Delineation Evaluation and Variation of Debris-Covered Glaciers Based on the Multi-Source Remote Sensing Images, Take Glaciers in the Eastern Tomur Peak Region for Example

Abstract

As a particular type of alpine glacier, debris-covered glaciers are essential for local water resources and glacial disaster warnings. The Eastern Tomur Peak Region (EPTR) is the most concentrated glacier in Tien Shan Mountain, China, where the glaciers have not been studied in detail. This paper evaluates the delineation accuracy of Landsat8 OLI, Sentinel-1A, and GF images for debris-covered glaciers in the EPTR. Each image uses the most advanced delineation method for itself to minimize the error of inherent resolutions. The results show that the accuracy of these images for delineating debris-covered glaciers is very high, and the F1 scores are expressed as 96.73%, 93.55%, and 95.81%, respectively. Therefore, Landsat images were selected to analyze the area change of EPTR from 2000 to 2022 over a 5-year time scale. The results indicate that glaciers of the EPTR decreased by 19.05 km² from 2000 to 2020, accounting for 1.9% (0.08% a⁻¹), and debris increased by 10.8%, which validates the opinion that the presence of debris inhibits glacier melting. The most varied time was 2010–2022, but it was much less than other Tien Shan regions. The lower glacier ablation rate in this area results from the combined effect of decreased bare ice and increased debris. The main reason for the change in debris-covered glaciers is the increase in temperature.

1. Introduction

Glaciers where the surface of the ablation zone is covered with a layer of debris are called debris-covered glaciers, which are widely distributed in arid alpine regions, such as the Cordillera in North and South America, the Alps in Europe, and the Himalayas, Karakorum Mountains, and Tien Shan mountains in Asia [1,2]. Under the effects of global warming, the majority of glaciers are experiencing different extents of ablation, which causes sea level rise, runoff circulation imbalance, and glacial disasters, such as ice avalanches and glacial mudflows [3,4], especially for debris-covered glaciers. It has an inevitable negative impact on the life of the population agriculture and economic development [5]. Debris affects glacier melting, mass balance, and glacier dynamics, significantly altering the rate of glacier melting [6]. It is influenced by the reflectivity of the subsurface, size, and thickness of grains and exhibits different characteristics from bare ice [7,8], which aids in its identification.

The development of remote sensing technology has enabled a significant step forward in glaciological research. High-resolution and time-continuous images can be obtained without going to the site in person [9,10,11]. For glaciological studies, the identification of glacier boundaries is the basis [12,13,14]. Currently, semi-automated methods for identifying glacier boundaries based on remote sensing include the Band Ratio method [15,16,17], Normalized Difference Snow Index (NDSI) method [18,19], Principal Component Analysis method [20], Object-oriented Analysis method [21], Synthetic Aperture Radar (SAR) interferometry technique [22], and Machine Learning [23]. However, none of the above methods can accurately identify and classify the boundary of debris-covered glaciers [24]. Due to the obvious differences in morphological structure between bare ice and debris, a method combining automatic classification of multispectral remote sensing data with terrain reconstruction analysis by the digital elevation model (DEM) and temperature is proposed, and good results were achieved [2,4]. It delineates glacial debris-covered areas with approximately 90% accuracy of the results [25]. In addition, Shukla presented the normalized difference debris index (NDDI) given the difference between the short-wave infrared band and thermal infrared bands, which was applied to Aster DEM with promising results [26]. The rapid development of deep learning techniques has made neural network-based image segmentation methods very effective for glacier delineation [27,28]. Synthetic Aperture Radar (SAR) is not influenced by weather and could be paired well with Deep Learning [29,30]. Some studies have combined Sentinel-1A and topographic data to compare the accuracy of debris identification between UNet and DeepLabV3+ networks for large debris-covered glaciers, such as Siachin and Baltok glaciers in the Southern Karakorum Mountains, and results had the highest accuracy of 93.34% [31]. Furthermore, UNet networks were utilized to extract Antarctic glaciers and ice shelf leading edges effectively and to study their deformation [32]. Visual interpretation is still the best method of interpreting debris-covered glacier boundaries [33]. However, it is time-consuming and labor-intensive [34]. If the accuracy is verified by sufficient visual interpretation, semi-automatic, or even fully automatic, delineation of boundaries based on remote sensing images is undoubtedly the best solution for glaciological data sources.

Previous studies have primarily focused on comparing different methods in delineating boundaries and comparing differences in deep learning networks [31,35]. Firstly, the basic difference between debris of glaciers and bare rocks is the presence of ice under the debris. Visual interpretation or higher resolution images cannot essentially solve the problem. Secondly, the sample data for deep learning are mainly derived from existing cataloged data, such as the second glacial catalogs of China or RGI 6.0 data, and the cataloged data itself is based on semi-automated delineation combined with human correction, meaning the data based on the deep learning UNet network cannot take maximum advantage of its 10 m spatial resolution [36,37]. In addition, remote sensing-based glaciological studies should not be limited by a single RS image type or a single research method because different images have different advantages for other research contents. For example, Landsat images are suitable for long-time series studies; SAR images are ideal for glacier flow rate and early warning of disasters, such as landslides; GF images with higher resolution are ideal for accurate mass balance studies and providing simulation parameters, etc. The purpose of using remote sensing technology to study regional glacier changes is to better understand the specific conditions of glacier indicators in the region, especially for areas such as EPTR, which have never been reviewed individually. It is worth mentioning that EPTR is the border zone between Kazakhstan, Kyrgyzstan, and China. The G219 Wensu-Zhaosu highway (West Tianshan extra-long tunnel section), which is to be built in China, passes through it. Considering the above aspects, our selected validation data are field measurements based on RTK-GPS (Real Time Kinematic-Global Positioning System) multi-point data positioning combined with GPR (Ground Penetrating Radar) to determine glacier thickness, which is very convincing [38,39]. The validation area is the glacier area of the EPTR, which not only has a large-scale debris-covered glacier, but has also not been studied as an independent study unit. This paper compares the three most advanced methods according to the obtained area index. In this study, not only was the method with the best debris delineation selected for the study of area variation, but the accuracy of other methods also provides a reference for the subsequent estimation of glacier reserves in the area, the simulation of material balance, calculation of the glacier flow rate, and the early warning of glacier disasters [40,41].

2. Study Region

The study area (42°N–42°30′N, 80°20′E–81°20′E) is located in the Eastern Tomur Peak Region (EPTR), which is the largest glacier aggregation center in the Tien Shan Mountains (Figure 1). Tomur Peak is the highest peak in Tien Shan Mountains (7435.5 m a.s.l.), and the glaciers in this area are specially named as Tomur-type glaciers because of the high debris coverage. The number of glaciers in this study is 275, with a total area of 996.67 km² [42]. The terrain is complex and varied, with large mountain drop-offs and extensive debris development. It is in the watershed between the Ili River Valley and the Tarim Basin, and the nearby river is the Muzart River. Glacial snow melt water mainly collects in the depressions of the study area. The climate type is a temperate continental climate. The climate can be characterized by severely cold winters, warm summers, and large annual and daily temperature differences.

3. Data and Methods

3.1. Data Source and Pre-Processing

To minimize the influence of snow and clouds, all remote sensing images were selected from June to September, with cloud coverage of less than 10. The optical images were pre-processed with radiation correction, atmospheric correction, and other related pre-processing. The improvement of spatial resolution and the retention of multispectral information can be achieved with the help of fused visible and panchromatic bands of Landsat OLI, with a resolution of 15 m after processing [43]. The VH polarization pattern of the IW mode GRD data of Sentinel-1A was adopted and pre-processed in SNAP. The field data were selected for the same time as the remote sensing images to minimize uncertainties as much as possible. Data from the national weather stations near the study area were collated to obtain the annual average temperature and precipitation.

Table 1. Data attributes, applications, and sources for study.

Category	Data	Date	Resolution	Application	Source
Optical Image	Landsat 7–9 ETM+/OLI	2000–2022	15 m/30 m	Boundary Delineation; Base Data for Area Change	http://earthexploorer.usgs.gov
	GF6_PMS	2022	2 m	Boundary Delineation;	——
	Sentinel-2A	2021; 2022	10 m	UNet Learning Label	https://scihub.copernicus.eu/
SAR Image	Sentinel-1A	2022	10 m	Boundary Delineation	https://scihub.copernicus.eu/
Field Data	RTK-GPS	2022		Terminus Position Base Data for Area Change	Field Survey
DEMs	ASTER GDEM V3	2000–2009	30 m	Topographic analysis	https://www.earthdata.nasa.gov/
	SRTM DEM	2000	30 m	Topographic supplement	http://gdex.cr.usgs.gov/gdex/
Glacier Inventory	Second Glacier Inventory of China	2017	30 m	Visual Reference	http://www.glims.org/download/
Meteorological data	Zhaosu/Aksu County	1960–2020	1 km	Climate Analysis	https://data.cma.cn/

shows all the data used in this paper:

3.2. Method

3.2.1. Delineation of Multi-Source RS Images

The details of this study include remote sensing-based glacier boundary delineation and accuracy evaluation based on field measurement data. The specific flow chart is shown in Figure 2.

1.

The TDSI method for Landsat Images

The TDSI (Temperature, NDDI, Slope, Ice) approach is a combination of surface temperature data, NDDI index, band ratio (NIR/SWIR1), and slope analysis to achieve the identification and delineation of the glacier and terminal debris [44]. The debris delineation is based on the intersection of slope analysis ≥ 24° [45] and surface temperature above the optimal threshold of 10 °C, where the slope analysis relies on the Toolbox in ArcMap, and the optimal threshold is selected by manual interaction. The numerical quantization values of GREEN, NIR/SWIR₁ bands in Landsat images were converted to reflectance; thermal infrared data were converted to radiance to calculate brightness temperature and blackbody radiance brightness. The surface temperature was calculated using Planck’s formula [26].

$$T_s=\frac{K_2}{\mathit{ln}\left(\frac{K_1}{B_s}+1\right)}$$

(1)

In Formula (1): K₁ and K₂ are constants, and Bs is the black body radiation brightness value.

It should be noted that if the surface temperature data combined with RGI 6.0 glacier boundary data could not delineate the debris area correctly, the NDDI index was used for delineation with a threshold range of [0.95, 0.99].

$$NDDI=\frac{{\rho }_{SWIR2}-{\epsilon }_{TIR}}{{\rho }_{SWIR2}+{\epsilon }_{TIR}}$$

(2)

In Formula (2): ${\rho }_{SWIR2}$ is the reflectance in the short-wave infrared band, and ε_TIR is the emissivity in the thermal infrared band.

The bare ice area delineation relies on the band ratio (NIR/SWIR₁) and manual interaction to take the optimal threshold (range with [0.0001, 0.0003]). The debris area and bare ice area were merged to obtain the debris-covered glacier extent, which is reclassified into a binary map and then denoised using a median filter (window size is 3 × 3, the unit is image size) to remove anomalies of the final result of the debris-covered glacier boundary [45], as shown in Figure 3. Since large debris-covered glaciers are widespread in the region, we used automatic ridge line extraction to segment and crop the glacier boundaries to obtain the vector boundary of the single glacier [33], which is convenient for subsequent glacier change studies.

2.

Deep Learning based on the UNet network for Sentinel-1A Images

UNet is a classical network for semantic segmentation, and VGG16 is a traditional convolutional neural network for feature extraction. The combination of UNet and VGG16 network can be effective for glacier identification. The UNet network has a U-shaped structure with two main components: the encoder and the decoder. The encoder reduces the dimensionality of the image and extracts features by convolution and pooling, while the decoder uses the same proportion of upsampled images as the feature part for stitching, combining shallow and deep elements to delineate debris more intuitively [31].

The VGG16 convolutional neural network model consists of 13 convolutional layers and 3 fully-connected layers, and it is used as the encoding part to extract glacier features, so only the first 13 convolutional layers are selected; the convolutional kernel size is 3 × 3, the step stride is 1, and the “SAME” method is used to fill in the excitation function. The activation function is the ReLU function, and its specific structure is shown in Figure 4.

The 4 parameters, backscatter coefficient, local incidence angle, DEM, and slope [46] under VH polarization of Sentinel-1A on 26 August 2022, were selected as the sample part for deep learning the sample data were restored to 10 m resolution. The visual interpretation of the Sentinel-2 glacier boundary with the same resolution was selected as the labeled part of the network for binary segmentation and the Sentinel-1A image was used for the evaluation of results. The sources of the data used to train the model are highlighted here. Due to the large memory of remote sensing images, we must segment and crop them to eliminate useless images to enhance the accuracy and speed of the study. Python can help to realize the massive processing of data; the main steps were as follows: change to binary classification → mask value is set to 0–255 → crop → eliminate the images that are all background → stitching → convert Tiff format to PNG format → get the final sample set. The samples and labels are divided into the training set, validation set, and test set, according to 6:2:2 [31]. Data augmentation is also required for the training and validation sets to improve the model’s generalization ability. Then, band combination and histogram equalization were performed to make the image rendering clearer.

3.

Visual Interpretation Delineation of GF images

GF6 is a low-orbit optical RS satellite with high resolution, comprehensive coverage, high quality, and efficient imaging capabilities. For this paper, it was chosen to be equipped with a 2-m panchromatic multispectral high-resolution camera (PMS). The visual interpretation of the debris-covered glaciers in GF6_PMS images is used as comparative data. Besides, the delineation of bare ice from the measured data is also based on manual visual interpretation in combination with Google Earth.

3.2.2. Field Measurement Based on the GRP and RTK-GPS Method

The thickness of the ice beneath the debris layer should be more than 0 to be called a glacier. Even in the field, binoculars cannot accurately identify the debris boundary. However, the combination of the pulsed EKKO PRO 100A Enhanced Ground Penetrating Radar (GPR; Sensors & Software Inc., Mississauga, ON, Canada) and RTK-GPS (Real Time Kinematic-Global Positioning System) was convincing. We traveled to the field site at about the same time as the RS (Remote Sensing) image acquisition in August 2022. Because the EPTR was so large, the thickness of the two most typical debris-covered glaciers in the region was selected for validation, to sum up 967 measured points, as shown in Figure 5c. The process is detailed; the ground-penetrating radar (GPR) is used to measure the edges of two glaciers point by point, and the outline of the glacier is measured at a fixed point with an accuracy of 0.1 m. The focus of GPR measurements is on the existence of subsurface ice layers. All uncertainties in ice thickness measurements are associated with the speed of electromagnetic wave propagation in snow, debris, and bare ice, inaccuracies in picking reflectors, and the resolution of the radar system. The estimation results show that the uncertainty of ice thickness measurements is within 1.2% in the Qingbingtan glacier region, which can be negligible for the following study.

3.2.3. Accuracy Evaluation

The errors caused by the spatial resolution of remote sensing images are systematic errors, which are inevitable [47]. The formula is as follows:

$$\epsilon =NA$$

(3)

where: ε is the glacier area error caused by the spatial resolution of the image, N is the perimeter of the glacier profile, and A is the side length of half image element.

Based on the premise of field observation data, we divided the regions of bare ice, debris layer, and others into three categories to evaluate the delineation accuracy of images in a balanced and uniform way. Evaluated indexes include accuracy, precision, recall, F1 scores, and kappa coefficient. Accuracy represents whether the nature of the glacier/debris extraction is correct. Precision metric expresses the accuracy of the delineated glacier area extent. Recall expresses the comprehensiveness of the glacier extent acquisition. An operation can be performed on the F1 score, which is the reciprocal of the average of the reciprocal of precision and Recall. The kappa coefficient is to characterize the results of the consistency test.

3.2.4. Analytical Methods for Glacier Area Change

The change in glacier area can be viewed in two aspects. One is the absolute amount of glacier change, which represents the amount of glacier ablation at the numerical level. The other is the relative amount of glacier change, which represents the change in glacier ablation as a ratio to its own area.

$$\mathrm{A}\mathrm{C}=\frac{{\mathrm{A}}_1-{\mathrm{A}}_0}{{\mathrm{T}}_1-{\mathrm{T}}_0}\times 100\mathrm{\%}$$

(4)

$$\mathrm{A}\mathrm{A}\mathrm{C}=\frac{{\mathrm{A}}_1-{\mathrm{A}}_0}{{\mathrm{A}}_0({\mathrm{T}}_1-{\mathrm{T}}_0)}\times 100\mathrm{\%}$$

(5)

AC is the rate of change of glacier area (km²/a); AAC is the average annual rate of change of glacier area (%/a); T₁ and T₀ denote the end and beginning of the study time interval; A₁ and A₀ represent the glacier area (km²) under T₁ and T₀, respectively.

4. Results and Analysis

The validation of remote sensing interpretation of glacier boundaries can rely on ground surveys or higher spatial resolution remote sensing data classification results as reference data [48]. The difficulty of debris-covered glacier identification lies in the definition of debris boundaries. Based on our field observations, we obtained the general pattern of debris cover distribution: (1) debris is mostly concentrated at the lower elevation of the whole glacier; (2) debris tends to be distributed along the mainstream line to the terminal glacier; and (3) the extent of debris coverage is continuous, along the edge of the ice tongue toward the center, which also facilitates our field measurements to determine the extent of debris. The debris extent obtained from the actual field measurements based on GRP and RTK-GPS methods, combined with the visual interpretation of Google Earth to determine the bare ice area, was collated to obtain the glacier boundary line. This result was used as the validation. The purpose of this paper is to evaluate the delineation accuracy of the debris-covered glacier extent by three methods: Landsat image delineation based on the TDSI method, Sentinel-1A deep learning delineation using the UNet network, and visual interpretation of the high-resolution image GF6_PMS. The delineated glacier area was chosen as an indicator to assess the accuracy of multi-source remote sensing images. The results are shown in

Table 2. Comparison of area extraction accuracy based on multi-source remote sensing images.

Item	Area of Delineated Glaciers		Area of Delineated Debris		Uncertainty of Delineation
	Muzart/km2	Uquir/km2	Muzart/km2	Uquir/km2	Glacier/km2	Debris/km2
Landsat ETM+/OLI	171.10 ± 4.56	176.01 ± 7.25	25.31 ± 0.93	26.33 ± 1.21	18.34	1.26
Sentinel-1A	177.01 ± 3.09	181.82 ± 4.92	26.90 ± 0.74	28.23 ± 0.97	24.06	1.75
GF6_PMS	160.23 ± 0.62	165.72 ± 0.98	21.83 ± 0.13	22.66 ± 0.17	1.18	−1.49
Field Data	163.46 ± 0.55	165.31 ± 0.93	23.49	24.89	——	——

.

The field debris boundary is obtained by comparing and filtering the positions of the points using RTK-GPS for multiple measurements, so the effect of error in the field data is negligible. The glacier area and debris area of Muzart and Uquir glaciers obtained based on field measurements are 163.46 km² and 165.31 km², and 23.49 km² and 24.89 km², respectively. In terms of the overall situation, the slightest error value of 1.18 km² was found for the visual interpretation of GF6_PMS, followed by the delineation of Landsat using the TDSI approach with a value of 18.34. The worst error value of 24.06 was not too significant for the delineation of Sentinel-1A images using the deep learning UNet network compared with the total area of 358.83. Surprisingly, for the case of debris overlay delineation, the slightest error is only 1.26 km² for Landsat delineation using the TDSI method, which is smaller than the error of the high-resolution GF6-PMS image. It is worth noting that the values of the glacier and debris areas obtained are on the large side except for the method of visual interpretation with GF6_PMS.

5. Discussion

5.1. Accuracy Analysis and Attribution of Different Delineation Methods

Firstly, systematic errors are always present and cannot be escaped by any improvement method. For remote sensing images, it is the spatial resolution. The spatial resolutions of the selected Landsat image, Sentinel-1A image, and GF_6PMS are 15 m, 10 m, and 2 m, respectively, and the corresponding error values of glacier area delineation are 11.81 km², 8.01 km², and 1.6 km², accounting for 3.4%, 2.29%, and 0.48% of the entire studied area; while the error values of debris delineation are 2.14 km², 1.71 km², and 0.3 km², accounting for 4.14%, 3.28%, and 0.61% of the entire studied area; the proportion is a little higher than the glacier area delineation.

Secondly, to evaluate these three most popular ways of delineating debris-covered glaciers in a more detailed and comprehensive manner, the area values of two specific areas covering two glaciers, whose names are Muzart and Uquir, were adopted as the evaluation index for classification. In this paper, the confusion matrix is used as the basis for the evaluation of the accuracy of image interpretation, as shown in Figure 6 and

Table 3. Comparison of accuracy of multi-source RS images.

Item		Accuracy	Kappa Coefficient	Precision	Recall	F1-Score
Landsat8-OIL	Glacier Area	98.18%	96.59%	94.83%	100.00%	97.92%
	Debris Layer			93.68%	100.00%	96.73%
Sentinel-1A	Glacier Area	97.32%	95.01%	91.74%	100.00%	95.69%
	Debris Layer			87.88%	100.00%	93.55%
GF6_PMS	Glacier Area	99.70%	99.43%	100.00%	99.01%	99.50%
	Debris Layer			100.00%	91.96%	95.81%

. Besides, accuracy, precision, recall, F1 score, and the kappa coefficient are used to characterize the evaluation in the glacier delineation process.

From the above table, the total accuracy of classification recognition based on visual interpretation of GF6_PMS is the highest with 99.70%, followed by the TDSI method applied to Landsat8 OLI images with 98.18%, and the worst is the result based on Sentinel-1A with 97.32%. The precision emphasizes the accuracy of the delineated glacier area extent. From the aspect of whether the glacier extent was comprehensively delineated, it was characterized by the recall. Since the area of GF6_PMS image visual interpretation is smaller than the actual value, the area which is delineated is considered correct, and the precision is described as 100%, but it also indicates that the delineated range is not comprehensive, and this error is reflected in the recall of 99.01% and 91.96% for glacier area and debris area. In contrast, the delineation values for Landsat8_OLI and Sentinel-1A were both larger than the valid values, and there is no incomplete delineation. The F1 scores can be averaged and evaluated for both precision and recall. The delineation method that chose Sentinel-1A as the base data with the most error was first excluded based on its results. The debris layer plays a vital role in this study, as the visual interpretation of the integrated GF6_PMS images is highly influenced by the time and energy of people. Therefore, the Landsat images with the highest debris F1 score of 96.73% were selected as the base data for the glacier change study, and its consistency test was perfect (kappa coefficient of 96.59%).

Thirdly, based on this paper, a brief attribution of glacier delineation accuracy is made for the three sources of RS images. First, the GF6_PMS image has the highest accuracy. On the one hand, due to its high resolution and the overlap with the bare ice delineation part of the validation data, the combination with 3D visual interpretation using Google Earth makes the results more accurate. However, the lower values of visual interpretation also illustrate that it is challenging to judge the ice under the debris layer, and the research in this area needs to be strengthened. Second, the uncertainty of Sentinel-1A image delineation relying on deep learning comes from the samples learned by the Unet network based on visual interpretation methods, which is inherently uncertain. In addition, some information is lost during the encoding and decoding process, which also causes some uncertainty. Combined with previous studies, Sentinel-1A’s deep learning semantic classification cut has a greater impact on delineation because SAR images are very suitable for glacier deformation analysis. There is no doubt that this way will improve the efficiency of the study within the range of accuracy allowed. Lastly, although the inherent resolution of Landsat images is not too high, the modified TDSI method based on thermal infrared band delineation is the most effective way to delineate the extent of the debris layer. The reason is that the principle of Landsat8_OLI delineation is based on the different reflectance of different wavelengths. Compared with GF6_PMS images, the Landsat8_OLI image is not only more accessible and preserved over a more extended period, but its spectral band classification is also more prosperous than that of GF6_PMS images.

5.2. 2000–2022 Glacier/Debris Change in the ETPR

Based on the importance of the study area selected for this paper, we explore the changes in debris-covered glaciers since the 21st century over a long time. Landsat series images have the advantage of a more extended period, lower price cost, and relatively minimal error, and it was chosen to study glacier change due to the uncertainty within acceptable limits. We compared debris and glacier changes in the ETPR based on the TDSI method, with a particular focus on two glaciers with an area of more than 150 km² in this region: the Muzart Glacier and the Uquir Glacier.

Figure 7 and

Table 4. Area variation of the bare ice and debris regions in the ETPR and typical glaciers from 2000 to 2022.

Item	Entire Studied Glaciers			Muzart Glacier			Uquir Glacier
	2000/km2	2022/km2	Change Ratio	2000/km2	2022/km2	Change Ratio	2000/km2	2022/km2	Change Ratio
Glacier Area	1002.38 ± 33.45	983.33 ± 33.43	−1.9%	173.24 ± 4.63	172.60 ± 4.63	−0.37%	177.04 ± 7.19	176.51 ± 7.18	−0.30%
Bare Ice Area	825.06 ± 27.53	786.86 ± 26.78	−4.63%	152.86 ± 3.63	147.39 ± 3.61	−3.58%	155.16 ± 6.04	150.08 ± 6.04	−3.27%
Debris Area	177.32 ± 5.92	196.47 ± 6.65	10.80%	20.38 ± 1.00	25.21 ± 1.02	19.16%	21.88 ± 1.15	26.43 ± 1.14	17.22%

show that the glaciers in the region show a weak retreating trend, while the overall proportion of the area occupied by debris increases. The entire studied glacier area of the EPTR decreased from 1002.38 km² to 983.33 km². The overall rate of change in area is −1.9%, with an annual average change rate of −0.08% a⁻¹, while the change rate in the large debris-covered glaciers Muzart Glacier and Uquir Glacier, is much lower than the overall rate, −0.37% and −0.3%, respectively. The debris coverage of the region is about 20%; the overall rate of debris increase in the region from 2000–2022 is about 10.8%, and the decrease of bare ice is −4.63%. The rate of rise of Muzart Uquir debris is significantly higher than the overall regional trend, 19.16% and 17.22%, respectively, but the decrease of bare ice is lower than the overall regional trend.

Furthermore, to explore in detail the specific changes in glaciers over the period 2000–2022, we chose 5 years as a scale for the study, relying on TDSI and manual revision means for semi-automated delineation. Based on the accuracy of >90% explored in the previous section, the results provide a more specific tendency of glacier change.

According to the average annual rate of change over a 5-year scale (as shown in Figure 8), the total area has smoothly and linearly decreased for the entire studied glaciers, with the greatest area loss occurring from 2010 to 2020 and the least from 2000 to 2005. The debris area increased synchronously as the glacier area decreased, with the greatest increase occurring in 2010–2015, with 1.36 km² annual growth, accounting for 0.14% of the total area. For the Muzart Glacier and Uquir Glacier, the trend is similar regardless of whether the variable is an area decrease or a debris increase. The area decrease trend is much slighter, with a roughly linear trend overall, while the growth trend in debris area has fluctuated.

5.3. Retreat of Glaciers in the EPTR Compared with the Typical Tien Shan Region

To study the EPTR’s glacier change, this study compared it with other typical basins or mountain areas of the Tien Shan Mountains [49,50,51,52,53,54,55,56]. Figure 9 showed that the relative changes of glaciers in the EPTR were much more minor than those in other regions. The reason may be that the debris coverage in this area is relatively high, and large glaciers are widely spread, so glacier ablation is inhibited. In general, the glacier retreat rate is different in different regions of the Tian Shan Mountains and previous studies have shown that glaciers with smaller areas are more affected by climate. From the perspective of spatial distribution, the glaciers in the Eastern Tianshan Mountains are small in scale and scattered in distribution. The retreat rate is the largest, followed by the Western Tianshan Mountains. The Central Tianshan Mountains have the largest scale of glacier, the most concentrated distribution, and the most debris, which retreated slowly. In addition, the glacier retreat rate on the north slope of the Tien Shan Mountains was significantly higher than that on the south slope.

5.4. Retreat of Muzart and Uquir Glaciers Compared with Other Typical Glaciers

Because of the particularity of the Muzart and Uquir glaciers, we selected different key glaciers in the Tianshan Mountains and other areas for comparison (

Table 5. Comparison of area change rates of typical debris-covered glaciers of different types and area sizes. ↑ represents an increase in debris coverage.

Glacier	Type	Region	Area/km2	Period/a	AC of Glacier/km2·a−1	Debris Coverage	Source
Qingbingtan NO.72	Continental	Tomur Region	7.27	2008–2013	−0.025	↑	[57]
Tomor Glacier	Continental	Tomur Region	310.14	1964–2009	−0.021	↑	[58]
Yinsugeti Glacier	Continental	Karakorum region	359.05	2011–2020	−0.045	↑	[44]
Muzart Glacier	Continental	Tomur Region	165.46	2000–2022	−0.029	↑	This Study
Uquir Glacier	Continental	Tomur Region	167.61	2000–2022	−0.024	↑	This Study
Hailuogou Glacier	Marine	Gonggar region	24.52	1974–2020	−0.010	↑	[44]
Xiaqu Glacier	Marine	Nyingchi Tanggula	133.58	2011–2020	−0.009	↑	[44]
Nalong Glacier	Marine	Ganjigab region	103.53	2011–2020	0	↑	[44]

). Debris-covered the selected glaciers, and the types are divided into continental glaciers and marine glaciers. The results show that the retreat of Muzart and Uquir glaciers is similar to that of the nearby Tomur glacier. These larger glaciers are shrinking at a lower rate than smaller glaciers in the same area, such as the Qingbingtan NO.72 Glacier. Even when compared with glaciers in different regions, large glaciers such as the Yinsugeti Glacier are losing less. The slower decline in glacier area is partly due to the increase in debris cover. Unlike the bare ice area, the attribution of the final debris variation is complex. Under the background of climate warming, the glacial end ablation causes the debris to melt together or calve into the nearby terminal moraine lake. At the same time, the accelerated melting of glaciers changes the dynamics of glaciers, making the upper glacier more likely to calve and collapse, forming new debris. As the glacier slows, the transported material accumulates further away from the end, causing debris to get toward the higher elevations of the upper glacier gradually. This part of accumulation made up for the reduction of debris area caused by terminal receding so that the final result was still an increase in the debris-covered area. Marine debris-covered glaciers also showed a slow retreat trend. However, unlike continental glaciers, marine glaciers have more precipitation, leading to more accumulation of glacier mass, and are more likely to be broken and calved at the top of the glacier.

5.5. Factor Analysis of Glacier Changes in the ETPR

The climate is the fundamental factor that influences glacier changes. Temperature mainly affects glacier ablation, while precipitation determines, to some extent, the amount of material recharged to the glacier. Many studies have demonstrated that temperature is the dominant factor influencing glacier change [59,60]. The melting of glaciers for every 1 °C increase in average summer requires 40~50% more precipitation to compensate [61]. Since the response of glaciers to climate has a lagging effect, climate data were selected since 1960 (as shown in Figure 10). Both Zhaosu weather station and Wensu weather stations show a significant increase trend in temperature over the past 60 years, but precipitation has risen more slowly.

Furthermore, the inter-annual variability was larger at Zhaosu meteorological station. The difference in precipitation between the two meteorological stations is mainly due to the altitude (Zhaosu 1851 m and Aksu 1108 m). From the linear fit, it is clear that the results from the Zhaosu weather station better characterize the response of the glaciers to climate, both in terms of temperature and precipitation. It can be concluded that temperature increase plays a decisive function, and precipitation increase is insufficient to counteract the effects of increasing temperatures for the significant glacier ablation. The changes of each indicator of the two large debris-covered glaciers show a substantial uniformity, which can characterize the changes of similar glaciers in the region to some extent.

6. Conclusions

The valley glaciers covered by debris in the ETPR were selected as the study area to analyze and evaluate the most advanced debris identification methods from multi-source remote sensing imagery. The area values used for evaluation, which included extracted debris layers, bare ice, and others that are not part of the glacier, were measured as the indicators for evaluating the accuracy of delineation of different RS images. Based on the above extraction means, the TDSI extraction method with Landsat8_OLI image as the base data, which is the best for debris extraction, was selected to analyze the area variation of EPTR and the two glaciers we measured. The results obtained were also compared with glaciers from different regions. Finally, the variation attribution of the glacier and debris was explored. The results are as follows:

(1) The field validation data measured 23.49 km² and 24.89 km² of debris for the Muzart and Uquir glaciers, respectively, corresponding to 163.46 km² and 165.31 km² of the glacier area. Values of landsat8 OLI and Sentinel-1A delineated area are both high, while the result of GF6_PMS image delineated based on visual interpretation is low. Combining the F1 scores obtained by precision and recall, the accuracy of glacier delineation by Landsat, Sentinel-1A, and GF images was 97.92%, 95.69%, and 99.50%, respectively, and for debris, the accuracy was 96.73%, 93.55%, and 95.81%, respectively. Since the glaciers in the EPTR are characterized by their debris, we chose Landsat series images for the area study of glaciers and debris.

(2) The glacier area of the EPTR decreased from 1002.38 km² to 983.33 km² (−0.08% a⁻¹), showing a weak retreating trend. This results from a combination of reduced bare ice (−4.63%) and increased debris (10.8%). For large debris-covered glaciers, the Muzart Glacier and Uquir Glacier, it was lower than the overall rate (−0.37% and −0.3%, respectively) and significantly higher than the general trend (19.16% and 17.22%, respectively). Furthermore, the greatest change period was in the glacier area from 2010–2020 and the debris area from 2010–2015. In general, the total glacier and debris area in the EPTR region show roughly linear trends, while the two large glaciers, Muzart Glacier and Uquir Glacier, have almost identical trends; their debris change trends are fluctuating but generally increasing.

(3) The glaciers in the EPTR area are the most concentrated and have a minor area change compared to other regions of The Tianshan Mountains. The glacier area in the Eastern Tianshan Mountains is decreasing the fastest, followed by the Western Tianshan Mountains. The area of the Central Tianshan Mountains, where the EPTR is located, has an average annual change rate of only −0.08% a⁻¹. In addition, the ablation rate is faster on the northern slopes of the Tianshan Mountains than on the southern slopes. For a single glacier study, the glacier’s size and the debris coverage affect the decrease rate of the area. In general, small glaciers are more sensitive to area reduction, indicating that debris slows down glacier ablation. In addition, marine debris-covered glaciers show less area change than continental debris-covered glaciers because more water contributes to more accumulation sources of the glacier. Although the rate of glacier area decreases in the EPTR is slower than in other regions, the rising temperature is the dominant factor in the change of debris and glacier area in the EPTR, while precipitation also contributes its share.

In this paper, the delineation and evaluation of the images have really promoted the prosperity of the glacier study in the Tomur region. However, there are still some shortcomings in the research process: first, there are few studies to compare multiple images using multiple methods, which is a bold attempt on our part, and the comparison process may be a little imperfect. Second, for the EPTR glacier study, only the area changes since the 21st century have been assessed. In the subsequent research work, the material balance, glacier flow rate, and ice thickness estimation are still urgent problems to be overcome. Glaciers are a gift from nature, and the most authentic source of validation data is necessarily from the field. However, field surveys need to break through multiple factors such as climatic factors, human physical ability, errors caused by snow cover, and sudden glacial disasters. In addition, glacial material balance is limited by the significant topographic drop of this glacial rock mass. Specifically, the material balance of this area cannot be studied in a traditional glaciological way due to the harsh environment of this area, thus, no splines have been deployed in this area in the past days. Additionally, the vertical drop would be too large to allow the high-resolution DEM to produce large errors in the alignment session. The investigation of glacier area change is the basis for glacier research, and it can serve as a more accurate parameter for model input, showing the most intuitive and clear understanding of the trend of glacier change in a region, and is a better study of the local hydroclimate and other content.