The study area is located in the permafrost region of the Qinghai-Tibet Plateau (QTP) (Ran et al., 2012). It extends from Xidatan in the north to Naqu in the south, covering a length of approximately 550 km and an area of 8 × 10⁴ km². Since 1950, the region has experienced significant warming (Li et al., 2023), resulting in permafrost degradation and a reduction in vegetation cover and species (Jin et al., 2020). These changes have contributed to increasing slope instability in the area (Jin et al., 2008; Wu et al., 2012).

The study area, primarily characterized by alpine meadows and alpine steppe, exhibits pronounced differences in color and texture in landslide-affected regions compared to the surrounding undisturbed landforms. Additionally, the progressive expansion traits of permafrost-related landslides can be distinctly identified through the analysis of multi-temporal high-resolution satellite imagery. The high-resolution images used are Planet images (https://www.planet.com/) of every July from 2016 to 2022 with a resolution of 3 m, which completely cover the study area. July was chosen as it is the ground surface of the study area was completely free of snow cover compared to May and June. Despite being time-consuming compared to automated identification, manual detection offers high accuracy and enables the addition of attributes during the identification process. In addition to assessing the whole study area, a typical landslide-prone sub-area was chosen for detailed spatiotemporal assessment (Figure 1). In addition to delineating landslide boundaries, a critical component of image interpretation involves categorizing landslide types. Landslides identified for the first time in a given image, when compared to the preceding temporal image, are classified as new landslides. Active landslides are those that show expansion and an increase in area relative to the most recent image (2022). It is noteworthy that for the earliest available image, such as the 2016 dataset for the entire study area, no prior image exists for comparative analysis. Consequently, landslides identified in this initial image are designated as pre-existing landslides without further classification. These pre-existing landslides serve as a foundational dataset for tracking and analyzing landslide occurrences and expansion in subsequent time periods.

To ensure the accuracy of the landslide inventory, field surveys were conducted in the study area in April and October 2023. Due to the largely uninhabited nature and limited accessibility of the region, surveys were restricted to a 3-km corridor on either side of the QTEC. On-site verification successfully confirmed all 5 landslides previously identified in satellite imagery. For instance, the Gushan landslide (Figures 1C, D) was initially identified in satellite imagery and subsequently confirmed through field verification. An unmanned aerial vehicle (UAV) was utilized to capture high-resolution aerial photographs of representative landslides, which were then processed to generate a digital orthophoto map (DOM) and a digital surface model (DSM) of these landslides. A comparative analysis of landslide areas derived from UAV and satellite imagery for selected landslides indicated that the area discrepancy between the two sources was less than 5%. This variance could be attributed to differences in the resolution of the imagery sources, or it may reflect ongoing landslide expansion between 2022 and 2023.

The air temperature and ground temperature data were provided by the National Tibetan Plateau Data Center (https://data.tpdc.ac.cn/). The meteorological stations indices include daily air temperature, precipitation, humidity, wind speed gradient observation, and radiation balance. However, it is important to note that certain data points may be missing due to variations in the construction time of each station or other factors. The daily ground temperature of the active layer observations was measured at different depths below the permafrost table from the ground surface between 2002 and 2018. The precipitation data was obtained from the China Meteorological Data Service Centre (http://www.cma.gov.cn/). There are 3 precipitation stations evenly distributed in the northern, middle, and southern parts along the QTEC.

To investigate the geological characteristics of the active layer, we excavated a 2-m-deep observation trench at the headwall of the Zilrama landslide (Figure 2A). To obtain undisturbed soil samples, the trench was positioned to avoid areas with pronounced fissures at the headwall and maintained a minimum distance of 0.5 m from any cracks. As shown in Figure 2B, the active layer, excluding the top organic-rich layer, was divided into three distinct sub-layers. Soil samples from each sub-layer were carefully collected using a ring cutter for laboratory testing of soil physical properties (Figure 2C). Additionally, approximately 50 kg of soil from each layer was gathered to prepare artificial samples for mechanical properties testing. Table 1 presents the physical properties of the soil layers, measured through a variety of tests, including sieving, Malvern particle size analysis, moisture content, specific gravity, and liquid and plastic limit tests. Figure 3 shows the grain size distribution curves. The results indicate that the active layer soil in permafrost landslide areas predominantly comprises fine-grained material, with particle size decreasing closer to the permafrost table. Field observations revealed substantial pore ice within the frozen soil of the active layer, while moisture content tests confirmed saturation of the soil in this layer (Figures 2D, E).

The landslide headwall height is approximately 2 m; hence, soil from the third layer was selected for low-temperature triaxial testing to evaluate the mechanical properties. Testing was conducted using a thermo-mechanical triaxial compression apparatus (UL200/60 SH) manufactured by Wille, Germany, with a maximum axial load capacity of 60 kN, confining pressure up to 4 MPa, and temperature control from −30°C to 60°C with 0.1°C precision (Figure 4). Sample dimensions were 70 mm in diameter and 140 mm in height. To ensure sample consistency, particle size was restricted to below one-tenth of the sample diameter. Particles exceeding 10 mm, comprising 1% of the sample mass, were excluded without impacting the mechanical analysis. Before preparation, soil samples were dried at 105°C for 24 h. The target dry density was set to 1.7 g/cm³, matching field values, with a 10% moisture content achieved by combining 916 g of soil particles with 91.6 g of water. The mixture was sealed and allowed to equilibrate for 24 h (Figure 5A). To ensure uniformity, the mixture was divided into 28 mm segments and compacted in 5 layers, with each layer compacted to a designated height and the surface leveled before adding the next layer. Once prepared (Figure 5B), the specimen was wrapped in a latex membrane and mounted onto the triaxial apparatus (Figure 5C). Saturation was achieved through back-pressure saturation, with gradual increases in confining and back pressure over 6 h to prevent soil structural damage. A pore water pressure coefficient (B-value) exceeding 0.95 indicated full saturation (Figure 5D). The temperature control system then lowered the coolant to the target level, maintaining it for 6 h to ensure thermal equilibrium within the sample. An unconsolidated undrained (UU) shear test was conducted to assess temperature effects on soil strength under shallow landslide conditions. The confining pressure was set to 50 kPa, and the temperature maintained between −6°C and 0°C. Loading was applied at a constant strain rate of 0.7 mm/min, ceasing at 20% axial strain. After testing, the failure characteristics were documented, and the sample was weighed, oven-dried, and reweighed to determine post-saturation moisture content.

The air temperature in the permafrost region of the central QTP has consistently risen from 2000 to 2020, as indicated by all meteorological stations (Figure 6A). On average, the temperature has been increasing at a rate of 0.5°C per decade. Among these stations, the NQ station, located at the lowest altitude of 4452 m above sea level, has experienced the fastest increase rate at an annual average rate of 0.7°C per decade (Table 2). On the other hand, the TGL station, situated at the highest altitude of 5,044 m above sea level, has observed the slowest increase rate, with an annual average rising rate of 0.3°C per decade. The annual change patterns are highly consistent across all meteorological stations. It is evident that the MAAT in 2016 was considerably higher compared to the historical average (Figure 6A). During 2010–2020, The MAAT reached its peak value in 2016, followed by a downward trend then. The MAAT from the four meteorological stations in 2016 increased by 0.59°C compared to 2015, surpassing the average rising rate observed in recent decades. This alignment with global climate change is in line with the fact that 2016 was recorded as the hottest year in some regions of the planet, partly due to a powerful El Niño (Witze, 2023). On the other side, the data from two northern and central stations along the QTEC indicate that the rainfall in 2016 was even lower than the average annual precipitation (Figure 6B). Interestingly, although the rainfall at the southern precipitation monitoring station in the study area increased significantly in 2016 compared to 2015, there were no significant occurrences of landslides in the southern study area. Furthermore, the precipitation at the northern station in the study area showed a significant increase after 2016, while the precipitation at the central and southern stations remained relatively stable after that year.

The thermal state of permafrost is primarily controlled by climate, with influences from land surface characteristics such as vegetation, snow cover, soil physical properties, and geological conditions including organic-layer thickness and ground ice content (Smith et al., 2022). When these surface and subsurface conditions remain stable over a short period of time, the thermal state of shallow permafrost is mainly determined by air temperature. During 2011–2018, the maximum mean monthly air temperature and the maximum mean monthly ground temperature exhibit an identical varying pattern (Figure 6C). Annually, the highest air temperature is observed in August, which corresponds to the peak ground temperature in the shallower layers. However, in the deeper layers, the maximum mean monthly ground temperature has a delayed response to air temperature, which emerged until October or November. This indicates that air temperature is the primary factor influencing the thermal state of permafrost in permafrost regions. The temperature of the near-surface ground closely follows the annual cycle of air temperature, while the temperature variation decreases with depth.

The range of temperature at different depths in various observations varies greatly. However, the mean annual ground temperature (MAGT) consistently showed an upward trend, except for the deep ground of CH06 (Figure 6D). On average, the surface ground (10 cm or 15 cm below the surface) at all observations experienced a rising rate of 0.87°C per decade, which is higher than the rate of air temperature increase. The fastest rate of surface ground temperature rise was observed at QT09, with a rate of 2°C per decade. On the other hand, the average rate of temperature increase near the permafrost table across all observations was 0.40°C per decade, which is lower than that of the surface ground. The highest rate of increase near the permafrost table was still observed at the QT09 site, with a rate of 0.9°C per decade. This can be attributed to the fact that QT09 is located at the edge of the continuous permafrost region. Table 3 presents the average rate of increase at different depths for each observation, as well as the magnitude of temperature rise in 2016 compared to 2015. Similar to the air temperature, the mean annual surface ground temperature in 2016 rose sharply by 0.41°C per year compared to 2015, which was significantly higher than the average rate of surface temperature increase.

The stress-strain curve is a widely used method for assessing soil deformation and strength. In the permafrost region of the Qinghai-Tibet Plateau, the failure behavior of saturated clay undergoes significant changes with temperature under low confining pressure. At soil temperatures of −3°C or lower, the soil exhibits pronounced brittle failure. On the other hand, at temperatures higher than −2°C, the soil shows significant plastic damage (Figure 7A). When the soil temperature is approximately −2°C, the transition from brittle to plastic failure occurs. In the case of plastic failure, the deformation increases with the applied stress. Therefore, the deviator stress corresponding to a strain of 20% is considered the strength or failure stress of the sample. At soil temperatures below −1.6°C, the soil strength exceeds 2 MPa and increases as the temperature decreases. However, when the soil temperature is slightly higher than −1.6°C but still below −1.0°C, the strength rapidly decreases to less than 500 kPa. As the soil temperature continues to rise to 0°C, the strength experiences a slight decrease to 200 kPa (Figure 7B).

The Figure 7C illustrates the variation of soil strength in the negative temperature range under low confining pressure, which can be divided into three zones. In the first zone, as the temperature rises, the strength gradually decreases. The second zone is characterized by a sharp decrease in strength as the temperature continues to increase. In the third zone, as the temperature approaches 0°C, the strength decreases at a slower rate. Typically, 0°C is considered the standard for defining frozen soil from a thermal state (Dobinski, 2011). However, from the viewpoint of frozen soil physics and mechanics, the definition of freezing temperature T_f should be lower than 0°C. For clay in the permafrost region of the Qinghai-Tibet Plateau, a standard of −1°C should be set. At this temperature, the failure mode is already similar to that at normal temperature, and the strength also decreases significantly, leading to a sharp decline in slope stability. This change in macroscopic physical and mechanical properties is attributed to the reduction in pore ice or the increase in unfrozen water content during the soil thawing process.

The active layer is a ground layer in permafrost regions that undergoes annual thawing and freezing, with its thickness (ALT) being an essential characteristic of permafrost environments (Dobiński, 2020). The freeze-thaw index is an important parameter associated with permafrost. It serves as a significant indicator for assessing the current state and long-term fluctuations of permafrost, as well as the spatial and temporal variations in active layer thickness (Riseborough, 2003; Smith et al., 2009). Thawing index refers to the cumulative value of positive temperature over time (Wu et al., 2015). ALT can be determined by applying a modified version of the Stefan solution that incorporates the thawing index (Nelson and Outcalt, 1987; Nixon and McRoberts, 1973). X = 416 k L TISG where X is ALT (m), and TISG is the annual thawing index of surface ground (°C·d) , k is the diffusivity of soil [cm²/s]; L is the volumetric latent heat of the soil [Cal/cm³]. ALT is an increasing function of TISG, and with the increase of TISG, ALT will also show an upward trend.

It is worth noting that the thawing index of surface ground may vary across different active layer observations. The average value of the thawing index at QT09 is 601.02°C·d, while at QT05 it is as high as 1,494.89°C·d. Although there is significant variation in the specific values of all the observations, the rising and falling trends of the thawing index in different years are generally consistent. The thawing index of the surface ground (TISG) in all active layer observations exhibited an upward trend. The index experienced a substantial increase and nearly reached its highest point in 2016, after which it began to decline in the following years (Figure 8A). Consequently, there was a significant deepening of the active layer thickness in 2016, nearly reaching the maximum value observed in recent years.

ALT is also estimated by using linear interpolation of ground temperatures above and below 0°C, calculated from semimonthly temperature data obtained through daily ground temperature measurements (Frauenfeld et al., 2004; Wu and Zhang, 2010b). ALT along the QTEC has shown a significant upward trend in the past 10 years. The average thickness of the active layer, based on all observations, is 197 cm. The maximum thickness of 309.5 cm was recorded at QT05, while the minimum thickness of 113.7 cm was observed at Ch04. The average rate of increase in the active layer thickness at the four stations was 1.5 cm/year, with slight variations among the different stations. In 2016, there was a significant increase in the thickness of the active layer at each station, which can be attributed to a considerable rise in temperature along the QTEC during that year. In comparison to 2015, the thickness of the active layer at all stations increased by an average of 18.5 cm in 2016, surpassing the average rate of increase (Figure 8B). Notably, Ch04 experienced an increase of 42.83 cm in 2016, reaching the highest value observed in recent decades. Interestingly, measurements from the Circumpolar Active Layer Monitoring network in 2016 indicated that the ALT at all Arctic sites was either at or close to the maximum level recorded in the past 18–21 years (Zhao et al., 2020).

The K3035W landslide (Figure 9A) occurred in late September 2010 (Niu et al., 2015). The QT03 observation, located 20 km away from the landslide, indicated that the active layer thickness (ALT) reached its highest value in 2010 (Figure 9B), with the deepest thawing occurring in late September (Figure 9C). A Google Earth image from October, 2010, clearly shows intact wheel ruts behind and on the sliding area of the landslide (Figure 9A), providing evidence that the landslide moved as a whole. This type of landslide is known as active-layer-detachment and is caused by the thawing of materials near the permafrost table. Active-layer-detachment typically occurs during late summer when the active layer reaches its greatest depth. The headwall of the landslide continued to retreat from 2011 to 2013 (Niu et al., 2015). Following the active-layer-detachment, the failure progressed into a retrogressive thaw slump due to the thawing of ice-rich permafrost or the exposure of ground ice.

The failures mainly initiated or expanded during the thawing season between August and November (Luo et al., 2022), so we collected satellite images of every July in the study area. Figure 10A illustrates that a single landslide took place in July 2016, suggesting that the landslide occurred prior to 2016. In July 2017, 6 landslides transpired within the same region, indicating that 5 new landslides emerged between August and October of the preceding warm season in 2016 (Figure 10B). The combined area of these 6 landslides in July 2017 amounted to 266,141 m². Over the subsequent years, the landslides continued to expand, resulting in an area of 430,189 m² by July 2022 (Figures 10C, D). This signifies a 61.6% increase in landslide area over the course of 5 years. Just like the K3035W, the typical landslides initially failed in the form of overall active layer detachment, and then the landslides continued to expand backward in the form of retrogressive thaw slump in the next few years.

The initiation of failures was concentrated in a specific year and month, rather than occurring steadily. Our inventory includes 1,298 landslides along the QTEC. Statistical analysis revealed that there were only 428 failures before 2016. However, in 2016, 621 new failures emerged, constituting 56.6% of the total number of landslides. In 2017 and 2018, only 55 and 130 failures formed, respectively. From 2019 to 2021, there were only 16 new landslides (Figure 11).