Introduction

The Qilian Mountains and adjacent areas constitute the primary geomorphic unit on the northeastern margin of the Qinghai-Tibet Plateau (NE-QTP). They are important transportation channels for the Belt and Road initiative, with many engineering corridors (ECs)/infrastructures that include railways, highways, and pipelines passing through the permafrost and loess area. In the Qilian Mountains, permafrost covers an area of 8.03 × 104 km2, with an average elevation of over 3400 m (Cheng and Wu 2007; Wu and Zhang 2008; Cao et al. 2019; Peng et al. 2021). This region is characterized by shallow thickness and relatively high temperature (close to 0 ℃) (Wu and Zhang 2008, 2010). Some studies have observed that the increasing trend in temperature and human activities have led to the rapid degradation of permafrost since 2011 in the Qinghai-Tibet Plateau (QTP) (Cao et al. 2018; Costard et al. 2021; Ran et al. 2022). Freeze-thaw-related slope failures are a common type of failure in periglacial landforms, where the soil often slides along the freeze-thaw surface with melted underground ice (Haeberli and Whiteman 2015). Therefore, freeze-thaw-related landslides can be considered a direct indicator of permafrost degradation. These failures typically consist of cracks in the crown, subvertical main scarps, radial cracks in the left and right flanks, and transverse ridges in the foot of the sliding body. They often occur on gentle slopes with few destroyed phenomena on the sliding body and thus are relatively intact on the images but with clear boundaries, which feature circular, elliptical, or strike-like shapes (Highland and Bobrowsky 2008). The triggering factors and mechanisms include high air temperatures, extreme precipitation, and human disturbance (Niu et al. 2016; Chen et al. 2022). Once sliding starts, melting of the exposed ice-rich permafrost can lead to rapid retrogressive failure from the crown and flanks, causing disruption of vegetation cover.

The area adjacent to the southeast Qilian Mountains is the transitional zone between the QTP and the Loess Plateau (LP). Loess covers an area of approximately 7.1 × 104 km2 with a thickness exceeding 50 m (Derbyshire et al. 1998), making it highly vulnerable to the impact of rainfall and human activities (Lade 1992; Xu et al. 2012) due to its fragmented distribution and sparse vegetation (Wang et al. 2018). Landslides related to loess are initially triggered by a partial collapse in the upper part of the loess slope (Wang and Wu 2002; Wang et al. 2020a). The strength of loess is further compromised by irrigation water, rainfall, and earthquakes, leading to enlarged crown cracks at the rear of landslides (Munõz-Castelblanco et al. 2011; Chen et al. 2017; Wang et al. 2019; Peng et al. 2019). Subsequently, the materials begin to slide along a steep surface with a slope gradient of more than 10° (Xu et al. 2011, 2013). The main scarps will regress and gradually develop into a larger landslide, or a secondary landslide will form (Huang 2007). Therefore, loess-related landslides can be characterized by semicircular, rectangular, and compound shapes. They are interconnected in some places, overlap with both the old and the new, and are distributed in clusters. Previous studies have shown that shallow ice can melt due to a rise in temperature in March–May, altering the stress state of the loess (Song et al. 2008; Ni and Shi 2014) and leading to more landslides (Zeng et al. 2016; Zhou et al. 2018). The structure of the loess can be damaged due to increased water infiltration caused by agricultural irrigation water and precipitation in July (Zeng et al. 2016), which can also result in an increase in the occurrence of landslides. Additionally, loess-related landslides can easily be triggered by human activities and extreme rainfall, leading to the destruction of roads, burial of farmlands, and endangerment of lives (Matsuura et al. 2008).

Another region of the study area is covered with seasonally frozen ground. The soil or rock mass is influenced by factors such as earthquakes, rainfall, river erosion, and human activities. It slides down as a whole or scatters along a certain weak surface or weak zone under gravity (Highland and Bobrowsky 2008). The slope ranges between 10 and 30°, exhibiting steep characteristics at the head and foot while being a gentler slope in the middle section (Dini et al. 2019; Wang et al. 2020b; Chang et al. 2021).

With the development of remote sensing earth observation systems, there are numerous new methods and techniques are available for creating landslide inventories. Considering the characteristic spatial, spectral, and temporal resolutions of landslides, the applicability of different remote sensing data (Landsat, SPOT) for landslide recognition was evaluated before 2000 (Brunsden 1985; Soeters and Van Westen 1996). Furthermore, field surveys, aerial photo interpretation, and analysis of historical records are conducted to provide a detailed overview of landslide zoning in EU member states (Van Den Eeckhaut and Hervás, 2012; Rosi et al. 2012). Moreover, internet news, media reports, and online databases can be considered as a constant source of information available in digital format. Researchers have developed a data-mining process to scan the continuous stream of data and identify news related to geohazards (landslides and floods) occurring in Italy (Battistini et al. 2013; Kirschbaum et al. 2015). Meanwhile, advances in interferometric synthetic aperture radar (InSAR) processing algorithms have been able to highlight the fastest motions at a regional scale (Frodella et al. 2014). Recently, the combination of Deep Learning and Sentinel-1 SAR amplitude data is a reliable method for mapping landslides in any weather conditions (Casagli et al. 2017; Bianchini et al. 2018; Rosi et al. 2019; Solari et al. 2020; Nava et al. 2022).

The landslides in the NE-QTP have various types (freeze-thaw and loess-related) and shapes (round, oval, long, semicircular, rectangular, and compound). They exhibit spectral and textural features that do not significantly differ from the surrounding environment in optical remote sensing images. Previous studies (Huang et al. 2020; Yin et al. 2021; Xia et al. 2022; Ju et al. 2022; Zhou et al. 2022) that utilized machine learning and InSAR technology to automatically identify landslides have shown promising results. However, these studies were limited to setting parameters and training models in specific study areas, which led to poor generalization and flexibility in the NE-QTP region. Therefore, high-resolution optical remote sensing imagery (GaoFen-6) and Google satellite imagery were selected for photo-interpretation in this study. We utilized optical images that were revisited multiple times to compensate for their vulnerability to overcast weather and obstruction by vegetation. The accuracy of the manual interpretation is verified through field investigations, which also help identify landslides with early deformation signs that may not be apparent in optical imagery.

This study is aimed at creating a comprehensive inventory of landslides with high accuracy along the infrastructures in the NE-QTP using visual interpretation or photo-interpretation and various supplementary datasets. The spatial distribution patterns, types, evolutionary characteristics, and influence on infrastructures of landslides were analyzed. Meanwhile, the triggering mechanisms of three types of landslides were also analyzed.

Study area

The study area is located on the northeastern margin of the QTP in China within the geographical coordinates of 95.142° E to 105.218° E and 35.195° N to 41.528° N (Fig. 1a, b). The Qilian Mountains, Loess Plateau, and Hexi Corridor are included. Notably, permafrost, seasonally frozen ground, and loess are the main types of ground in the study area (Fig. 1b). The terrain gradually slopes downward from west to east, and the altitude ranges between 5000 and 1500 m. The study area covers approximately 30 × 104 km2, with approximately 1250 km of constructed railways, highways, pipelines, and other national roads. Furthermore, this area serves as a transition zone between the high, cold region of the QTP, and the northwest desert area, characterized by a typical plateau continental climate. The western part is dominated by westerly winds, while the eastern part is influenced by the East Asian and South Asian summer monsoons (Gou et al. 2015). Precipitation is mainly concentrated in the summer, and the annual precipitation gradually decreases from the eastern region to the western region, ranging from 371 to 62 mm/year (Geng et al. 2017). In winter, the precipitation is mainly snowy at high altitudes. The intensity and frequency of seismic activity are high (Fig. 1c), with 29 historical earthquakes of magnitude 6 or above recorded. In recent years, there have been several earthquakes, including the Menyuan M6.4 earthquake in 2016 and the Menyuan M6.9 earthquake in 2022 (Wang et al. 2017; Zhang et al. 2020). Among them, the Menyuan M6.9 earthquake occurred on January 8, 2022, causing severe damage to a section of the railway (Xu et al. 2022). Figure 1d shows a photograph of the landslide taken during a field investigation.

Fig. 1
figure 1

Environmental settings of the study area. a Schematic diagram showing the location of the study area in China. b Coverage of the study area and distribution of ground types. The permafrost (Obu et al. 2019) and seasonally frozen ground, Loess Plateau, and the desert region constitute the background. The locations of our field investigation are indicated by the red triangles. c Elevation of the study area, along with maps showing active faults and earthquakes. Earthquakes with a magnitude greater than 4.5 occurred between 2000 and 2022. d A photograph of the landslide taken during a field investigation (center: 100.196° E, 38.137° N)

Full size image

Material

Many data sources have been used to identify and detect landslides. The parameters and sources of each type of data are listed in detail in Table 1.

Table 1 List of datasets used in landslide mapping and analysis
Full size table

Optical imagery. A total of 144 scenes of GaoFen-6 (GF-6) satellite imagery, with a spatial resolution of 2 m, were acquired from 2021 to 2022 for interpreting landslides in the study area. Satellite imagery from Google Earth and Esri World Imagery was also utilized as supplementary reference data. UAV images and photographs were captured using the DJI Inspire 2 multispectral flight platform at 11 sites located near roads. The resolution of the images and photographs ranged from approximately 40 to 100 cm.

Digital elevation model (DEM). We collected 23 scenes of DEM data from ALOS-PALSAR to calculate elevation, slope, and aspect. The spatial resolution of the DEM data was 12.5 m.

Normalized difference vegetation index (NDVI). The NDVI data at a 30 m resolution published by Yang et al. (2019) were selected, and the threshold range was set at [− 0.2, 0.84]. It is used to analyze the extent of vegetation cover where the landslide occurred. Higher values indicate greater vegetation coverage.

Other available datasets. Permafrost, seasonally frozen ground, and loess maps have been harvested. Data on active faults, earthquakes, loess thickness map, and soil moisture have been downloaded. For the precipitation and air temperature, the annual average point data recorded by 48 meteorological stations in the study area since 2000 were utilized.

Methods

Photo-interpretation for landslide mapping and spatial analysis

Photo-interpretation is used to identify landslides in optical satellite imagery (GF-6) based on four characteristics: (1) exposed accumulative material with a sliding/flowing direction, (2) well-defined curved head scarps, (3) sparse vegetation, and (4) variation in color compared to the surrounding area (Fig. 2). When the image is affected by atmospheric conditions, other images from the same period or auxiliary data from the same location will be selected for reconfirmation to eliminate the influence of clouds. All visible landslides on georeferenced imagery were digitized into polygons using QGIS. When a landslide is observed in an image in 1 year, it is assumed to have persisted in subsequent years. Panchromatic photographs provide a high spatial resolution, equal to 2 m, which allows for inventorying only landslides larger than 1000 m2. Specifically, we used a fishnet with a mesh size of 1.5 km × 1.5 km to divide the satellite imagery of the study area for the purpose of quantifying landslides, mesh by mesh.

Fig. 2
figure 2

The GF-6 imagery displays examples of landslides with identified features for photo-interpretation. a, b, and c represent freeze-thaw-related landslides, loess-related landslides, and general landslides, respectively. The FIDs are assigned in the inventory using QGIS. The white polygons highlight the boundary changes in landslides in the ESRI World Imagery, Google Earth, and UAV imagery from 2020 to 2022. To minimize these errors, Esri Wayback Imagery, Google Earth, and UAV imagery with a spatial resolution of less than 1 m were utilized to assist in visual interpretation and ensure accurate results

Full size image

To quantify the spatial distribution characteristics of landslides, the landslide area percentage (LAP) has been calculated using a regular grid of cells measuring 20 km × 20 km. Subsequently, the extent of the infrastructure affected by landslides was ultimately determined by calculating the proportion of the landslide area in each grid. LAP in each grid is categorized into five types: 0–10%, 10–20%, 20–30%, 30–60%, and > 60%. Further analysis is conducted on the grids where the proportion of landslides exceeds 60% in consecutive multiple grids. An influence area around the infrastructure of 5 km and 20 km is established, and the size of the landslide area and their distance from the infrastructure within the buffer zone are recorded.

Landslide classification and statistical analysis

Many case studies have classified landslide types according to the material composition, formation mechanisms, age, sliding velocity, and thickness of sliding (Borrelli and Gullà, 2017; Li and Mo 2019; Pourghasemi et al. 2020). In this study, based on the ground types and triggering factors, we distinguish three types of landslides: freeze-thaw landslides (FTLs), mainly triggered by the freeze-thaw cycle in the permafrost region; loess landslides (LLs), caused by rainfall and irrigation in the loess region; and general landslides (GLs), influenced by multiple triggers, such as earthquakes, extreme rainfall, and freeze-thaw cycles in the seasonally frozen ground region (Varnes 1978; Cruden and Varnes 1996; Highland and Bobrowsky 2008; Hungr et al. 2014). Additionally, we used DEM data to analyze the area, elevation, slope, and aspect of the three types of landslides by reclassifying them in QGIS to identify their spatial distribution patterns. Moreover, the characteristics of the landslides were analyzed by overlaying the data on NDVI.

Landslides triggering mechanism analysis

The selected environmental factors consist of natural factors (such as precipitation, temperature, distance from rivers, and proximity to earthquakes) and human activity factors near infrastructures. The geographic detector was utilized to determine the influential environmental factors associated with different types of landslides (Wang et al. 2019). The q-statistic was employed to evaluate the spatial relationship between deformation and relevant controlling variables, which varies between 0 and 1, indicates the extent to which a factor influences the spatial distribution of deformation.

The Kriging interpolation method was employed to model surface data for precipitation and air temperature. Additionally, the study utilizes the near analysis method to determine the Euclidean distance between the landslide event and three additional environmental variables: proximity to the river, distance from seismic activity, and distance from infrastructure. The seismic data considered in the analysis encompasses earthquakes of magnitude 4.5 or higher occurring between 2000 and 2021.

Accuracy evaluation of inventory

The accuracy of the landslide inventory is evaluated through field investigations, complemented by Google satellite images and ESRI World Imagery. Field investigations were carried out in the winter of 2020, as well as in the spring and summer of 2021, and the summer and fall of 2022. Landslides that align with field survey findings are marked on the Ovital Map (Version 9.7.1). Newly discovered landslides, known as uninterpreted landslides, are pinpointed and labelled with GPS coordinates (latitude, longitude, and altitude). They are then documented using digital photography and UAV photogrammetry. Given the high occurrence of landslides in the study area, we selected four regions for accuracy assessment that are easily accessible for field investigation. We defined a landslide as a true positive (TP) when the interpretation results match the field investigations. The true positive rate is calculated as follows:

$$\text{TP }= {N}_{tp}/N \times 100{\%}$$
(1)

A false negative (FN) occurs when a landslide is found by field investigation but not interpreted. The false negative rate is:

$$\text{FN }= {N}_{fn}/N \times 100{\%}$$
(2)

A false positive (FP) is when a landslide is interpreted but not found in the field investigation. The false positive rate is:

$$\text{FP }= {N}_{fp}/N \times 100{\%}$$
(3)

where N represents the number of landslides (Zhang et al. 2019).

Results

Landslide inventory and spatial characteristics

In the study area, 11,914 landslides have been mapped, totaling 720.12 km2 (Fig. 3). The number of landslides gradually increases from the west to the east of the study area. They appear to be more densely distributed in the east and sparsely distributed in the west. There are more landslides occurring in the south than in the north (see histogram in Fig. 3). Additionally, most of the infrastructures are flanked by landslides.

Fig. 3
figure 3

The spatial distribution of landslides in the NE-QTP. The legend for the infrastructures and ground types is shown in Fig. 1. The concentration of landslides is depicted through histograms showing the frequency (%) of latitudinal and longitudinal distribution, along with their corresponding curve fits

Full size image

Figure 4a illustrates that the grids with the largest LAP (0.3–2.4%) are primarily situated in the southeastern part of the study area in patchy formations, with sporadic distributions in the central and western regions. Over 32% of the grids exhibit LAP values ranging from 0.3 to 2.4%, indicating that the infrastructures passing through these grids are located in areas significantly impacted by landslides. Additionally, we define the region where the LAP accounts for more than 0.1% per grid as the landslide-affected area with infrastructures passing through. We deduce that around 44.85% of the infrastructures are influenced by landslides. Notably, highways and other roads (mainly national roads) are the most affected, with 53.96% and 53.28%, respectively. Railways followed with 44.66%. Pipelines were the least affected, at 27.5% (Fig. 4b).

Fig. 4
figure 4

Distribution of landslide area percentage (LAP) and its influence on infrastructures. a The legend for the infrastructures is shown in Fig. 1. Colored grids represent the areas affected by landslides and the LAP on the 20 km × 20 km grid. The red boxes indicate the locations in Fig. 5. b A comparison between the total length of the four engineering corridors in the study area and the length of these corridors affected by landslides is presented on the left, and the ratio of the length of engineering corridors affected by landslides to their total lengths is shown on the right

Full size image

In order to further explore the landslide spatial distribution characteristics of landslides around infrastructures within 5 km and 20 km buffer zones (Fig. 5a), including the size of the landslide area and their distance from infrastructure, we selected the zones most severely affected by landslides for analysis. Within the zone of less than 5 km from the infrastructures, most of the landslide areas are less than 113 ha, concentrated between 50 and 100 ha. The landslide area in the 5–20 km buffer zone is smaller than that in the 5 km buffer zone, and they are concentrated between 50 and 100 ha (Fig. 5b). Furthermore, to clarify the distance between the landslide and the infrastructures, we analyzed the proximity of the landslide to four types of infrastructures. The results indicated that the average minimum distance between the infrastructures and the landslide was approximately 3000 m. Among them, the closest landslide to the railway is less than 2000 m away (Fig. 5c).

Fig. 5
figure 5

The regional distribution of landslides along the infrastructures. a The region with the highest density of landslides was chosen to demonstrate the distribution characteristics of landslides along the infrastructures within 20 km and 5 km. The size of the circle represents the area of the landslide. Orange circles indicate landslides within 5 km of the infrastructures, while the blue circles indicate landslides within 20 km. b Area statistics of landslides distributed within 5 km and 20 km of the infrastructures. c Box plot showing the distance between the landslide and the infrastructure in the 5 km influence area. The positions are indicated in Fig. 4a. The legend for the infrastructures is displayed in Fig. 1

Full size image

Classification of landslides and their characteristics

Three types of landslides have been classified and harvested as follows: 7930 LLs (66.5%), 2889 GLs (24.2%), and 1095 FTLs (9.2%). FTLs are mainly found in the transitional areas between permafrost and seasonally frozen ground, with fewer occurrences in the permafrost interior, displaying a linear distribution pattern as depicted in Fig. 6a. Analysis of the elevation, slope, and aspect of these landslides revealed that the majority of FTLs are distributed at relatively higher altitudes of 3400–3800 m (Fig. 6b). Figure 6c shows that the slopes of the FTLs are mostly between 7 and 25° with a very gentle slope. Particularly, between 10 and 20°, nearly 60% of landslides are distributed. In terms of aspect, FTLs are predominantly concentrated on the northern slopes, particularly in the northeast and northwest directions, with fewer instances on the southern slopes (Fig. 6d). Furthermore, the NDVI index indicates two distinct peaks, ranging from 0.05 to 0.2 and 0.3 to 0.5, respectively.

Fig. 6
figure 6

a The spatial distribution of freeze-thaw landslides along the infrastructures. Statistical summaries of the distribution patterns and characteristics of landslides. b–e The histograms and radar charts show the frequency distribution statistics of FTLs on elevation, slope, aspect, and NDVI, respectively

Full size image

The LLs are characterized by a clustered distribution with planar shapes, as depicted in Fig. 7a. Their spatial distribution density surpasses that of FTLs, with numerous infrastructures densely situated in this region being impacted by landslides. The Yellow River and the Taohe River traverse through this area. Additionally, the majority of LLs are found at altitudes ranging from 1800 to 2100 m, as illustrated in Fig. 7b. Figure 7c shows that the slopes of the LLs are mostly between 15° and 30°. In particular, between 20° and 30°, nearly 50% of landslides are distributed. Regarding landslides’ aspect, they are primarily concentrated on the south slope, specifically in the west and southwest directions, while few occur on the northeast (Fig. 7d). Furthermore, NDVI datasets indicate that the NDVI index varies between 0.2 and 0.3, as demonstrated in Fig. 7e.

Fig. 7
figure 7

a The spatial distribution of loess landslides along the infrastructures. Statistical summaries of the distribution patterns and characteristics of landslides. b–e The histograms and radar charts represent the frequency statistics of LLs on elevation, slope, aspect, and NDVI, respectively

Full size image

A significant feature is observed between 103.5° E and 104° E in LLs: a “landslide empty area” with a width of approximately 50–60 km and a length of about 150 km. This region is narrow in the middle and widens towards the north and the south. The morphological structure is depicted in Fig. 8a. To gain a better understanding of the topography of this region, we conducted a detailed analysis of the elevation distribution. Our investigation revealed that the “landslide empty area” exhibits low altitudes and relatively gentle terrain, as shown by three profile lines in Fig. 8b. A comparison of the slope distribution between the loess landslide area and the “landslide empty area” is presented in Fig. 8c and d, highlighting a distinct contrast in slope distribution between the two regions.

Fig. 8
figure 8

Landslide empty area in LLs’ distribution area. a Map showing the elevation of this area, with black triangles indicating the altitude of some of the peaks. Three profile lines along the LLs’ distribution area and the landslide empty area. b The surface elevation across the landslide empty area along the profile lines. c Slope distribution map of this area. d Slope distribution statistics of the landslide area and the landslide empty area

Full size image

Furthermore, a comprehensive examination of the thickness of loess was carried out. Through statistical analysis, it was determined that the loess thickness within the area prone to loess landslides was notably greater, with the most prevalent thickness falling between 100 and 200 m. In contrast, the distribution of loess in the region designated as the “landslide empty area” was found to be less dense (refer to Fig. 9b). Additionally, an assessment of soil moisture distribution was conducted. As depicted in Fig. 9d, the soil moisture levels in the “landslide empty area” were relatively low, primarily concentrated within the range of 0.06 to 0.08. Conversely, the soil moisture content in the loess landslide area was relatively high, typically falling between 0.09 and 0.11.

Fig. 9
figure 9

a Distribution map of the thickness of loess in this area. b The distribution of loess thickness in the landslide area and the landslide empty area. c Soil moisture map in this area. d The distribution of soil moisture in the landslide area and the landslide empty area

Full size image

The spatial distribution of GLs is characterized by both freeze-thaw landslides and loess landslides (Fig. 10a). Most of the GLs are distributed at an altitude of approximately 2400–2600 m, situated between the FTLs and the LLs (Fig. 10b). Figure 10c shows that the slopes of GLs are mostly in the range of 20–30° with a relatively steeper slope compared to the FTLs. The GLs are primarily concentrated on the northern and southern slopes, particularly in the northeast and southwest directions (Fig. 10d). They exhibit common characteristics of LLs and FTLs in the aspect. Additionally, the NDVI index is moderate, mainly dominant in the range of 0.2–0.35 (Fig. 10e). In summary, the characteristics of GLs in terms of altitude, slope, aspect, and NDVI are between FTLs and LLs.

Fig. 10
figure 10

a The spatial distribution of general landslides along the infrastructures. Statistical summaries of the distribution patterns and characteristics of landslides. b–e: The histograms and radar charts represent the frequency distribution statistics of GLs on elevation, slope, aspect, and NDVI, respectively

Full size image

Landslides triggering mechanisms

The results calculated by the geographic detector indicate that the primary environmental factor influencing FTLs is the temperature (Fig. 11a). Additionally, the distance from the river and precipitation also play a significant role. Figure 12b demonstrates that FTLs are frequently located in areas where there is a transition from lower to higher temperatures. Figure 12a, c depict the relationship between FTLs and rivers, as well as FTLs and precipitation. FTLs are commonly found near river pathways and are more prevalent in regions with higher levels of precipitation.

Fig. 11
figure 11

The q-statistic of environmental factors on controlling three types of landslides occurred. a is the q-statistic of FTLs, b is for the LLs, and c is for the GLs

Full size image
Fig. 12
figure 12

a Maps of landslides, infrastructures, rivers, earthquakes, and partial seismogenic faults in the study area. Earthquakes have occurred between 2000 and 2021 with a magnitude greater than 4.5. The size of the circle represents the magnitude of the earthquake. b, c The air temperature interpolation results are obtained based on temperature data recorded from 48 spatially evenly distributed meteorological stations in the study area since 2000. The correlation analysis between the interpolation results and the observed values was conducted, revealing a correlation coefficient of R2 = 0.92, indicating a strong correlation. d, e The same method was used to obtain the precipitation interpolation results for the study area. The correlation coefficient between the interpolation results and the observed values was R2 = 0.92, which also showed a good correlation

Full size image

The main environmental factors controlling LLs are related to water (rivers and precipitation) (Fig. 11b). The region where landslides are prevalent is intersected by numerous major rivers, as depicted in Fig. 12a. Additionally, the average level of precipitation in this area is relatively high, as illustrated in Fig. 12d.

The main environmental factor influencing GLs is seismic activity, particularly earthquakes, as depicted in Fig. 11c. This is evidenced by the concentrated presence of earthquakes and seismic faults in the region, as shown in Fig. 12a. In contrast to FTLs and LLs, the environmental conditions for GLs differ. The q-statistic analysis revealed that other factors such as human engineering activities, air temperature, precipitation, and proximity to rivers also play a significant role in influencing GL occurrences. The combined impact of these various environmental factors contributes to the formation of GLs, as they are typically located in the transitional zone between the QTP and the LP.

Inventory accuracy

Four regions were selected from the study area (Fig. 13a) for accuracy evaluation analysis, utilizing the defined metrics of TP, FN, and FP. The results show that there is a false negative for landslides (FID5) but no false positives. The rate of TP was 98.8%, the rate of FN was 1.2%, and the rate of FP was 0 (Fig. 13b). From the UAV pictures, we observed that the crown cracks of the landslide (FID5) are being developed. However, the main scarp is still not clearly visible in the imagery, and the sliding deformation of the main body is not obvious. In region (c), there is 1 false negative landslide and no false positives (Fig. 13c). The TP rate was 98.8%, the FN rate was 1.2%, and the FP rate was 0. The subfigures show UAV pictures of landslides taken during the 2022 field investigations. In region (d), there are 2 false negatives (FID726 and FID739) and no false positives (Fig. 13d). The rate of TP is 98.1%, the rate of FN is 1.9%, and the rate of FP is 0. The UAV pictures show that the crown cracks of landslides (FID726) and landslides (FID739) are developing, but the main scarps and sliding features of the main body are not obvious, thus causing misinterpretations. There were no false negatives or false positives in region (e), and the rate of TP was 100% (Fig. 13e). The deformation characteristics of landslides are more pronounced, as can be observed from the UAV photographs. These include subvertical main scarps, sliding main bodies, and accumulation features/ridges at the foot of the sliding body. The average TP rate of the four regions is greater than 95%.

Fig. 13
figure 13

Four regions were selected for an evaluation of inventory accuracy. TP represents true positive; FN represents false negative; FP represents false positive. a Locations of the selected four regions. b, c, d, and e represent the spatial distribution of landslides in the four regions and their UAV images, respectively

Full size image

Discussions

Comparison with previous landslide inventory

Our inventory is the first complete one in the entire infrastructure region of NE-QTP. Previous studies have not been able to compile a comprehensive inventory of landslides in this crucial area due to difficulties in accessing landslides in remote and harsh permafrost and loess regions, as well as limitations in automatically mapping them using remote sensing imagery (such as poor generalization). Some studies have identified retrogressive thaw slumps (RTSs) in specific subregions of the Qilian Mountains and loess-related landslides in the Heifangtai region. For example, Jiang (2019), Mu et al. (2020), and Gao et al. (2021) used satellite imagery to extract RTSs in the Qilian Mountains and eastern Tibetan Plateau from 1969 to 2019. Liu et al. (2020) utilized InSAR, unmanned aerial vehicle (UAV), and Google imagery data to identify landslide distributions in the Heifangtai region during 2007–2011 and 2015–2018. In comparison to existing landslide datasets in the subregions, our inventory stands out for its comprehensiveness, originality, and being open source. We not only identified landslide locations, provided information on areas, elevations, slope, aspect, NDVI, and LAP for infrastructures but also categorized landslides into different types, analyzed their triggers, and assessed the accuracy of the inventory based on multiple field investigations to offer a more precise inventory.

Inventory applications

This inventory presents a valuable dataset for quantifying the assessment of landslides affecting infrastructures exposed to climate change, seismic activity, and human interventions. Understanding the effects of landslides in these areas is essential for informing policymakers about infrastructure safety. The inventory also supports the validation of automated landslide identification models, aiding in the development of a fully automated method for detecting landslides in the NE-QTP region. Various potential applications of this landslide inventory have been identified by researchers (Guo et al. 2021; Nguyen and Kim 2021; Wang et al. 2021), and it is expected that the research community will explore these uses over time. Importantly, this inventory will assist researchers in quickly identifying potential locations for studying individual landslides. Road managers in Northwest China should also pay attention to this inventory, as the size and location of these landslides could be crucial in preventing hazards in road operations. Additionally, by analyzing historical landslides’ triggering factors, mechanisms, and spatial distribution characteristics, a landslide prediction model has been created to forecast potential landslide events in the future.

Conclusions

We have visually interpreted 11,914 landslides along the infrastructures in the NE-QTP of China using GF-6 images (2 m resolution) captured from 2021 to 2022. The spatial distribution of landslides and their impact on infrastructures have been analyzed. Furthermore, these landslides were classified into FTLs, LLs, and GLs based on the ground type and triggering factors. The main conclusions drawn are as follows:

  1. 1.

    The number of landslides gradually increases from the west to the east of the study area. Approximately 44.85% of the infrastructures are affected by landslides.

  2. 2.

    FTLs are mainly distributed in the boundary regions between permafrost and seasonally frozen ground. They are less common in the permafrost hinterland and exhibit strip distribution characteristics. LLs exhibit characteristics of clustered distribution with planar shapes and higher spatial distribution density. GLs have significant transitional characteristics and commonalities between FTLs and LLs.

  3. 3.

    Some predisposing factors have been identified through the analysis of the triggering mechanisms of landslide events. In this study, temperature, proximity to rivers, and seismic activity are among the factors that influence the geomorphology of the area.