-automatic and automatic extraction methods based on Landsat 8-OLI, sentinel-1B and sentinel-2A RSD.

FT-SL Maps (/Method/RSD/(Filter))	Density_Min (km/km2)	Density_Max (km/km2)	Density_Mean (km/km2)	Density_Ratio
Method_1/OLI_data	0	0.793	0.064	11%
Method_2/OLI_data	0	5.079	1.284	222%
Method_3/OLI_B7_data/Sobel_filter	0	2.838	0.416	72.0%
Method_3/OLI_B7_data/Laplacian_filter	0	1.662	0.355	61.5%
Method_3/OLI_B7_data/Kuan_filter	0	1.88	0.362	62.7%
Mean/OLI_data	0	2.450	0.496	86%
Method_1/S1B_data	0	1.66	0.31	54%
Method_2/S1B_data	0	7.26	1.90	329%
Method_3/S1B_VV_data/Sobel_filter	0	5.82	2.268	393.5%
Method_3/S1B_VV_data/Laplacian_filter	0	6.02	2.363	409.9%
Method_3/S1B_VV_data/Kuan_filter	0	8.97	1.484	257.4%
Mean/S1B_data	0	5.946	1.665	289%
Method_1/S2A_data	0	0.79	0.07	12%
Method_2/S2A_data	0	10.15	0.83	140%
Method_3/S2A_B12_data/Sobel_filter	0	1.21	0.216	36.7%
Method_3/S2A_B12_data/Laplacian_filter	0	1.18	0.219	37.2%
Method_3/S2A_B12_data/Kuan_filter	0	1.21	0.2038	34.5%
Mean/S2A_data	0	2.908	0.308	52%
Geostructures/Geological maps	0	3.676	0.578

; Figure 13). These maxima are predominantly located within the Internal Rif Domain, particularly in the Dorsale Calcaire Complex (Figure 12A,F,K and Figure 13). Notably, the Sentinel-1B radar-derived density map (Figure 12F) is the only one displaying additional high-density zones within the Semtides and Gomarides Complexes. The mean SL densities remain low across all datasets: 0.06 km/km² (Landsat 8-OLI), 0.31 km/km² (Sentinel-1B), and 0.07 km/km² (Sentinel-2A) (Figure 12A,F,K and Figure 13; Table 8). Furthermore, when compared to the average density of pre-existing geological maps (Figure 13), the ratios are notably low—11% (Landsat 8-OLI), 54% (Sentinel-1B), and 12% (Sentinel-2A)—underscoring the limitations of manual extraction for FT-SL mapping in the geologically complex Rif Domain.

In contrast, SL density maps derived from semi-automatic supervised extraction reveal extensive high-density zones, with maxima reaching 5.1 km/km² (Landsat 8-OLI), 7.3 km/km² (Sentinel-1B), and 10.1 km/km² (Sentinel-2A) (Figure 12B,G,L; Table 8). These zones are concentrated in the central, northwestern, and southeastern sectors of the study area, exhibiting a NW-SE alignment consistent with the major thrust systems of the Moroccan Rif Belt. The mean densities remain elevated: 1.3 km/km² (Landsat 8-OLI), 1.9 km/km² (Sentinel-1B), and 0.8 km/km² (Sentinel-2A) (Figure 12B,G,L and Figure 13; Table 8). This heightened density likely reflects the pronounced topographic relief and intense tectonic deformation within the Rif Domain, where surface-exposed faults, fractures, and thrusts are prevalent. However, the density ratios relative to established geological maps are exceptionally high—222% (Landsat 8-OLI), 329% (Sentinel-1B), and 140% (Sentinel-2A)—exceeding field-mapped densities by over 100% (Figure 13; Section 2). These discrepancies suggest a substantial proportion of false lineaments introduced during the semi-automatic extraction process, even after enhancement post-processing steps to remove hydrographic and road network artifacts.

SL density maps generated through automatic extraction coupled with spatial filtering exhibit significant variations depending on both RSD source and the specific filtering technique applied. In this study, three spatial filters were employed: Sobel, Laplacian, and Kuan (see Section 3.2.3). For Landsat-8 OLI data, the resulting SL density maps display multiple high-density zones, with maximum values of 2.8 km/km² (Sobel), 1.7 km/km² (Laplacian), and 1.8 km/km² (Kuan). Among these, the Sobel directional filter proves most effective in highlighting geologically significant features, particularly in the central, northwestern, southeastern, and southwestern sectors of the study area, where SL align predominantly NW-SE (Figure 12C–E; Table 8). The density ratios relative to pre-existing geological maps (Figure 13) remain within acceptable limits: 72% (Sobel), 62% (Laplacian), and 63% (Kuan), demonstrating the efficacy of automatic extraction with spatial filtering for SL detection in optical RSD.

When applied to Sentinel-1B radar data, the automatic extraction method yields markedly higher SL densities: 5.8 km/km² (Sobel), 6 km/km² (Laplacian), and 9 km/km² (Kuan). The Sobel and Laplacian filters produce extensive high-density zones covering nearly the entire study area, whereas the Kuan filter delineates more discrete, NNW-SSE-oriented zones (Figure 12H–J; Table 8)—an alignment consistent with the major thrust systems of the Moroccan Rif Belt. Despite the Kuan filter’s superior performance in isolating SL, the density ratios relative to geological references remain excessively high: 393% (Sobel), 410% (Laplacian), and 257% (Kuan) (Figure 13). This discrepancy suggests significant ambiguity in radar-derived SL, as the extracted lineaments likely correspond to ridge lines rather than FT-SL, a limitation attributable to the complex topography of the Rif Domain.

The SL density maps derived from Sentinel-2A data closely resemble those from Landsat-8 OLI, exhibiting similar spatial distributions (Figure 12M–O; Table 8). Mean densities are consistently low (~0.21 km/km² across all filters), though localized high-density zones (~1.2 km/km²) are evident (Figure 13). These maxima are concentrated in the Internal Rif Domain (notably the Dorsale Calcaire Complex) and the southwestern sector near the boundary between the Flysch and External Domains. The results underscore the comparability of Sentinel-2A and Landsat-8 OLI for SL mapping when processed with identical automatic extraction and filtering methods.

4.4. Spatial-Correlation Between SLM Densities and Geological Map Densities

Spatial analysis and correlation techniques serve as fundamental tools for examining relationships between geological variables, enabling pattern identification, model refinement, and analytical validation. Geographic Information Systems (GIS) provide robust capabilities for autocorrelation analysis and spatial visualization, which have been extensively employed in fault- and trust-based distribution studies through lineament analysis tools [14,142,145]. Our fourth-stage SLM multicriteria analysis integrates manual, semi-automatic, and automatic extraction methods applied to Landsat 8-OLI, Sentinel-1, and Sentinel-2 RSD, focusing specifically on spatial correlations between SLM densities and geological map densities (Figure 9). This approach utilizes density maps to compare fault-and-thrust SLM distributions (Figure 8 and Figure 12) with existing geological mappings (Figure 2 and Figure 3) within the Oued-Laou watershed.

The spatial correlation work is focused on zones A, B and D of the study area, due to the availability of geological data. In fact, the basic geological map of the Oued-Laou watershed region, which served as the foundation for the validation of this work, is based on the 1:50,000 geological maps of: Bab Taza, Talambote, Souk Larbaa Beni Hessane, corresponding to zones A, B and D, respectively, of the study area. For zone C, no detailed geological map is currently available; the geological setting work is based on the 1:200,000 geological map of the Moroccan Rif (Figure 2 and Figure 3).

The third step involves the implementation of spatial correlation, in its true sense, between the two populations at the centre of each grid cell (algorithm of extraction multivalues to point), between the points on the pre-existing geological structure density map and the points on each of the 15 SL density maps detected for each zone. The resultant dataset comprises 15 correlation plots per zone, wherein the x-axis denotes the points or population of previously mapped geological structures and the y-axis denotes the points or population of SLM detected by each method on the basis of each RSD item (Figure 14). As illustrated in Figure 15, the correlation plots for the three zones (A, B and D) are grouped. The correlation graphs for each zone are presented in Figure A6, Figure A7, Figure A8 (Appendix F) and correspond to the spatial correlation of densities in zones A, B and D, respectively, of the Oued-Laou watershed region. The correlation coefficient corresponding to each plot is called the DCC (Density Correlation Coefficient), and all 60 DCC resulting for each zone and for each method are listed in

Table 9. DCC (Density Correlation Coefficient) between the FT-SL of the different density maps and those of the geological maps densities for the three zones A, B and D of the Oued-Laou watershed area.

FT-SL Maps (/Method/RSD/(Filter))	DCC Zone A	DCC Zone B	DCC Zone D	DCC Zones ABD
Method_1/OLI_data	0.271	0.171	0.245	0.229
Method_2/OLI_data	0.299	0.374	0.414	0.363
Method_3/OLI_B7_data/Sobel_filter	0.242	0.077	0.183	0.167
Method_3/OLI_B7_data/Laplacian_filter	0.041	0.041	0.142	0.048
Method_3/OLI_B7_data/Kuan_filter	0.188	0.066	0.221	0.158
Mean_Methods/OLI_data	0.208	0.146	0.241	0.193
Method_1/S1B_data	0.387	0.250	0.585	0.407
Method_2/S1B_data	0.694	0.728	0.679	0.700
Method_3/S1B_VV_data/Sobel_filter	0.343	0.266	0.180	0.263
Method_3/S1B_VV_data/Laplacian_filter	0.323	0.280	0.183	0.262
Method_3/S1B_VV_data/Kuan_filter	0.281	0.264	0.287	0.278
Mean_Methods/S1B_data	0.406	0.358	0.383	0.382
Method_1/S2A_data	0.321	0.301	0.251	0.291
Method_2/S2A_data	0.324	0.401	0.401	0.376
Method_3/S2A_B12_data/Sobel_filter	0.267	0.186	0.083	0.179
Method_3/S2A_B12_data/Laplacian_filter	0.258	0.167	0.01	0.138
Method_3/S2A_B12_data/Kuan_filter	0.225	0.214	0.029	0.156
Mean_Methods/S2A_data	0.279	0.254	0.155	0.228

.

The Densities correlation results are integrated into our object-oriented knowledge-based system, specifically through its dedicated “SLM_Correlation” class package. The outputs of the densities correlation analysis serve as input data for our object-oriented knowledge-based system, specifically for the “SLM_Correlation” class package.

The Pearson correlation coefficient (r) quantitatively measures the strength and direction of linear relationships between variables. In our spatial analysis, consistently positive DCC values (ranging 0.1–0.728) indicate direct proportional relationships between SLM densities and geological map densities (Figure 15, Table 9). These coefficients demonstrate that:

• r ≈ 1 signifies strong positive correlation;
• r > 0 indicates direct relationships (increasing SLM density corresponds to increasing geological feature density);
• r ≈ 0 suggests no linear relationship.

Analysis reveals significant DCC variation based on the following (Figure 15; Table 9): (1) Remote Sensing Data (RSD) type: Sentinel-1 radar data yielded superior correlations (DCC 0.679–0.728), while optical data (Landsat-8 OLI/Sentinel-2A) showed weaker correlations (DCC 0.1–0.3); (2) Extraction methodology: Semi-automatic methods achieved highest DCC values, while automatic methods with spatial filtering showed minimal variation between filter types. (3) Spatial zones (A, B, D): Zone B demonstrated peak correlation (DCC 0.728) for semi-automatic/Sentinel-1 combination.

The semi-automatic extraction method applied to Sentinel-1B data (Method_2/S1B_Data) consistently produced optimal results across all zones (DCC: 0.694 Zone A, 0.728 Zone B, 0.679 Zone D). This superior performance suggests radar data’s enhanced capability for fault-and-thrust lineament detection compared to optical alternatives. The minimal filter-dependent variation in automatic methods (Figure 15A–E,K–O) indicates spatial filtering choice has negligible impact on correlation strength when using standardized extraction parameters.

4.5. Knowledge Based MCDA

As illustrated in Figure 9, this study adopts a holistic approach. A comparison of results from the four phases of structural lineament (SL) analysis—using manual, semi-automatic, and automatic extraction methods applied to OLI, Sentinel-1, and Sentinel-2 sensor data—reveals that selecting a single method or remote sensing data (RSD) type is non-trivial. The selection process must account for multiple criteria, including SL statistics (e.g., length), orientation, spatial distribution, density, and correlation with existing geological structures.

However, these criteria often exhibit trade-offs. For instance, manual extraction applied to radar (Sentinel-1) RSD yielded satisfactory SL orientation results but detected only 18% of lineaments based on length statistics. Similarly, automatic extraction methods coupled with Sobel, Laplacian, and Kuan spatial filters (applied to OLI data) produced nearly identical results for SL length and density. Yet, only the Sobel-filter-based method achieved robust SL orientation outcomes. Such multi-criteria discrepancies complicate method selection, motivating the integration of a knowledge-based system (KBS) with multicriteria decision analysis (MCDA) to systematize decision-making.

While MCDA offers advantages (e.g., explicit trade-off analysis and consensus-building), its efficacy hinges on precise criterion definition and weighting—a non-trivial task, particularly in geospatial modeling of complex geological structures in inaccessible mountainous terrain. Here, traditional field mapping is impractical, necessitating RSD-based solutions. However, the four integrated analyses yielded divergent results, raising two key questions: Which RSD type is optimal for SL extraction? Which extraction method is most suitable?

To address these challenges, we developed an integrated Knowledge-Based Multi-Criteria Decision Analysis (KB-MCDA) framework. MCDA enhances decision quality by quantifying trade-offs, while KBS standardizes objective/criterion definitions and weight heuristics via rule-based logic. This synergy bridges quantitative and qualitative insights, improving analytical rigor and practical applicability in SL modeling.

The KBS contribution operates at two levels: (1) Rule-based system: Clarifies objectives, criteria, and criterion weights for the study context. (2) Object-oriented fact base: Represents the four core objectives of KBMCDA—Good SL Statistics, Good SL Orientation, Good SL Density, and Good Correlation Coefficient—as class packages. For example, the Good SL Statistics objective maps to the SL_Statistics class package, which includes superclasses (e.g., SL_Statistics_SC) and subclasses (e.g., SL_Count, SL_Length, SL_Confidence_Interval_1). Attributes like Min_Length, Max_Length, and Mean_Length aggregate data for criteria under this objective (see Appendix E, Figure A5, for full KBS architecture).

The final KB-MCDA scores (method detailed in Section 3.4) are presented in

Table 10. KB-MCDA score results grouped by criteria and by objective for SLM obtained using manual, semi-automatic and automatic extraction methods, based on Landsat 8-OLI, sentinel-1B and sentinel-2A RSD.

FT-SL Maps (/Method/RSD)			Objective 1						Objective 2		Objective 3		Objective 4			MCDA Score/Method
			SL_Count	SL_Sum_Length	SL_Length_Min	SL_Length_Max	SL_Length_Mean	SL_Length_CI	Vector_Mean	Vector_CI	Density_Max	Density_Mean	DCC_Zone-A	DCC_Zone-B	DCC_Zone-D
Method_1/OLI_data																10
Method_2/OLI_data																30
Method_3/OLI_B7/Sobel_f																22
Method_3/OLI_B7/Laplac_f																15
Method_3/OLI_B7/Kuan_f																16
Method_1/S1B_data																35
Method_2/S1B_data																24
Method_3/S1B_VV/Sobel_f																17
Method_3/S1B_VV/Laplac_f																17
Method_3/S1B_VV/Kuan_f																21
Method_1/S2A_data																17
Method_2/S2A_data																29
Method_3/S2A_B12/Sobel_f																15
Method_3/S2A_B12/Laplac_f																14
Method_3/S2A_B12/Kuan_f																15
MCDA Score/Criteria			5	5	18	13	9	14	54	54	33	33	22	18	16
MCDA Score/Objective			64						108		66		56
		Very satisfactory criteria				Slightly satisfactory Criteria						Not satisfactory Criteria

, which aggregates results by RSD type, extraction method, criterion, and objective. For instance, the highest-scoring method (KB-MCDA Score = 35) was manual extraction from Sentinel-1 data, underscoring radar’s potential to detect surface and subsurface structures—albeit with labor-intensive, expertise-dependent workflows.

In scenarios lacking expert manual extraction, the top-performing automatic method (either Method_2/OLI_data, MCDA Score = 30, or Method_2/S2A_data, Score = 20) is recommended (Table 10). For spatial filter selection, Sobel (OLI) and Kuan (Sentinel-1) filters proved most effective. Querying by objective revealed that most methods achieved the Good SL Orientation target (validated via rose-diagram analysis). However, scores for Good SL Statistics, Good SL Density, and Good Correlation Coefficient objectives were comparatively low (mean scores: 64, 66, 56), likely due to the study area’s complex geomorphology, which introduces false lineaments alongside genuine geological features (Table 10).

5. Discussion and Recommendations

A more detailed examination of this approach was required to evaluate its potential for mapping structural lineaments (SLs), including thrusts and faults, in the Alpine Geological Domain. This generalization was based on several geological similarities between the Rif and Alps mountain ranges: (i) Alpine Orogeny [109,110,111,112,113]: both geological domains originated during the same tectonic period, the Alpine orogeny, which began in the Mesozoic era and continued into the Cenozoic era. This period was characterized by the collision of tectonic plates, particularly the African and Eurasian plates; (ii) Fold Structures: both domains contained characteristic fold structures consisting of thrusts and thick, overlapping rock layers—typical features of collision-formed mountain ranges; (iii) Molasse and Olistostromes: both domains exhibited common sedimentary deposits including Molasse deposits (sedimentary rocks formed by mountain erosion) and olistostromes (chaotic rock formations deposited by mass movements).

Remote Sensing had demonstrated its effectiveness in various branches of geology, particularly in structural geology and structural lineament mapping. SL were defined as geological features including subvertical and subhorizontal faults and thrusts. These structures could be extracted from various RS data sources and GIS techniques. Conventional approaches to mapping SL were often time-consuming, costly, and difficult to apply due to challenging geomorphological/geological conditions in mountainous regions like the Rif belt in Morocco. The utilization of RS and GIS facilitated the identification of faults, thrusts and their tectonic implications, providing an alternative to conventional methodologies.

The primary objectives of this study were (1) to introduce commonly encountered RSD and their properties for geological mapping and structural lineament extraction, (2) to compare various manual, semi-automated, and automated structural lineament extraction methods and algorithms, and (3) to evaluate the challenges and accuracy of these techniques in detecting structural lineaments within rugged mountainous regions.

While Multi-Criteria Decision Analysis (MCDA) offered advantages (e.g., explicit trade-off analysis and consensus-building), its effectiveness depended on precise criterion definition and weighting—a non-trivial task, particularly in geospatial modeling of complex geological structures in inaccessible mountainous terrain. Traditional field mapping was impractical in such areas, necessitating RSD-based solutions. However, the four integrated analyses yielded divergent results, raising two key questions: Which RSD type was optimal for SL extraction? Which extraction method was most suitable?

To address these challenges, we developed an integrated Knowledge-Based Multi-Criteria Decision Analysis (KB-MCDA) framework. MCDA enhanced decision quality by quantifying trade-offs, while the Knowledge-Based System (KBS) standardized objective/criterion definitions and weight heuristics through rule-based logic. This synergy bridged quantitative and qualitative insights, improving analytical rigor and practical applicability in SL modeling.

The preliminary analysis compared four categories of result sets, designated ‘KB-MCDA-objective’: (1) SL statistics analyzed through statistical evaluation of each SLM; (2) SL orientation analyzed through rose diagrams of length and frequency; (3) SL densities studied through density maps; (4) Density correlation assessed through spatial correlation between geological maps and SLMs in zones A, B and D of the Oued-Laou watershed region.

Statistical Analysis of Structural Lineament Maps (KB-MCDA Objective 1): the statistical evaluation of the SLMs incorporated several key parameters: the number of identified SL, their minimum, maximum, total and mean lengths, standard deviation, and confidence intervals. Analysis of the 15 generated SLM revealed that the detected lineaments in the Oued-Laou Watershed exhibited strong correlation with the structural framework of the Rif Domain, showing notable variation across different geological units. The majority of SL were concentrated in the Internal Rif Domain, particularly within the Gomarides and Septides Complexes and the Dorsale Calcaire Complex.

Comparative analysis demonstrated that radar RSD (Sentinel-1 SAR) yielded optimal results, outperforming other data sources in terms of total SL count (ranging between 116 and 239 m), cumulative length, and size distribution. However, elevated standard deviation values in semi-automatic and automatic extraction outputs suggested significant false positive detection rates. This phenomenon was attributed to the inherent sensitivity of radar sensors to topographic features, which frequently led to misidentification of ridgelines as structural lineaments.

Orientation Analysis Using Rose Diagrams (KB-MCDA Objective 2): Rose Diagrams analysis provided valuable insights into SL orientation patterns within the study area. Manual interpretation of Landsat 8-OLI, Sentinel-1B and Sentinel-2A data consistently revealed a dominant NW-SE orientation trend, particularly evident in Sentinel-1B outputs (showing narrow confidence intervals). These findings showed strong agreement with existing geological maps, reaffirming the value of expert interpretation in structural mapping.

Semi-automatic and automatic processing of optical data (Landsat-8 OLI and Sentinel-2A) produced more variable orientation patterns, with notable concentrations in N-S, NW-SE and E-W directions. However, these results required cautious interpretation due to wide confidence intervals (reaching 150° in some cases). Radar data (Sentinel-1B) exhibited a pronounced N-S orientation bias with exceptionally narrow confidence intervals, reflecting the system’s tendency to preferentially detect ridgelines in the rugged terrain.

Density Distribution Patterns (KB-MCDA Objective 3): Density mapping revealed distinct spatial patterns across different extraction methods. Manual methods produced localized high-density zones (0.8 km/km² for optical, 1.6 km/km² for radar RSD), while semi-automatic approaches identified extensive high-density areas aligned NW-SE, paralleling the principal thrust directions of the Rif belt. Automatic extraction with Sentinel-2A data yielded density distributions remarkably similar to Landsat-8 outputs, with peak values concentrated in the Internal Rif Domain (particularly the Dorsale Calcaire complex) and along the southwestern boundary between the Flysch and External Rif domains.

Density Correlation Validation (KB-MCDA Objective 4): Spatial correlation analysis between SLM densities and mapped geological structures yielded Density Correlation Coefficients (DCC) ranging from 0.1 to 0.728. Sentinel-1 data achieved the highest validation scores (DCC ≈ 0.7 across zones A, B, and D), indicating 70% detection accuracy of verified structures. In contrast, optical sensors (Landsat-8 and Sentinel-2) showed consistently lower correlation values (DCC = 0.1–0.3).

The comprehensive KB-MCDA evaluation identified manual SAR-based extraction as the most accurate approach for structural lineament mapping, though this method demanded significant labor investment and required expert interpretation. To enhance operational efficiency while preserving data quality, we recommended adopting supervised automated extraction with optical RSD, implemented alongside rigorous false-positive filtering protocols. Furthermore, we developed two optimized automated workflows: Sobel filter-based processing for Landsat data and Kuan filter-enhanced extraction for Sentinel-1 SAR data, with both approaches specifically designed for the RSD types and filtering methods examined in this study.

Looking forward, our findings suggest three critical research priorities: (1) spectral band optimization to improve SL detection across different RSD categories, (2) development of advanced spatial filtering algorithms tailored to mountainous terrain applications, and (3) integration of UAV-LiDAR technology, which offers exceptional potential through its high-resolution topographic data acquisition (100–1000 points/m²), superior vegetation penetration capabilities, and rapid deployment in challenging terrains. The demonstrated methodology shows particular promise for extension to other inaccessible mountainous regions exhibiting similar geological characteristics to our study area.

6. Conclusions

This research successfully integrates remote sensing data and advanced analytical techniques to map structural lineaments in the geologically complex Rif Belt. The study demonstrates that Sentinel-1 SAR data, combined with manual extraction, offers the highest accuracy in detecting faults and thrusts, despite its labor-intensive nature. Semi-automatic and automatic methods, particularly when enhanced by Sobel and Kuan filters, provide viable alternatives for large-scale mapping. The KBS-MCDA framework proves effective in evaluating multi-source data and extraction methods, offering a scalable solution for similar mountainous regions. Future work should focus on optimizing band selection and spatial filtering algorithms to further improve automated lineament detection. The methodology’s applicability to the Alps underscores its potential for global geological studies, contributing to resource exploration, hazard assessment, and environmental monitoring in challenging terrains.