Abstract

Pixel-based classification struggles to accurately identify glacier changes in areas with similar spectral characteristics, particularly in debris-covered regions where spectral features closely resemble surrounding mountains and rocks, resulting in low extraction accuracy. This study investigates the Yinsugaiti and Yarong Glaciers using Google Earth Engine to integrate spectral indices, microwave texture features, and topographic data. An object-based (OB) machine learning algorithm is applied for automated glacier extraction and compared with pixel-based (PB) classification methods. The results demonstrate the following: (1) The OB classification approach integrating multi-feature fusion significantly improved glacier extraction accuracy. The OB_RF classifier achieved an overall accuracy of 98.1%, a Kappa coefficient of 0.97, and an F1 score of 98.67%, outperforming the OB_CART and OB_GTB classifiers. Compared with PB_RF, the overall accuracy, Kappa coefficient, and F1 score increased by 1.7%, 0.024, and 5.57%, respectively. (2) Between 2001 and 2022, the Yinsugaiti and Yarong Glaciers retreated at average annual rates of 0.08% and 0.13%, respectively. (3) Supraglacial debris was primarily distributed below 5000 m and 4800 m on the Yinsugaiti and Yarong Glaciers, respectively, with debris-covered areas on both glaciers expanding upward during this period.

Keywords: glacier boundary extraction; object-based; pixel-based; machine learning; multi-feature fusion

Introduction

Glacier change serves as an important indicator of climate change. Glacier formation involves the processes of snow accumulation, compaction, and refreezing \cite{1}, whereas debris-covered glaciers form when glacier ablation, movement, rock weathering, and collapse cover the glacier surface with debris material \cite{2}. As a vital freshwater resource, glacier meltwater profoundly impacts regional hydrology \cite{3} and ecological environments \cite{4}. Therefore, rapidly and accurately obtaining glacier boundary change information not only reveals dynamic area changes but also supports ice volume estimation \cite{5} and monitoring of glacier mass changes \cite{6}, which holds significant scientific meaning and practical value for understanding glacier change mechanisms and developing response strategies.

Currently, the Global Land Ice Measurements from Space (GLIMS) system provides only single-class vector data without distinguishing different glacier categories, limiting research on dynamic changes of different glacier types and their environmental impacts. With advances in satellite remote sensing technology, the increasing number of satellites and sensors provides richer data support for glacier dynamic change studies. Automated and semi-automated methods based on multispectral satellite imagery can extract glacier area dynamics, including the Normalized Difference Snow Index (NDSI) \cite{7}, band ratio methods \cite{8}, synthetic aperture radar interferometry \cite{9}, and object-based approaches \cite{10}. Although these methods perform excellently in bare ice region extraction, debris-covered areas remain challenging due to their spectral similarity to surrounding mountains and rocks.

To further improve debris-covered glacier extraction accuracy, various methods have been proposed, including the Normalized Difference Debris Index (NDDI) \cite{11} utilizing differences in shortwave infrared and thermal infrared bands to identify debris-covered areas, incorporating morphometric parameters \cite{12} to further improve extraction accuracy, and using coherence changes between SAR images \cite{13}. Semantic segmentation \cite{14}, machine learning \cite{15}, and deep learning \cite{16} have also been applied to debris-covered glacier extraction, achieving accuracies of 89%–96%. However, fusing multi-sensor data for long-term glacier monitoring requires addressing inconsistencies in data format and resolution while processing massive volumes of remote sensing imagery, creating significant economic and technical barriers for large-scale glacier change monitoring and analysis \cite{17}.

In recent years, the Google Earth Engine (GEE) platform has enabled efficient access to massive remote sensing datasets \cite{18}, allowing rapid pixel-level data processing and analysis. GEE has been widely used for monitoring climate change \cite{19}, land use change \cite{20}, and snow-ice changes \cite{21}. Considering that pixel-based classification typically focuses on individual pixels without simultaneously incorporating non-spectral information such as texture and topography, it performs poorly in complex terrain or areas with similar spectral characteristics \cite{22}. Therefore, this study proposes an object-based approach on the GEE platform that combines spectral indices, microwave texture, and topographic features using machine learning algorithms (RF, GTB, CART) for glacier automatic classification. This approach is compared with pixel-based classification methods to explore differences in glacier extraction accuracy and analyze spatiotemporal glacier changes from 2001 to 2022, providing new insights and techniques for large-scale, long-term glacier dynamic monitoring.

1. Study Area and Data

1.1 Study Area Overview

To better evaluate classification method effectiveness under different debris-covered environments, two debris-covered glaciers were selected as study areas (Fig. 1): the Yinsugaiti Glacier in the Karakoram Mountains and the Yarong Glacier in the Gangrigabu Range. The Yinsugaiti Glacier (36°50′N, 76°70′E) is China's largest glacier with a total area of approximately 359 km², featuring thick and continuous debris cover. The Yarong Glacier (29°20′N, 96°40′E) has a total area of approximately 179 km², with debris concentrated primarily at the glacier tongue, lateral moraine development, and thin, ambiguous debris distribution.

1.2 Data Sources

This study utilized Landsat series imagery with 30 m spatial resolution spanning 2001–2022, with missing data supplemented by adjacent years. Since cloud masking algorithms often misclassify high-reflectance pixels such as snow, ice, and salt lakes as clouds, this study manually identified cloud-free Landsat images after temporal and cloud filtering (Table 1). Microwave data were derived from PALSAR-2 and Sentinel-1 datasets; elevation data from NASADEM; and reference glacier boundaries from the Randolph Glacier Inventory (RGI).

1.3 Research Methods

This study integrated spectral indices, microwave texture, and topographic features on the GEE platform, employing machine learning algorithms for object-based glacier classification. The technical workflow is illustrated in Fig. 2.

1.3.1 Feature Extraction

Various index features suitable for glacier identification were calculated and fused with original image bands to create multi-band feature images for classification input. The Normalized Difference Snow Index (NDSI) is widely used for snow and ice identification, as snow/ice exhibits high reflectance in visible bands and very low reflectance in shortwave infrared. NDSI leverages this strong difference to identify snow and ice. The Normalized Difference Vegetation Index (NDVI) does not directly identify glaciers, but snow's high sensitivity in the red band yields negative NDVI values, helping distinguish seasonal snow from permanent ice. The Normalized Difference Water Index (NDWI) highlights water bodies, preventing misclassification of moraine-dammed lakes at glacier tongues as glaciers.

Since debris-covered glaciers spectrally resemble surrounding bare land, identification based solely on these indices remains challenging. The cooling effect of underlying ice on supraglacial debris typically makes debris-covered surfaces cooler than surrounding bare land, providing an important basis for distinguishing debris-covered areas from non-glacier regions. As debris spatial distribution is influenced by terrain and elevation, slope was calculated from NASADEM and combined with elevation data to improve classification accuracy. Given SAR's ability to penetrate clouds and debris, providing additional internal structural information, polarization differences and ratios were extracted from PALSAR-2 and Sentinel-1 images to highlight glacier texture differences from surrounding features and incorporated into the original imagery.

The spectral indices are calculated as follows:

$$
\begin{align}
NDVI &= \frac{\rho_{Nir} - \rho_{Red}}{\rho_{Nir} + \rho_{Red}} \
NDSI &= \frac{\rho_{Green} - \rho_{Swir}}{\rho_{Green} + \rho_{Swir}} \
NDWI &= \frac{\rho_{Green} - \rho_{Nir}}{\rho_{Green} + \rho_{Nir}} \
NDDI &= \frac{NDVI - NDSI}{NDVI + NDSI}
\end{align}
$$

where $\rho_{Green}$, $\rho_{Red}$, $\rho_{Nir}$, and $\rho_{Swir}$ represent green, red, near-infrared, and shortwave infrared bands, respectively.

While adding spectral indices, texture, and topographic features significantly improves classification accuracy, excessive input bands create redundancy and reduce computational efficiency. Therefore, after integrating spectral indices, texture features, and topographic factors, feature importance scores were calculated. Features with the lowest contribution were iteratively removed, with model accuracy and performance evaluated after each iteration. Through repeated iterative combination, 15 bands with the highest importance scores were selected for image fusion (Fig. 3).

1.3.2 Simple Non-Iterative Clustering Superpixel Segmentation

Using objects as the minimum study unit leverages spatial relationships between objects, enhancing classification coherence and reducing noise in optical remote sensing imagery for better results. Segmentation is a key step in Object-Based Image Analysis (OBIA), helping reduce isolated misclassified pixels ("salt-and-pepper noise") in pixel-based classification. This study employed the Simple Non-Iterative Clustering (SNIC) algorithm for image segmentation, which aggregates similar pixels into superpixels based on color, texture, brightness, and spatial location information \cite{35}. Key parameters include compactness, connectivity, seed/grid size (Seeds), and neighborhood size (NeighborSize). Compactness defines cluster shape, with values closer to 0 producing more square-like pixel shapes; connectivity defines the direction for merging neighboring superpixels, with 4 indicating orthogonal adjacency and 8 including diagonal adjacency; seed size determines the initial position or spacing of cluster centers, with segmentation scale directly impacting results \cite{36}.

1.3.3 Glacier Extraction Based on Machine Learning

Three commonly used machine learning algorithms (RF, GTB, CART) were selected for automated glacier classification on both segmented and unsegmented images. Random Forest (RF) is a supervised machine learning algorithm that combines outputs from multiple decision trees to produce a single result. RF performance primarily involves two parameters: number of trees (Ntree) and number of features (Mtry). Through iterative testing and accuracy recording at each step, model accuracy stabilized and reached optimum when Ntree was approximately 100. Mtry was set as the square root of the total number of features in the training samples.

Gradient Tree Boosting (GTB) improves model predictive capability by sequentially optimizing a series of weak learners, with each new decision tree correcting prediction errors from the previous round to continuously enhance overall accuracy. The Classification and Regression Tree (CART) method builds a single decision tree model through recursive data partitioning.

Machine learning classifiers rely on labeled data for training. Therefore, this study visually interpreted and selected samples including bare ice, debris-covered areas, and non-glacier regions. For 2001 imagery, 1,500 sample points were selected (500 each for bare ice, debris-covered, and non-glacier areas). For 2010 imagery, 400 samples were selected for each class. For 2022 imagery, 500 samples were selected for each class. Samples were randomly divided into 70% for training and 30% for validation.

1.3.4 Accuracy Evaluation

A confusion matrix was used to evaluate classification model performance, with primary metrics including overall accuracy, Kappa coefficient, and F1 score. Overall accuracy represents the proportion of correctly classified samples in the validation set. The Kappa coefficient measures agreement between predicted and actual categories, with higher values indicating greater consistency. The F1 score calculates the harmonic mean of precision and recall.

1.3.5 Area Uncertainty Assessment

To quantify uncertainty in glacier boundary extraction, this study employed a buffer analysis method. This approach assumes maximum area determination error within half a pixel, buffering the generated glacier boundary by half a pixel to estimate error range. A 15 m buffer (half of Landsat 5/7/8 pixel size) was applied to 2001–2022 classification results. Results showed area calculation uncertainties of ±1.09%, ±1.04%, and ±1.02% for Yinsugaiti Glacier, and ±1.46%, ±1.57%, and ±1.27% for Yarong Glacier, with mean overall area uncertainty of ±1.47%, which is within reasonable limits \cite{37}.

2. Results

2.1 Optimal Superpixel Seed Selection

The most important parameter in the SNIC algorithm is superpixel seed (Seeds) segmentation size (Table 2). When seed size is too small, superpixels with the same attributes may be segmented into different classes; however, when segmentation scale is too large, superpixels with different attributes may be incorrectly segmented into the same class, reducing classification accuracy. In this study, a superpixel seed segmentation scale of 100 pixels yielded optimal classification results.

2.2 Accuracy Comparison

As shown in Table 2, the multi-feature fusion OB_RF classification accuracy outperformed PB_RF. For the Yinsugaiti Glacier region, OB_RF achieved overall accuracy (98.1%) and Kappa coefficient (0.97) of 96.51% and 0.95, respectively, representing improvements of 1.7% and 2.86% over PB_RF. The OB_RF overall accuracy and Kappa coefficient increased by 1.58% and 0.024, respectively, demonstrating the highest classification accuracy. For Yarong Glacier, OB_RF overall accuracy and Kappa coefficient improved by 3.47% and 0.06, respectively.

To further validate model accuracy, F1 scores were included as supplementary verification (Table 3). OB_RF achieved the optimal classification accuracy with an F1 score of 98.67%, confirming the reliability of the OB_RF method. Pixel-based classification methods showed poor spatial continuity, with PB_RF, PB_GTB, and PB_CART algorithms producing substantial "salt-and-pepper noise," particularly at mixed pixels between bare ice and debris where pixel-level classification struggled to differentiate effectively, leading to numerous omission and commission errors. In contrast, OB_RF, OB_GTB, and OB_CART methods fully considered spatial information, texture, and spectral features, effectively improving commission/omission errors and salt-and-pepper noise.

As illustrated in glacier classification mapping results (Fig. 5), OB_RF-extracted glacier boundaries conform better to actual glacier extents with fewer small, dispersed patches, effectively reducing post-processing manual correction time and making the method more suitable for practical glacier data statistical updates. This demonstrates the feasibility of our approach for accurate, rapid glacier change mapping at regional scales.

2.3 Glacier Change Characteristics

2.3.1 Glacier Area Changes from 2001 to 2022

Statistical analysis revealed overall glacier retreat from 2001 to 2022 (Table 4). For bare ice areas, Yinsugaiti Glacier retreated from 308.70 km² in 2001 to 298.52 km² in 2022, a total retreat of 10.18 km² (3.29% retreat rate, 0.16% annual average). Conversely, debris-covered areas expanded from 10.18 km² in 2001 to 20.36 km² in 2022, a total expansion of 10.18 km² (0.48 km² annual average). Glacier retreat magnitude showed spatiotemporal variation, with the 2001–2010 period showing maximum annual bare ice area change (4.53 km² retreat, 0.25% annual average). During 2010–2022, bare ice area retreated 37.38 km² (0.21% annual average), while corresponding debris-covered area expanded 9.61 km² (0.80 km² annual average, 9.61% growth rate).

Yarong Glacier's bare ice retreated from 169.21 km² in 2001 to 160.44 km² in 2022 (5.18% retreat rate, 0.25% annual average). Debris-covered area expanded from 9.79 km² in 2001 to 18.56 km² in 2022 (89.48% expansion rate, 0.42 km² annual average). The proportion of debris-covered area to total glacier area increased from 5.47% in 2001 to 10.36% in 2022.

2.3.2 Glacier Area Variations with Elevation

To further investigate glacier area changes with elevation, variations in bare ice and debris-covered areas were statistically analyzed at 200 m intervals from 4600–5400 m (Fig. 7). Yinsugaiti Glacier is distributed between 4000–7000 m, with maximum area at 5200–5800 m, while debris-covered areas are distributed between 4000–5000 m. From 2001–2022, bare ice and debris-covered area changes primarily concentrated in the 4600–5200 m range, with maximum changes at 4600–5200 m. Yarong Glacier is distributed between 4000–6400 m, with debris primarily concentrated at 4000–5000 m. Bare ice areas showed slight expansion in the 4200–6400 m range.

2.3.3 Spatial Change Characteristics of Debris-Covered Areas

Spatial change results for debris-covered areas on Yinsugaiti and Yarong Glaciers (Fig. 8) show upward expansion trends at eastern glacier tongues. At the junction between bare ice and debris-covered areas, heat conduction from debris to adjacent bare ice raises temperatures and accelerates bare ice melting \cite{38}, likely causing greater retreat magnitude in bare ice adjacent to debris-covered areas compared to other regions.

3. Discussion

This study confirms the significant advantages of OB_RF classification for debris-covered glacier boundary extraction, a conclusion validated in previous research \cite{15,21}. The core strength lies in using the SNIC algorithm to cluster multi-band feature images into homogeneous regions with similar attributes, effectively overcoming salt-and-pepper noise caused by spectral mixing effects in PB classification. Particularly in debris-covered areas, glacier retreat enhances surface heterogeneity (e.g., supraglacial lake development, debris thickness gradient changes), exacerbating uncertainties in single-band classification methods.

OB_RF classification enhances feature separability by fusing SAR texture features (backscatter coefficient differences) with topographic factors, differing from traditional methods relying on optical texture features (gray-level co-occurrence matrices) \cite{25}. PALSAR-2 backscatter coefficients provide stronger penetration capability for cloud-shadow areas and debris-bedrock mixed pixels, improving classification accuracy by 6.9%–10.5% compared to traditional Landsat-based methods \cite{14}, further validating the universality of microwave data for debris-covered glacier monitoring. Algorithmically, RF's capability to handle high-dimensional nonlinear features and resist overfitting enables superior performance in both bare ice and debris-covered areas compared to GTB and CART. GTB's sensitivity to sample distribution reduces its classification stability in areas with uneven debris cover compared to RF, a conclusion similar to existing glacier classification studies \cite{22}, further verifying RF's robustness for debris-covered glacier classification.

Therefore, the OB_RF classification method constructed on the GEE platform in this study not only achieves high mapping accuracy but also rapidly obtains debris-covered glacier change information in less than one minute, greatly solving problems of traditional glacier mapping methods that rely on desktop software like ENVI for coordinate conversion, radiometric correction, index calculation, and feature fusion with high computational costs \cite{26}, a critical advantage for large-scale glacier change research. Studies show that Karakoram and Hengduan Mountains glaciers are retreating overall \cite{33,39}. During 2001–2010 and 2010–2022, glacier area annual retreat rates in the Karakoram's Keleqing River basin were 0.29% and 0.23%, respectively, while Hengduan Mountains glaciers retreated at 0.15% and 0.28% annually \cite{40}. In terms of elevation changes, Kunlun Mountains glacier area peaked at 5600–5800 m, with retreat primarily below 4800 m \cite{41}. Southeastern Tibet glacier area changes mainly concentrated at 3900–5400 m, decreasing with elevation \cite{42}. This study's Yinsugaiti and Yarong Glaciers showed total annual retreat rates of 0.08% and 0.13% from 2001–2022, with maximum area changes at 4600–5200 m, similar to neighboring regions in showing retreat trends. Although both glaciers are in alpine zones, their retreat rates differ, with Yarong Glacier showing greater change than Yinsugaiti Glacier. Analysis reveals that Yinsugaiti Glacier experienced three glacier advances during 2001–2022 \cite{43}, which slowed overall retreat rates. Additionally, Yinsugaiti Glacier has thicker debris cover than Yarong Glacier, indicating that thicker debris cover inhibits glacier ablation \cite{44}.

SNIC segmentation effectively improves overall model accuracy. However, classification performance is highly sensitive to superpixel seed segmentation scale; this study determined 100 pixels as the optimal segmentation scale. Vector-based glacier classification results show that classification based on segmented images yields boundaries closer to actual glacier extents, greatly reducing post-processing manual correction workload.

4. Conclusions

1. This study combined spectral indices, texture, and topographic features for object-based glacier extraction. OB_RF classification achieved overall accuracy, Kappa coefficient, and F1 scores of 98.1%, 0.97, and 98.67%, respectively, representing 1%–4% improvement in overall accuracy and 4.1%–6.2% improvement in F1 score compared with PB_RF classification.

2. Analysis of Yinsugaiti and Yarong Glacier changes from 2001–2022 revealed overall retreat trends, with cumulative retreat areas of 30.56 km² and 18.32 km², respectively. Bare ice area retreat rates were 3.19% and 5.18%, while debris-covered area expansion rates were 89.48% and 89.48%, respectively. Glacier changes primarily occurred in mid-low elevation zones at the bare ice-debris interface.

3. Although this method achieved high accuracy in both bare ice and debris-covered area extraction, reflectance characteristic differences between thin and thick debris may affect classification accuracy, and lack of measured sample data increases uncertainty. Future research will focus on thick debris-covered regions and incorporate measured sample data to evaluate model reliability and applicability under varying debris thicknesses.

References

\cite{1} Cogley J G, Hock R, Rasmussen L A, et al. Glossary of glacier mass balance and related terms[J]. Arctic Antarctic and Alpine Research, 2012, 44(2): 256-258.
\cite{2} Hewitt K. Rock avalanches that travel onto glaciers and related developments, Karakoram Himalaya, Inner Asia[J]. Geomorphology, 2009, 103(1): 66-79.
\cite{3} Cauvy Fraunié S, Andino P, Espinosa R, et al. Ecological responses to experimental glacier runoff reduction in alpine rivers[J]. Nature Communications, 2016, 7(1): 12025.
\cite{4} Millan R, Mouginot J, Rabatel A, et al. Ice velocity and thickness of the world s glaciers[J]. Nature Geoscience, 2022, 15(2): 124-129.
\cite{5} Zemp M, Huss M, Thibert E, et al. Global glacier mass changes and their contributions to sea level rise from 1961 to 2016[J]. Nature, 2019, 568(7752): 382-386.
\cite{6} Hall D K, Riggs G A, Salomonson V V. Development of methods for mapping global snow cover using moderate resolution imaging spectroradiometer data[J]. Remote Sensing of Environment, 1995, 54(2): 127-140.
\cite{7} Zhang M, Wang X H, Shi C L, et al. Automated glacier extraction index by optimization of Red/SWIR and NIR/SWIR ratio index for glacier mapping using Landsat imagery[J]. Water, 2019, 11(6): 1223.
\cite{8} Du Weibing, Li Junli, Bao Anming, et al. Information extraction method of alpine glaciers with multitemporal and multiangle remote sensing[J]. Acta Geodactica et Cartographica Sinica, 2015, 44(1): 59-66.
\cite{9} Huang L, Li Z, Tian B S, et al. Classification and snow line detection for glacial areas using the polarimetric SAR image[J]. Remote Sensing of Environment, 2011, 115(7): 1721-1732.
\cite{10} Walter V. Object based classification of remote sensing data for change detection[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2004, 58(3-4): 225-238.
\cite{11} Huai Baojuan, Li Zhongqin, Sun Meiping, et al. Discussion on RS methods for glacier outline detection——A case study in headwaters of the Kanas River[J]. Arid Zone Research, 2013, 30(2): 372-377.
\cite{12} Shukla A, Arora M K, Gupta R P. Synergistic approach for mapping debris covered glaciers using optical thermal remote sensing data with inputs from geomorphometric parameters[J]. Remote Sensing of Environment, 2010, 114(7): 1378-1387.
\cite{13} Lei Saiyue, Fang Li, Li Chende, et al. Improving the algorithm for extracting debris covered glaciers in optical satellite images[J]. Computer Engineering, 2025, 2(51): 269-277.
\cite{14} Xue Jiao, Yao Xiaojun, Zhang Cong, et al. Extraction method and change of debris covered glaciers[J]. Journal of Glaciology and Geocryology, 2022, 44(5): 1653-1664.
\cite{15} Yousuf B, Shukla A, Arora M K, et al. On drivers of subpixel classification accuracy—An example from glacier facies[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020, 13(1): 601-608.
\cite{16} Aftab A K, Akhtar J, Dostdar H, et al. Machine learning algorithms for mapping debris covered glaciers: The Hunza Basin case study[J]. IEEE Access, 2020, 8(1): 12725-12734.
\cite{17} Xie Z Y, Asari V K, Haritashya U K. Evaluating deep learning models for debris covered glacier mapping[J]. Applied Computing and Geosciences, 2021, 12(1): 100071.
\cite{18} Yang S J, Wang F T, Xie Y D, et al. 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[J]. Remote Sensing, 2023, 15(10): 2575.
\cite{19} Racoviteanu A, Williams M W. Decision tree and texture analysis for mapping debris covered glaciers in the Kangchenjunga area, Eastern Himalaya[J]. Remote Sensing, 2012, 4(10): 3078-3109.
\cite{20} Zhao Q, Yu L, Li X C, et al. Progress and trends in the application of Google Earth and Google Earth Engine[J]. Remote Sensing, 2021, 13(18): 3778.
\cite{21} Amani M, Mahdavi S, Afshar M, et al. Canadian wetland inventory using Google Earth Engine: The first map and preliminary results[J]. Remote Sensing, 2019, 11(7): 842.
\cite{22} Yu Caili, Lu Jianqiang, Dou Xufeng, et al. Extraction of greengage planting information from multi source remote sensing images based on GEE[J]. Jiangsu Journal of Agricultural Sciences, 2024, 40(8): 1455-1463.
\cite{23} Ma Dongling, Liu Weixing, Zhang Chunhong. Study on land use spatial structure based on Google Earth Engine and fractal theory: A case study of ecological function reserve in Yellow River Delta[J]. Soils, 2024, 56(3): 655-665.
\cite{24} Bevington A R, Menounos B. Accelerated change in the glaciated environments of western Canada revealed through trend analysis of optical satellite imagery[J]. Remote Sensing of Environment, 2022, 270(1): 112862.
\cite{25} Hu M C, Zhou G S, Lv X M, et al. A new automatic extraction method for glaciers on the Tibetan Plateau under clouds, shadows and snow cover[J]. Remote Sensing, 2022, 14(13): 3084.
\cite{26} Huang L, Li Z, Zhou J M, et al. An automatic method for clean glacier and nonseasonal snow area change estimation in High Mountain Asia from 1990 to 2018[J]. Remote Sensing of Environment, 2021, 258(1): 112376.
\cite{27} Ali A, Dunlop P, Coleman S, et al. Glacier area changes in Novaya Zemlya from 1986-89 to 2019-21 using object based image analysis in Google Earth Engine[J]. Journal of Glaciology, 2023, 69(277): 1305-1316.
\cite{28} Paul F, Huggel C, Kääb A. Combining satellite multispectral image data and a digital elevation model for mapping debris covered glaciers[J]. Remote Sensing of Environment, 2004, 89(4): 510-518.
\cite{29} Östrem G. Ice melting under a thin layer of moraine, and the existence of ice cores in moraine ridges[J]. Geografiska Annaler, 1959, 41(4): 228-230.
\cite{30} Zhuang L C, Ke C Q, Cai Y, et al. Measuring glacier changes in the Tianshan Mountains over the past 20 years using Google Earth Engine and machine learning[J]. Journal of Geographical Sciences, 2023, 33(9): 1939-1964.
\cite{31} Xu Aiwen. Monitoring Glacier Change Based on Remote Sensing in China Karakoram for the Last Four Decades[D]. Lanzhou: Lanzhou University, 2017.
\cite{32} Pei Huan, Sun Tianjiao, Wang Xiaoyan. Object oriented land use/cover classification based on texture features of Landsat 8 OLI image[J]. Transactions of the Chinese Society of Agricultural Engineering, 2018, 34(2): 248-255.
\cite{33} Zhao Jinbiao. Study on Characteristics of Glacier Change and Movement in Southeast Xizang[D]. Anhui: Anhui University of Science and Technology, 2024.
\cite{34} Holobâc H, Tielidze L G, Ivan K, et al. Multi sensor remote sensing to map glacier debris cover in the Greater Caucasus, Georgia[J]. Journal of Glaciology, 2021, 67(264): 685-696.
\cite{35} Achanta R, Susstrunk S. Superpixels and polygons using simple iterative clustering[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE Computer Society, 2017: 4651-4660.
\cite{36} Liu Shiyin, Yao Xiaojun, Guo Wanqin, et al. The contemporary glaciers in China based on the Second Chinese Glacier Inventory[J]. Acta Geographica Sinica, 2015, 70(1): 3-16.
\cite{37} Zhang Yong, Liu Shiyin, Wang Xin. Debris cover effect in the Tibetan Plateau and surroundings: A review[J]. Journal of Glaciology and Geocryology, 2022, 44(3): 900-913.
\cite{38} Chen Yiqing, Duan Keqin, Shi Peihong, et al. Change trend of debris free glaciers in the Pamir based on Google Earth Engine from 1990 to 2020[J]. Journal of Glaciology and Geocryology, 2024, 46(3): 742-753.
\cite{39} Huang Xiaoran, Bao Anming, Guo Hao, et al. Change of typical glaciers and its response to meteorological factors in the Eastern Tianshan Mountains in China in recent 20 years[J]. Arid Zone Research, 2017, 34(4): 870-880.
\cite{40} Li Ruonan, Li Junli, Li Shuangshuang, et al. Monitoring the glacier changes in Yilian Habirga Mountain using Sentinel-2 data[J]. Arid Zone Research, 2024, 41(6): 940-950.
\cite{41} Zhou Jianmin, Li Zhen, Xing Qiang. Deriving glacier border information based on analysis of decorrelation in SAR interferometry[J]. Journal of Glaciology and Geocryology, 2010, 32(1): 133-138.
\cite{42} Zhang Yong, Liu Shiyin, Wang Xin. Debris cover effect in the Tibetan Plateau and surroundings: A review[J]. Journal of Glaciology and Geocryology, 2022, 44(3): 900-913.
\cite{43} Chen Yiqing, Duan Keqin, Shi Peihong, et al. Change trend of debris free glaciers in the Pamir based on Google Earth Engine from 1990 to 2020[J]. Journal of Glaciology and Geocryology, 2024, 46(3): 742-753.
\cite{44} Huang Xiaoran, Bao Anming, Guo Hao, et al. Change of typical glaciers and its response to meteorological factors in the Eastern Tianshan Mountains in China in recent 20 years[J]. Arid Zone Research, 2017, 34(4): 870-880.