Against the background of climate warming and the consequent widespread mass loss of glaciers in alpine regions, increasing volumes of glacial meltwater are being released. This results in glacial lake expansion and extended areas of low-lying terrain (e.g. depressions and troughs) left behind by retreating glaciers in which water can accumulate and new glacial lakes can form (Clague and Evans, 2000; Mool et al., 2001; Song et al., 2016). As both a water resource and a source of flash flood and debris flow hazards, glacial lakes participate in several natural processes, e.g. regional energy and water cycles (Slemmons et al., 2013); act as both indicators and containers of environmental information (Wang et al., 2016, 2019b; Zhang et al., 2019); and drive hillslope erosion and landscape evolution (Cook et al., 2018) in the alpine cryosphere. On the one hand, glacial lakes act as temporary storage for the meltwater resource because a considerable amount of meltwater is retained by glacial lake expansion; e.g. approximately 0.2 % a-1 of the total glacial meltwater was reserved in glacial lakes from 1990 to 2010 in the Tien Shan in central Asia (Wang et al., 2013). On the other hand, given the worldwide expansion in lake area in recent decades, the potential will increase for glacial lakes to develop into glacial lake outburst floods and related debris flows that could threaten downstream residents, infrastructure, and regional ecological and environmental security (Huggel et al., 2002; ICMOD, 2011; Bolch et al., 2012; Haeberli et al., 2016). Thus, glacial lakes perform important roles both in the meltwater cycle and in glacier hazard evolution in the cryosphere.

Following the rapid development of remote sensing technology and computer science, remote sensing imagery acquired by various satellites and sensors has been used widely in glacial lake research. In particular, Landsat imagery has become the most important data source for the dynamic investigation of glacial lakes because of its wide coverage, continuous and long-term temporal sequence, and accessibility. Based on remote sensing data, both the distribution and the characteristics of change in glacial lakes in the mountains and watersheds in the high-mountain Asia (HMA) region have been widely reported (Supplement Table S1). For example, multi-source remote sensing imagery has been used to compile glacial lake inventories for regions of the Tibetan Plateau (Zhang et al., 2015), the Tien Shan (Wang et al., 2013), the Himalaya (Gardelle et al., 2011; Nie et al., 2017), the Hengduan Mountains (Wang et al., 2017), Uzbekistan (Petrov et al., 2017), Pakistan (Senese et al., 2018), and HMA, excluding the Altai and Sayan (Chen et al., 2020). These inventories have proved an important data resource both for revealing the spatio-temporal characteristics of glacial lakes and for understanding the response of glacial lakes to the effects of climate change in these regions.

Automatic and semi-automatic glacial lake boundary vectorization approaches have been used most widely in regional glacial lake investigations because of their higher efficiency and objectivity in comparison with manual visual vectorization. In such research, water bodies are usually determined based on the characteristics of different remote sensing bands and computer-dependent algorithms, e.g. the normalized difference water index (NDWI), band ratio, support vector machine, decision tree, spectral transformation, object-oriented classification, global–local iterative scheme, active contour model, and random forest (Gardelle et al., 2011; Huggel et al., 2002; Li et al., 2011; Veh et al., 2018; Zhang et al., 2018). However, manual post-processing is often required to calibrate the uncertainties that could easily be produced by the above approaches. Furthermore, the labour costs associated with rectification of lake boundary errors increase sharply with increasing complexity of study area terrain (Yang et al., 2019). With consideration of the accuracy, efficiency, and time overheads associated with the various vectorization approaches, a manual vectorization approach was adopted for investigation of the glacial lakes on the Tibetan Plateau (Zhang et al., 2015) despite the labour requirements and the anticipated additional errors produced by individual subjectivity (Nie et al., 2017; Yang et al., 2019; Song et al., 2014).

Controversies and knowledge gaps remain regarding available glacial lake inventories for different alpine cryosphere regions, which present certain obstacles to extensive utilization of glacial lake inventory data. The main problems relate to regional differences in lake development, inconsistent specifications of lake definition, and the adoption of various approaches regarding lake interpretation (Yao et al., 2018). There is no existing comprehensive glacial lake inventory for the entire HMA, and knowledge regarding the spatio-temporal characteristics of glacial lakes in this region remains incomplete. The objectives of this study were to fill this knowledge gap by producing a glacial lake inventory data set for HMA derived from Landsat images and to provide fundamental data for water resource evaluation, assessment of glacial lake outburst floods, and glacier hydrology research in the mountain cryosphere region.

The HMA area mainly comprises the Tibetan Plateau and surrounding alpine ranges. The area is divided into 13 subregions in version 6 of the Randolph Glacier Inventory (RGI 6.0), i.e. the Himalayan area (western Himalaya, central Himalaya, and eastern Himalaya), the Hengduan Mountains, southern and eastern Tibet, Inner Tibet, the Karakoram and western Kunlun, the Qilian Shan and eastern Kunlun, the Hindu Kush, the Pamir, the Hissar–Alay and western Tien Shan, the eastern Tien Shan, and the Altai and Sayan (Arendt et al., 2015; Pfeffer et al., 2014). The boundaries of the 13 subregions and outlines of the glaciers in HMA derived from RGI 6.0 are shown in Fig. 1. This region (26–54∘ N, 67–104∘ E) is characterized by tremendously complex topographic conditions with a widespread distribution of mountain glaciers. According to the Climatic Research Unit Time Series v4.02 data set (http://data.ceda.ac.uk/badc/cru/data/cru_ts/cru_ts_4.02/, last access: 5 September 2020), the air temperature of the different subregions in HMA increased at an average annual rate of 0.002–0.054 ∘ a-1 during 1990–2018 (Fig. 1). The annual rate of change in precipitation in HMA during 1990–2018 varied from -9.9 to 4.2 mm a-1 with a small average rate of increase of 0.3 mm a-1.

The HMA area has the largest surviving glaciers of any region other than the polar regions. As reported in RGI 6.0, there were 97 973 modern glaciers in our study area, covering a total area of approximately 98 768.86 km2. Together, these glaciers produced an average negative mass balance of -150±110 kg m-2 a-1 (Hock et al., 2019), which was the primary source of water supply for the development of glacial lakes. Over recent decades, glaciers in most areas of HMA appear to have experienced widespread mass wastage and area shrinkage (Bolch et al., 2012; Yao et al., 2012; Kääb et al., 2012; Brun et al., 2017). However, the so-called “Karakoram Anomaly” refers to a region that is a prominent exception, which is characterized by glaciers with stable or positive mass balance (Hewitt, 2005; Gardner et al., 2013; Kääb et al., 2015).

We developed our glacial lake inventory of HMA based on 668 high-quality images selected from more than 1800 Landsat images with a 30 m spatial resolution derived from the websites of the United States Geological Survey (https://www.usgs.gov/, last access: 5 September 2020) and Geospatial Data Cloud (http://www.gscloud.cn/, last access: 5 September 2020). To ensure the accuracy of glacial lake boundary extraction, the following criteria were applied to imagery selection. First, the cloud coverage in an image had to be <10 %. Second, for areas with no eligible or only low-quality imagery (because of snow or shadows) in the given year, acceptable images from years closest to the given year were chosen as replacements (Fig. 2). Third, images acquired in summer or autumn (June–November), when lake areas were believed to be near to or at their maximal extent, were set as optimal choices to minimize the impact produced by seasonal area changes in the glacial lakes (Fig. 2). Based on the above criteria, 394 and 274 Landsat images were selected to represent circa 1990 and circa 2018, respectively, which completely covered the buffer area within 10 km of glacier extent acquired from the second Chinese glacier inventory (http://westdc.westgis.ac.cn, last access: 5 September 2020) and RGI 6.0 (https://www.glims.org/RGI/rgi60_dl.html, last access: 5 September 2020). Among the selected images, those acquired during summer and autumn (June–November) accounted for 82.0 % of the total number of selected images, while those acquired during autumn (September–November) accounted for 56.9 % of the total number. In addition, a Shuttle Radar Topography Mission digital elevation model with a spatial resolution of 1 arcsec (http://imagico.de/map/demsearch.php, last access: 5 September 2020) was used to derive the elevation of the glacial lakes.

The errors associated with glacial lake extraction from remote sensing imagery using manual visual delineation are generally related to components of the quality of the images (e.g. spatio-temporal resolution, cloud coverage, and mountain shadows), experience, operative subjectivity, and the threshold area of the inventory (Gardelle et al., 2011; Hall et al., 2003; Paul et al., 2004; Salerno et al., 2012; Zhang et al., 2015). It has been reported that the area error in glacial lake boundary extraction based on remote sensing images can be approximately ±0.5 pixels depending on the quality of the imagery (Fujita et al., 2009; Salerno et al., 2012). Furthermore, the area error in glacial lake delineation attributable to manual delineation can be assumed to follow a Gaussian distribution (Hanshaw and Bookhagen, 2014). Hence, the theoretical maximum area error in glacial lake boundary extraction is the half area of the edge pixels because pure lake water body pixels are usually surrounded by mixed pixels (Fig. 4). The lake area error of a single glacial lake within 1 standard deviation (1σ) can be expressed as follows (Hanshaw and Bookhagen, 2014): 4Error(1σ)=PG×G22×0.6872,5E=Error(1σ)A×100%, where P is the perimeter of the glacial lake (m), G is the spatial resolution of the remote sensing imagery (30 m in this data set), 0.6872 is the revised coefficient under 1σ (i.e. approximately 69 % of peripheral pixels are subjected to errors), E is the relative error of the glacial lake, and A is the total area of the glacial lake. Then the accumulation of errors in all the lake areas for the entire study region or subregions can be calculated using the following formula based on error propagation theory: 6ET=∑i=1nai2, where ET is the area error of the entire study region or subregions, i is the lake of no. i in the entire study region or subregions, and a is the error area of a single lake.

The resulting calculated error indicated that the total absolute area error in HMA glacial lakes was approximately ±2.11 and ±2.28 km2 and the average relative error was ±13.5 % and ±13.2 % in 1990 and 2018, respectively. The relative area errors of each lake varied from 1 % to 79 %, and a significant power exponential relationship was found between the relative area error and the sizes of the glacial lakes (E=0.050A-0.45,R2=0.96,α<0.001; Fig. 5a). Small-sized lakes (i.e. area ≤0.01 km2, which accounted for 2 % of the total lake area in HMA) had the largest average relative area error of 44.6 % (Fig. 5b). Medium-sized lakes (i.e. area of 0.01–0.1 km2, which accounted for 34 % of the total lake area in HMA) had an average relative area error of 22.0 % (Fig. 5c). Large-sized lakes (i.e. area ≥0.1 km2, which accounted for 64 % of the total lake area in HMA) had the smallest average relative area error of 7.6 % (Fig. 5d). In summary, smaller glacial lakes in the HMA region had larger relative area errors and vice versa.

To further verify the accuracy of the manual delineation of glacial lake boundaries, nine lakes located within the HMA region were surveyed using a portable GPS device (Trimble GeoXH6000) with decimetre accuracy during July–August 2018 (Fig. 6). The lakes selected for field survey covered areas of 0.01–2.97 km2. The field-based lake boundaries were compared with those obtained via manual delineation (i.e. derived from Landsat OLI imagery acquired during 2018). It was found that the area error (i.e. the percentage difference in the absolute area encircled by the manually delineated lake boundary and that derived by the GPS survey) varied from 5.5 % to 16.3 %. Moreover, it was determined that the average horizontal distance deviation between the two types of boundary varied from 3.2 to 15.3 m (Table 1). Overall, the horizontal deviations were largely confined to 1 pixel, and the average accuracy of the delineation of glacial lake boundaries was within ±0.5 pixels (±15 m).

As indicated by the achieved HMA glacial lake inventory, 30 121 (2080.12±2.28 km2) glacial lakes were identified in 2018 and their distribution had considerable spatial heterogeneity (Fig. 7). The greatest concentration of glacial lakes was in the Altai and Sayan (335.42±0.88 km2, which accounted for 16.1 % of the total area of glacial lakes) and the eastern Himalaya (310.37±0.89 km2, which accounted for 14.9 % of the total area of glacial lakes). Relatively few glacial lakes were found distributed in the eastern Kunlun and Qilian Shan (38.85±0.29 km2, which accounted for 1.9 % of the total area of glacial lakes) and the eastern Tien Shan (40.55±0.32 km2, which accounted for 2.0 % of the total area of glacial lakes). The HMA glacial lakes were located within the elevation range of 1357–6247 m in 2018. An approximate normal distribution was presented both for the lakes of the entire HMA region and for the lakes in most subregions. More than 43 % of the HMA lake area has survived within the vertical range of 4500–5400 m, with the peak lake area of 241.89±0.80 km2 (accounting for 11.6 % of the total area) in the range of 5100–5300 m. The elevation band of peak lake area in the different subregions varied from 2300–2500 m in the Altai and Sayan to 5300–5500 m in the central Himalaya, Karakoram, and western Kunlun.

The HMA glacial lakes experienced widespread areal expansion during 1990–2018 with an average rate of increase in area of 15.2 % (Fig. 7). The rate of change in area varied widely between different subregions and different 200 m elevation bands. The glacial lakes in eastern Kunlun and Qilian Shan experienced the most rapid expansion in area during 1990–2018 with an average rate of increase of 45.6 %, whereas the rate of change was only 7.5 % in Altai and Sayan. Glacial lakes have tended to develop to higher elevations during recent decades with the maximum distribution elevation of 6078 m in 1990 rising to 6247 m in 2018. The rate of change of glacial lakes in the different 200 m elevation bands presented a large average trend against elevation, rising as a whole during 1990–2018 (Fig. 7). The lake area expanded most from the elevation of approximately 5000 m, and the rate of expansion reached approximately 28 % at 5700–5900 m in the entire HMA region, although it differed between different subregions. Lake area showed a notable rate of increase with elevation in most subregions, e.g. the Hissar–Alay and western Tien Shan, Hindu Kush, eastern Himalaya, Hengduan Mountains, eastern Tien Shan, and Altai and Sayan. The rate of expansion varied markedly and no observable trends in the rate of increase or decrease with elevation were discovered in Karakoram and western Kunlun, western Himalaya, and Inner Tibet. The rate of expansion in the central Himalaya and southern and eastern Tibet was found to have seemingly decreased with increasing elevation (Fig. 7).

There are at least 34 published reports or data sets on the regional extent of glacial lakes in the HMA area, which are based on various lake boundary extraction methods and different data sources (Supplement Table S1). This previous research work examined glacial lakes from as early as 1962 up until 2017. However, it is difficult to evaluate any discrepancies comprehensively because different extents of glacial lake distribution were examined and inconsistent thresholds of minimum lake area were used. Glacial lake inventory data of the Third Pole region in 1990 (Zhang et al., 2015) and of the HMA (Chen et al., 2020) in 2017 were used for comparison because both recorded glacial lakes in the same buffer zone (i.e. within 10 km of the modern glacier extent) and over similar periods. For the comparison, the same thresholds and regions were adopted for the inventory data. Marked discrepancies were found to exist between the different data sets in terms of both the number and the area of the glacial lakes. In 1990, only 4601 glacial lakes (≥0.0054 km2) with a total area of 554.33 km2 were recorded by Zhang et al. (2015), whereas 20 410 glacial lakes with a total area of 1376.23 km2 were catalogued in the Third Pole region in this study. In 2017, 14 477 glacial lakes with a total area of 1635.94 km2 were recorded by Chen et al. (2020), whereas we recorded 22 727 glacial lakes (≥0.0081 km2) with a total area of 1726.41 km2 in 2018 in HMA (excluding Altai and Sayan). We consider the discrepancies attributable to three primary factors. (1) The buffer zone within 10 km of the modern glacier extent was inconsistent between the data sets because different glacier inventories were used. (2) Different operatives catalogued the glacial lakes using different remote sensing data covering different periods. (3) Many glacial lakes were possibly missed because of the comparatively less manual vectorization effort involved in the work of Zhang et al. (2015) and Chen et al. (2020). Overall, our glacial lake inventory catalogued glacial lakes throughout the entire HMA more comprehensively and with more careful error assessment when compared with available glacial lake data sets from regional or river-basin-based studies.

Several limitations deserve proper consideration when using the glacial lake inventory data. First, a degree of uncertainty resulted from using Landsat image data that covered different periods, i.e. both interannually and seasonally. Although images acquired in summer or autumn (June–November) were set as optimal choices, the selected images covered most seasons of the year; e.g. the images selected in June–November accounted for only 72.3 % and 88.8 % of the total number in 2018 and 1990, respectively. Interannually, images were selected from a span of 10 a (1986–1995) and 4 a (2016–2019) to obtain sufficient high-quality images of the HMA area. Second, this study recorded all lakes located within the 10 km buffer area of glacier extent as glacial lakes. Therefore, certain lakes that have no relation to glaciers or to glaciation (i.e. non-glacial lakes) in the alpine cryosphere were potentially catalogued in error because of the difficulty in distinguishing non-glacial lakes from glacial lakes based on remote sensing data. Third, we identified water bodies related to glaciers or to glaciation in the alpine cryosphere as glacial lakes. However, in many cases, it was difficult to determine whether such bodies should be recorded as glacial lakes, e.g. cases of long narrow water bodies on rivers and cases where the number of pure water body pixels was small. Thus, some errors and inconsistences were inevitable because of having different operatives performing the lake boundary vectorization and inspection. In future, this glacial lake inventory will be updated and shared on the National Special Environment and Function of Observation and Research Stations Shared Service Platform (China).

The data set developed in this study comprises two .shp file documents containing the glacial lake inventory of the HMA region in 1990 and 2018. The data set can now be accessed via the website of the National Special Environment and Function of Observation and Research Stations Shared Service Platform (China): 10.12072/casnw.064.2019.db (Wang et al., 2019a).

A glacial lake inventory of the HMA region was realized based on satellite remote sensing data and GIS techniques. Eight attribute items were recorded in the glacial lake inventory data set of the HMA region. Lake area error was assessed carefully with respect to theoretical analysis of lake boundary pixels and actual boundaries derived by GPS field-based surveys. On average, the deviations between the delineation of lake boundaries derived using the two methods were within ±0.5 pixels (±15 m). The relative area errors of each lake in 2018 varied from 1 % to 79 %, and the average relative area errors of ±13.2 % in the entire HMA region were characterized by an increase in the relative area error with decreasing lake size.

Overall, 30 121 glacial lakes with a total area of 2080.12±2.28 km2 were catalogued in 2018 in the HMA region. Glacial lakes survived in all 13 subregions of HMA from the elevation of 1357 to 6247 m. Glacial lakes were found to be concentrated in the subregions of Altai and Sayan and the eastern Himalaya and at elevation bands of 4500–5400 m. The HMA glacial lakes have experienced widespread expansion with an average rate of increase in area of 15.2 %. Lake area expanded most in the higher-elevation bands during 1990–2018. The data set is expected to provide basic data to support cryosphere hydrology research, water resources utilization and management, and assessment of glacier-related hazards in the HMA region.