Active Landslide Mapping Along the Karakoram Highway Alternate Route in North Pakistan; Implications for the Expansion of China−Pakistan Economic Corridor

Abstract

Slowly moving active landslides threaten infrastructure, particularly along highway routes traversing active mountainous ranges. Detecting and characterizing such landslides in highly elevated mountainous terrains is challenging due to their inaccessibility, wide area coverage, limited approaches, and the complex nature of mass movements. In this study, we processed Sentinel-1 Synthetic Aperture Radar data acquired from 2015 to 2024 to detect active landslides along the Karakoram Highway alternate route (Chitral-Gilgit) and the Karakoram Highway part (Gilgit-Khunjerab). We detected 1037 active landslides in the study region using phase gradient stacking and a deep learning network. Based on the detection, we applied time series InSAR analysis to reveal the velocity and deformation series for some large-scale landslides, revealing high displacement rates with line-of-sight velocities reaching up to −81 mm/yr. We validated our detections by comparing them with Google Earth imagery and the previously published landslide inventories along the Karakoram Highway. This study reveals the spatial distribution of active landslides along the uplifted mountainous terrain, highlighting potentially unstable zones, and offers insights into hazard mitigation and risk analysis, especially for less monitored economic roads in orogenic zones.

1. Introduction

Inspired by the ancient Silk Road, China established the Belt and Road Initiative (BRI) to promote global political collaboration, cultural diversity, and financial cooperation through six economic corridors [1,2]. A key component of the BRI is the China−Pakistan Economic Corridor (CPEC), which was officially launched on 20 April 2015. With an approximate length of 3000 km from Kashgar in China to the Gwadar region of Pakistan, this route passes through complex geological and topographical environments, including the world’s highest mountain ranges: the Karakoram, Hindukush, and Himalayas (KHH) [3,4,5].

Landslides have become more common in recent years because of increased interactions between human activities and natural systems, exacerbated by climate change and population growth [6]. Landslide inventory maps serve various functions, including documenting landslide extents at different scales [7], assisting with susceptibility, hazard, and risk assessments [8], and studying landscape evolution from mass-wasting events [9,10]. The Karakoram region presents a highly complex scenario in which multiple interacting factors govern landslides. Situated on a collisional boundary, the region is tectonically active, with retreating glaciers, deeply dissected valleys, steep topography, elevations exceeding 7000 m, extreme weather variations, alluvial terraces, and disturbed rock formations. These conditions create a unique environment for the initiation and development of landslides, contributing to the diverse types that make the KHH region an essential yet challenging region for developing a comprehensive landslide inventory.

The province of Gilgit Baltistan in Pakistan, through which the Karakoram Highway (KKH) passes, is notably susceptible to catastrophic events like landslides and rock falls. Factors contributing to this vulnerability include complex geological structures, seismic activity, snowmelt, challenging climates, and anthropogenic activities. The KHH, the world’s youngest and most prominent mountain system, accounts for a substantial percentage (~30%) of global landslides [11], with Pakistan’s regions among the worst affected [12]. Significant events, such as the co-seismic landslides during the 2005 Kashmir Earthquake [13] and the Attabad landslide, resulted in substantial human and infrastructure losses, including damage to the KKH, thereby disrupting land connectivity between China, northern areas, and the rest of Pakistan [14]. The passage of the CPEC route through this area underscores its strategic importance, highlighting the need for detailed landslide inventories to assess the potential hazards and mitigate the risks to infrastructure.

In the 6th Joint China Corridor (JCC) of the China−Pakistan Economic Corridor (CPEC) held in 2016, it was agreed to include the Chitral Corridor Road as the KKH alternate route (KKH-AR) from Gilgit to Chitral in the CPEC portfolio (Figure 1). The road starts from Gilgit town, located on KKH-AR, and ends at Chitral town, located on the Nowshera-Chitral Highway, spanning a length of 366 km (more details are available at www.cpec.gov.pk). Similar to the KKH, KKH-AR also passes through similar geologic and topographical environments and is susceptible to landslides. Researchers have worked on landslide mapping along the KKH in various aspects, including traditional susceptibility mapping [15,16,17,18,19,20] and GPR and UAV for specific landslides [18,21]. In contrast, some researchers have focused on InSAR technology to map and recognize landslide deformation [22,23,24,25,26,27,28]. These studies mapped the KKH from Khunjerab to Gilgit and identified potential active landslides. However, similar studies have not yet been carried out along the KKH-AR, leaving a gap in the reliable landslide inventory.

Traditional approaches for mapping landslides have utilized optical remote sensing images with high spatial resolutions [29,30]. Remote sensing provides an efficient method for mapping landslides due to its broad coverage and ability to capture data at different times. Advancements in remote sensing technologies like LiDAR and InSAR have improved map development and quality evaluation. LiDAR and optical remote sensing can identify landslides based on their size, shape, and surface roughness [31,32,33]. InSAR detects landslides by identifying slight displacements on sliding hill slopes to distinguish between unstable and stable slopes across extensive areas [34,35,36,37,38,39,40,41,42].

Recent developments in landslide detection have used a combination of InSAR and deep learning approaches to enhance accuracy and efficiency. In particular, phase-gradient maps and deep learning have recently been demonstrated to strengthen the detection reliability for identifying localized deformations [43,44,45]. Effective deep learning models, such as YOLOv8, have been used to analyze InSAR-derived velocity maps, thereby advancing the field [46,47,48]. Multiple data sources, including optical imagery and time series InSAR, have been investigated to improve the identification of slow-moving landslides [49,50,51,52]. These developments highlight the potential of combining phase-gradient approaches with deep learning for effective and precise detection of landslides.

Despite the growing landslide activity along the KKH, which has been widely documented, this study aims to assess the landslide inventory along the KKH-AR region. Here, we employ a highly effective approach of phase-gradient stacking in conjunction with the YOLOv3 deep learning network [25,43] to detect and map the first comprehensive landslide inventory along the KKH-AR (Gilgit to Chitral) and update the existing inventory along the KKH between Khunjerab and Gilgit. By integrating InSAR with deep learning, our approach improves and strengthens landslide mapping and detects more landslides that were missed in previous studies.

2. Study Area

The study area is marked by a 10 km buffer along the KKH-AR and KKH (Gilgit-Khunjerab section), connecting northern Pakistan with Western China and a critical segment of the CPEC under the BRI (Figure 1). The study area runs across the hilly regions of the Karakoram and Himalayan ranges, which are sculpted by the continuing collision of the Indian and Eurasian tectonic plates, including the Kohistan Island Arc [53]. The topographic altitude of the region ranges from 1211 to 7831 m above sea level. The area receives roughly 154 mm of precipitation yearly, with glaciated peaks throughout the year. Further, this region relies significantly on rivers and streams, fed mainly by snowmelt and glacial runoff from the surrounding mountains, to provide agricultural irrigation. The complex geographical features and climate contribute to frequent landslides and avalanches in this area [54].

Tectonically, the study area is part of the orogenic processes of the Indo-Eurasian (continent-island arc-continent) collision around 50 million years ago, with an ongoing convergence rate of 5 cm annually [55]. This process sculpted the whole region and is driven by crustal shortening and active faulting. The Main Karakoram Thrust (MKT), Tirich Mir fault (TF), and Main Mantle Thrust (MMT) (Figure 2) are the major tectonic features representing the Indo-Eurasian tectonic blocks [56] and pose significant brittle deformation along the KKH-AH and KKH [57]. The TF marks the boundary between the western Karakoram and eastern Hindukush, whereas the Reshun fault (RF) separates the western Karakoram into northern and southern units [58]. Sedimentary, igneous, and metamorphic rocks constitute the main geology along the route. The rock masses exposed along KKH-AR and KKH are extensively jointed and fractured due to the brittle deformation of regional tectonic elements [59,60], classified as the major lithology along the Karakoram Kohistan Suture between Hunza and Drosh, and interpreted as an olistrosome-type mélange containing various geological units of limestone red shales, conglomerates, quartzite, volcanic greenstone, and serpentinites.

The majority of the rocks covered in the research region comprise the southern Karakoram metamorphic complex, which consists of paragneisses, garnet amphibolites, basalt dikes, and interbedded marbles (Figure 2). The Hunza plutonic complex comprises granodiorite, tonalite, quartz diorite, and migmatite rocks (Figure 2). The Chalt Volcanics comprise greenschist, gneisses, amphibolites, andesite metabasalt, and the Kohistan batholith comprises granite, granodiorite, and dioritic rocks. The Yasin group comprises bottom sedimentary and upper volcanic units, whereas Iskuman Valley Ghizer comprises volcanic and limestone units. The sediments are slaty, silty quartzite, and conglomeratic near the Main Karakoram Thrust [61]. Further to the Chitral region, Devonian carbonates, high-grade metamorphic rocks, Reshun marble, Chitral slates, and the Kohistan batholith represent the lithological strata [62]. The region’s lithology displays diverse rock types with a deformed nature, reflecting the complex tectonic history that has shaped the area’s landform.

Figure 2. Geological map of the study area: multiple colors represent the lithological units along the study area, with black fault lines representing regional tectonic boundaries. The Zigzag black color line represents the CPEC KKH route. In the figure, MKT represents the Main Karakoram Thrust; MMT, Main Mantle Thrust; TF, Tirich Mir fault; RF, Reshun fault; HFS, Hunza fault system; NP, Nanga Parbat massif, map modified after [63].

Figure 2. Geological map of the study area: multiple colors represent the lithological units along the study area, with black fault lines representing regional tectonic boundaries. The Zigzag black color line represents the CPEC KKH route. In the figure, MKT represents the Main Karakoram Thrust; MMT, Main Mantle Thrust; TF, Tirich Mir fault; RF, Reshun fault; HFS, Hunza fault system; NP, Nanga Parbat massif, map modified after [63].

3. Materials and Methods

3.1. Data Set

In this study, we used 1073 Sentinel-1 SAR images, including 537 images for the ascending track and 540 images for the descending track, acquired from 1 January 2015, to 1 August 2024, by the European Space Agency (https://search.asf.alaska.edu/) accessed on 16 July 2024. The tempospatial baseline was organized within 48 days and 150 m, respectively (Figure 3). The Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (DEM) with 30 m resolution was acquired from [64]. The geological map of the area was modified from a previously published map of northern Pakistan [63].

3.2. Phase-Gradient Stacking

We use the Sentinel-1 Interferometry Processor [65] to process two sets of ascending and descending Sentinel-1 imagery. To mitigate the topographic influence, we used the Shuttle Radar Topography Mission (SRTM) Digital Elevation Model with a resolution of 30 m [64]. A 2D convolution filter with a 10-by-10 pixel window is applied to the wrapped interferograms and weighted by coherence. A total of 1412 and 1377 interferograms were generated for the ascending and descending orbits, respectively, with multilook parameters of 2 and 8, representing the azimuth and range directions, resulting in a ground resolution of roughly 40 m. This study focuses on landslides with dimensions larger than 40 m, while smaller landslides are difficult to detect due to the low signal-to-noise ratio in the interferograms. Phase-gradient stacking combined with a deep learning model [43] automatically detects active landslides, as illustrated in Figure 4. The stacked gradient maps and detection results are then geocoded and imported into the Google Earth platform for validation using optical imagery. Phase gradients are calculated in both the azimuth and range directions by contrasting the phases of adjacent pixels in each direction. The gradients are then accumulated over time for every burst. The interferometric phase gradients are wrapped within the range of (−π, π) to adequately represent phase shifts between nearby pixels.

Figure 4. Methodological framework (a) represents the interferogram formation with phase gradient stacking and deep learning for detection of landslide, adopted from [43] (b) a detailed overview of the LiCSBAS workflow from phase unwrapping to velocity estimation.

Figure 4. Methodological framework (a) represents the interferogram formation with phase gradient stacking and deep learning for detection of landslide, adopted from [43] (b) a detailed overview of the LiCSBAS workflow from phase unwrapping to velocity estimation.

This method assumes that the unwrapped phase difference between neighboring pixels is usually less than π/2, which is a common criterion for phase unwrapping algorithms [66]. The wrapped phase-gradient maps accurately depict the unwrapped phase changes between neighboring pixels, and coherence is used as a weight throughout the stacking procedure to mitigate the noisy outcomes. In places with complete decorrelation, the stacked phase gradient values cancel out, and there is no identifiable red-blue coupling pattern, which has no impact on landslide detection. While decorrelation can make deformation identification more difficult, phase-gradient stacking is still effective since it only fails to recognize deformation signals when there is complete decorrelation.

After stacking the phase gradients, the atmospheric phase effects are reduced since the atmospheric phase differences between neighboring pixels are insignificant and shift randomly over time. For localized deformation, the phase gradients reflect a recognizable sign change along the calculation direction. In contrast, landslide deformation is characterized by a typical sequence of a shift from positive to zero and subsequently to negative values. This pattern is visually represented by the red-blue coupling in the stacked phase-gradient maps that deep learning structures can easily recognize (Figure 3).

Phase-gradient maps serve to identify areas of localized deformation, which are represented by red and blue patterns. However, these maps only indicate signs of deformation and do not reveal specific landslide boundaries. In Zhang et al. [67], landslide boundaries are demarcated using deep learning and phase gradient investigation to recognize deformation zones where sharp changes in the phase gradient correspond to the boundaries of the landslides. However, the main focus of this study was to detect the location of landslides and study their spatial distribution. As a follow-up study, we will conduct an absolute gradient with velocity maps to delineate the boundaries of individual landslides.

3.3. Deep Learning Detection

In this study, we used the deep learning system developed by Fu et al. [43] to detect localized deformation associated with slow-moving landslides in stacked phase-gradient maps. The model uses an implemented YOLOv3 architecture comprising 53 layers. It was trained using a dataset of 5692 manually assigned samples gathered from various locations in Western China [43]. The training method produced a precision of 0.95 and a recall of 0.87 on the validation dataset, indicating the model’s capacity to detect landslides effectively. The strong performance of the pre-trained YOLOv3 network, as demonstrated in previous research [43], allows for the automatic identification of landslides despite the study region, indicating that this method is both applicable and adaptable for future applications in other areas. Unlike typical InSAR algorithms that employ reference points to estimate hillslope deformation, this approach identifies landslide borders by examining the spatial variations in phase gradients [47]. In the phase-gradient maps, distinct red-to-blue transitions indicate areas of localized deformation. These color variations correspond to changes in the phase gradient values (Figure 4a), where red represents positive phase shifts and blue represents negative phase shifts, highlighting potential landslide activity. To enhance the visual contrast of these variations, we scale the phase gradient values to [−4, 4] radians, making the red-blue deformation patterns more prominent. This pattern, observed in both the azimuth and range phase-gradient maps (Figure 4a), serves as a key indicator of active landslides in various areas. The pre-trained YOLOv3 network automatically detects slow-moving landslides (Figure 4a) without employing gradient-velocity criteria, mainly focusing on the change from positive to negative phase values.

The network recognizes gradient patterns in the azimuth and range directions by creating rectangular bounding boxes around deformation spots, with the center of each box revealing the slow-moving landslide region. Each box represents the detection of a considerable distortion within the phase-gradient maps. To merge the deformation signals from the azimuth and range gradient maps, we use a Generalized Intersection over Union (GIoU) technique [43]. The GIoU value, which ranges from −1 to 1, represents the overlap of two bounding boxes. A GIoU value of 1 suggests a perfect overlap, whereas values near −1 indicate a complete separation between the boxes.

By testing different values, we set the GIoU threshold to 0.1, which means that there is sufficient overlap between the two detection boxes to merge the detections from the azimuth and range gradients. This ensures that the determined deformation zones are properly documented, reduces false positives, and increases the spatial reliability of landslide detection. The technique was applied to both ascending and descending track data, providing a comprehensive detection of slow-moving landslides across the whole research area using information from both tracks.

3.4. SBAS Analysis

After detection, we apply the SBAS (Small Baseline Subset Analysis) approach to compute the deformation velocity and time series. The SBAS technique captures multi-temporal ground motion and maps the deformation velocity and time series with millimeter-level accuracy [68], which involves integrating unwrapped interferograms spatially isolated based on the baseline between each orbital acquisition at short intervals [69,70]. Furthermore, ref. [69] proposed an improvement to the inversion process of unwrapped interferograms, focusing on enhancing the deformation extraction over time. We use LiCSBAS v1.3.6 software, an open-source application built on Python 3 and Bash, to analyze cropped interferograms of the targeted landslides [71]. The multi-temporal assessment includes a quality check of the interferograms and a closed-loop correction to address unwrapping problems. After the evaluation, the remaining small baseline network was inverted to determine the surface velocities. The final steps involve calculating the velocity standard deviation, masking noisy pixels, and applying spatial-temporal filtering. The final result reflects the surface movement velocity and indicates time series deformation over time [72,73] (Figure 4b).

4. Results

4.1. Detection of Active Landslide

We detected 1037 targets with localized deformation after combining data from ascending and descending tracks within a 10 km buffer along the Khunjerab to the Chitral route. These detections are scattered widely across the KKH and KKH-AR (Figure 5). We import the phase-gradient maps into the Google Earth platform and carefully validate our detection with the geomorphological features of landslides. As the KKH passes through a high-altitude region with significant snow coverage, our area of interest lies within a 10 km buffer zone that excludes the snow-covered region. Regarding this, a small region coincides with the snow peaks. Among those, we detected 22 cases, which were considered false detections.

We compared our detections with previously detected active landslides in the eastern Karakoram region along the KKH, as documented in previous studies [22,26]. We showed these previous detections (Mayoon landslide, Miacher landslide, and Budalas landslide) on our gradient maps (Figure 6) to cross-validate the robustness of our approach (Figure 6). The phase gradient in azimuth and range clearly displays the red and blue pattern for the published landslides with newly detected active landslides (Figure 6).

Figure 6. Detection of major active landslides in eastern Karakoram by phase-gradient stacking and deep learning network (blue boxes) obtained from track AT100, with a black circle representing the previously published landslide by [26], Mayoon and Budalas landslide in phase gradient azimuth and range maps (a,b), Maicher landslide in (c,d).

Figure 6. Detection of major active landslides in eastern Karakoram by phase-gradient stacking and deep learning network (blue boxes) obtained from track AT100, with a black circle representing the previously published landslide by [26], Mayoon and Budalas landslide in phase gradient azimuth and range maps (a,b), Maicher landslide in (c,d).

4.2. Landslide Density

Based on the detected active landslides along the KKH Khunjerab to Chitral road, we compute the landslide density. We overlay the derived density map on the regional geological structures in our study region. In the western Karakoram, the area between Chitral and Reshun displays a high-density zone bounded by three major fault systems, namely (the MKT, RF, and TF). Meanwhile, in the eastern Karakoram, the Hunza Attabad region also shows the highest density, followed by the region from Sost to Khunjerab (Figure 7). A series of fault systems cluster in the Attabad Hunza Valley, creating a high-density zone. These regions display the highest density (2.49), highlighting deep-seated landslides and signifying greater instability in the area. The region between Ghizer and Mastuj and Gilgat’s surrounding areas show a moderate density zone, reflecting shallow landslide activity.

4.3. Significant Cases

We present three representative cases of active landslides detected in this study. The positive and negative gradients represent the maximum deformation in response to the phase-gradient computation. All three cases indicate that the maximum deformation in the red and blue patterns highlight the subsidence and uplifting parts of active landslides (Figure 8).

• Riri Oweer Landslide

Riri landslide is located in the Riri Oweer village in the upper Chitral district, approximately 45 km NE of the Chitral city. It is accessible via a 5 km unpaved road southwest of Reshun Town at a latitude of 36.141° and longitude of 72.044°. This landslide hosts Riri village at the toe of the sloping terrain near the junction of the Riri River, which flows into the Mastuj River to the west. We detected this landslide in the gradient maps and further ingested the gradient maps into Google Earth to correlate with the surface features of the landslide (Figure 8). The LOS ascending velocity of this landslide was recorded as −72 mm/yr (Figure 8).

• Nishku Landslide

The Nishku landslide in the Torkhow Valley of the upper Chitral district is about 108 km NNE of Chitral City. It is accessible 20 km north of Booni town at a latitude of 36.399° and a longitude of 72.365°. The main landslide body is north of Nishku village, and the northern part of Nishku village is also affected by this landslide. The Mulkhow River flows in a valley and is the toe cutter for this landslide, whereas the head region hosts multiple, old scarps. We detected this landslide in the gradient maps and further overlaid the gradient maps on Google Earth to verify and match them with the morphological features of the landslide (Figure 8). The LOS ascending velocity of this landslide is recorded as −81 mm/yr (Figure 8).

• Hassis landslide

The Hassis landslide is located in the Punial vVlley of the district Ghizer, approximately 100 km west of Gilgit City. It is accessible 15 km north of Gahkuch town at a latitude of 36.282° and a longitude of 73.826°. The landslide outcrop threatens two villages on the foothill side. The LOS velocity of this landslide is −67 mm/yr. (Figure 8). We also detected this landslide in gradient maps and further ingested the gradient maps into Google Earth to verify the spatial distribution and correlate with the morphological features of the landslide (Figure 8).

4.4. Deformation Analysis of Significant Cases

InSAR is a powerful tool for detecting small-scale features of ground surface deformation and provides details on the characteristics of landslide movement. The barren terrain of the Karakoram region offers a conducive environment for the InSAR analysis. We implemented the SBAS-InSAR approach to analyze the SAR data and compute the landslide deformation rate for selected bodies along the CPEC route from 2015 to 2024. SBAS-InSAR maps deformation along the LOS direction for both the ascending and descending tracks. The deformation of three representative landslides (Hassis, Rirri Owir, and Nishku Torkhou landslides) was examined in terms of displacement velocities and temporal variation. These landslides are a serious threat to local villages standing at their toe. The displacement rate of the detected representative landslide reveals ground deformation, indicating variations in landscape features. The time series accumulated deformation from 2015 to 2024 is computed for three significant cases, namely, Hassis landslide (−140 mm), Rirri Oweer (−400 mm), and Nishku Torkhow (−250). Notably, the Nishku Torkhow landslide displays an uplifted rate at the toe side, representing the accumulation of sliding material from the upside of the landslide over time (Figure 8). The time series deformation and moving rate of the downslope materials reflect the active nature of these landslides. The observed high deformation InSAR velocities in our study are aligned with regional deformation patterns found in similar settings, supported by cross-validation with previous landslide InSAR-driven inventories.

4.5. Validation

We used a two-phase approach to validate the results. We first used manually annotated detailed Google Earth images to detect the geomorphological features of landslides, including scarps, fractures, and slope trends. The manual annotation procedure involves marking the locations and extent of landslides from optical imagery (Figure 8). Although manual annotation involves some subjectivity, it remains a commonly accepted technique for validating remote sensing results, especially in challenging terrain [46] (Figure 8). In addition to manual annotation and visual comparison, we conducted a spatial overlap analysis using GIS techniques. This required a comparison of the points of our detected landslides with those in a previously published landslide inventory, which helped to validate the reliability of our detections. Specifically, we compare our detections with the previous 29 landslides from the published PS-InSAR-derived inventory in eastern Karakoram along the KKH [26] (Figure 9). Among the 29 cases, five are located outside our study buffer zone, whereas 23 landslides positively matched our detections, resulting in an approximately 95% match rate. This spatial correlation of InSAR approaches between our detections and previously published landslide inventories reveals a high level of reliability in the distribution pattern of landslides.

5. Discussion

The number of landslide inventories along the KKH has significantly increased over the past few decades, particularly after the official announcement of the CPEC project route. The increase in inventory data has helped in understanding landslide risk and has significantly contributed to the hazard assessment of the region. Previous studies [15,19,20,59,74,75,76] have mentioned the complex combination of geological, hydrological, and anthropogenic factors that triggered landslide activities along the KKH route in northern areas. These landslides cause transportation, traffic congestion, and communication problems, complicating the region’s overall management. In light of the urgent need for better infrastructure management and hazard mitigation along the CPEC alternate route, our study developed a detailed landslide inventory for the newly defined KKH alternate route.

The landsliding phenomenon along the KKH is controlled by various geological, climatic and tectonic factors. The use of multi-temporal InSAR techniques over a long duration considerably enhanced the detection and evaluation of slow-moving landslides in this challenging terrain [19,22,24,77]. However, these studies highlight issues in places with dense vegetation and steep topography, where decorrelation affects displacement measurements, requiring long-wavelength SAR data to enhance their accuracy [19,23]. The geological composition, particularly Quaternary deposits and severely deformed igneous and metamorphic complexes, is essential in slope instability because these units are more prone to weathering activities and structural weakness [17,59,78]. Furthermore, seasonal precipitation, mainly during monsoons, causes debris flows and reactivates pre-existing landslides, resulting in regular infrastructure disruptions [15,74]. The region’s intense seismic activity exacerbates slope failures, with fault zones contributing to significant deformation triggers in the eastern Karakoram along the KKH [26,27,77].

The integration of phase-gradient stacking with LiCSBAS improves the sensitivity and resolution of ground displacement data, resulting in more effective landslide monitoring and risk assessment. This method is beneficial for detecting slow-moving active landslides, where traditional InSAR approaches can be challenging due to decorrelation effects. Phase-gradient stacking enhances vulnerability assessments by refining displacement signals and capturing small deformations, particularly in regions with complex topographies and dynamic environmental conditions.

In this study, phase-gradient stacking integrated with LicSBAS significantly highlights the sensitivity and resolution of ground displacement, particularly for slow-moving active landslides [26,27,77] within the 10 km buffer of the KKH alternate route. The derived landslide density map reflects deep-seated landslide activity, whereas the LiCSBAS-based velocity and time series of three representative cases reflect a highly linear deformation pattern. Such types of deep-seated landslides with higher linear deformation patterns are strongly controlled by tectonic activity, where active faulting causes intense disintegration and deformation of rocks and creates slope instability [78,79,80]. Despite being predominantly controlled by tectonic activity, precipitation, seismic, and anthropogenic activities may have some degree of influence, as found in other regional studies [18,28]. These landslides are scattered along the MKT, Reshun fault, and Tirich Mir fault in the western Karakoram. The KKH route passes through two primary suture zones: the Shyoke Suture Zone (MKT) and the Indus Suture Zone (MMT). The influence of multiple stages of tectonic activity along these two suture zones has resulted in a series of fault systems, joints, and fractures in a region crucial for altering the landscape and terrain slope.

The stacking azimuth and range phase gradients improves the landslide deformation by focusing on the magnitude and extent of the displacement signal. Phase-gradient maps can only reveal the deformed area and do not capture the boundaries of slow-moving, active landslides. By combining gradient maps with a deep learning network, the phase gradient method locates and marks the approximate boundary based on these maps. However, the phase gradient only provides qualitative evidence with no specific details regarding the velocity or temporal changes in deformation.

In the Karakoram region, where towns and villages are limited to hillslope areas due to the lack of plains in the valleys, most settlements are placed on hillslopes and alluvial fans, making them highly prone to landslide hazards. High displacement velocities and instability of active slopes pose a considerable threat to communities residing at the toe of these landslides. The proximity of settlements to these active hillslopes increases their susceptibility to abrupt slope failures, which may result in catastrophic situations for local settlements. By examining these critical hillslopes, this study underscores the serious need to implement mitigation measures and proper monitoring to ensure the safety of local communities.

The landslide inventory constructed in this work has vital implications for landslide risk analysis, management, and future planning for the CPEC route and communities residing in such risky regions. Detecting and mapping landslides along the KKH alternate highway provides valuable information for focusing on mitigation efforts and developing infrastructure strategies for reducing landslide hazards. Regions with a high scattering of landslides require more sophisticated engineering solutions, such as slope stabilization and drainage networks, to protect highways and nearby communities from slope failures.

6. Conclusions

The Karakoram region in northern Pakistan was investigated to develop an active landslide inventory for the KKH-AR part of the CPEC route. This route is essential for the economic partnership between China and Pakistan, which serves as a vital corridor for trade and infrastructure expansion. Our findings show that 1037 active landslides are distributed along the Khunjerab to Chitral CPEC economic route. The obtained inventory reflects the presence of active landslides that potentially affect the highway corridor. The gradient maps and InSAR-derived velocities highlighted the deep-seated nature of the three representative active landslides, all showing velocities exceeding 60 mm/yr. The approach employed in this study provides a promising overview of potential landslides along the KKH-AR. It can be regarded as an effective awareness tool for thoroughly investigating and mitigating hazards and helping administrative authorities enhance infrastructure and transportation safety along highways.