Risk of glacier collapse in the Southeast Tibetan basin
Introduction
The Qinghai Tibet Plateau is an important water tower area, affecting about 2 billion life1. As an amplifier of global change, the Tibetan Plateau has a warming rate of 0.42 °C/10a, about twice the global average2. Therefore, glaciers on the Qinghai Tibet Plateau have experienced accelerated mass loss3,4, thermal change5 and glacier flow change6, which have increased glacier ablation and glacier instability7,8. This situation leads to frequent glacier collapses9, causing huge infrastructure damage and even life loss. For example, in 2016, two large-scale glacier collapses occurred in the Aru region in the west of the Qinghai Tibet Plateau. Nine people were killed and dozens of square kilometers of grassland were damaged10,11. On February 7, 2021, glacier collapse in Chamoli region of Uttarakhand, causing a serious impact on the area within 5 km of the downstream12,13, causing more than 200 deaths or disappearances and serious damage to infrastructure.
Southeast Tibet (SETP) has the largest area of maritime glaciers14. These glaciers are usually more sensitive to climate change than continental glaciers on the plateau15, and the terrain is higher, which is more conducive to the occurrence of glacier collapses. In recent years, the disasters caused by glacier collapses in SETP have been continuously recorded. In October, 2017, November, 2017, December, 2017, January, 2018 and July, 2018, glacier collapses occurred many times in the Sedongpu basin16,17,18. Debris blocked the river channel, forming a barrier lake, and caused huge disasters and potential threats to the residents and upstream and downstream traffic lines in Paizhen, Motuo County and along the river bank19,20. In 2020, the 1.14 Mm3 ice debris mixture from the collapse of the zelongnong glacier formed a high-speed debris flow, partially blocking the Yarlung Zangbo River and damaging the Zhibai bridge21. On March 22, 2021, a large-scale glacier collapse of 50 Mm3 occurred in the Sedongpu basin, blocking the Yarlung Zangbo River and flooding the road to Gyala Village22. In fact, SETP is one of the hot spots in the study of the cryosphere disaster on the Qinghai Tibet Plateau.
Due to the lack of observation data, only a few glacier collapses have been deeply studied13. Some studies focus on the river blocking event caused by glacier collapse, and describe the process and disaster characteristics of glacier collapse by remote sensing10,23. Other studies have classified different types of glacier collapses, established remote sensing interpretation markers, and recently identified hidden spots on the Qinghai Tibet Plateau24,25,26, but the understanding of the risk of glacier collapses is still developing. In addition, due to the suddenness of the glacier collapse disaster and the huge observation difficulty, it is challenging to accurately analyze its disaster process and driving factors17,27,28. Where is the dangerous area of glacier collapse and how big are the risks in different areas? Further quantitative analysis is needed. In addition, the glacier collapse risk assessment scheme is mainly qualitative or semi quantitative, and generally uses experience or simple models to judge the risk of glacier collapse. Although these first-order studies help to determine the scope of large-scale priority areas, they are not enough to determine the detailed glacier collapse risk and glacier risk areas to meet the needs of local development planning and disaster prevention11,19. With the development of artificial intelligence technology, some scholars have applied machine learning methods to the risk assessment of avalanches, landslides and other disasters29,30,31. The high risk area of avalanche is identified by machine learning method, and different risk levels are divided. This proves that the machine learning method is feasible to evaluate the complex disaster risk affected by multiple factors.
In view of these research gaps and the need to integrate and upgrade the current research points, we conducted this study to summarize the driving factors, development characteristics and risk assessment of glacier collapse in SETP. Finally, we identified hazardous area and systematically evaluated and quantified the hazards and risks of glacier collapses in the SETP basin.
The study area is located in the glacier area of the Linzhi and Changdu mountainous areas (Fig. 1a), with an average elevation of over 4000 meters, characterized by high mountains and deep basins. The highest peak of SETP is Namjagbarwa (located in the eastern Himalayas), with an altitude of 7782 meters, while the basin in the lower reaches of the Yarlung Zangbo River is less than 500 meters. Glaciers are mainly concentrated in the Yarlung Zangbo River Basin and its adjacent areas, with longitude between 92.5° and 98°E and latitude between 28.5° and 31.5 °N. Due to the combined impact of the Indian Ocean warm and humid monsoon and complex terrain, SETP is a region with abundant precipitation, with an annual peak precipitation of more than 1000 mm32. The abundant precipitation and low temperatures in high-altitude areas result in SETP having a large number of maritime glaciers. Compared with continental glaciers, these glaciers have experienced intense glacier melting, and their ice bodies are more unstable8. This phenomenon provides sufficient material source for the occurrence of glacier collapse disaster. The study area is located in the Himalayan tectonic fault zone. The crust is unstable, and frequent geological activities exacerbate the instability of glaciers. Therefore, the distribution, topographic characteristics and frequent geological activities of glaciers in this area make it a key area for glacier collapse disasters.
a Geographical scope and location of the study area. b Characteristics of annual average temperature and annual precipitation change in Linzhi station from 1960 to 2019. c Elevation and glacier distribution in SETP. d Slope and glacier collapse distribution in Sedongpu basin.
Sedongpu basin (Fig. 1c) is located on the western flank of the Jialabailei peak. The highest altitude of the basin is 7294 m, and the lowest point is 2746 m, the average slope is about 35° and the basin area is about 68 km2 20. Since 2017, there have been many glacier collapses in Sedongpu, resulting in more samples in the basin. Therefore, the basin is the area where we conduct model training, and the sample points are shown in the figure (Fig. 1d). In the Sedongpu basin, there are 12 glaciers, covering an area of approximately 19 km2. The type of glacier is maritime glacier. Compared with Continental glacier, it is more vulnerable to the impact of global temperature rise and accelerated melting. At the same time, maritime glaciers have the characteristics of fast speed, long distance and relative fragmentation. Sedongpu is rich in debris flow material source, and the glacier surface is covered with abundant loose materials, which together with a large number of moraines constitute the main material source of debris flow in the basin.
Results
Analysis of factors affecting glacier collapse and factor collinearity test
Understanding historical disasters is of great significance for selecting evaluation factors. Therefore, we analyzed the severe glacier collapse events that occurred in October and November 2017 in Sedongpu basin. Topographically, the elevation difference from the top to the bottom of the glacier in the basin is 3500 m, and the average slope is as high as 48° (Fig. 1d). Steep conditions are also conducive to the rapid flow of glacier, resulting in instability of ice and snow, which will lead to more glacier collapses. From the perspective of meteorological elements, from 1960 to 2017, the annual temperature of Linzhi station fluctuated between 7.4 ~ 10.3 °C, the annual precipitation fluctuated between 452 ~ 985 mm, the annual average temperature was 8.8 °C, and the annual average precipitation was 680 mm (Fig. 1b). The average temperature in Sedongpu showed an obvious warming trend, and the increasing trend of precipitation was not obvious, but the annual precipitation increased significantly before the occurrence of ice avalanche. All these have changed the physical properties of glaciers and become more unstable.
The earthquake is considered to be the direct trigger for this collapse disaster. On November 18th, 2017, an earthquake of magnitude 6.9 occurred 7.9 km southeast of Sedongpu basin, with a focal depth of 10 km, followed by six aftershocks. The largest aftershock reached magnitude 5, with a focal depth of 6 km, and the straight-line distance from Sedongpu basin was only 4.7 km (Table 1). In the year after Linzhi 6.9 Mw earthquake, four collapse debris flows occurred in the Sedongpu basin, indicating that the earthquake had obvious destructive effect and direct impact on glaciers.
Through extensive research by scholars23,28,33,34 and analysis of historical disasters, it is believed that the collapse of Sedongpu glacier may be the result of the synergy of temperature, precipitation and seismic activity under the background of steep terrain. Specifically, the rise of temperature exacerbated the melting and instability of glaciers, while precipitation affected the collapse of glaciers by increasing the process of mass accumulation and sliding.
The extraction of evaluation factors should take into account the representativeness of factors, whether it can comprehensively reflect the formation conditions of glacier collapse, and whether it can be expressed quantitatively. Therefore, on the basis of previous research and analysis of historical events, nine quantifiable extraction factors, such as altitude, slope, aspect, curvature, annual average temperature, annual precipitation, seismic risk, glacier velocity, glacier thickness, were selected. This study suggests that these factors are very important for understanding the complex phenomenon of glacier collapse.
In addition, the potential multicollinearity between assessment factors should be considered in factor selection. If there is a high or accurate correlation between variables in the model, multicollinearity will occur, which may lead to objectively inaccurate results. In order to solve this problem, this study uses variance inflation factor (VIF) as a tool to check the multicollinearity between factors. The results are shown in Table 2. The VIF values of the selected evaluation factors are less than 10, the maximum value is 3.26, and the minimum value is 1.14, which meets the requirements of multicollinearity analysis. This validates the selected evaluation factors, which are suitable for further study.
Model training
The assessment unit is the most basic unit for the extraction of risk assessment factors and quantitative risk assessment of glacier collapse. At present, risk assessment units include grid units, slope units and administrative units. Grid division is simple, objective and accurate. Therefore, the grid units are selected as the evaluation and analysis unit in this study. Considering the actual situation of the study area, 90 m grid units were selected, with a total of 770200 grid units. A total of 279 glacier collapse grid samples were obtained in the study35. In order to create a balanced dataset, GIS was used to randomly generate the same number of non-glacier collapse samples, resulting in a comprehensive dataset containing 558 samples. This study uses three different machine learning techniques36,37: support vector machine (SVM), convolutional neural network (CNN) and K nearest neighbor (KNN) to train glacier classification model, and uses the tenfold cross validation method to evaluate the accuracy of the model. Accuracy is an important evaluation index in machine learning, which is used to measure the accuracy of classification model prediction. The accuracy is calculated as the proportion of correctly predicted samples to the total number of samples. The higher the accuracy, the more samples the model correctly predicts, and the better the performance of the model. The calculation formula of accuracy is as follows:
The model classification accuracy illustrated in Table 3, reveal that the CNN classifier achieved the highest classification accuracy of 82.8%, thereby demonstrating its superiority. Consequently, CNN model was selected for further risk assessment.
However, during the training process, a notable misclassification issue was observed in the non-glacier collapse samples. As shown in the classification results in Fig. 2 approximately 0.6% (KNN), 0.6% (SVM) and 1.2% (CNN) of these samples were mistakenly classified as glacier collapse samples. This discrepancy can be attributed to the limited training area, which encompassed geomorphic units and glacier features that share significant similarities, thus increasing the likelihood of misclassification during the training phase.
Model classification results (1 represents glacier collapse, 0 represents non-glacier collapse).
The importance score of variables in the process of machine learning training is shown in Fig. 3. According to Fig. 3, the top five assessment factors with the highest importance score are precipitation, aspect, annual average temperature, earthquake and slope. We believe that these factors are also the key characteristics to distinguish areas susceptible to glacier collapse. Specifically, the increase of the liquid precipitation increases the potential glacier sliding and instability, resulting in greater sliding force and increased possibility of glacier collapse. Aspect and slope represent the local topographic changes in the area where glacier collapse occurs, and aspect affects the distribution of glaciers. The steep slope constitutes the necessary topographic conditions for glacier movement along the slope. The change of annual average temperature reflects the cold storage conditions in areas where glacier collapse may occur. Under the action of frequent earthquakes, the integrity of glaciers and rock masses is destroyed, and the mechanical strength is reduced, which provides favorable preconditions for the occurrence of debris flow of glacier collapse. The thickness of the glacier is the characteristic of the material conditions that lead to the formation of glacier collapse.
Factor importance results in machine learning classification model.
Model validation
The model is applied to glaciers in other basins for assessment and verification. For the red area identified by the model, we refer to it as a hazardous region, otherwise as a non-hazardous glacier region (Fig. 4b). We selected six glaciers in different basins in SETP for assessment and verification, and the classification results are shown in (Fig. 4c–h). Among them, the glacier in Sedongpu, Zelongnong and RGI60-13.01430 identified a large area of hazardous region, which is also consistent with the observation results of Yang Wei and others22,38 in this region (Fig. 4a). In Gongpu glacier, Kuangzha glacier and RGI60-15.12925 glacier has not identified to a large area of hazardous region. Through inquiry with local people and investigation by government departments, there has been no record of glacier collapse on these three glaciers for many years (there may be slight glacier collapse that has not been found or ignored), so our identification results for non-hazardous areas have also been confirmed. Therefore, we believe that the model is applicable to the identification of hazardous glaciers in the SETP basin.
a Spatial distribution of mean surface elevation change (m/yr) during 2010–2020 and locations of the three glaciers with abnormal surface thickening (Yang, 2023). b Identification of hazardous glaciers in SETP. c–h Verification of glacier identification results in different basins. i Statistics of hazardous glacier distribution aspect. j Elevation distribution of hazardous glacier. k Statistic of hazardous glacier velocity. l Statistics of hazardous glacier distribution slope.
Characteristics of hazardous glaciers
Through model identification, we divided the hazardous area and non-hazardous area, and analyzed the distribution characteristics of hazardous area. (Fig. 4i–l). The results indicate that the altitudes at which glacier collapses develop in SETP are predominantly concentrated between 4500 m and 6000 m, with a significant number of glaciers distributed within this range. In terms of slope, the development zones for collapses are between 35° and 50°. Slope gradient determines the stress distribution characteristics of the slope, and it also influences the accumulation of loose materials on the slope covered by snow and ice. Generally, the greater the ground slope gradient, the larger the scale and probability of collapse occurrence. Since very steep areas are not conducive to the accumulation of glacial material, gradients exceeding 50° are not conducive to the development of glacier collapses. Additionally, statistics have found that the northwest aspect is prone to collapse risk. Analysis suggests that southern aspect, being exposed to direct sunlight, are not conducive to the storage and accumulation of glacial material. Therefore, there is a northerly trend in the development of glacier collapses. Furthermore, glaciers in hazardous areas have relatively slower speeds. This is relatively well understood; if an ice mass moves down a slope in a stable flow pattern towards a lower elevation, the probability of its sudden collapse will be smaller.
Hazard of glacier collapse
The area is divided according to the geographical location of the concentrated distribution of glaciers. SETP is divided into seven large regions, and the smaller basins inside are numbered more carefully according to A to G (Fig. 5a). We have mapped 480 basins in total, and counted the area and proportion of hazardous glaciers in seven large regions: the total area of hazardous glaciers in SETP is about 946 km2. The largest area is region B in the southeast, with an area of about 320 km2. The highest proportion is in the southern region D, where the glacier area is 193 km2, and the proportion of hazardous areas reaches 55.6%.
a Hazard degree of glacier collapse in SETP basin, as well as the area and proportion of hazardous glaciers (The ring represents the proportion of hazardous glacier, and the size of the circular inside the ring represents the area of hazardous glaciers). b The hazardous glacier area in high and extremely high hazard degree basins.
According to the area of hazardous glaciers in the basin, the hazard degree of glacier collapse in each basin is evaluated. Using the Natural breaks (Jenks) method, the basin is divided into four different hazard degrees: extremely high, high, medium and low (Fig. 5a). Extremely high degree basins mainly occur in the south and southeast of the study area, and 45 basins in the whole study area are facing high and extremely high hazards. According to statistics, the area of dangerous glaciers in 45 basins exceeds 8 km2 (Fig. 5b). Among them, the dangerous area of B36 basin in the southeast is the largest, reaching 33.7 km2. The basin is facing the possibility of great glacier collapse.
Risk assessment of glacier collapse
The SETP basin uses the natural fracture method to classify the basin risk assessment into four degrees (Fig. 6a): low risk, medium risk, high risk and extremely high risk. We have divided 8 extremely high risk basins, and D30 basin has the highest risk, which is consistent with the research results of other scholars in SETP38.
a Risk degree of glacier collapse in SETP basin and distribution of residential areas, rivers, and roads. b Statistics of risk factors in extremely high risk degree basins. The gray column represents the number of residential area, the blue column represents the number of roads, the purple column represents the number of rivers, the orange column represents the area of dangerous glaciers, and the red line represents the value at risk.
In terms of disaster bearing bodies, 85 residential areas, 131 roads and 52 rivers are threatened by potentially extremely high glacier collapses (Fig. 6b). The results also show that there are regional differences in the degree of risk. In high-altitude areas, the types of affected bodies are relatively homogeneous, with low economic value, fewer threat objects, and relatively low risk of glacier collapse. However, in areas with high population density, complex terrain types and frequent human activities, the risk is higher. In addition, high-risk areas are concentrated in areas with high hazard of glacier collapse.
Discussion
In this work, the selection of the nine model factors is based on the consideration of large-scale glacier collapse risk assessment23. These nine factors are more representative on a large-scale, and previous study also show that these factors have a relatively significant impact on glacier collapse. In addition, other factors, such as the maximum monthly temperature, have been considered but it performed badly in model factor collinerarity test. Therefore, combined the previous study and model collinerarity test, we choose these nine factors in this work.
Disaster early warning is an important part of disaster prevention and mitigation. We have mapped the high-risk basin of glacier collapse, but the prevention of disasters still needs long-term real-time observation to achieve the purpose of early warning. At the same time, monitoring a complete glacier collapse process is also very important for its mechanism interpretation. Remote sensing is a common monitoring method in Glacier research, but the avalanche source usually occurs at very high altitude10,12,38. The rich monsoon moisture and cloud formation in high altitude areas make it impossible to obtain real-time images without cloud interference39, so it is impossible to determine the location and scale of glacier collapse. At present, the early warning systems has been implemented in the Sedongpu basin in the SETP, and the glacier collapse event has been successfully warned through monitoring22,38. However, in other high-risk areas identified by us, such as Yigong Zangbu, there is a lack of monitoring for glacier collapses. The reason is that in these glacier areas, the installation and maintenance of the monitoring system are generally faced with many challenges, such as equipment transportation in high altitude mountainous areas, severe extreme weather conditions, power supply and data transmission40. However, it cannot be ignored in high-risk areas, and relevant departments and governments should work on monitoring hazardous glaciers.
In view of the hot spot of glacier collapse research in recent years, SETP is still the main area of concern. Recent studies have shown that due to the deglaciation of glaciers caused by temperature rise in these areas2,41, glacier collapse activities have intensified. Although the region A in the east of SETP, which we divided, currently shows a limited area and proportion of hazardous areas, climate change and geological activities also cause concern about the possibility of small, highly destructive outbreaks in the future. SETP is a key area invested by the government, including the construction of Sichuan Tibet railway, the development of subsequent economic corridors, and the construction of large-scale hydropower projects, which are highly sensitive to natural disasters. Our findings highlight the major challenges posed by the large number of potential disasters in these economically disadvantaged and highly vulnerable areas. Considering that the risk of glacier collapse will further increase under the background of climate change in the future7,32, the government recognizes the urgency of responding to the threat of glacier collapse, and it is essential to speed up the submission of response plans.
In conclusion, this study reveals the main factors affecting the occurrence of glacier collapse in the mountains of SETP, as well as the distribution characteristics and potential risk of glacier collapses. By combining machine learning modeling and vulnerability calculation, the risk of glacier collapse is quantified. It is found that climate change, steep terrain, and seismic activity are the main influencing factors of glacier collapse. The hazardous area of SETP basin covers an area of 946 km2. The distribution slope is 35°−50°, and the altitude is concentrated in 4500 m–6000 m, and its slope direction has certain northwest characteristics. In terms of the distribution of high-risk basins, the high-risk areas are mainly located in region B and D. The hazardous area in region B reaches 320 km2, and region D has the highest proportion of hazardous areas, accounting for 55.6%, while region A in the East is the least, accounting for only 2.7%. In addition, 8 high-risk areas, 85 residential areas, 131 roads of all sizes, and 52 rivers are at high risk of glacier collapse.
At present, the risk assessment results are mainly based on historical data and machine learning model analysis. In view of the harsh natural environment of the Qinghai Tibet Plateau, it is not possible to carry out field investigation one by one. It is believed that the risk assessment of glacier collapse will be more perfect with the improvement of monitoring and in-depth study of mechanism in the future.
Methods
Datasets
The temperature and precipitation data used in this study are from China 1 km Resolution Monthly Average Temperature Dataset (1901–2022) and the China 1 km Resolution Monthly Precipitation Dataset (1901–2022) (https://data.tpdc.ac.cn/)42. In the study, the digital elevation model (DEM) data is SRTM1 data with a resolution of 90 m. DEM data is obtained from Geospatial Data Cloud (https://www.gscloud.cn/). The elevation, slope, aspect and curvature data used in the study were obtained through the above DEM calculation.
The data of the second China Glacier Inventory (v1.0) were used for the glacier boundary43. The glaciers list represents the latest and most comprehensive catalog, effectively minimizing the impact of cloud, seasonal snow and moraine cover on glacier range identification. Glacier velocity and volume data are from the data published by Millan. The team has produced high-resolution maps of global glacier surface velocity based on Landsat 8, sentinel-2 and sentinel-1 data44. The seismic risk data is from the pan tertiary seismic zoning data set of the National Tibetan Plateau Data Center (https://data.tpdc.ac.cn/). According to the downloaded global seismic zoning data45, the data were obtained by using the boundary clipping of the pan tertiary basin. Vulnerability data including residential areas, roads and rivers are from the Open Street Map (https://openmaptiles.org). The Information of all the data used in this study is listed in Table 4.
The above data was subjected to multicollinearity testing and a total of nine influencing factors were extracted for the training of machine learning classifiers. The selection of assessment factors is introduced in Analysis of factors affecting glacier collapse and factor collinearity test. Figure 7 shows the assessment process of basin glacier collapse risk based on machine learning methods used in this study.
Glacier collapse risk assessment process based on machine learning method.
Risk assessment method
The risk assessment involves determining the hazard of glacier collapse in the basin and the potential losses resulting from such disasters. Ultimately, through risk categorization, the degree of glacier collapse risk in each basin is determined, providing early warnings for residents in high-risk areas. The risk assessment is based on the results of the hazard degree. The vulnerability of the hazard-affected bodies is taken as an impact factor, and finally, the risk degree in different basins is obtained. The calculation method is as follows:
In this formula, R denotes risk, H denotes the hazard of glacier collapse, and V denotes the vulnerability46.
The vulnerability analysis of hazard-affected bodies necessitates the extraction of data pertaining to objects endangered by glacier collapse in the research area47. These objects primarily encompass population, properties, public facilities, land resources, and more. Notably, there are substantial disparities in the economic value in different locations. However, given the vast research area and limited data availability, a comprehensive classification of hazard-affected bodies in the research area is not feasible. Consequently, this article focuses primarily on three key hazard-affected bodies: residents, transportation, and river system. The Open Street Map may cause uncertainty between the results and the actual situation, which is also inevitable and the limitation of this study. The vulnerability was evaluated through a weighted density mapping method. This method considered the weighted superposition of residential area, transportation, and the river system with assigned weights of 0.5, 0.3, and 0.2 to derive the overall vulnerability48.
Responses