Introduction

The Third Pole of the Earth, where includes the Tibetan Plateau and surrounding Himalayas, Hindu Kush, and Tianshan Mountains, is the world’s most important water tower1,2, Due to climate warming, this region has experienced widespread downwasting and the retreat of over 10,000 glaciers over the past three decades3,4,5. Consequently, the increased melting and exposed depressions have facilitated the development of glacial lakes6. From 1990 to 2018, a significant number of new and rapidly expanding glacial lakes have been observed in the Third Pole, with a relative area increase of 15.2%7. Although glacial lakes capture a small portion of the meltwater contributing to sea level rise and have the potential for hydroelectric power development8,9, more research interest has been devoted to the appraisal of dangerous glacial lakes and the analysis of destructive glacial lake outburst floods (GLOFs). When a glacial lake is impacted by external forces, such as snow/ice avalanche, landslide, rockfall, etc., or is destabilized by continuous melting of the underlying buried ice in the moraine dam, it can suddenly release a large volume of water. The resulting flood is rapid and forceful, causes intense erosion along the channel, and is devastating to infrastructure in the GLOF’s path10,11. A typical Gongbatongshaco GLOF in the eastern Himalayas was documented in 2016. It had a 40-km runout distance and destroyed a hydropower plant in Nepal12. Since 1900, more than 110 GLOFs originating from such moraine-dammed lakes have been recorded in the Third Pole and have claimed ~7000 lives13,14. The Himalayas, Nyainqentanglha, and Tianshan Mountains serve as the main sources of GLOFs in the Third Pole.

Understanding the historical features of disasters and evaluating their trends are crucial for local disaster response and sustainable development15. However, the current knowledge of glacial lake changes and GLOF activities and risks in the Third Pole is still under development. For instance, several recently established independent inventories of glacial lakes in the Third Pole have yielded various quantities due to the different glacial lake definitions used7,16. The number of glacial lakes reported ranges from 10,000 to 30,000 during 2015–2020. This significant discrepancy hampers our ability to analyze the in-depth change signals and makes the effective assessment of glacial lake hazards and risks challenging. Moreover, the incomplete GLOF dataset makes it challenging to accurately analyze their trends, magnitudes, and drivers17. Whether GLOF activity is intensifying, stable, or diminishing requires further quantitative analysis18,19. Additionally, existing glacial lake risk assessment schemes in the Third Pole are primarily qualitative or semi-quantitative, and empirical or simple models are generally used to simulate GLOF paths and are combined with a relatively rough downstream exposure20,21. Although these first-order studies help to identify priority areas of concern on a large scale, they are insufficient to determine detailed GLOF risks and specific exposure locations for individual glacial lakes or local development planning22.

Given these research gaps and the need to integrate and upgrade the current research essentials, we conduct a study to re-overview the glacial lake changes, GLOF characteristics, and risk assessment in the Third Pole. Ultimately, we map an inventory of glacial lakes and compile a GLOF dataset to reveal their changes and mutual relationships. Through numerical modeling and the construction of a detailed exposure dataset, we systematically assess and quantify the hazards, exposure, and risks of the glacial lakes in the Third Pole.

Results

Heterogeneous changes in glacial lakes

All of the glacial lakes with areas ≥0.02 km2 and that are primarily fed by contemporary glacier meltwater were identified and included in the inventory using Sentinel-2A/B images acquired in 2018, 2020, and 2022. Classification was performed based on the topological positions of the lakes relative to their glaciers (see “Methods”). A total of 5894 glacial lakes were mapped in the Third Pole in 2022, with a combined area of 748.79\(\pm\)41.16 km2 and a combined volume of 20.13\(\pm\)17.12 km3 (Fig. 1a). Among them, 869 lakes were classified as proglacial lakes (207.33\(\pm\)8.2 km2; 10.37\(\pm\)9.41 km3), 2735 as periglacial lakes (222.24\(\pm\)15.91 km2; 4.84\(\pm\)4.25 km3), 1929 as extraglacial lakes (290.22\(\pm\)14.74 km2; 4.47\(\pm\)3.123 km3), 113 as supraglacial lakes (6.92\(\pm\)0.71 km2; 0.07\(\pm\)0.05 km3), and 248 as ice-dammed lakes (22.08\(\pm\)1.6 km2; 0.38\(\pm\)0.27 km3). Periglacial or extraglacial lakes are the main types of glacial lakes in most of the regions, accounting for an average proportion of 79%, except in East and West Kunlun where ice-dammed lakes predominate. Proglacial lakes directly connected to glaciers contribute an average of 17% of the total number of glacial lakes in all of the regions. Glacial lakes with a small size (≤0.1 km2) account for 74% of the total number of glacial lakes in the Third Pole (Fig. 1b). There are two altitude concentration intervals of 3400–3900 m and 4700–5800 m based on the number of glacial lakes (Fig. 1c), which are controlled by the increasing glaciation limits from southeast to northwest in the Third Pole.

Fig. 1: Glacial lake distribution and changes in the Third Pole.
figure 1

a Maps highlighting the glacial lake expansion rate between 2018 and 2022. The pie charts with different colors and sizes illustrate the number and types of glacial lakes based on the Global Terrestrial Network for Glaciers (GTN-G) regions. The histograms present the frequency of the glacial lakes in terms of their size (b) and elevation (c). The smoothed density distribution of the area change rate for all of the glacial lakes and lakes with detectable changes (error > change area) in 2018–2022 (d) and 1990–2018 (e).

Full size image

Overall, the expansion of glacial lakes in the Third Pole has continued in recent years. Between 2018 and 2022, the area of the glacial lakes increased by 22.68\(\pm\)8.84 km2 (3.03% relative area change), and 83% of this total increase was contributed by proglacial lakes. The mean expansion rate of individual proglacial lakes (0.022\(\pm\)0.002 km2/5 a) was an order of magnitude higher than that of the other lake types. To trace and compare the glacial lake changes since 1990, we combined the previously available inventories7,16,7,16,23,47,48. These differences can be attributed to varying area thresholds of 0.05–0.003 km2, as well as different definitions of glacial lakes. Initially, glacial lakes were identified as those within a 10 km buffer of a glacier and with a hydraulic connection to a glacier23. However, recent studies have included all lakes within the buffer regardless of glacier connections, resulting in inflated inventories. In this study, we needed to not only integrate previous research to analyze the state of the glacial lakes in the Third Pole but also to conduct a detailed risk assessment. We focused on glacial lakes with areas ≥0.02 km2 that were primarily fed by contemporary glacier meltwater within a 10 km glacier buffer. Numerous thermokarst lakes and lakes without parent glaciers were excluded. Three time windows, 2018, 2020, and 2022, were selected to create new inventories and were combined with other available datasets to reveal the short-term and long-term glacial lake changes. A total of 878 Sentinel-2A/B images (10-m resolution) were used to manually delineate the glacial lakes (Fig. S5). To ensure a sufficient storage period for the glacial lakes during the year and to minimize the presence of mountain shadows, priority was given to images captured between July and November with less than 10% cloud coverage. The images were processed using a false-color composition of bands 4, 3, and 2 to highlight the water bodies. The initial locations for the lake extraction were based on the 2018 glacial lake inventory created by Wang et al.7. Throughout the workflow, each glacial lake underwent thorough review at least six times, ensuring the inventory’s completeness according to our standards.

The glacial lakes were classified into five types based on their topologic positions relative to their glaciers (for example, see Rick et al.49): (1) proglacial lake, located in contact with the glacier terminus50; (2) periglacial lake, decoupled from its glacier but situated at the glacial moraine51; (3) extraglacial lake, located far from the glacier terminus and often with a landslide dam or without any dams52; (4) supraglacial lake, located on the surface of a glacier; and (5) ice-dammed lake, formed when a glacier surge blocks a downstream valley or when meltwater fills a depression between a retreating tributary and the main glacier (Fig. S6). This classification method emphasizes the connections between lakes and glaciers, facilitating the analysis of glacial lake dynamics within the context of glacial changes, compared to categorizing them based on the dam materials. Supraglacial lakes are typically small, exhibit significant fluctuations within and between years, and frequently appear and disappear53,54. In comparison, numerous, occasionally concealed supraglacial lakes are another cause of the discrepancies in glacial lake datasets47. Our inventory only includes a small number of stable lakes. Additionally, our inventory included ice-dammed lakes that were overlooked by previous studies. These lakes are distributed across several ice caps in Inner Tibet and the Kunlun Mountains. Because they are often frozen, automatic identification approaches based on the normalized difference water index face challenges in extracting them The glacial lake volume was estimated using a set of empirical equations proposed by Zhang et al.55, and its uncertainty was calculated by using 95% confidence intervals of the regression curves. The uncertainty in the glacial lake area (\(\sigma\)) was estimated as follows (Fig. S7):

$$\sigma=0.5P{\cdot }I$$
(1)

where \(P\) is the lake perimeter, and \(I\) is the pixel resolution of the imagery. The absolute area change rate (\(R\)) was used to reflect the glacial lake expansion or shrinkage, and its uncertainty was estimated as follows:

$$R=({A}_{2}-{A}_{1})/T$$
(2)
$$\sigma R=\sqrt{{\left(\frac{{\sigma A}_{1}}{{A}_{1}}\right)}^{2}+{\left(\frac{{\sigma A}_{2}}{{A}_{2}}\right)}^{2}}$$
(3)

where \({A}_{1}\) and \({A}_{2}\) are the lake area in the first and second periods, respectively; \(T\) is the time difference; and \({\sigma A}_{1}\) and \({\sigma A}_{2}\) are the area uncertainties. We calculated the glacial lake change error (\(\sigma A\)) as follows:

$$\sigma A=\sqrt{{{\sigma A}_{1}}^{2}+{{\sigma A}_{2}}^{2}}$$
(4)

If error > |area change|, the lake change is not significant; otherwise, the lake exhibits a positive or negative change. To accurately calculate and interpolate the lake change rate over the past decades, especially for newly formed glacial lakes, we utilized the dataset for 1990 created by Wang et al.7, that for 2000 created by Zhang et al.66, we added an additional indicator, namely, the (6) horizontal distance between the glacier terminus and lake, to further highlight the impacts of glaciers. This indicator has been widely used and is highly effective in susceptibility assessment67,68,69. It is generally assumed that ice avalanches do not affect glacial lakes when the distance exceeds 1 km70,71. Extraglacial lakes, located far from the parent glacier and developed in a consolidated moraine, were excluded from the consideration of the dam metrics. Since earthquake-triggered GLOFs are rare in the Third Pole, indicators for assessing the impacts of earthquake were not included in the selection process. Furthermore, a sensitivity experiment conducted in the Himalayas and Southeast Tibet demonstrated the ambiguous role of the influence of climate changes on GLOF susceptibility. As a result, indicators related to these factors were not incorporated into the assessment.

To calculate the slope and determine the watershed area, we utilized the 30 m Advanced Land Observing Satellite Global Digital Surface Model (AW3D30). Each indicator was normalized to a range of 0–1 and was assigned a weight using the analytical hierarchy process (AHP). While like AHP and some other multi-criteria decision analysis methodologies are prevalent in Earth Science, for relatively complex assessment cases, it should be applied cautiously, with input from experts in operations research to ensure the use of the most current and appropriate methodology. Based on the natural Jenks approach in ArcGIS, the final composited indexes of the GLOF susceptibility were classified into five categories: very low, low, medium, high, and very high. Seventy-two glacial lakes that had previously produced GLOFs were included in our assessment samples and were used to validate the evaluation accuracy of the GLOF susceptibility.

GLOF simulation and hazard map**

A GLOF simulation scheme was implemented specifically for glacial lakes with high or very high susceptibility levels (i.e., a total of 1499 glacial lakes with such a high outburst potential) without considering all of the glacial lakes. The HEC-RAS 2-dimensional hydraulic model was selected to simulate the potential flood propagation. This model has successfully been used to reconstruct and evaluate many GLOF events worldwide72,73,74, demonstrating its practicality and efficiency. By utilizing the 2-dimensional flow surface, as well as a dam-breach hydrograph and high-resolution digital elevation model (DEM), an unsteady flow simulation was conducted for each glacial lake with a high outburst potential.

It should be noted that the complete drainage of a glacial lake is size-dependent. The statistical data revealed that since 1980, there have been 29 small outburst glacial lakes (<0.1 km2) with an average failure volume of 85% of the total lake volume, and as the lake size decreases, the frequency of complete drainage increases. Conversely, 36 medium/large glacial lakes exhibited an average failure volume of 58% of the total lake volume, with larger lakes experiencing relatively smaller drainage volume portions. To maximize the GLOF simulation within a reasonable range, we utilized the upper limit curve of the scatter value of the drainage volume and glacial lake volume (\({V}_{d}=2.01{V}_{t}^{0.65}\)) to determine the potential maximum drainage volume of each glacial lake (Fig. S9a), and we designated all of the small glacial lakes as completely drained. Due to the limited applicability range of this curve, we set the maximum drainage volume of the glacial lakes as 20 × 106 m3, resulting in restrictions for a total of 76 large glacial lakes. Subsequently, based on the estimated drainage volume and peak discharge, we adopted a linear assumption for the increase/decrease in the flood discharge to generate a dam-breach hydrograph75. Previous case studies have demonstrated that this hypothetical approach does not significantly affect the results72,73,76. Additionally, we established a relationship between the GLOF impact distance and drainage volume25 to determine the length of the two-dimensional flow surface, ensuring that the flood remained contained (Fig. S9b).

Regarding the DEM used, we developed a composited scheme using two available DEMs for the Third Pole. The High Mountain Asia (HMA) DEM (8 m) was given priority as it was generated from very high-resolution commercial satellite imagery77. The HMA DEM provides detailed ground features but has a limited coverage78. Therefore, in areas where the HMA DEM is incomplete or missing, we utilized the Advanced Land Observing Satellite (ALOS) Phased Array type L-band Synthetic Aperture Radar (PALSAR) DEM (12.5 m) either for mosaicking with the HMA DEM or for direct use. All of the DEMs were subjected to depression filling and surface smoothing processes to warrant unobstructed water flow. Given the need for extensive GLOF simulations in the topographically complex Third Pole region, the model requires a robust and efficient set of parameters that allow the simulated flow to adapt to narrow valleys such as those in the Himalayas, as well as relatively flat surfaces such as Inner Tibet. Through preliminary sensitivity analysis, we determined the crucial parameters of the model to be as follows: the minimum computational unit is 30 m, Manning’s coefficient is 0.06, and the computation interval is 5 s.

By employing the HEC-RAS model, we obtained the distributions of the inundation extent, maximum water depth and flow velocity, and the arrival time for each potential GLOF. We defined an indicator, namely, the GLOF probability (\({{{{{{\rm{P}}}}}}}_{{{{{{\rm{GLOF}}}}}}}\)), as the degree of threat posed by the GLOF at specific locations. This indicator was calculated as follows:

$${{{{{{\rm{P}}}}}}}_{{{{{{\rm{GLOF}}}}}}}=\frac{\sum {n}_{i}}{N}$$
(5)

where \({n}_{i}\) represents the number of GLOF events in a specific site, and N corresponds to the maximum number of GLOF occurrences in a specific site across the Third Pole.

Furthermore, we integrated the water depth, flow velocity, and hazard index to create a hazard map for each glacial lake. The map** approach, based on the concept proposed by Zhang et al.79, allows us to inform ground-level planning and response actions. The improved formula is utilized in the computation:

$${H}_{s}=S \, * \, D \, *\, V$$
(6)
$${H}_{r}=\mathop{\sum }\limits_{i=1}^{n}{S}_{i} * {D}_{i} * {V}_{i}$$
(7)

where \({H}_{s}\) is the hazard for a single glacial lake; \(S\) is the value of the GLOF susceptibility index; \(D\) is the simulated flood depth; and \(V\) is the flow velocity. For regional hazard map**, various single GLOF hazards are overlain and calculated. Similarly, each hazard map is divided into five classes (very low, low, medium, high, and very high).

Exposure and risk

In various GLOF events, the impacts commonly reported include damage to buildings, hydropower projects, farmland, roads, and bridges27. These indicators were employed to quantify the GLOF exposure. All of the infrastructure features, except for hydropower information, were obtained from OpenStreetMap (https://www.openstreetmap.org/). However, OpenStreetMap has limited coverage of features throughout the Third Pole80. We checked these structures along our simulated GLOF paths, resulting in a significant number of buildings, roads, and farmland not being mapped, particularly in villages situated in gullies, such as those in Inner Tibet, Hindu Kush, and Central Asia. With the exception of Nepal, which had better coverage, the completeness of the rest of the Third Pole was less than roughly 5% (Fig. S10). Hence, we manually supplemented the structure features for the entire Third Pole. Considering the uncertainty of the modeled floodplain boundaries, a buffer distance of 200 m around the potential inundation areas of the GLOFs was incorporated to the extract exposure elements from high-accuracy Maxar Premium Imagery. Finally, we identified a total of 520,610 buildings, 1996 km2 of farmland, 98,630 km of roads, and 6035 bridges within these buffers. The reconstructed exposure elements can accurately reflect the construction status of the Third Pole between 2016 and 2020.

We normalized each indicator to a range of 0–1 and employed the AHP method to allocate weights to the exposed buildings, hydropower projects, farmland, roads, and bridges. This ranking takes into account their varying levels of impacts on humans. By integrating these factors, we derived the exposure index for glacial lakes. The GLOF risk is the result of multiplying their normalized hazard by the exposure. Both the exposure and risk were also classified into five categories, i.e., very low, low, medium, high, and very high, utilizing the natural Jenks approach. To investigate the characteristics of the GLOFs at both the single lake and regional scales, we defined the level of exposure per unit of inundated area as the potential disaster intensity (PDI), which was calculated as follows:

$${{{{{\rm{PDI}}}}}}=\frac{E}{{A}_{i}}$$
(8)

where \(E\) is the GLOF exposure, and \({A}_{i}\) is the normalized inundation area of the GLOF.

Robustness of assessments and simulation

The indicators used to assess the glacial lake susceptibility have been carefully selected to correspond to various GLOF triggers, and weights have been assigned based on their importance. This contributes to a more scientifically grounded assessment. In the GLOF simulation scheme, glacial lakes without a high outburst potential were excluded. Previous studies have verified that over 90% of drained lakes can be identified as having a high or very high susceptibility20,35. Moreover, in comparisons of different assessment cases, glacial lakes with a high outburst potential can be robustly identified even using different combinations of evaluation indicators and weighting schemes63. These phenomena indicate that scaling down the sample for further GLOF simulations would not significantly impact the results as it covers all of the interesting glacial lakes and ensures the accurate determination of future GLOF sources. For example, our simulation samples include all eight important glacial lakes in Nepal identified by Rounce et al.81 and ICIMOD82, and our assessment of the GLOF threat to Almaty City is in agreement with the results of Bloch et al.83.

Due to the nature of conducting a maximum potential GLOF simulation, quantifying uncertainties is challenging. There are two key sources of uncertainty to consider. The first is the assignment of model parameters. For example, we selected a relatively smaller Manning’s value to enhance the water flow efficiency. A sensitivity analysis performed in the reconstruction of the 1981 Cirenmaco GLOF in the Poiqu River Basin indicated that variations in the Manning’s value had no decisive impact on the flood inundation area and mean water depth72. This conclusion aligns with our modeling requirements. Second, the simulation accuracy of GLOFs is primarily determined by the high-resolution DEM as it serves as the main input data. There are essential differences between the results obtained from the commercial imagery produced HMA DEM and freely available PALSAR DEM. GLOF propagation is significantly hindered by extensive depression filling on the surface of the PALSAR DEM, leading to underestimations of the flood inundation area, arrival time, and downstream exposure. Of the 1499 GLOF propagation simulations, the HMA DEM was utilized for 52.6%, which guarantees the fundamental credibility of our results. The mosaic production was used in 30.7% of the simulations, and the ALOS PALSAR DEM was only used in 16.5% of the simulations. Regionally, the HMA DEM was less frequently applied in the Altai and Southeastern Tibet regions, limiting the reliability of their simulation results. Overall, it should be noted that our maximum potential GLOF simulation and exposure analysis may still underestimate the actual risk due to data limitations in various areas.