Analysis of Glacial Morphological Characteristics in Ányêmaqên Mountains Using Multi-Source Time-Series High-Resolution Remote Sensing Imagery

Abstract

Since the 1990s, glaciers in the Ányêmaqên Mountains of the Qinghai–Tibet Plateau have exhibited anomalous retreat and thinning. This persistent deglaciation has triggered secondary disasters including glacial debris flows, ice collapses, and glacial lake outburst floods, posing significant threats to regional ecological security and sustainable socioeconomic development. To address this issue, we conducted a comprehensive analysis of glacial morphological characteristics using multi-source time-series high-resolution remote sensing imagery spanning 2013–2024. Glacier boundaries were extracted through integrated methodologies combining manual visual interpretation, band ratio thresholding, three-dimensional geomorphic analysis, and an optimized DeepLabV3+ convolutional neural network with adaptive activation thresholds. Extraction accuracy was rigorously validated using quantitative metrics (Accuracy, Precision, Recall, Loss, and F1-score). Key findings reveal the following: dominant glacier types include ice caps, valley glaciers, and hanging glaciers distributed at mean elevations of 5200–5600 m; total glacial area decreased from 102.71 km² to 81.10 km², yielding an average annual decrease rate of −1.93%; glacier count increased from 74 to 86, corresponding to a mean relative change rate of 1.18% per annum; and thirty-eight geohazard sites were identified predominantly on upper slopes (30–50°) of north-facing terrain, with elevations ranging from 4500–5400 m (base) to 5120–6050 m (crest). These results provide critical data support for enhancing ecological resilience, strengthening disaster mitigation capabilities, and safeguarding public safety and infrastructure against climate change impacts in the region.

1. Introduction

Glaciers constitute critical freshwater reservoirs on Earth and serve as indispensable indicators of climate change, playing a pivotal role in global change research [1]. Across the Qinghai–Tibet Plateau and its surrounding mountain ranges, extensively distributed glaciers significantly regulate regional water resources, stabilize ecosystems, and underpin socioeconomic development. Under persistent global warming since the mid-20th century, these glaciers have experienced intensified surface ablation and accelerated thinning, driving a pronounced decline in ice mass with marked spatial heterogeneity [2]. Moreover, deglaciation processes have triggered cascading hazards including ice avalanches, debris flow–dammed lake–outburst flood chains, thermokarst slumping, and active-layer detachment failures, posing escalating threats to local communities and ecological integrity. Consequently, quantifying glacial dynamics [3] across the plateau is imperative for developing evidence-based climate adaptation frameworks and sustainable development strategies.

The Ányêmaqên Mountain glaciers in the headwaters of the Yellow River (Qinghai–Tibet Plateau) exhibit distinct vertical zonation in distribution patterns [1,3]. Documented as retreating since the Little Ice Age with accelerated mass loss during 1966–2000 [2,3,4], these glaciers entered a phase of reduced yet persistent recession post-2000. Contemporary observations confirm continuous surface lowering across major outlet glaciers [3,4], accompanied by an area shrinkage rate of 0.41% yr⁻¹. This sustained deglaciation has precipitated cascading geohazards—including glacial debris flows, ice collapses, and glacial lake outburst floods—that increasingly threaten regional ecological security and socioeconomic resilience. To address these challenges, this study leverages multi-source time-series high-resolution remote sensing data integrated with advanced intelligent interpretation models. Our systematic analysis of glacial morphological characteristics elucidates cryospheric response mechanisms to climate forcing, providing critical insights for sustaining hydrological cycles and ecosystem stability in this vital water tower region.

Rapid and high-precision glacier monitoring is therefore imperative for tracking cryospheric dynamics. The integration of multi-source remote sensing data has consequently emerged as a cornerstone in contemporary glaciological research. Pioneering studies on Qinghai–Tibet Plateau glaciers established critical baselines, exemplified by China’s Second Glacier Inventory (2015)—a systematic record documenting 48,571 glaciers covering 51,480 km² through field investigations and remote sensing analyses [5]. This inventory reveals consistent area reduction across the Himalayan glaciers during 1999–2015, aligning with global trends in negative mass balance under climate warming [6,7]. Advanced methodologies have since enhanced monitoring capabilities: Ref. [5] deployed Sentinel-1A SAR coherence analysis for glacier boundary delineation in major ice-covered regions (Kunlun, Karakoram, Himalaya, Qilian, Tanggula, Nyainqêntanglha, and Transhimalaya ranges). Bhambri et al. (2017) [8] identified 172 surge-type glaciers in the Karakoram (1840–2017) using terminus positions and velocity fields from Landsat/ASTER and field data. Multi-temporal DEM comparisons (Landsat, Corona KH-4, Hexagon KH-9, SRTM, AW3D30, ASTER GDEM, and TanDEM-X) enabled the authors of [9,10] to catalog 202 and 206 surging glaciers in the Pamirs (1960s–2018), respectively. Lv et al. (2022) [11] further detected 362 active glaciers across High Mountain Asia (HMA), with Pamirs (37%) and Karakoram (35%) constituting primary clusters. This collective evidence substantiates widespread glacier surge activity, particularly within the anomalous Karakoram sector.

Conventional rule-based or statistical remote sensing approaches face significant limitations—including suboptimal accuracy and feature extraction challenges—when delineating glacier boundaries, monitoring ice flow velocities, or quantifying area changes across the topographically complex Qinghai–Tibet Plateau. Recent advances in artificial intelligence have established deep learning as a transformative paradigm for remote sensing analysis, leveraging its superior nonlinear modeling capacity and automated feature extraction capabilities. Integrated with time-series analysis and transfer learning strategies, these models enable dynamic monitoring and predictive modeling of long-term glacial evolution. The integration of multi-source data (optical imagery, synthetic aperture radar [SAR], and digital elevation models [DEMs]) has further propelled the development of advanced deep learning frameworks. Architectures including convolutional neural networks (CNNs), U-Net variants, and Transformer models now achieve unprecedented precision in glacial feature extraction. Chen et al. (2022) [12] deployed an enhanced LandsNet architecture to extract glacier contours in the Three-River Headwaters region (1986–2021) using multi-temporal remote sensing data, systematically analyzing spatiotemporal changes with Gaofen-2 (GF-2) 1 m resolution imagery for debris-covered ice mapping. Yang et al. (2024) [13] integrated GF-2 reflectance, topographic features, and land surface temperature (LST) data through attention-enhanced U-Net, FCNN, and DeepLabV3+ ensembles to delineate debris-covered glaciers in the Central Karakoram. These innovations demonstrate deep learning’s capacity to overcome historical methodological constraints in cryospheric remote sensing.

Davari et al. (2022) [14] developed three automated post-processing techniques for calving front position (CFP) extraction—statistical thresholding, conditional random fields (CRFs), and optimized U-Net architectures—with the latter achieving superior contour delineation accuracy. Zhang et al. (2020) [15] demonstrated exceptional generalization across glacier types using multi-sensor fusion (Landsat-8/Sentinel-2 optical + Envisat/ALOS/TerraSAR-X/Sentinel-1 SAR) processed through ResNet/DRN/MobileNet-based U-Net and DeepLabv3+ frameworks. Peng et al. (2023) [16] attained 0.972 overall accuracy in Qilian Mountains glacier mapping by integrating Sentinel-1/2 data, HMA DEM, and SRTM topography within a novel U-Net variant featuring local–global transformer encoders and CNN blocks. Concurrently, multi-source approaches have enhanced hazard assessment. Wanget al. (2024) [17] characterized recurrent ice-rock avalanches in the Ányêmaqên Mountains through synergistic analysis of remote sensing, meteorological records, and field observations, revealing critical failure mechanisms for disaster prevention. An et al. (2021) [18] established an early warning system (EWS) for Brahmaputra Glacier hazard cascades (collapse → debris flow → river damming → outburst flood), successfully predicting GCRB events. Agarwal et al. (2023) [19] quantified debris-cover impacts on Himalayan glacier response using thickness modeling combined with multi-temporal Corona/Hexagon/Landsat imagery. These integrated methodologies provide transformative insights into cryospheric hazard dynamics.

Building upon these methodological advancements, this study employs multi-source time-series high-resolution remote sensing imagery integrated with a synergistic approach: manual visual interpretation, band ratio thresholding, 3D geomorphic analysis, and an optimized lightweight DeepLabV3+ convolutional neural network [20,21,22] incorporating activation threshold optimization. This integrated framework systematically quantifies spatiotemporal changes across the Ányêmaqên Mountain glaciers. Our principal contributions are threefold: 1. Innovative boundary extraction: We pioneer the application of activation threshold-optimized DeepLabV3+ architecture for multi-decadal glacier boundary delineation in the Ányêmaqên Mountain region, significantly enhancing long-term change detection accuracy. 2. Comprehensive cryospheric database: A foundational geodatabase documenting glacier area, inventory, and elevation dynamics (2013–2024) is established, enabling quantitative analysis of mass balance trends. 3. Geohazard mapping validation: Field-verified spatial distribution analysis of 38 geohazard sites provides empirical evidence for slope failure mechanisms, delivering critical scientific support for regional ecological security and climate-resilient development.

2. Materials and Methods

2.1. Study Area

The Ányêmaqên Mountain glacial region (34.81° N, 99.49° E) occupies a critical sector of the Qinghai–Tibet Plateau within the Ányêmaqên eugeosyncline belt of the Songpan-Garzê Indosinian fold system (Figure 1). This tectonically active domain exhibits well-developed fold-thrust structures, traversed by a 1000 km SW-dipping fault system (50–60° dip angle) along its northern margin [23]. According to China’s Second Glacier Inventory, 85 modern glaciers covering 87.33 km² persist above 5000 m elevation, predominantly comprising ice caps (42%), valley glaciers (35%), and hanging glaciers (23%). This concentration represents the largest cryospheric reservoir in the Yellow River headwaters. Notable features include: The Halong Glacier (NE slope): 7.7 km length, 23.50 km² area, 1800 m vertical relief; The longest ice tongue (5963.45 m) with 945 m elevation differential. The southwestern slope hosts numerous debris-prone valley glaciers, with Xiaomagou valley experiencing four major ice avalanche-debris flows (2004, 2007, 2016, 2022) involving the Yehelong, Weigeledangxiong, and Halong glaciers. Climatologically characterized by a continental regime, the area records: Mean annual temperature: −10.2 °C. Annual precipitation: ≈300 mm. Warm season (May–September): Synchronous precipitation and temperature maxima driving rapid meltwater generation. Cold season (October–April): Persistent snow/ice cover with ground freezing.

2.2. Utilization of Time-Series Multisource Remote Sensing Imagery in the Study Area

Current glacier research in the Ányêmaqên Mountain region predominantly relies on Landsat TM/OLI remote sensing data [24]. However, the spatial resolution of this satellite data is notably lower than that of contemporary Chinese high-resolution optical satellites, such as the GF (Gaofen) Series, and the Resource Series satellites (see

Table 1. Optical remote sensing imagery data sources.

Data Source	Sensor Type	Revisit Period (Days)	Number of Bands	Spatial Resolution		Scene Count (Scene)	Data Timeliness (Year)	Spatial Coverage (%)
				Panchromatic Band	Multispectral Bands
GF-1	PMS/WFV	4	4	2 m	8.0 m	4	2014, 2023, 2024	100%
GF1B/C/D	PMS	4	4	2 m	8.0 m	6	2020, 2021, 2022	100%
GF2	PMS	5	4	0.8 m	3.2 m	4	2015, 2017	100%
GF6	PMS/WFV	4	4	2 m	8.0 m	5	2019	100%
ZY3-02A	NAD/MUX	5	4	2.1 m	5.8 m	5	2016, 2018	100%
ZY1-02C	PMS/HRC	3	4	2.3 m	5.0 m	6	2013	100%
DEM		-	-	12.5 m		-	2011	100%
Glacier Inventory Compilation		-	-	-		-	2007	100%

Note(s): Within the table, PMS denotes the Panchromatic Multispectral Sensor, WFV represents the Wide Field-of-View multispectral sensor, HRC stands for the High-Resolution Camera, NAD refers to the Nadir Panchromatic Sensor, and MUX indicates the Multispectral Sensor.

). Consequently, this study employs data from the GF Series and Resource Series satellites. Due to variations in the acquisition cycles, overpass times, and spatial coverage of valid images (i.e., cloud-free, snow-free, and temporally proximate to the end of the ablation season) across different satellites within the study area, the selection of imagery sources was optimized. This approach ensures the chosen data maximizes both the temporal span and the spatial coverage necessary to effectively support the objectives of this regional investigation.

Following the interpretation area delineation outlined in Section 2.1, multi-temporal remote sensing data from the periods 2013–2024 were selected. The dataset comprises imagery acquired by GaoFen-1 (GF-1), GaoFen-6 (GF-6), the 2 m/8 m optical satellites (GF-1B, C, D), Ziyuan-3-01 (ZY3-01), and Ziyuan-02C (ZY02C). Data selection prioritized high spatial resolution (1 m, 2 m) and acquisition timing (June to September) [25], As displayed in Figure 2. The selected imagery is characterized by generally low noise levels, an absence of bad pixels, minimal cloud and shadow cover (<5%), and high interpretability [26].

The optical satellite imagery was sourced from the High-Resolution Earth Observation System of the Qinghai Provincial Remote Sensing Center for Natural Resources. Glacier inventory data were obtained from the National Tibetan Plateau Data Center (http://data.tpdc.ac.cn/zh-hans/ accessed on 15 January 2025). The Digital Elevation Model (DEM) was derived from ALOS PALSAR data with a spatial resolution of 12.5 m, accessed via the United States Geological Survey (USGS) Earth Explorer platform (https://earthexplorer.usgs.gov accessed on 15 January 2025). The specific optical remote sensing imagery data sources employed in this study are summarized in Table 1.

2.3. Methods

2.3.1. Processing of Multi-Temporal Remote Sensing Imagery and Visual Interpretation

For the acquired twelve epochs of high-resolution remote sensing imagery (2013–2024), the following processing sequence was implemented: First, panchromatic and multispectral images from various optical satellites [27] were subjected to rigorous preprocessing. This included RPC parameter resolution, bundle block adjustment with aerial triangulation for densification, geometric correction of both panchromatic and multispectral bands, image fusion, enhancement, and subsequent mosaicking and clipping. These steps generated foundational true-color and false-color Digital Orthophoto Maps (DOMs) for analysis [28]. Subsequently, the derived true/false-color DOMs were utilized for glacier boundary delineation: Pixel offset vectors between consecutive images were calculated. These displacement fields aided in discriminating glacier boundaries obtained via band ratio methods and initial manual visual interpretation (the corresponding band ratio map is presented in Figure 3).

Key processing parameters—including initial and final window sizes, masking thresholds, step sizes, and iteration counts—were optimized specifically for different sensors. Noise mitigation involved: removal of decorrelation noise by applying signal-to-noise ratio (SNR) thresholds and filtering; elimination of orbital errors through polynomial trend surface fitting; correction of striping artifacts using along-track pixel averaging; and suppression of residual noise via quadtree filtering. Median filtering was applied to binary images to minimize shadow noise. Finally, refined glacier boundaries were extracted: Boundaries initially identified through displacement tracking and band ratio methods were fused. To achieve the highest possible boundary resolution, these fused boundaries were visually inspected at a large scale within a GIS platform. Expert-guided manual interpretation was performed, involving cross-verification across multi-temporal images and revision against the Second Glacier Inventory data. This comprehensive process enabled the extraction of all pertinent glacier feature information [29].

2.3.2. Deep Learning-Based Intelligent Interpretation of Glacier Boundaries in the Ányêmaqên Mountain Region

This study employed the DeepLabV3+ deep learning model (proposed by Google in 2018), which integrates an Atrous Spatial Pyramid Pooling (ASPP) module with an encoder–decoder architecture. Model training was conducted using annotated samples derived from manual visual interpretation and band ratio techniques. Utilizing multi-source optical remote sensing data (ZY02C, ZY302A, GF1, GF2, GF1B/C satellite imagery) acquired as described in Section 2.3.1, automated image segmentation and classification were performed. An optimized activation threshold strategy was implemented to enhance the precision of glacier extraction. Post-processing involved noise removal using morphological filtering (opening and closing operations) [30] and terrain-based correction of misclassified regions using slope and shadow data. This methodology enabled high-precision interpretation of glacier area and count within the Ányêmaqên Mountain region.

The DeepLabV3+ model was selected primarily due to its superior performance in semantic segmentation, particularly for glacier boundary delineation tasks. This effectiveness stems from its core Atrous Spatial Pyramid Pooling (ASPP) module, which employs parallel atrous convolutions with varying dilation rates. This architecture enables efficient aggregation of multi-scale contextual information—a critical capability for identifying glaciers. Glacier features in remote sensing imagery (e.g., accumulation basins, ice tongues, crevasses) manifest across diverse scales and exhibit spectral signatures frequently confounded by topographic shadows, snow cover, and exposed bedrock, necessitating robust contextual understanding for accurate discrimination. Furthermore, DeepLabV3+ utilizes an encoder–decoder structure. The encoder, typically based on a powerful backbone network like ResNet, extracts high-level semantic features. The decoder subsequently refines target boundary localization precision by progressively upsampling feature maps and integrating them with lower-level, spatially detailed features from the encoder’s shallower layers. This capability is essential for accurately delineating the complex and irregular boundaries characteristic of glaciers. The model also demonstrates significant robustness in complex terrains. The use of atrous convolutions expands the receptive field without resorting to pooling operations, thereby mitigating excessive loss of spatial detail. This design ensures accurate identification of glacier bodies and boundaries even within challenging high-mountain and canyon environments [31]. Consequently, DeepLabV3+ was chosen as the optimal model for glacier boundary extraction in this study, leveraging its strengths in multi-scale feature fusion, precise boundary recovery, and adaptability to complex geographical scenes.

Activation threshold optimization constitutes a critical post-processing step for refining the final segmentation accuracy of the DeepLabV3+ model. Its core objective is identifying the optimal threshold within the model’s output probability map that best discriminates between “glacier” and “non-glacier” pixels. In this study, to transform the continuous probability map into a discrete binary mask delineating glacier boundaries, we optimized the conventional practice of binarization using a fixed threshold (e.g., 0.5). This optimization is essential because spectral confusion (e.g., between glaciers, accumulated snow, topographic shadows, and exposed bedrock), ambiguous boundaries, and thin ice zones often lead to uncertain model predictions (typically exhibiting probability values fluctuating between 0.3 and 0.7) in these critical areas. Applying a fixed threshold of 0.5 invariably introduces significant noise or loss of fine detail. To address this, we systematically evaluated segmentation accuracy across a range of thresholds on the validation set. Threshold values were exhaustively traversed within a reasonable range (e.g., 0.1 to 0.9, with increments of 0.01 or 0.05). For each probability map in the validation set, corresponding binary masks were generated by applying each candidate threshold. The optimal threshold was then determined by maximizing the mean Intersection over Union (IoU) across the entire validation set. Finally, this optimized threshold was applied to binarize the probability maps of all test images, ensuring precise and robust extraction of glacier boundaries. The detailed technical procedure is depicted in Figure 4.

This study implemented a DeepLabV3+-based deep learning framework for glacier boundary classification, following this core workflow: Regions of Interest (ROIs) encompassing target glaciers were extracted from annual remote sensing imagery. Each scene, comprising four spectral bands (Red (R), Green (G), Blue (B), and Near-Infrared (NIR)), was uniformly cropped to 572 × 572 pixels to ensure spatial scale consistency across samples. To enhance model generalization and robustness for glacier boundaries under complex terrain and varying illumination, systematic data augmentation techniques were applied, including random rotation, horizontal flipping, contrast adjustment, and Gaussian noise injection. DeepLabV3+ was selected as the core model architecture. During training, an activation threshold optimization strategy (via traversal of candidate threshold parameters) was implemented alongside cross-validation techniques to effectively mitigate overfitting risk. Model performance was evaluated with the primary optimization objective of maximizing the mean Intersection over Union (mIoU) on the validation set [31]. Furthermore, considering the feature space diversity of the dataset, the optimal number of training epochs was experimentally determined and ultimately set to 25.

The converged model was then employed to perform automated pixel-level classification and identification of glacier boundaries on the independent test set imagery. To ensure result reliability and statistical validity, a stratified random sampling strategy (stratified by acquisition year) partitioned the dataset into training and test sets at a 3:1 ratio. Specifically, 1200 images per year served as the base training samples. Following data augmentation, the cumulative effective training dataset across all 12 years reached 14,400 images. Concurrently, 400 images per year were reserved, forming an independent test set totaling 4800 images for objective assessment of model generalization capability.

To enhance the model’s generalization capability and mitigate the risk of overfitting, particularly given the limited initial number of manually interpreted samples, a comprehensive data augmentation strategy was employed during the training phase. Each training image-mask pair was dynamically transformed in real-time during each epoch using the following operations: (1) random rotation within a range of ±30°; (2) horizontal and vertical flipping with a 50% probability; (3) random brightness and contrast adjustments with a variation factor of up to ±15%; and (4) additive Gaussian noise with a mean of zero and a standard deviation of 0.01 applied to the input pixel values. These techniques artificially expanded the diversity of the training dataset, encouraging the model to learn invariant features of glaciers under varying orientations, illumination conditions, and minor noise perturbations. Furthermore, to ensure a rigorous and unbiased evaluation of the model’s performance, the dataset was meticulously partitioned into independent training, validation, and test sets using a stratified random sampling strategy. The stratification was performed based on the image acquisition year to guarantee that all sets contained representative samples from the entire temporal range (2013–2024), preventing any temporal bias from influencing the performance metrics. The dataset was split in an approximate ratio of 3:1:1. Specifically, from the total annual pool of samples, 60% were allocated for training, 20% for validation (used for hyperparameter tuning and early stopping), and the remaining 20% were held out as a completely independent test set. Crucially, the test set was strictly isolated from the training process at all stages; its images and corresponding ground truth masks were never used for data augmentation, model training, or parameter optimization. This protocol ensures that the reported accuracy metrics (e.g., >96%) reliably reflect the model’s true generalization performance on unseen data from novel years and conditions [31,32]. Through this systematic experimental design, rigorous model optimization, and stringent validation protocol, this study provides an efficient and robust deep learning solution for long-term, high-precision boundary mapping of glaciers within the Ányêmaqên Mountain region.

2.3.3. Accuracy Evaluation Metrics

Glacier Interpretation Accuracy Metrics

Manual visual interpretation of glacier boundaries from remote sensing imagery inherently introduces uncertainties in boundary delineation [32,33]. These inaccuracies, arising from factors such as image quality and interpreter subjectivity, inevitably propagate errors into calculated glacier area estimates. Studies by [34] demonstrate a strong correlation between glacier area error and the spatial resolution of the employed imagery. To quantify this interpretation-induced area error, we adopted the established methodology of [35].

$$\mathrm{Error}\left(1\sigma \right)={\displaystyle \frac{\mathrm{P}}{\mathrm{G}}}\ast {\displaystyle \frac{{\mathrm{G}}^2}{2}}\ast \sigma$$

(1)

$$\mathrm{E} ={\displaystyle \frac{\mathrm{Error}\left(1\sigma \right)}{A}}\ast 100\%$$

(2)

where

P denotes the glacier perimeter (m). G represents the spatial resolution of the remote sensing imagery (0.7–2.0 m). σ is the weighting factor for stochastic error, typically assigned a value of 0.6872, assuming a one standard deviation confidence interval. A signifies the total glacier area. E indicates the relative area error. Building upon this formula, the cumulative area error for all glaciers within the study region was derived based on error propagation theory, calculated as follows:

$${\mathrm{E}}_{\mathrm{T}} =\sqrt{{{\displaystyle \sum }}_{i=1}^{\mathrm{n}}{\mathrm{a}}_i^2}$$

(3)

where

${\mathrm{E}}_{\mathrm{T}}$ represents the total area error for all glaciers within the study region. $i$ denotes the number of glaciers. $\mathrm{a}$ signifies the error in the area for an individual glacier i.

Deep Learning Performance Metrics

This study employed established deep learning performance metrics—Pixel Accuracy (PA), Loss, Precision, Recall, and F1-Score [36,37]—to comprehensively evaluate the DeeplabV3+ model’s performance in glacier boundary extraction. Pixel Accuracy (PA) quantifies the proportion of correctly classified pixels relative to the total number of pixels, reflecting the model’s overall classification accuracy across all classes (glacier and non-glacier). Loss is a dimensionless scalar value representing the discrepancy between model predictions and the validation data during training. A Loss of 0 indicates perfect agreement; higher values signify greater prediction error. Precision (User’s Accuracy) measures the proportion of correctly classified glacier pixels among all pixels predicted as glacier by the model. Recall (Producer’s Accuracy) measures the proportion of actual glacier pixels correctly identified by the model. F1-Score, the harmonic mean of Precision and Recall, provides a balanced assessment of model performance. A higher F1-Score indicates robust performance in minimizing both false positives (misclassification) and false negatives (omissions), reflecting a more balanced overall capability for the glacier extraction task. The formulas for these metrics are defined as follows [38]:

$$\mathrm{Pa} ={\displaystyle \frac{{\sum }_{c=1}^C{\mathrm{T}}_{\mathrm{c}}}{\mathrm{M}}}$$

(4)

$$\mathrm{Loss}=-{\displaystyle \frac{1}{\mathrm{N}}}{\displaystyle \sum }_{\mathrm{i}=1}^{\mathrm{N}}\left[{\mathrm{y}}_{\mathrm{i}}\log \left(\widehat{{\mathrm{y}}_{\mathrm{i}}}\right)+\left(1-{\mathrm{y}}_{\mathrm{i}}\right)\log \left(1-\widehat{{\mathrm{y}}_{\mathrm{i}}}\right)\right]$$

(5)

$$\mathrm{Precision}={\displaystyle \frac{{\mathrm{T}}_{\mathrm{c}}}{{\mathrm{T}}_{\mathrm{c}}+{\mathrm{F}}_{\mathrm{c}}^{+}}}$$

(6)

$$\mathrm{Recall}={\displaystyle \frac{{\mathrm{T}}_{\mathrm{c}}}{{\mathrm{T}}_{\mathrm{c}}+{\mathrm{F}}_{\mathrm{c}}^{-}}}$$

(7)

$$\mathrm{F}1-\mathrm{Score}={\displaystyle \frac{2·\mathrm{Precision}·\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}}$$

(8)

where

$C$ denotes the total number of classes. ${\mathrm{T}}_{\mathrm{c}}$ represents the number of correctly classified pixels for class c. $\mathrm{M}$ signifies the total number of pixels in the image. $\mathrm{N}$ indicates the number of samples. ${\mathrm{y}}_{\mathrm{i}}$ is the true label value (0 or 1). $\widehat{{\mathrm{y}}_{\mathrm{i}}}$ is the predicted probability output by the model. ${\mathrm{F}}_{\mathrm{c}}^{+}$ (False Positives for class c) is the number of pixels incorrectly classified as class c. ${\mathrm{F}}_{\mathrm{c}}^{-}$ (False Negatives for class c) is the number of pixels belonging to class c that were incorrectly classified as other classes.

Glacier Change Accuracy Metrics

To quantify glacier dynamics, we employed two key metrics: Glacier Area Change Rate and Relative Glacier Area Change Rate, calculated as follows:

$$V_{GAC}={\displaystyle \frac{G{A}_s-G{A}_f}{Y_{f-s}}}\sigma$$

(9)

$$P{V}_{GAC}=\left[{\left({\displaystyle \frac{G{A}_s}{G{A}_f}}\right)}^{{\displaystyle \frac{1}{Y_{f-s}}}}-1\right]\times 100\%$$

(10)

$$Y_{f-s}={\displaystyle \frac{{\sum }_{i=1}^m{A}_i\cdot Y_i}{{\sum }_{i=1}^m{A}_i}}-{\displaystyle \frac{{\sum }_{j=1}^n{A}_j\cdot Y_j}{{\sum }_{j=1}^n{A}_j}}$$

(11)

where

$V_{GAC}$ denotes the glacier area change rate (km²/a). $P{V}_{GAC}$ represents the relative glacier area change rate (%/a). $G{A}_s$ and $G{A}_f$ are the glacier areas (km²) derived from the second and first glacier inventories, respectively. $Y_{f-s}$ is the time span (years) between the acquisition dates of the data sources used for the two inventories. $A_i$ signifies the area (km²) of the i glacier in the second glacier inventory. $A_j$ signifies the area (km²) of the j glacier in the first glacier inventory. $Y_i$ indicates the source data year for the i glacier in the second inventory. $Y_j$ indicates the source data year for the j glacier in the first inventory. m is the total number of glaciers within the basin for the second inventory. n is the total number of glaciers within the basin for the first inventory.

3. Results

3.1. Glacier Boundary Delineation Data Preprocessing

3.1.1. Glacier Boundary Preprocessing: Band Ratios and Manual Delineation

Glacier boundaries within the Ányêmaqên Mountain region were delineated for the period 2013–2024 (spanning 12 years) using semi-automated and manual classification techniques. These methods leveraged the spectral, texture, and shape characteristics of high-resolution remote sensing imagery. Primary approaches employed included manual visual interpretation, band ratio analysis, 2D visualization (Figure 5a), and 3D stereoscopic visualization (Figure 5b). The delineation process referenced the Second Chinese Glacier Inventory Dataset and incorporated topological checks and boundary refinements. Key results derived from this process encompass: annual glacier boundaries, glacier area, glacier count, glacier morphology, glacier affiliation, and glacier distribution across different elevation zones [39]. Furthermore, an interactive manual quality control procedure was implemented. This involved scrutinizing the vector boundaries of each individual glacier to identify and verify instances of advancing glaciers, disappearing glaciers, divided glaciers, and glaciers affected by terrain shadows or seasonal snow cover. The resulting vector and raster datasets serve as the foundational data for constructing the labeled sample dataset required for subsequent deep learning applications.

3.1.2. Subsubsection Deep Learning Model Input Preparation for Glacier Boundary Delineation

Prior to deep learning model training, orthorectified base imagery for all years within the study area was generated using optical remote sensing satellite data. Subsequently, a labeled raster sample dataset spanning the 12-year period from 2013 to 2024 was constructed [40]. This dataset utilized the annual glacier boundary vectors derived in Section 3.1.1 and their corresponding high-resolution images as the ground truth labels. As evident from Figure 6, variations in the image sources selected for different years result in corresponding variations in the glacier boundary representations within the labeled raster samples for the Ányêmaqên Mountains region. Crucially, however, the labeled raster samples for each individual year achieve maximal spatial coverage of the study area. This comprehensive annual coverage provides high-precision foundational data essential for the development and training of subsequent deep learning models.

3.2. Accuracy Analysis

3.2.1. Accuracy Assessment of Band Ratio/Manual Delineation

During the glacier boundary delineation process, the resolution-induced error in glacier area interpretation was relatively minor, quantified at ±0.42 km². This error typically represents only 0.58% to 2.46% of individual glacier area. Crucially, the absolute magnitude of this error is significantly lower than the substantial changes in glacier area observed across the study region over the 2013–2024 period. Furthermore, the area error exhibits a statistically significant positive correlation with glacier size [41]. Given its small magnitude and the fact that it is substantially smaller than the scale of actual glacier changes observed, this resolution-induced error has a negligible impact on the principal conclusions of this study concerning the overall trend and magnitude of glacier area change. It does not fundamentally alter the core understanding of regional glacier change dynamics derived from our analysis. Consequently, this error source is thus considered negligible within the context of the present study.

3.2.2. Accuracy Assessment of the Deep Learning Model

To comprehensively evaluate the performance of the DeepLabV3+ deep learning model for glacier boundary delineation in the Ányêmaqên Mountain region, the following standard accuracy metrics were employed after model training: Accuracy, Precision, Recall, F1 Score, and Intersection over Union (IoU). Model training and analysis were conducted using different classification thresholds. Figure 7 presents the model’s accuracy metrics under three representative threshold values. The model loss decreased progressively from an initial value of 0.186 to 0.152, indicating improved alignment with the validation data throughout the iterative training process. User’s Accuracy (equivalent to Precision) increased significantly from an initial 39.8% to a final 92.3%. This demonstrates a substantial increase in the proportion of correctly classified glacier pixels within the model’s predicted glacier class for the Ányêmaqên Mountain region, reflecting enhanced prediction reliability. Producer’s Accuracy (equivalent to Recall) rose markedly from an initial 10.2% to a final 98.3%. This indicates a strong increase in the proportion of actual glacier pixels within the Ányêmaqên Mountain region that were correctly identified by the model, signifying improved detection capability of glacier extent [42]. The F1 Score, representing the harmonic mean of Precision (User’s Accuracy) and Recall (Producer’s Accuracy), increased substantially from an initial 0.62 to a final 0.96. This significant rise confirms a major overall improvement in the model’s balanced classification accuracy. Collectively, these results provide robust accuracy support for the subsequent application of this model in extracting and classifying annual glacier boundaries within the Ányêmaqên Mountain region.

3.3. Deep Learning Implementation and Results

Based on the delineated glacier boundaries, binary masks were generated for each year’s imagery. Within these masks, pixel values of 1 represent glacier area, while 0 represents non-glacier terrain. These annual masks were chronologically sequenced to form a continuous temporal record, enabling the analysis of glacier area change trends. To enhance both the quantity and diversity of the training samples, the original binary masks underwent data augmentation employing spatial transformations (rotation, scaling) and noise addition techniques. Furthermore, multi-year data integration was performed by strategically combining labeled data across different years. This approach ensures the model captures long-term glacier change dynamics. Adhering to the temporal nature of the dataset, the data were partitioned chronologically into training, validation, and test sets. Specifically:

Data from the initial years constituted the training set. Data from intermediate years formed the validation set. Data from the most recent years comprised the test set [43].

This temporally segmented dataset was used to train the image segmentation model. During the training phase, the validation set was utilized for hyperparameter tuning and monitoring model performance to prevent overfitting. Multiple rounds of iterative optimization were conducted to refine both the model architecture and hyperparameter configurations, ultimately enhancing the model’s overall generalization capability.

The DeepLabV3+ model was trained for glacier boundary delineation within the Ányêmaqên Mountain region. During the training process, input images (comprising four spectral bands) underwent tile partitioning into patches of 572 × 572 pixels. These patches were then processed through the model’s hierarchical architecture, leveraging iterative feature extraction via multiple convolutional layers. All models were trained for 25 epochs. To address the limitations of the original training dataset size and enhance the model’s ability to recognize features (glaciers) of varying orientations and scales, data augmentation was implemented. This involved spatial transformations (scaling and rotation) [44]. Concurrently, the model underwent continuous optimization throughout the training phase. The final prediction outputs (class activation maps) for all years are represented as single-band grayscale rasters. Each pixel value within these rasters signifies the predicted probability of that pixel belonging to the glacier boundary class.

Analysis of the annual class activation probability rasters (Figure 8) reveals interannual variations in their grayscale value distributions. These variations arise from differences in satellite data sources and image quality across the years. The grayscale values (representing pixel-wise probability) for the 2013–2024 period range between 0.20 and 0.99, while the corresponding confidence level distribution ranges from 0.83 to 0.92. Comparative assessment against manual delineation results (Figure 9) indicates that the accuracy of the deep learning-based glacier boundary extraction is primarily influenced by spatial resolution of the imagery, temporal resolution (acquisition timing), extent of cloud and snow cover [45]. Notwithstanding these influencing factors, the overall classification accuracy consistently exceeds 96% throughout the 2013–2024 period. This high accuracy demonstrates the model’s reliability in extracting glacier boundary features and effectively characterizing the spatial extent of glaciers within the Ányêmaqên Mountain region.

4. Discussion

4.1. Glacier Area and Changes Across Different Periods

Glaciers within the Ányêmaqên Mountains, situated at the headwaters of the Yellow River, have experienced sustained retreat since the Little Ice Age. This retreat exhibited an accelerating trend between 1966 and 2000 [2]. Although the rate of retreat moderated somewhat during the subsequent decade (post-2000), the glaciers continued to shrink overall. During this period, significant surface lowering was observed on representative large glaciers. This persistent retreat is primarily driven by rising temperatures. Analysis of twelve epochs of high-resolution remote sensing imagery (2013–2024) confirms that the glaciers in the Ányêmaqênregion are predominantly continental-type valley glaciers. The results demonstrate a continued net reduction in total glacier area throughout this recent period [46]. Remote sensing interpretation indicates that this areal loss is largely attributable to the retreat of glacier tongues. Detailed quantification of these changes is provided in

Table 2. Statistics of Glacier Area and Changes Across Different Periods.

Year	(2009)	2013	2014	2015	2016	2017	2018	2019	2020	2021	2022	2023	2024
Glacier Count (individual)	74	80	80	81	82	82	83	84	85	84	84	85	86
Relative Area Change (km2)	0.00	6.00	0.00	1.00	1.00	0.00	1.00	1.00	1.00	−1.00	0.00	1.00	1.00
Mean Annual Change Rate (%)	0.00	8.11	0.00	1.25	1.23	0.00	1.22	1.20	1.19	−1.18	0.00	1.19	1.18
Glaciated Area (km2)	102.71	98.46	98.31	94.64	91.63	91.79	91.01	90.71	87.33	86.98	86.26	81.20	81.10
Annual Net Change (km2)	0.00	−4.25	−0.15	−3.67	−3.01	0.16	−0.78	−0.30	−3.38	−0.35	−0.72	−5.06	−0.10
Change Rate (%)	0.00	−4.14	−0.15	−3.73	−3.18	0.17	−0.85	−0.33	−3.73	−0.40	−0.83	−5.87	−0.12

Note(s): The 2009 dataset in this table corresponds to the official compilation results of China’s Second Glacier Inventory.

.

Between 2009 and 2024, the total glacier area decreased from 102.71 km² to 81.10 km², representing a net loss of 21.61 km². This corresponds to an average annual decrease rate of −1.93% a⁻¹. Periods of Accelerated Retreat: The most substantial annual area losses occurred during 2009–2013 (−4.25 km², retreat rate: −4.14% a⁻¹) and 2022–2023 (−5.06 km², retreat rate: −5.87% a⁻¹). Periods of Reduced Retreat: Comparatively minor annual losses were observed during 2013–2014 (−0.15 km², retreat rate: −0.15% a⁻¹) and 2023–2024 (−0.10 km², retreat rate: −0.12% a⁻¹). While the annual area change and retreat rate exhibited interannual variability, the overall magnitude of year-to-year fluctuations was relatively small [47]. Crucially, the dominant trend throughout the 2009–2024 period was one of persistent areal reduction. Comparative image analysis confirmed that this areal loss was primarily driven by the retreat of glacier tongues (Figure 10a).

Concurrently, the number of individual glaciers increased from 74 in 2009 to 86 in 2024, representing a net gain of 12 glaciers. This corresponds to an average annual relative change rate of +1.18% a⁻¹. The most significant increase in glacier count occurred during 2009–2013 (+6 glaciers, relative change rate: +8.11% a⁻¹). No change in glacier count was recorded between 2021 and 2022 (count remained at 84 glaciers). Despite exhibiting year-to-year fluctuations, the overall trend in glacier count was one of gradual increase. However, the absolute magnitude of the annual changes remained relatively modest (Figure 10b).

Analysis of the Ányêmaqên Mountain glaciers over the 15-year period from 2009 to 2024 reveals discernible morphological changes [48], characterized primarily by alterations in total area and number. As synthesized from Table 2 and Figure 10, the glacierized area exhibited a marked reduction trend, while the number of distinct glaciers showed a concomitant increase. The interannual variability in glacier shrinkage rates can be attributed to a combination of climatic anomalies and glacier-specific responses. For instance, the accelerated retreat during 2009–2013 aligns with a period of persistently above-average summer temperatures and reduced snowfall in the region, leading to enhanced ablation. Conversely, the notably slow retreat during 2013–2014 coincided with cooler summer conditions and higher accumulation in the preceding winter, temporarily mitigating mass loss. The pronounced retreat in 2022–2023 may be further exacerbated by increased debris cover and proglacial lake development, amplifying melt-through thermal regulation and calving processes. Such year-to-year fluctuations highlight the complex interplay between climate forcing and non-climatic factors such as glacier geometry, hypsometry, and debris cover, which modulate the sensitivity of individual glaciers to atmospheric warming.

Examination of meteorological data (precipitation, temperature, solar radiation, humidity) indicates that this pattern is attributable to rising mean annual air temperatures in the region, consistent with the effects of global warming. This warming has driven accelerated melting and glacier snout retreat, leading to the fragmentation of larger ice bodies into smaller, discrete glaciers. Consequently, the overall glacier area decreased, while the count increased due to this fragmentation process. To further investigate the drivers of this observed change pattern, the distribution of glaciers across different elevation zones was analyzed.

Analysis of

Table 3. Distribution characteristics of glaciers across different elevation zones.

Area (km2)	Observation Year	Mean Elevation Range (m)
		<5100	5100–5200	5200–5300	5300–5400	5400–5500	5500–5600	>5600
Glaciers in the Ányêmaqên Mountains	2009	1.4	24.9	35.2	1.7	28	6.1	5.4
	2013	0.67	5.13	51.25	2.3	27.83	5.98	5.3
	2014	0.64	5.11	51.88	2.02	27.66	5.93	5.07
	2015	0.42	4.65	50.85	1.71	26.41	5.63	4.97
	2016	0.53	4.02	50.15	1.65	25.01	5.41	4.86
	2017	0.53	4.17	50.1	1.65	25.02	5.42	4.9
	2018	0.61	4.31	49.46	1.73	24.70	5.99	4.21
	2019	0.57	5.54	49.65	2.29	23.44	4.59	4.63
	2020	0.45	3.45	49.46	1.55	22.71	5.04	4.67
	2021	0.02	3.53	48.96	1.99	22.49	4.85	5.14
	2022	0.3	3.27	49.02	1.37	21.38	4.99	5.93
	2023	0.27	3.01	48.34	1.87	17.29	5.25	5.17
	2024	0.25	3.01	48.04	1.96	17.32	5.25	5.27

reveals distinct characteristics and spatial heterogeneity in glacier changes and distribution across different zones and time periods within the Ányêmaqên Mountains, varying significantly with elevation [49]. Glaciers in this region are predominantly concentrated within the 5200–5400 m elevation band. This zone exhibited substantial overall area loss, accounting for 57.26% (mean area: 50.47 km²) of the total glacierized area during 2009–2024. Notably, interannual area fluctuations within this elevation band were minor, with values generally oscillating near 51 km², indicating relatively stable glacier behavior. Conversely, glacier coverage below 5200 m is limited. This lower zone, corresponding to the primary snout distribution area, represented only 5.12% (mean area: 6.21 km²) of the total area. However, this elevation band experienced more pronounced interannual area variations [50], suggesting heightened activity associated with glacier termini dynamics.

The 5400–5600 m elevation band also exhibits significant glacier concentration, albeit reduced compared to the 5200–5400 m zone. During the 2009–2024 period, this band contained a mean glacierized area of 29.26 km², accounting for 31.65% of the regional total. Functioning primarily as a glacier accumulation zone, this elevation band displayed minimal interannual area variability, with values generally fluctuating near 30.12 km², indicative of stable glacier behavior (Figure 11a). Above 5600 m, constrained by the topographic limits of the Ányêmaqên Mountains [51], glacier coverage diminishes substantially. This highest zone maintained a mean area of only 5.04 km² (5.89% of the total) during the study period. Characterized as a perennial accumulation zone, glaciers in this altitudinal range exhibit lower responsiveness to climatic fluctuations (Figure 11b). Notably, this zone demonstrated a net expansion trend over time. Collectively, these trends reveal that glacier retreat rates progressively diminish with increasing elevation, with net area gain observed in some years at higher altitudes. However, the expansion rates observed at higher elevations remain substantially lower than the retreat rates dominating lower-elevation glaciers.

4.2. Development of Glacial Hazards

Current classification systems for ice avalanche hazards are limited and lack standardization. Given the significant differences in prevention and mitigation strategies required for different ice avalanche types, a unified framework is essential. Building upon prior research [52], this study proposes a novel, comprehensive, operational, and standardized classification system for ice avalanche hazards on the Qinghai–Tibet Plateau, based on failure mechanisms and hazard manifestation: (1) Direct Ice Avalanche Hazard: Characterized by the detachment and collapse of slope glaciers under gravity, generating high-velocity debris flows incorporating ice and rock. These flows cause severe damage to settlements and infrastructure in their path. (2) Ice Avalanche-Triggered Glacial Lake Outburst Flood (GLOF) Hazard: Initiated by glacier collapse entering a proglacial or supraglacial lake, generating displacement waves that breach the lake dam, resulting in catastrophic flooding and downstream loss of life and property. (3) Ice Avalanche-Damming and Outburst Flood Chain Hazard: Occurs when debris from a glacier collapse blocks a river channel, forming an ice/rock avalanche-dammed lake. Subsequent dam failure leads to outburst flooding, causing downstream loss of life and property. This tripartite classification encompasses the spectrum of hazards triggered by glacier collapse, establishing a foundational framework for systematic research into ice avalanche hazards on the Qinghai–Tibet Plateau.

Multi-temporal analysis of high-resolution remote sensing imagery within the Ányêmaqênkey monitoring area enabled the identification of 38 glacial hazard sites (Figure 12). This inventory was compiled based on the following criteria: (1) proximity of potential impact zones (e.g., flow paths, depositional areas) to vulnerable assets such as settlements and infrastructure [53]; (2) presence of clear indicators of past ice avalanches or debris flows triggered by glacial melt and (3) immediate threat proximity to at-risk objects. The identified hazards comprise 34 direct ice avalanche sites and 4 ice avalanche-damming and outburst flood chain sites. Spatially, these hazards are clustered within distinct sectors: five ice avalanche sectors (totalling 21 sites) are located on north-facing slopes, while six sectors (totalling 17 sites) occur on south-facing slopes. One sector is classified as medium-scale; the remaining ten are classified as large-scale. All hazard sectors originate within the mid-to-upper sections of mountain slopes (gradient: 30–50°), spanning elevations from 4500–5400 m (minimum) to 5120–6050 m (maximum). The relative relief within these sectors ranges from 250 to 1,300 m. Individual ice avalanche bodies exhibit the following dimensions: length (250–1300 m), width (300–3300 m), thickness (15–30 m), and estimated volume (7.5 × 10⁴ m³ to 1.7 × 10⁸ m³). Owing to the high altitudes of these features, vegetation cover is virtually absent.

Post-failure debris from ice avalanches consistently deposited within valley outlets at the base of glacial slopes. These deposits exhibit triangular or irregular geometries, often forming overlapping lobes, with thicknesses ranging from 20 to 40 m. The predominantly unvegetated surfaces consist of exposed glacial till. These hazards pose significant threats to human life, transportation infrastructure (roads), dwellings, and livestock. Observed horizontal runout distances range from 500 to 2300 m. While no direct property damage has been recorded to date, the assessed hazard magnitude classifies these events as moderate to severe. A representative case is site ANMQ11-1, which in February 2004 exhibited a runout distance of approximately 5500 m, destroying roughly 2000 m of a provincial highway. Current assessments indicate that the ice avalanche bodies remain in an unstable state, with a projected trend of accelerating deformation. Primary triggering mechanisms include seismic activity, freeze–thaw cycles, and intense rainfall [54].

4.3. Glacial Collapse Chain Disasters at the Maqên Gangri Glacier, Xiaoma Valley

Since 2004, three distinct chain-reaction disasters—involving glacial collapse, debris flows, landslide dam formation, and subsequent dam-break flooding—have occurred on the south slope of the Ányêmaqên Mountains at the Maqên Gangri Glacier in Xiaoma Valley. These events were recorded in February 2004, 8 October 2007, and 6 October 2016. Additionally, four separate glacial collapse–debris flow–damming events occurred in August 2022. The glacier is currently highly fractured, indicating a significant probability of recurrent chain-disaster events. The February 2004 event was the most catastrophic and established conditions facilitating subsequent hazards. A large-scale ice avalanche originating at ~5900 m on the main peak plunged 1150 m vertically. Transforming into a debris flow that entrained fragmented weathered bedrock, it inundated the valley outlet at ~4300 m elevation. This deposit blocked a 1500 m section of the channel, impounding a lake. Continued water inflow created a significant impoundment breach risk. The deposit, oriented 251° azimuth, measured 2.4 km (E-W) by 1.5 km (N-S), covered an area of 2.4 km², averaged 10 m in thickness, and had an estimated volume of ~24 million m³. This event extensively damaged grassland ecosystems and completely blocked the Qinglong Valley channel. Impoundment of upstream flow formed a rapidly enlarging lake with rising water levels, representing a critical breach hazard.

The debris flow obstructed the perpendicularly flowing Qinglong Valley, forming a debris dam within the main channel approximately 200 m wide, 1500 m long (along-channel), 30 m high, with a volume of ~5.6 million m³. Widespread deformation features, including subsidence cracks, sinkholes, and slumping, developed due to post-depositional ice melt and compaction. Since 2004, meltwater has progressively drained through incised melt channels in the mid-to-lower sections of the ice-debris dam into the Qinglong Valley downstream. The glacial collapses on 8 October 2007 and 6 October 2016 generated debris flows that partially obstructed the Qinglong Valley to varying degrees. However, differences in the kinetic conditions, volumes, and velocities of the debris flows between events [55], combined with variations in the resulting dam dimensions, volumes, ambient temperatures, and river discharges, led to significant differences in the number, size, and longevity of the impounded lakes formed [56].

Based on these findings, we conducted field verification at the glacier. The direct impact zone of the Maqên Gangri Glacier collapse is located in Xiaoma Valley, Xueshan Township, Maqên County, Golog Tibetan Autonomous Prefecture (Figure 13), at geographic coordinates 99°26′33.0″ E, 34°48′45.1″ N. The glacier is cataloged as G099452E34793N (primary and secondary inventory codes). Key characteristics include an area of 1.71 km², a maximum elevation of 6250 m, a minimum elevation of 4881 m, an elevation range of 1369 m, and classification as a valley glacier. Well-developed transverse arcuate crevasses are present in the mid- and lower ablation zones [57,58]. The site exhibits well-developed moraine deposits. Crucially, no infrastructure or settlements exist downstream. The debris flow runout zone extends approximately 4.812 km. Analysis of 12 epochs of high-resolution satellite imagery reveals a consistent retreat trend: the glacier area decreased from 1.83 km² in 2014 to 1.71 km² in 2020, and the glacier terminus retreated approximately 335 m during this period.

Based on field investigations and analysis of historical remote sensing imagery, the formation process of the glacial collapse chain disaster in Xiaoma Valley, Ányêmaqên Mountains, is reconstructed as follows: Glacier retreat due to melt led to the development of numerous densely spaced, deep crevasses. Under gravitational loading, a section of the glacier detached from the mountain flank. During its descent, it entrained and plucked underlying bedrock, primarily phyllitic slate and sandy slate (Figure 13a1), while rapidly sliding towards the center of Xiaoma Valley [59]. The moving mass further scoured and incorporated ice-rich permafrost within the valley, composed of previous glacial collapse deposits, moraines, fluvioglacial sediments, and ground ice. Subsequently, the mass underwent extensive fragmentation, transforming into a high-velocity, long-runout debris flow that traveled 5.5 km down the middle and lower reaches of Xiaoma Valley (Figure 13a2,a3). It ultimately deposited within the alluvial-proluvial fan area and the confluence zone of Qinglong Gully and Qianlong Gully.

Concurrently, the debris flow deposits obstructed Qinglong Gully, forming an unconsolidated dam. Impoundment of upstream runoff led to the formation of a landslide-dammed lake. Subsequent dam failure, triggered by processes including ice melt within the dam body, seepage-induced deformation, and fluvial erosion [60], released a destructive outburst flood, causing significant economic damage downstream. The fundamental characteristics of this glacial collapse chain disaster are detailed in the following five sections: (1) Collapse Source Zone, (2) Plucking and Transport Zone, (3) Main Debris Flow Deposition Zone, (4) Landslide Dam and Impounded Lake, and (5) Outburst Flood [61]. Field investigations confirm that while this specific site is not currently classified as a potential hazard point, the event itself was a major disaster. The current state of the glacier is unstable. Although there are no immediate assets or infrastructure directly threatened at this location, the primary triggering factors identified are thaw processes and seismic activity. Notably, significant glacier retreat is observed, which significantly enhances the likelihood of glacier meltwater triggering debris flows (Figure 13a4). These potential future events pose a tangible threat to downstream infrastructure, particularly roads.

4.4. Limitations and Uncertainty Analysis

While this study quantified errors associated with glacial interpretation and model accuracy, several other sources of uncertainty deserve discussion. Firstly, the ALOS PALSAR DEM (12.5 m resolution) used for topographic analysis and hazard mapping is known to have increased vertical errors in high-altitude, steep terrain due to radar shadowing and layover effects. This introduces uncertainty into derived parameters such as slope angle, which is a critical factor in our stability assessment [62]. Secondly, the climate data used to contextualize glacier melt potentially suffers from spatial representation errors. Meteorological station data are point measurements that may not fully capture heterogeneity across the study area, while reanalysis data can exhibit biases over complex topography like the Ányêmaqên Mountains. These uncertainties in foundational data propagate into our final hazard inventory. However, we adopted a conservative identification threshold to ensure that the 38 sites reported are highly likely to be true hazards. Despite these limitations [63], the consistent patterns observed across multiple data sources and the clear mechanistic understanding of the disaster chains support the overall robustness of our conclusions.

4.5. Implications for Risk Assessment and Management

The identification of 38 potential hazard sites is a crucial first step, but translating this into actionable risk reduction requires further analysis. Although the immediate threat to existing infrastructure at the Maqên Gangri site is low, many of the other identified sites are proximate to critical transportation corridors, such as National Highway G214, and pastoral grazing lands. A logical next step is to conduct a formal risk assessment by overlaying our hazard inventory with data on exposed elements (e.g., population, infrastructure, livestock) and their vulnerability. We propose a preliminary prioritization of the 38 sites based on their potential volume, runout distance, and proximity to downstream assets to guide urgent monitoring efforts [63]. For high-priority sites, we recommend implementing cost-effective monitoring systems, such as time-lapse cameras, seismic sensors, or InSAR analysis, to detect precursor movements and provide early warning. Furthermore, our findings on the disaster chain mechanism underscore the necessity of integrated risk management plans that account for cascade effects—from ice collapse to outburst floods. Land-use planning should restrict development in high-risk runout zones and debris flow fans [64]. Finally, the methodology applied here, combining multi-source remote sensing and geomorphic analysis, provides a replicable framework for proactive glacial hazard assessment across the Tibetan Plateau, enabling stakeholders to mitigate future risks in a warming climate.

5. Conclusions

This study addresses anomalous glacier conditions in the Ányêmaqên Mountains. Utilizing multi-source, high-resolution time-series remote sensing imagery spanning 2013–2024, we analyzed regional glacier morphometric characteristics. Glacier boundaries for the period 2012–2023 were delineated primarily through semi-automatic and manual classification techniques, referencing the Second Chinese Glacier Inventory Dataset. Methods employed included manual visual interpretation, band ratio thresholding, 2D analysis, and 3D stereo modeling, followed by rigorous topological checks and boundary refinement. Delineation accuracy was quantitatively assessed using the metrics Accuracy, Precision, Recall, Loss, and F1-score.

Key findings reveal that glaciers within the Ányêmaqên Mountain region are predominantly ice caps, valley glaciers, and hanging glaciers. Glacier area exhibits an overall retreat trend, largely driven by glacier terminus recession. While total glacier area displays inter-annual variability, a net reduction is evident. Between 2009 and 2024, total glacier area decreased from 102.71 km² to 81.10 km², a reduction of 21.61 km², corresponding to an average annual decrease rate of −1.93%. The number of distinct glacier entities increased from 74 to 86, a net gain of 12 glaciers, representing an average annual relative change rate of +1.18%. This divergent trend—decreasing area yet increasing glacier count—is attributed to enhanced glacier disintegration under rising regional annual temperatures driven by global warming. Progressive melting and terminus retreat have caused larger glaciers to fragment into smaller, discrete entities.

Remote sensing interpretation identified 38 geohazard sites in the study area, concentrated on shaded (north-facing) and sunlit (south-facing) slopes: 34 sites correspond to direct glacial collapse hazards, while 4 sites involve glacial collapse–damming–outburst flood chain disasters. Hazards predominantly occur on mid- to upper mountain slopes with gradients of 30–50°, at elevations ranging from 4500–5400 m (minimum) to 5120–6050 m (maximum). Shaded (north-facing) slopes host 5 glacial collapse zones encompassing 21 hazard points. Sunlit (south-facing) slopes host 6 glacial collapse zones encompassing 17 hazard points. One zone is classified as medium-scale; the remaining ten are large-scale.

Field verification at the representative Xiaoma Valley Maqên Gangri Glacier site confirmed a direct glacial collapse hazard. The glacier exhibits well-developed transverse arcuate crevasses in its mid- and lower ablation zones and mature moraine deposits. Its area is 1.71 km², with a terminus retreat of approximately 335 m. This site is currently assessed as unstable, though no immediate assets are directly threatened. Primary triggering factors are thaw processes and seismic activity. However, significant glacier retreat substantially increases the likelihood of meltwater-triggered debris flows, posing a tangible threat to downstream infrastructure, particularly roads.

These findings provide critical insights for understanding glacier change trends on the Tibetan Plateau, informing adaptation strategies and sustainable development policies, revealing glacier response mechanisms to regional environmental and climatic changes, assessing impacts on regional water cycles and ecological balance, strengthening ecological security barriers, enhancing disaster prevention and mitigation capabilities, and safeguarding lives, property, and infrastructure. This research holds significant scientific merit and practical value for cryospheric science, hazard management, and climate change adaptation in high-mountain Asia.