Glacier Area and Surface Flow Velocity Variations for 2016–2024 in the West Kunlun Mountains Based on Time-Series Sentinel-2 Images

Abstract

The West Kunlun Mountains (WKL) gather lots of large-scale glaciers, which play an important role in the climate and freshwater resource for central Asia. Despite extensive studies on glaciers in this region, a comprehensive understanding of inter-annual variations in glacier area, flow velocity, and terminus remains lacking. This study used a deep learning model to derive time-series glacier boundaries and the sub-pixel cross-correlation method to calculate inter-annual surface flow velocity in this region from 71 Sentinel-2 images acquired between 2016 and 2024. We analyzed the spatial-temporal variations of glacier area, velocity, and terminus. The results indicate that, as follows: (1) The glacier area in the WKL remained relatively stable, with three glaciers expanding by more than 0.5 km² and five glaciers shrinking by over 0.5 km² from 2016 to 2024. (2) Five glaciers exhibited surging behavior during the study period. (3) Six glaciers, with velocities exceeding 50 m/y, have the potential to surge. (4) There were eight obvious advancing glaciers and nine obvious retreating glaciers during the study period. Our study demonstrates the potential of Sentinel-2 for comprehensively monitoring inter-annual changes in mountain glacier area, velocity, and terminus, as well as identifying glacier surging events in regions beyond the study area.

1. Introduction

The Western Kunlun Mountains (WKL), situated on the northwestern Tibetan Plateau, is home to one of the highest concentrations of glaciers. The meltwater from these glaciers is a crucial water source for Central Asia, particularly for southern Xinjiang [1]. In the context of global warming, glaciers across the Tibetan Plateau are experiencing a negative mass balance [2], with those in the Himalayas showing a significant mass loss [3,4,5]. In contrast, glaciers in the Pamirs, Karakoram, and WKL have exhibited a relatively limited retreat, with some even advancing or maintaining a positive mass balance [2,4,6,7].

Previous studies on glacier dynamic variations in the WKL have mainly focused on elevation, area, velocity, and mass balance over long-time intervals or individual subjects in isolation. Zhang et al. [8] analyzed changes in glacier area and surrounding lake areas using Landsat images from 1991 to 2009, along with NECP/NCAR reanalysis climate data. Zhang et al. [1] analyzed the outlines and length variations of glaciers in the WKL at 5-year intervals from 2000 to 2020, using Landsat images and a combination of the band ratio method with manual interpretation. Wang et al. [9] analyzed the glacier velocity changes for 2014–2023 in the WKL using Sentinel-1 Synthetic Aperture Radar (SAR) images and the Feature-Tracking method. Guan et al. [10] updated the inventory of a surge glacier in the WKL using Landsat images, Sentinel-1 SAR images, and Digital Elevation Models (DEMs) spanning from 1972 to 2020. Zhou et al. [2] used multi-temporal DEMs to analyze glacier mass balance in the Tibetan Plateau and found that glaciers in the WKL have remained nearly stable or experienced slight mass gains over the past four decades (from the mid-1970s to the mid-2010s). Wang et al. [11] studied the glacier anomaly over the WKL through analyzing glacier length, area, and DEM changes between the mid-1970s and the mid-2000s. Luo et al. [12] investigated climate-driven glacier anomalies in the WKL by analyzing ASTER DEMs and ERA5 reanalysis data. However, there has been a lack of comprehensive time-series knowledge of inter-annual variations in glacier area, flow velocity, and terminus, which are influenced by factors such as regional precipitation, temperature, and basal friction [9]. Such monitoring is essential for the early identification of glacier-related disaster risks, long-term water security management, and the study of ecological impacts [13,14].

Glacier boundaries are generally extracted through manual visual interpretation or the Red/SWIR (shortwave infrared) band ratio method, which relies on spectral differences in satellite images [15]. While a manual interpretation can provide relatively accurate results, it is time-consuming and labor-intensive, especially over large regions [16]. A global threshold of the Red/SWIR band ratio results is challenging to determine, and the accuracy of the glacier extraction results requires further improvement. Deep learning algorithms have been widely used for landform identification. Among open-source deep learning models, the DeepLab V3+ algorithm has demonstrated exceptional performance in identifying glacial and periglacial landforms [17,18,19].

Traditional approaches to monitoring glacier flow velocity primarily rely on in situ measurements. Although these ground-based methods offer a high accuracy, they are costly, hard to deploy, and constrained in terms of spatial coverage [20,21]. Satellite remote sensing technology offers an effective solution for monitoring the large-scale, repeated monitoring of glacier surface flow velocity, especially for remote regions. Although SAR image pairs, combined with Interferometric SAR (InSAR) and offset-tracking techniques, offer advantages for monitoring glacier surface velocity in all weather conditions [9,22,23], the side-view imaging geometry in the slant-range direction can lead to issues such as foreshortening, layover, and radar shadows in rugged terrains, like mountain glaciers [24]. Optical satellites can avoid the afore-mentioned challenges, and provide a significant alternative method for monitoring glacier flow velocity [25]. The Sentinel-2 satellites, offering 10-day repeat observations with a single satellite and 5-day repeat observations with the satellite constellation, provide a valuable data source for monitoring glacier area and surface velocity [26,27,28,29,30,31].

The objective of this study is to utilize Sentinel-2 optical images to comprehensively analyze variations in glacier area, flow velocity, and terminus in the WKL from 2016 to 2024, identifying ongoing and potential glacier surge events.

2. Study Area

The WKL is one of the highest concentrations of large-scale mountainous glaciers, situated on the northwestern Tibetan Plateau (Figure 1). Elevations in the WKL range from approximately 4800 m to 7167 m, with the Kunlun Peak as its summit [1,32]. The WKL can be divided into southern and northern slopes by its ridge. The southern slope is relatively gentle and mainly features wide-tailed valley glaciers, cirque glaciers, and ice caps, while the northern slope has longer and steeper glaciers, primarily composed of compound and dendritic valley glaciers [11,33]. Based on the glacier inventories derived in this study, Duofeng Glacier, with a length of 25.7 km and an area of 243.7 km² in 2024, is the largest glacier in the region. These glaciers play a crucial role in supplying meltwater to the rivers and lakes in the surrounding areas. Sentinel-2 satellite images (tiles T44SME, T44SNE, and T44SMD) provide full coverage of this glacier-rich region.

The region is predominantly influenced by the mid-latitude westerlies, resulting in a cold and semi-arid climate [11]. Based on short-term field observation data conducted in 1987 and 1989, the snow line was located at an elevation of 5930 m, where the annual average temperature was −13.9 °C and total annual precipitation was ~300 mm [33,34].

3. Data and Methods

3.1. Data

3.1.1. Sentinel-2

Sentinel-2 satellite images, used to obtain boundaries and surface flow velocities of the glaciers in this study, are available for free download from “Copernicus Browser” (https://browser.dataspace.copernicus.eu (accessed on 9 January 2025)). The Sentinel-2 satellites provide high-resolution multispectral images from its sun-synchronous orbit at 786 km altitude. Equipped with a Multispectral Imager (MSI), it captures data across 12 spectral bands, ranging from visible and near-infrared to short-wave infrared. The ground resolutions are 10 m, 20 m, and 60 m, respectively. The near-infrared band (Band 8), red band (Band 4), green band (Band 3), and blue band (Band 2) have a resolution of 10 m. The first satellite, Sentinel-2A, was launched on 23 June 2015, followed by its twin, Sentinel-2B, on 7 March 2017. A single satellite has a revisit cycle of 10 days, while a constellation of two satellites reduces the revisit period to 5 days. This high revisit frequency allows Sentinel-2 to collect a vast amount of observational data over time.

To minimize snow-related interference, we selected images from the ablation season (primarily July to September) between 2016 and 2024 to delineate glacier boundaries and calculate interannual flow velocities. The information of images used in our study is presented in

Table 1. The selected Sentinel-2 images for the delineation of glacier boundaries.

Year	Main Images (Tile Id-YYYYMMDD)	Auxiliary Images (Tile Id-YYYYMMDD)
2016	T44SME-20160721, T44SNE-20160721, T44SMD-20160721	T44SME-20160919, T44SNE-20160919, T44SNE-20160522, T44SMD-20160522
2017	T44SME-20171009, T44SNE-20170815, T44SMD-20171009	T44SME-20170815, T44SME-20170919, T44SNE-20171009, T44SMD-20170815
2018	T44SME-20180711, T44SNE-20180830, T44SMD-20180830	T44SME-20180919, T44SME-20181004, T44SNE-20180711, T44SMD-20180522
2019	T44SME-20190904, T44SNE-20190904, T44SMD-20190904	T44SME-20190706, T44SME-20190909, T44SME-20190919, T44SNE-20190914, T44SNE-20190919, T44SMD-20190909
2020	T44SME-20200903, T44SNE-20200903, T44SMD-20200903	T44SME-20200630, T44SME-20200824, T44SNE-20200814, T44SNE-20200824, T44SMD-20200814, T44SMD-20200630
2021	T44SME-20210819, T44SNE-20210819, T44SMD-20210819	T44SME-20210710, T44SME-20210908, T44SNE-20210908, T44SMD-20210908
2022	T44SME-20220913, T44SNE-20220804, T44SMD-20220804	T44SME-20220625, T44SME-20220705, T44SNE-20220819, T44SMD-20220725, T44SMD-20220819, T44SMD-20220913
2023	T44SME-20230913, T44SNE-20230913, T44SMD-20230913	T44SME-20230625, T44SME-20230715, T44SME-20230903, T44SNE-20230903, T44SMD-20230903, T44SMD-20230715
2024	T44SME-20240917, T44SNE-20240907, T44SMD-20240907	T44SME-20240922, T44SNE-20240813, T44SMD-20240729, T44SMD-20240719

and

Table 2. The selected Sentinel-2 images for the calculation of glacier surface flow velocities.

Glacier Velocity	Pre-Phase Images (Tile Id-YYYYMMDD)	Post-Phase Images (Tile Id-YYYYMMDD)
2016–2017	T44SME-20160721 T44SNE-20160721 T44SMD-20160721	T44SME-20171009 T44SNE-20170815 T44SMD-20171009
2017–2018	T44SME-20171009 T44SNE-20170815 T44SMD-20171009	T44SME-20180711 T44SNE-20180830 T44SMD-20180830
2018–2019	T44SME-20180711 T44SNE-20180830 T44SMD-20180830	T44SME-20190904 T44SNE-20190904 T44SMD-20190904
2019–2020	T44SME-20190904 T44SNE-20190904 T44SMD-20190904	T44SME-20200903 T44SNE-20200903 T44SMD-20200903
2020–2021	T44SME-20200903 T44SNE-20200903 T44SMD-20200903	T44SME-20210819 T44SNE-20210819 T44SMD-20210819
2021–2022	T44SME-20210819 T44SNE-20210819 T44SMD-20210819	T44SME-20220913 T44SNE-20220804 T44SMD-20220804
2022–2023	T44SME-20220913 T44SNE-20220804 T44SMD-20220804	T44SME-20230913 T44SNE-20230913 T44SMD-20230913
2023–2024	T44SME-20230913 T44SNE-20230913 T44SMD-20230913	T44SME-20240917 T44SNE-20240907 T44SMD-20240907

. Band 8 has a significantly larger bandwidth (145 nm for S2A and 133 nm for S2B) compared to Band 4 (38 nm for S2A, 39 nm for S2B), Band 3 (45 nm for S2A, 46 nm for S2B), and Band 2 (98 nm for both S2A and S2B). This broader bandwidth enhances spatial detail, improves image clarity, and enables the detection of finer variations [35]. Therefore, we chose Band 8 for the calculation of glacier surface flow velocity. To enhance the effective use of spectral and spatial information from Sentinel-2 images, we employed the near-infrared, red, green, and blue bands to identify the glacier boundaries. This study utilized Level 1C and Level 2A products from both Sentinel-2A and Sentinel-2B. All images were orthorectified during preprocessing prior to delineating glacier boundaries and calculating glacier surface flow velocities.

3.1.2. SRTM DEM

The Shuttle Radar Topography Mission (SRTM) 1 arc-second C-band DEM, provided from NASA’s Earth Data platform (https://search.earthdata.nasa.gov (accessed on 9 January 2025)), was utilized to acquire glacier divides, glacier centerlines, and slopes in this study. The SRTM mission was conducted from 11 February 2000 to 22 February 2000. A high-quality DEM dataset for over 119 million km² coverage (more than 80% of global land) between 60° N and 60° S was generated after more than two years of data processing [36,37]. The spatial resolution of the SRTM 1 arc-second C-band DEM is around 30 m [38]. The SRTM DEM dataset has been publicly available and free for global access since 2003.

3.1.3. ERA5-Land Data

The ERA5-Land dataset was used to analyze temporal variations in temperature and precipitation in the WKL. This high-resolution global atmospheric reanalysis dataset, provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), possesses a spatial resolution of 0.1° (approximately 9 km) and an hourly temporal resolution [39]. The ERA5-Land dataset is well suited for in-depth analysis and provides valuable insights for climate change studies [40]. The dataset is freely available through the Copernicus Climate Data Store (CDS) platform (https://cds.climate.copernicus.eu (accessed on 9 January 2025)).

3.2. Methods

3.2.1. Glacier Boundary Delineation

The glacier delineation process is shown in Figure 2. Interpreted glacier records were obtained by manually delineating glaciers in Sentinel-2 images. These records, combined with four-band (Bands 8, 4, 3, and 2) composited Sentinel-2 images, were then converted into a sample dataset with an image size of 256 × 256 pixels (Figure 3). A total of 2503 labelled images were generated, with 80% (2002) allocated for training and the remaining 20% (501) for validation. To expand and diversify the training dataset, methods such as horizontal flipping, vertical flipping, diagonal flipping, 90° clockwise rotation, and 90° counterclockwise rotation were applied.

The glacier outlines within the WKL were extracted using the DeepLab V3+ neural network [41]. The network was trained with an initial learning rate of 0.002, a batch size of 32, and 100 epochs to ensure effective feature learning from the sample dataset. The trained model demonstrated a superior performance, achieving high accuracy and low error rates, as evidenced in its accuracy and loss values (Figure 4). The validation accuracy reached 96.1% at the 100th epoch. The binary raster images, generated as direct predictions by the model, were then vectorized to extract the outlines of the identified glaciers. These glacier outlines were subsequently smoothed and manually revised to better align with the actual glaciers using multi-temporal Sentinel-2 images. Glacier divides, extracted from SRTM DEM, were used to segment the glacier outlines, and finally the individual glacier boundaries were obtained [42].

3.2.2. Glacier Velocity Estimation

This study employed frequency a domain cross-correlation algorithm with the Co-registration of Optically Sensed Images and Correlation (COSI-Corr) package to estimate the glacier movements. The COSI-Corr package was developed by Leprince et al. [43] at the California Institute of Technology (Caltech) and released in 2015. Initially designed for ground deformation extraction with optical images, the package has been used to monitor glacier movements [44,45,46]. The accuracy of the estimated glacier surface movements depends on the image quality, typically achieving sub-pixel precision [46,47].

The data processing workflow for glacier velocity extraction is shown in Figure 5. The frequency-domain cross-correlation algorithm determines the relative position relationship of corresponding points between two images by applying the inverse Fourier transform to the normalized cross-power spectrum. Assuming there is a relative displacement $({\Delta }_x,{\Delta }_y)$ between the two image windows:

$$i_2(x,y)=i_1(x-{\Delta }_x,y-{\Delta }_y)$$

(1)

where $i_1$ and $i_2$ represent the reference image and the secondary image, respectively. After performing a Fourier transform on them, Equation (1) can be expressed as follows:

$$I_2({\omega }_x,{\omega }_y)=I_1({\omega }_x,{\omega }_y)e^{-j({\omega }_x{\Delta }_x+{\omega }_y{\Delta }_y)}$$

(2)

where $I_1$ and $I_2$ are the Fourier transform of $i_1$ and $i_2$, and their normalized cross-power spectrum is denoted as follows:

$$C_{i_1{i}_2}\left({\omega }_x,{\omega }_y\right)={\displaystyle \frac{I_1({\omega }_x,{\omega }_y)I_2^{\ast }({\omega }_x,{\omega }_y)}{\left|I_1({\omega }_x,{\omega }_y)I_2^{\ast }({\omega }_x,{\omega }_y)\right|}}=e^{j({\omega }_x{\Delta }_x+{\omega }_y{\Delta }_y)}$$

(3)

After performing the Fourier transform, we can obtain the following:

$$F^{-1}=\left\{e^{j({\omega }_x{\Delta }_x+{\omega }_y{\Delta }_y)}\right\}=\delta \left(x+{\Delta }_x,y+{\Delta }_y\right)$$

(4)

By taking the inverse Fourier transform, we can obtain the peak points of the surface to determine the displacement between ${\Delta }_x$ and ${\Delta }_y$.

Cross-correlation was initially performed using the 128-pixel window to derive approximate glacier movements. This was followed by a refined cross-correlation using the 32-pixel window to estimate glacier movements at the sub-pixel level. The step length was set to 2 pixels × 2 pixels (with a ground resolution of 20 m × 20 m), and the number of iterations was set to 2. The east/west horizontal displacement $D_{E-W}$ and south/north horizontal displacement $D_{S-N}$, as well as the signal-to-noise ratio (SNR), were obtained between the two images’ acquisition time $t$.

Due to variations in optical conditions between the two images’ acquisition time, such as changes in solar altitude and azimuth angles, cloud coverage, and satellite attitude, the glacier surface velocity estimated by the COSI-Corr package contained noises and errors. To effectively improve the preliminary velocity data, it is necessary to employ post-processing, including decorrelation noise elimination, striping errors elimination, a non-local means filtering, and a mask with an SNR greater than 0.9. Finally, the glacier surface displacement $D$ and velocity $V$ were achieved using Equations (5) and (6):

$$D=\sqrt{D_{E-W}^2+D_{S-N}^2}$$

(5)

$$V=D/t$$

(6)

3.2.3. Uncertainty Assessment

Assuming that the glacier boundaries passed through the center of each pixel, the uncertainty of each glacier area was estimated using the image spatial resolution and the length of the glacier boundary [3]. The computation formula is as follows:

$${\delta }_{area}={\displaystyle \frac{1}{2}}\times l\times r$$

(7)

where $l$ is the length of a glacier boundary, and $r$ is the satellite image resolution.

Errors in the glacier surface flow velocity results may arise from factors such as image quality, image registration, the cross-correlation algorithm, and terrain-associated offsets [46,48]. However, due to the remote location of the study area, in situ measurements of glacier flow velocity are not available [9], and directly evaluating these errors is challenging. Assuming that no displacement occurs in the flat (slope < 30°) ice-free areas during the acquisition time interval between the reference image and the secondary image, the deformation error in the off-glacier region can be used to assess the velocity error in the glacier region [46]. The calculation formula is as follows:

$$e_{off}=\sqrt{{\left({\displaystyle \frac{ST{D}_{off}}{\sqrt{N_{off}}}}\right)}^2+Mea{n}_{off}^2}$$

(8)

where $e_{off}$ represents the motion velocity error in the flat off-glacier region, $ST{D}_{off}$ denotes the standard deviation of the flow velocity in the flat off-glacier region, and $Mea{n}_{off}$ is the mean motion velocity in the flat off-glacier region. $N_{off}$ is the number of effective measurements required to eliminate the influence of autocorrelation in the flat off-glacier region, and the calculation formula is as follows:

$$N_{off}={\displaystyle \frac{n_{total}\times r}{2d}}$$

(9)

where $n_{total}$ represents the total number of pixels in the flat off-glacier region, $r$ stands for the pixel resolution, and $d$ is the spatial autocorrelation distance used to eliminate the influence of autocorrelation, typically set to 20 times the pixel resolution [49].

4. Results

4.1. Characteristics of Glacier Area Change

From 2016 to 2024, inter-annual change rates in the glacier area were less than 0.3%, exhibiting a relatively stable pattern (Figure 6). Three glaciers exhibited an area increase exceeding 0.5 km² and were classified as “increasing”, whereas five glaciers experienced an area reduction greater than 0.5 km² and were classified as “decreasing” (Figure 7). Glaciers with area changes of less than 0.5 km² were categorized as “others”.

4.2. Characteristics of Glacier Velocity Change

The average inter-annual glacier surface flow velocity of the WKL was from 0 m/y to 188.8 ± 0.84 m/y during 2016–2024 (Figure 8). Eleven glaciers were more active than the others, with maximum values of the average inter-annual surface flow velocity exceeding 50 m/y (

Table 3. Maximum average inter-annual surface flow velocities of the eleven active glaciers during 2016–2024.

Glacier Name	Surface Flow Velocity (m/y)	Glacier Name	Surface Flow Velocity (m/y)
Xikunlun West Glacier	115.2	Xikunlun East Glacier	69.1
N1 Glacier	72.4	The b2 (branch 2) of the Kunlun Glacier	75.2
N2 Glacier	101.8	Duofeng Glacier	92.4
N3 Glacier	57.5	N4 Glacier	72.4
N6 Glacier	188.8	Alakesayi Glacier	63.6
The b4 of the Zhongfeng Glacier	92.4

). The subplot presents velocity statistics calculated from 11,294,231 pixels across the flat ice-free areas. The results reveal an average bias of 0.84 ± 1.61 m in inter-annual glacier surface flow velocity across these areas during 2016–2024, which indicate the reliability of the velocity measurements in this study.

To understand the changes of annual glacier movement, the time-series inter-annual glacier flow velocity maps during 2016–2024 were generated for the WKL (Figure 9). Using inter-annual glacier flow velocity data, we analyzed velocity variations across 11 glaciers with high surface flow rates.

The surface flow velocity of the Xikunlun West Glacier exhibited a fluctuating pattern during 2016–2024. The velocity reached the first peak of 259.1 m/y during 2019–2020, and then slowed down by 2021–2022, as can be seen in Figure 10a. During 2022–2023, the velocity reached the second peak of 214.2 m/y. The b2 of the Xikunlun West Glacier showed more obvious changes in surface flow velocity compared to the b1 (branch 1) of the Xikunlun West Glacier, as can be seen in Figure 9. We believe that the b1 of the Xikunlun West Glacier experienced surging events during 2019–2020 and during 2022–2023 twice.

The surface flow velocity of the Xikunlun East Glacier, N1 Glacier, and Alakesayi Glacier generally experienced a similar pattern, with an acceleration during 2016–2017 and then deceleration, with a peak velocity of around 168.2 m/y, 202.7 m/y, and 178.1 m/y, respectively, as can be seen from Figure 10b–d. We believe that the Xikunlun East Glacier, N1 Glacier, and Alakesayi Glacier likely underwent surging events during 2016 or prior to 2016. Particularly, in the Alakesayi Glacier, the rapid movements of the upstream high-velocity zone shifted to the downstream, leading to a substantial increase in downstream velocity, as can be seen in Figure 10d.

The b4 of the Zhongfeng Glacier consistently maintained a high velocity, exhibiting a rapid acceleration during 2022–2023 and peaking at 224.4 m/y in 2023–2024, as can be seen in Figure 10e. We believe that the b4 of the Zhongfeng Glacier experienced surging events during 2022–2024.

The b2 of the Kunlun Glacier, N2 Glacier, Duofeng Glacier, N3 Glacier, N4 Glacier, and N6 Glacier exhibited similar surface flow velocity patterns, maintaining consistently high and relatively stable velocities throughout the study period (2016–2024), as can be seen from Figure 10f–k. The b2 of the Kunlun Glacier, N2 Glacier, Duofeng Glacier, N3 Glacier, N4 Glacier, and N6 Glacier had the maximum inter-annual velocities greater than 64.4 m/y, 79.8 m/y, 102.7 m/y, 71.1 m/y, 69.6 m/y, and 167.1 m/y, respectively. Spatially, the velocity of the N6 Glacier exhibited an interesting radial pattern, with the middle part of the glacier significantly being more active than the upstream and the terminus of the glacier, as can be seen in Figure 10k.

4.3. Glacier Terminus Change

In the WKL, there were eight obvious advancing glaciers including the Xikunlun West Glacier, Xikunlun East Glacier, N1 Glacier, N2 Glacier, N4 Glacier, N6 Glacier, Alakesayi Glacier, and Zhongfeng Glacier, with nine obvious retreating glaciers including the Xiyulong Glacier, Kunlun Glacier, Duofeng Glacier, N5 Glacier, Yulong Glacier, Chongce Glacier, S1 Glacier, Gongxing Glacier, and Duota Glacier, of which we would perform a detailed analysis.

The terminus of the Xikunlun West Glacier remained relatively stable from 2016 to 2019, with no significant advance or retreat. However, a notable advance began during 2019–2020, peaking at 369.5 m/y during 2020–2021 before gradually decelerating (Figure 11a). Notably, the rapid terminus advanced during 2020–2021 may be triggered by a surging event during 2019–2020. By 2022, the glacier’s terminus began to compress the Xikunlun East Glacier, with the compression distance reaching 187.6 m by 2024. Meanwhile, the area of the Xikunlun West Glacier experienced a slight retreat from 2016 to 2019, followed by a more significant expansion beginning in 2019. A sharp increase in area occurred during 2020–2021, after which the rate of expansion slowed. The sharp expansion of the Xikunlun West Glacier during 2020–2021 may be also caused by the surge event during 2019–2020.

From 2016 to 2024, both the Xikunlun East Glacier and N1 Glacier exhibited similar advancing trends characterized by progressively declining rates. During this period, the average advancing rates for the Xikunlun East Glacier and N1 Glacier were 40.6 m/y and 28.9 m/y, respectively (Figure 11b). The continuous terminus advancing of the two glaciers may be related to the earlier surging events. Meanwhile, both glaciers exhibited an overall increase in area, although the rate of expansion slowed over time. Notably, the Xikunlun East Glacier experienced a decrease in area between 2021 and 2024, primarily due to the compression from the Western Kunlun West Glacier.

The terminus of the Xiyulong Glacier showed a slight retreat from 2016 to 2020, but a more significant retreat during 2020–2021 and 2022–2023 (Figure 11c). Overall, the area of the Xiyulong Glacier exhibited a shrinkage trend from 2016 to 2024, with a notable peak in 2020–2021, during which the decreasing area reached 0.33 km².

From 2016 to 2024, the Kunlun Glacier experienced an overall retreating trend, with an average annual shrinkage area of about 0.06 km². However, the b2 of the glacier consistently maintain an advancing trend, with an average annual advancing rate of 11.6 m/y. During this process, the b2 of the glacier exerted pressure on the b1, causing a significant deformation of the b1 (Figure 11d).

From 2016 to 2024, the N2 Glacier consistently maintained an expansion trend, with an average annual increasing rate of 0.03 km²/y. The advancing rate of the glacier terminus remained relatively stable, with an averaging advancing rate of 35.6 m/y (Figure 11e). The Duofeng Glacier showed an overall shrinkage trend, with a pattern of acceleration and deceleration (Figure 11f). Between 2019 and 2021, the glacier area decreased significantly, with an average shrinkage rate of 0.05 km²/y. The average retreating rate of the glacier terminus was 33.8 m/y.

The N4 Glacier advanced continuously from 2016 to 2024, with an average advancing rate of 16.3 m/y (Figure 11g). The glacier area exhibited a significant increase during 2016–2017, after which it stabilized in an oscillatory pattern. The terminus of the N5 Glacier continuously retreated from 2016 to 2024, with an average retreating rate of 43.2 m/y (Figure 11h). The glacier area decreased consistently, with its shrinkage rate accelerating to 0.12 km²/y by 2023–2024.

The terminus of the N6 Glacier exhibited a steady forward movement from 2016 to 2024, advancing at an average rate of 22.4 m/y (Figure 11i). Throughout this period, the glacier area remained relatively stable. Meanwhile, the Yulong Glacier underwent a continuous retreat, with an average retreating rate of 30.5 m/y. The area consistently shrank, with an acceleration pattern. By 2023–2024, the shrinkage rate of the glacier area had reached 0.18 km²/y.

From 2016 to 2024, the terminus of the Alakesayi Glacier consistently advanced. The advancing rate was 320.9 m/y during 2016–2017, but it gradually slowed, reaching just 9.6 m/y by 2023–2024 (Figure 11j). The glaciers’ initial significant terminus advance could be related to the earlier surging events. The glacier area continued to expand, with an increasing rate of 1.29 km² during 2016–2017, although the increasing rate gradually declined thereafter.

The terminus of the Chongce Glacier consistently retreated, with an average retreating rate of 22.3 m/y from 2016 to 2024 (Figure 11k). The glacier area showed an overall shrinkage trend, with an accelerating pattern, reaching 0.56 km²/y during 2023–2024. The terminus of the S1 Glacier also experienced a continuous retreat, with an average retreating rate of 9.7 m/y from 2016 to 2024. The glacier area remained relatively stable, with a low shrinkage rate.

Between 2016 and 2024, the terminus at the b1 of the Zhongfeng Glacier showed a continuous retreat, while the terminuses at the b2, the b3, and the b4 experienced an ongoing advance (Figure 11l). Particularly, the b4 of the Zhongfeng Glacier exhibited a significant advance during 2022–2023. Overall, the glacier area maintained a relative stability, with the exception of an obvious expansion during 2022–2023. We believe that the obvious terminus advancing and area increasing was related to the surging event.

The terminus of the Gongxing Glacier exhibited a sustained retreat at an average rate of 64.1 m/y between 2016 and 2024 (Figure 11m). The glacier area generally showed a shrinkage trend, with an average annual loss of 0.17 km²/y. Meanwhile, the Duota Glacier experienced a sustained terminus retreat at an average rate of 20.1 m/y (Figure 11n), while its area showed gradual shrinkage with an annual loss of 0.05 km².

5. Discussion

5.1. A Comparison with Previous Glacier Area and Velocity Studies

To validate the reliability of our glacier terminus and area change analysis, we compared our results with previously published results derived from similar time periods (

Table 4. A comparative analysis of glacial changes: area fluctuations, terminus positions, and surface velocities.

Aspect	This Study (2016–2024)	Previous Studies	Similarities	Differences
Area/terminus	Derived from Sentinel-2 optical images (2016–2024)	Zhang et al. [1]: derived from Landsat series optical images (2015–2020)	Gongxing Glacier, Xiyulong Glacier, Yulong Glacier, Chongce Glacier, Kunlun Glacier, b1 of the Zhongfeng Glacier, Duofeng Glacier	N5 Glacier, S1 Glacier, Duota Glacier
Velocity	Derived from Sentinel-2 optical images (2016–2024)	Wang et al. [9]: derived from Sentinel-1 SAR images (2014–2023)	Alakesayi Glacier, b4 of the Zhongfeng Glacier, Xikunlun West Glacier, N2 Glacier, Duofeng Glacier, b2 of the Kunlun Glacier	Void regions in SAR-based velocity maps

). Prior to our study, Zhang et al. [1] reported the terminus and area changes of the Gongxing Glacier, Xiyulong Glacier, Yulong Glacier, Chongce Glacier, Kunlun Glacier, b1 of Zhongfeng Glacier, and Duofeng Glacier. These glaciers all exhibited a shrinkage trend in area and experienced terminal retreats between 2015 and 2020. We derived the area and terminus changes of these glaciers for the period of 2016–2024, which generally agreed with the results of Zhang et al. [1], supporting the reliability of our glacier area and terminus changes. Our study found that the N5 Glacier, S1 Glacier, and Duota Glacier exhibited retreating trends, which were not reported in the results of Zhang et al. [1]. This discrepancy is likely because each area of these three glaciers is less than 100 km², which fell below the size threshold considered in the analysis of Zhang et al. [1]. Additionally, slight differences in retreat rates and area changes may arise from differences in the time periods analyzed (2016–2024 vs. 2015–2020) and the resolution of the images used (Sentinel-2 optical images vs. Landsat series optical images). Our extended observational period enables the detection of recent glacier dynamics, revealing updated patterns of area change, the terminus position, and the flow velocity that were previously undocumented.

To validate the reliability of our glacier velocity estimation results, we compared them with previously published results obtained during similar time periods (Table 4). Before our study, Wang et al. [9] reported the velocities of the Alakesayi Glacier, the b4 of the Zhongfeng Glacier, the Xikunlun West Glacier, the N2 Glacier, the Duofeng Glacier, and the b2 of the Kunlun Glacier. The Alakesayi Glacier reached its peak velocity in 2016, after which it declined. The upstream high-velocity zone shifted to the downstream, significantly increasing the downstream velocity. The b4 of the Zhongfeng Glacier entered a fast-moving phase in 2020, peaked in 2022, and then decelerated in 2023. The Xikunlun West Glacier exhibited a pattern of initial acceleration and subsequent deceleration. The N2 Glacier, Duofeng Glacier, and the b2 of the Kunlun Glacier showed consistently high and stable velocities. The velocity of these glaciers for 2016–2024, derived in our study as described in Section 4.2, generally agreed with the results of Wang et al. [9], supporting the reliability of our glacier velocity estimations. Notably, the glacier surface flow velocity maps obtained from Sentinel-2 optical images in our study show no void regions, whereas those generated from Sentinel-1 SAR images exhibit many void regions [9]. The absence of void regions in our Sentinel-2-based velocity maps, compared to the void regions in Sentinel-1 SAR-based maps, highlights the advantage of optical images in mountainous regions. Therefore, SAR imaging issues (e.g., foreshortening, layover, and radar shadows) are avoided in optical data, providing a more complete and reliable coverage.

5.2. Surge Glaciers During the Study Period

Previous studies have identified numerous surging glaciers in the WKL, including the Xikunlun West Glacier, Xikunlun East Glacier, N1 Glacier, Xiyulong Glacier, Kunlun Glacier, N2 Glacier, Duofeng Glacier, N3 Glacier, N4 Glacier, N5 Glacier, N6 Glacier, Yulong Glacier, Xiezhi Glacier, Alakesayi Glacier, Chongce Glacier, Zhongfeng Glacier, Gongxing Glacier, Duota Glacier, and Quanshui Glacier [10,50]. Some of these glaciers, such as the Xiyulong, N5, Yulong, Xiezhi, Chongce, Gongxing, Duota, and Quanshui glaciers, exhibited low velocities, suggesting they may have been in their quiescent phase during the study period. In contrast, other glaciers showed significant changes in annual velocity, indicating that they may have been in their active phase.

Based on the analysis of the glacier area, inter-annual velocity, and terminus variations in Section 4.1, Section 4.2, and Section 4.3, five glaciers exhibited surging behavior during the study period: the b2 of the Xikunlun West Glacier, the Xikunlun East Glacier, N1 Glacier, Alakesayi Glacier, and the b4 of the Zhongfeng Glacier. This observation aligns with the conclusions of Wang et al. [9], further supporting the reliability of the estimations and analysis presented in our study. The b2 of the Kunlun Glacier, N2 Glacier, Duofeng Glacier, N3 Glacier, N4 Glacier, and N6 Glacier exhibited higher velocities (greater than 50 m/y) compared to the five surging glaciers in their quiescent phase (less than 30 m/y), suggesting that these glaciers may have the potential to surge.

5.3. The Relationship Between Climate Change and Glacier Dynamics

We analyzed long-term trends in precipitation and temperature in the WKL using the ERA5-Land dataset, revealing an increase of approximately 15.8% in annual precipitation and a rise of about 7.1% in temperature, both exhibiting fluctuating patterns over the past 43 years. We observed that changes in the total glacier area in the region were closely correlated with variations in temperature and precipitation during the period 2016–2024. The total glacier area in the WKL increased during the periods of 2016–2017, 2018–2019, and 2022–2023 (Figure 6). Correspondingly, sharp drops of the annual average temperature were observed during these periods, while a similar change pattern in the annual total precipitation lagged behind, showing no significant drop yet (Figure 12). In contrast, during other periods when temperatures either rose rapidly or remained stable, the total glacier area in the WKL experienced varying degrees of shrinkage.

The glaciers in the WKL appeared to have maintained near-equilibrium or slightly positive mass balance conditions from at least the mid-1970s until 2015, likely due to the significant precipitation brought by westerlies [2,4,6,7,51,52,53]. Between 2000 and 2020, glacier area changes in the WKL were relatively unobvious on the whole, with a slight increase observed in north-facing glaciers but decreases in glaciers facing other orientations [1]. Despite this stability, the Tibetan Plateau is projected to experience sustained warming in the near-term, mid-term, and long-term future, as indicated by CMIP6 (Coupled Model Intercomparison Project Phase 6) models [54]. Precipitation variations are closely linked to temperature changes [55], and glacier surges in the study area appear to be influenced by a combination of hydrological and thermal mechanisms [9]. Therefore, based on historical and projected temperature and precipitation trends, surge events are likely to become more frequent in the coming decades in the study area, while the glacier mass may remain relatively stable.

6. Conclusions

In this study, we utilized Sentinel-2 images to generate annual glacier boundaries using a deep learning model and to estimate inter-annual glacier surface flow velocities through a sub-pixel cross-correlation method in the West Kunlun Mountains (WKL) from 2016 to 2024. Our analysis revealed that the total glacier area in the WKL remained relatively stable, with change rates of less than 0.3% over the study period. However, notable variations were observed at the individual glacier level: three glaciers experienced an area increase exceeding 0.5 km², while five glaciers experienced a comparable degree of shrinkage.

We identified several glaciers exhibiting dynamic behavior, including surging, advancing, and retreating trends. For instance, the branches b2 of the Xikunlun West Glacier, Xikunlun East Glacier, N1 Glacier, Alakesayi Glacier, and the b4 of the Zhongfeng Glacier displayed surging behavior. Additionally, the branches b2 of the Kunlun Glacier, N2 Glacier, Duofeng Glacier, N3 Glacier, N4 Glacier, and N6 Glacier showed higher flow velocities (exceeding 50 m/y) compared to surging glaciers in their quiescent phase (less than 30 m/y), suggesting their potential to surge in the future. In terms of terminus changes, eight glaciers, including the Xikunlun West Glacier, Xikunlun East Glacier, N1 Glacier, N2 Glacier, N4 Glacier, N6 Glacier, Alakesayi Glacier, and Zhongfeng Glacier, exhibited a clear advancing behavior, while nine glaciers, including Xiyulong Glacier, Kunlun Glacier, Duofeng Glacier, N5 Glacier, Yulong Glacier, Chongce Glacier, S1 Glacier, Gongxing Glacier, and Duota Glacier, showed a significant retreat during the study period.

A key advantage of our approach is the use of Sentinel-2 optical images, which enables the generation of glacier flow velocity maps without the void regions commonly present in maps derived from Sentinel-1 SAR data. Future research could focus on integrating multi-source remote sensing data, such as combining optical images, SAR data, and ICESat-1/2 elevation measurements, to further improve the accuracy and coverage of glacier monitoring.