Using the Improved YOLOv5-Seg Network and Sentinel-2 Imagery to Map Glacial Lakes in High Mountain Asia

Abstract

Continuously monitoring and mapping glacial lake variation is of great importance for determining changes in water resources and potential hazards in alpine cryospheric regions. The semi-automated glacial lake mapping methods used currently are hampered by inherent subjectivity and inefficiency. This study used improved YOLOv5 strategies to extract glacial lake boundaries from Sentinel-2 imagery. These strategies include using the space-to-depth technique to identify small glacial lakes, and adopting the coordinate attention and the convolution block attention modules to improve mapping performance and adaptability. In terms of glacial lake extraction, the improved YOLOv5-seg network achieved values of 0.95, 0.93, 0.96, and 0.94 for precision (P), recall (R), mAP_0.5, and the F1 score, respectively, indicating an overall improvement in performance of 12% compared to that of the newest YOLOv8 networks. In High Mountain Asia (HMA), 23,108 glacial lakes with a total area of 1847.5 km² were identified in imagery from 2022 using the proposed method. Compared with the use of manual interpretation for lake boundary extraction in test sites of HMA, the proposed method achieved values of 0.89, 0.87, and 0.86 for P, R, and the F1 score, respectively. Our proposed deep learning method has improved accuracy in glacial lake extraction because it can address the challenge represented by frozen or high-turbidity glacial lakes in HMA.

1. Introduction

Glacial lakes are formed by glaciation or glacial meltwater recharge and are generally located in areas of glaciation, including glacial depression lakes, lateral moraine dammed lakes, ice-dammed lakes, etc. [1,2]. High Mountain Asia (HMA) contains the largest area of glaciers outside the polar regions [3,4]. Over past decades, global warming has caused substantial glacier retreat and the acceleration of negative glacier mass in HMA [5,6]. Meanwhile, the total area and the number of glacial lakes in the HMA region have shown a consistent trend of increase [2,7,8,9,10]. The continuous increase in both the number and the size of glacial lakes might impound glacier meltwater and aggravate the risk of glacial lake outburst floods [11,12,13]. Therefore, the rapid and accurate mapping of glacial lakes is crucial for water resource management and disaster assessment in alpine cryospheric regions [2,14,15,16].

Recent advances in remote sensing and computer vision have accelerated the widespread use of multisource remote sensing images for glacial lake mapping. Semiautomatic extraction methods, based on the geometric characteristics of glacial lakes, include the normalized difference water index [17], normalized difference snow index [18], and modified normalized difference water index [19]. However, these methods need manual postprocessing to fine-tune results in areas with complex terrain and extreme climatic conditions, such as those of the HMA region [20].

Machine learning frameworks (e.g., the Random Forest method) have been used previously for glacial lake segmentation [14,21]. These approaches, which excel at extracting glacial lakes characterized by spectral reflectance, are challenged when segmenting areas characterized by diminished spectral reflectance [14]. In recent years, deep learning methods have shown substantial potential for automatic segmentation in glacial lake mapping, and in overcoming the constraints encountered in traditional machine learning methods [16,20,22,23,24,25,26,27]. Neural networks exhibit strong segmentation capabilities; however, certain limitations remain, e.g., the challenge in extracting the boundaries of frozen or partially frozen glacial lakes [28], the misleading errors caused by mixed pixels of ice and water and fragmented wet ice when classifying lake edges [27], and the technical challenges associated with handling training datasets at inconsistent temporal and spatial scales [23]. Typically, machine learning methods require large numbers of images for training neural networks. Although such methods have been applied to several subregions of HMA, they were often found to be inadequate for the boundary extraction of small-area glacial lakes, and lacked applicability to the large-scale HMA region.

In this paper, we introduce a new deep learning method based on the improved YOLOv5 network and Sentinel-2 imagery to extract the glacial lakes boundaries in HMA in 2022 (Figure 1). The proposed method is a reproducibly automated glacial lake boundary extraction framework that integrates comprehensive technical strategies in data preprocessing and model training using multiband Sentinel-2 imagery.

2. Materials and Methods

2.1. Data Sources

Overall, 1787 high-quality Sentinel-2 images (143 in 2018 and 1644 in 2022) with 10 m spatial resolution were obtained from the Copernicus Open Access Hub (https://scihub.copernicus.eu/dhus/#/home/; last accessed on 5 May 2023) for use in this study (Figure 2). The selected images were captured in summer and autumn (June–November) with cloud coverage of <10% in 2022. To train the glacial lake extraction models, we used 143 Sentinel-2 images from 2018 and the High Asia Glacial Lake dataset [2]. The Second Chinese Glacier Inventory and RGI 6.0 were used to determine a 10 km buffer area from glaciers [29,30]. The Shuttle Radar Topography Mission digital elevation model with spatial resolution of 1 arcsecond (http://imagico.de/map/demsearch.php; last accessed on 5 September 2023) was used to extract glacial lake elevation information.

2.2. Methods

2.2.1. Data Preprocessing

Preprocessing techniques included upsampling, slicing, image enhancement, culling and replacement, and automatic labeling (Figure 3). First, a band 11 image (shortwave infrared) was upsampled to 10 m resolution and synthesized with bands 8, 4, 3, and 2 (near-infrared, red, green, and blue). Second, original and corresponding mask images were sliced to a size of 640 × 640 pixels. All missing values were converted to 0. Automatic labeling using the glacial lake inventory and randomly selected pairs according to a ratio of 7:3 was used to create the training set and the validation set. Finally, the images of the training set were flipped and rotated, and pretzel noise was added randomly.

2.2.2. Improved Convolutional Neural Network

The basic network used for the segmentation of glacial lake is the YOLOv5-seg v7.0 model. The coordinate attention (CA) module, convolution block attention module (CBAM), and the small object detection layer were integrated with the YOLOv5-seg network [31,32,33]. The C3 module was also replaced by the C2f module with the purpose of obtaining additional information on gradient flow. The space-to-depth convolution (SPD-Conv) module was introduced to retain more spatial information [34].

(1)

Attention mechanism

This study focused on glacial lakes and considered other ground features such as glaciers, bare ground, and mountain shadows as background information. It has been reported that the YOLOv5-seg algorithm is prone to interference [35,36]. Therefore, we adopted the CA module and the CBAM to enhance the learning of glacial lake characteristics by the YOLOv5-seg network (Figure 4 and Figure 5). By adding different attention mechanisms to two similar networks, the network can flexibly and independently learn the characteristics of glacial lakes and consequently improve extraction accuracy. The algorithms facilitate the calculation of the weighted aggregation of other information by eliminating background details. By specifying the name of the glacial lake information as an input for the attention feature, the model directs its attention to the relevant information, thereby improving the extraction of glacial lake features [31,32].

The CA module encodes channel relationships and long-term dependencies in precise location information through coordinate embedding and coordinate attention generation [31]. The process of coordinate embedding overcomes global pooling in the channel and enhances the retention of positional information. To capture cross-channel information and long-term dependencies in one spatial direction and to retain accurate positional information in another, the features are considered along the two spatial directions to yield a pair of direction-aware feature maps that facilitate the accurate segmentation of glacial lakes in remote sensing images by the neural network. The process can be expressed as follows:

$${\mathrm{z}}_{\mathrm{C}}^{\mathrm{H}}\left(\mathrm{h}\right)=\frac{1}{\mathrm{W}}\sum _{0\le \mathrm{i}<\mathrm{H}}{\mathrm{x}}_{\mathrm{C}}\left(\mathrm{H}, \mathrm{i} \right)$$

(1)

$${\mathrm{z}}_{\mathrm{C}}^{\mathrm{W}}\left(\mathrm{h}\right)=\frac{1}{\mathrm{H}}\sum _{0\le \mathrm{j}<\mathrm{H}}{\mathrm{x}}_{\mathrm{C}}\left(\mathrm{j}, \mathrm{W}\right)$$

(2)

$${\mathrm{z}}_{\mathrm{C}}=\frac{1}{\mathrm{H} \times \mathrm{W}}\sum _{\mathrm{i}=1}^{\mathrm{H}}\sum _{\mathrm{j}=1}^{\mathrm{W}}{\mathrm{x}}_{\mathrm{C}}\left( \mathrm{i}, \mathrm{j}\right)$$

(3)

where ${\mathrm{x}}_{\mathrm{C}}\left(\mathrm{H}, \mathrm{i} \right)$ and ${\mathrm{x}}_{\mathrm{c}}\left(\mathrm{j}, \mathrm{W}\right)$ represent the C channel at height h for the i-th value and the C channel at height H and width W for the j-th value, respectively. The output of ${\mathrm{z}}_{\mathrm{C}}^{\mathrm{H}}$ is the C channel at height h and that of ${\mathrm{z}}_{\mathrm{C}}^{\mathrm{W}}$ is the C channel at width W. The information derived in the two dimensions needs to be combined via concatenation, convolution, and normalization. This step can be expressed as follows:

$$\mathrm{f}=\mathrm{\delta }\left({\mathrm{F}}_1\left(\left[{\mathrm{z}}^{\mathrm{H}},{\mathrm{z}}^{\mathsf{\omega }}\right]\right)\right)$$

(4)

where $\mathrm{\delta }$ is a nonlinear activation function, and $\mathrm{f}\in {\mathrm{R}}^{\frac{\mathrm{c}\times (\mathrm{H}+\mathrm{W})}{\mathrm{r}}}$ is the intermediate feature map that encodes spatial information in both the horizontal and vertical directions.

Convolution is performed for W and H separately, and the final output can be expressed using the following formulas:

$${\mathrm{g}}^{\mathrm{H}}=\mathrm{\sigma }\left({\mathrm{F}}_{\mathrm{h}}\left({\mathrm{f}}^{\mathrm{H}}\right)\right)$$

(5)

$${\mathrm{g}}^{\mathrm{W}}=\mathrm{\sigma }\left({\mathrm{F}}_{\mathsf{\omega }}\left({\mathrm{f}}^{\mathsf{\omega }}\right)\right)$$

(6)

$${\mathrm{y}}_{\mathrm{C}}\left(\mathrm{i},\mathrm{j}\right)={\mathrm{x}}_{\mathrm{C}}\left(\mathrm{i},\mathrm{j}\right) \times {\mathrm{g}}_{\mathrm{C}}^{\mathrm{H}}\left(\mathrm{i}\right) \times {\mathrm{g}}_{\mathrm{C}}^{\mathsf{\omega }}\left(\mathrm{j}\right)$$

(7)

where F_h and F_w are 1 × 1 convolutional transformation used to transform f^H and f^W, respectively, to tensors with the same channel number as the input X, and f^H and f^W are the horizontal and vertical components of f, respectively.

The CBAM contains two submodules: the channel attention module (CAM) and the spatial attention module (SAM) [32]. The input feature map F($\mathrm{F}\in {\mathrm{R}}^{\mathrm{c}\times \mathrm{h}\times \mathrm{w}}$) is passed through the maximum pooling layer and the average pooling layer in the CAM. The two summed 1D vectors become the intermediate feature map ${\mathrm{M}}_{\mathrm{c}}$(${\mathrm{M}}_{\mathrm{c}}\in {\mathrm{R}}^{\mathrm{c}\times 1\times 1}$) through the fully connected layer. The channel attention is multiplied by the input element F to obtain the adjusted feature map F′. Similarly, the feature map F is passed through the SAM to obtain 2D convolution ${\mathrm{M}}_{\mathrm{s}}$ (${\mathrm{M}}_{\mathrm{s}}\in {\mathrm{R}}^{1\times \mathrm{h}\times \mathrm{w}}$), which is then multiplied by F′ to obtain the feature map F″. The process of generating attention by the CBAM can be expressed as follows:

$${\mathrm{F}}^{\prime }={\mathrm{M}}_{\mathrm{c}}\left(\mathrm{F}\right)\otimes \mathrm{F}$$

(8)

$${\mathrm{F}}^{″}={\mathrm{M}}_{\mathrm{c}}\left({\mathrm{F}}^{\prime }\right)\otimes \mathrm{F}$$

(9)

where $\otimes$ denotes the weighted multiplication, F′ denotes the result obtained after passing through the CAM, and F″ denotes the result obtained after passing through the SAM.

(2)

Space-to-depth (SPD) module

In this study, downsampling was replaced by the SPD-Conv module to reduce the misjudgment of small-area glacial lakes. A non-spanning convolutional layer retains more detailed information about the glacial lake by rearranging the pixel blocks of the input feature map, and it reduces the number of parametric quantities to a certain extent [34]. The SPD-Conv module acquires four 2-fold downsampled sub-maps, each containing the spatial information, through mapping and slicing the input feature map. The submaps can splice along the channel dimensions and adjust them through the non-spanning convolutional layer. Additionally, the SPD-Conv module preserves global spatial feature information in the channel dimensions, as illustrated in Figure 6.

(3)

Small target layers

The target detection layer of the YOLOv5 network has an output range of 80 × 80, 40 × 40, and 20 × 20 image pixels, with a minimum acceptance range of 8 × 8. To enhance the capability of the algorithm in segmenting small-area glacial lakes, the detection header for small targets is integrated into the original network. The aim here is improving the accuracy of extracting glacial lakes smaller than 8 × 8 pixels in size and refining the precision of boundary extraction. The glacial extraction network structure is shown in Figure 7.

2.2.3. Postprocessing and Accuracy Assessment

To evaluate the results produced by the improved neural network algorithm, an Intersection over Union (IOU) threshold of 0.5 was used to extract glacial lake boundaries. To eliminate misclassification, slopes and shadow relief that were larger than 10° and 0.25, respectively, were removed based on the digital elevation model data [10,37]. Precision (P), recall (R), and mean average precision (mAP) were used to assess the performance of the algorithm. The R and P can be expressed as follows:

$$\mathrm{R}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$

(10)

$$\mathrm{P}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$$

(11)

where TP is true positive, FP is false positive, and FN is false negative. If TP is positive and the true value is also positive, the positive samples are correctly identified. If FP is positive and the true value is negative, the negative samples are incorrectly identified. IF FN is negative and the true value is positive, positive samples are missed. The mAP_0.5 (i.e., mAP with an IOU threshold of 0.5), used as an evaluation index, can express the recognition accuracy and the number of effective recognitions. The F1 score effectively represents the overall capability in identifying glacial lakes. The formulas for calculation of the mAP, IOU, and F1 score can be expressed as follows:

$$\mathrm{mAP}=\frac{1}{\mathrm{n}}\sum _{\mathrm{i}=1}^{\mathrm{n}-1}\mathrm{A}{\mathrm{P}}_{\left(\mathrm{i}\right)}\times 100\%$$

(12)

$$\mathrm{IOU}=\frac{\mathrm{a}\cap \mathrm{b}}{\mathrm{a}\cup \mathrm{b}}$$

(13)

$$\mathrm{F}1 \mathrm{score}=\frac{2 \times \mathrm{P} \times \mathrm{R}}{\mathrm{P}+\mathrm{R}}$$

(14)

We used S_accuracy as a measurement of the accuracy of glacial lake extraction, which can be expressed as follows:

$${\mathrm{S}}_{\mathrm{accuracy}}=1 -\frac{\left|{\mathrm{S}}_{\mathrm{TP}} - {\mathrm{S}}_{\mathrm{recall}}\right|}{{\mathrm{S}}_{\mathrm{recall}}}$$

(15)

where S_TP is the network-mapped glacial lake area and S_recall is the glacial lake area detected by manual vectorization mapping.

3. Results

3.1. Performance of the Proposed Method

The segmentation algorithm exhibits promising results after 1000 training epochs. The accuracies of the improved YOLOv5-seg and other commonly used instance segmentation models in relation to our constructed dataset, listed in

Table 1. Comparison of mainstream instance segmentation algorithms (data for YOLOv5-seg and YOLOv8-seg were obtained from Ultralytics Inc. (Los Angeles, CA, USA); data for YOLOv7 and YOLOR were obtained from [38,39,40,41]), respectively).

Algorithm	Precision	Recall	mAP_0.5	F1 Score	Bands
YOLOv5-seg	0.6	0.64	0.56	0.62	8   4   3
YOLOv7-seg	0.651	0.71	0.78	0.70	8   4   3
YOLOR-seg	0.583	0.731	0.60	0.790	8   4   3
YOLOv8-seg	0.916	0.75	0.77	0.82	8   4   3
Improved YOLOv5-seg	0.702	0.711	0.75	0.71	4   3   2
YOLOv5-seg	0.673	0.701	0.724	0.687	4   3   2
Improved YOLOv5-seg(CA)	0.87	0.801	0.88	0.84	8   4   3
Improved YOLOv5-seg(CBAM)	0.95	0.928	0.96	0.94	8   4   3
Improved YOLOv5-seg(CA)	0.916	0.734	0.80	0.81	11   4   3

, demonstrate that our improved YOLOv5-seg network exhibits the best performance. The F1 score is 0.94 and 0.81 for the synthesized group of bands 8, 4 and 3 and 11, 4 and 3, respectively.

The YOLOv8 network was used as a benchmark for our ablation experiment to compare the effects of different modules on accuracy (

Table 2. Results of ablation experiments.

Index	CA	CBAM	Small-Object Detection Layer	SPD-Conv	C2f	mAP_0.5	F1 Score
YOLOv5-seg	-	-	-	-	-	0.56	0.62
1	√	-	-	-	-	0.62	0.65
2	√	-	√	-	-	0.73	0.70
YOLOv8-seg	-	-	-	-	√	0.77	0.82
4	√	-	-	-	√	0.88	0.84
5	-	√	-	-	√	0.88	0.85
6	√	-	-	√	√	0.87	0.83
Improved YOLOv5-seg (CBAM)	-	√	-	√	√	0.96	0.94

). The YOLOv5-seg (CBAM) network, refined through ablation experiments, achieves the optimal F1 score of 0.94, which is 12% higher than that of YOLOv8-seg. Additionally, the F1 score of the improved YOLOv5-seg (CA) is 2% higher than that of the YOLOv8-seg. The combination of the improved algorithms (bands 11, 8, 4, 3, 2) achieves at least 4% higher F1 scores in the test sites relative to the other algorithms(

Table 3. Results of convolutional neural network in relation to the test sites.

Algorithm	Recall	Precision	Saccuracy	F1 Score	Bands
Improved YOLOv5-seg	0.70	0.80	0.95	0.75	(4,3,2)
Improved YOLOv5-seg (CA)	0.76	0.86	0.95	0.81	(8,4,3)
Improved YOLOv5-seg (CA)	0.81	0.84	0.94	0.82	(11,4,3)
Improved YOLOv5-seg (CBAM)	0.82	0.89	0.92	0.85	(8,4,3)
Combined the improved algorithms	0.87	0.89	0.96	0.89	(11,8,4,3,2)

).

3.2. Mapping of Glacial Lakes in HMA

In 2022 imagery, 23,108 glacial lakes covering a total area of 1847.5 km² were identified in HMA (Figure 8). The eastern Himalayas and Inner Tibet have the highest numbers of glacial lakes. The eastern Himalayas have 3675 lakes covering 322.5 km² (17.5% of the total area), while Inner Tibet has 3568 lakes covering 303.4 km² (16.4% of the total area). Meanwhile, only 100 lakes with an area of 7.5 km² (0.4% of the total area) and 213 lakes with an area of 14.9 km² (0.8% of the total area) are found in the Qilian Mountains and the Hissar–Alay region. The elevations of the glacial lakes in HMA in 2022 are in the range 1700–6400 m. A bimodal distribution pattern is observed for the entire HMA region and most of its subregions. The main reason for this phenomenon lies in the bimodal distribution of unconnected glacial lakes and proglacial lakes at different altitudes [42]. In the HMA region, the primary peak in glacial lake area occurs at 5000–5500 m, with a secondary peak at 4000–4500 m. In the different subregions, the peak in the glacial lake area varies from 2000–2500 m in the Hissar–Alay region to 5000–5500 m in the Central Himalaya, Karakoram, and Western Kunlun regions (Figure 8).

4. Discussion

4.1. Advantages of the Improved Strategies

The rapid changes in glacial lakes caused by climate change demand a fast, reliable, reproducible, and automated mapping method [43]. Manual mapping and semi-automated methods demand extensive and labor-intensive post-processing [14,19,28]. We developed a deep learning method by combining several improved YOLOv5-seg networks. The proposed method can independently assimilate the merits of multiple networks in relation to images with different band combinations, and analysis results reveal that our method can automatically map glacial lakes in HMA. The SPD-Conv module included in the network of our proposed method can rearrange the pixels of input feature maps and improve the capability of lake boundary extraction. The combination of the CA module and the CBAM can improve the attention degree of glacial lake boundaries [31,32]. The small target layer is used to extract small-area glacial lakes [33]. Transfer learning and large spatial scale glacial lake datasets were used in the proposed deep learning method to improve network generalization. Our proposed method enhances the network’s learning ability, showing improved performance in extracting glacial lake boundaries (Figure 9). It is particularly effective for lakes with high turbidity, frozen surfaces, shallow water, and small areas (<0.01 km²), which have raised challenges in deep learning-based detection [14,26]. The proposed method offers several advantages: (1) The single algorithm merged with the SPD-Conv and attention modules can retain more detailed information and focus attention on glacial lake information [31,32,34]. (2) The combination of the improved algorithms can compensate for misjudgements caused by a single network. (3) The use of the large volume of glacial lake data on different temporal and spatial scales allows for the synthesis of images as a dataset based on the reflectance of glacial lakes in different bands. Supported by this dataset, the network can learn the different spectral characteristics of glacial lakes on different temporal and spatial scales [2,10,15,44].

The false positive detection of glacial lakes could be largely excluded in the postprocessing step in our proposed method (Figure 9); however, certain limitations remain. First, a lake with an area covering fewer than 8 × 8 pixels might be misclassified as background information owing to the increase in the number of layers and the minimum recognition size of 8 × 8 pixels in the output layer. Second, the uncertainty is often derived from glacial lakes with extreme spectral characteristics, even with the assistance of multiband input data, because of the similarity with other background features (e.g., seasonal snow and rock) [28]. Additionally, the combination of several networks demands a large amount of memory storage and computer processing time; thus, it is still expected for the algorithms to improve in terms of increasing the identification accuracy and the calculation speed of glacial lake boundary extraction.

4.2. Reliability of the Present Glacial Lake Inventory

A number of glacial lake inventory datasets have been published in previous years, most of which were produced using long-term Landsat images of the HMA region [2,8,10,42,45,46,47]. First, the glacial lake inventory of the China–Pakistan Economic Corridor published in 2018 [2] and that published in 2020 [15], produced by visual examination, were selected as references to appraise the lake boundary extraction reliability of our improved deep learning method. Seven Sentinel-2 satellite images acquired in 2018 and 2020 were used to produce the glacial lake inventories in the China–Pakistan Economic Corridor using our proposed method. The results indicate that 91.0% and 84.0% of individual glacial lakes identified by our extraction method coincided with the glacial lakes in the inventories from [2,15], respectively, and the differences in total glacial lake area were 2.2% and 4.7%, respectively. Second, 80.2% of the glacial lakes (area ≥ 0.0064 km²) in the present glacial lake inventory coincided with those in the inventory in HMA [2], despite the use of different source images obtained at different times. Moreover, given the P, R, and F1 score values of 0.886, 0.872, and 0.86, respectively, in the present glacial lake inventory, the numbers and areas of glacial lakes in HMA in 2022 were determined to be in the ranges 20,473–26,500 and 1636–2118 km², respectively. Overall, the present glacial lake inventory produced by the improved deep learning method is considered reliable.

The area of glacial lakes varies seasonally owing to variations in inflow from the melting of the parent glacier and precipitation, although most source images were obtained in summer [2,10]. The available glacial lake inventories usually exclude uncertainties resulting from the use of imagery from different months. Here, 200 Sentinel-2 satellite images with less than 10% cloud coverage were collected (June to October, 2022) to detect monthly glacial lake changes. In this case, only glacial lakes in nine subregions of HMA were derived for analysis, specifically located in the Hengduan Shan, Southeast Tibet, Inner Tibet, Center Himalaya, Western Himalaya, Western Kunlun, Karakoram, Pamir, and Western Tien Shan regions. The relative anomaly of monthly change in glacial lake area was used to depict the lake area variation in different months of 2022. During June–October, notable monthly variation in glacial lake area was found across different subregions of HMA (Figure 10). In the subregions of HMA, 75% of the relative anomaly in monthly glacial lake area variation during June–October was within ±2–4%, and the maximum monthly variation was ±6%.

5. Conclusions

This study proposed an improved deep learning method using the YOLOv5-seg network to extract glacial lake boundaries based on Sentinel-2 imagery. The SPD-Conv module, CA module, CBAM attention module, and small target detection layer, integrated to extract the boundary of glacial lakes, achieved values of 0.95, 0.93, 0.96, and 0.94 for P, R, mAP_0.5, and the F1 score, respectively. The results confirm that the improved machine learning method has evident advantages in overcoming the problems of extracting glacial lakes with high turbidity and a frozen surface through the improved capability of network generalization. Due to its higher accuracy and reproducibility, the proposed method significantly outperforms existing glacial lake mapping techniques.

We selected 1644 Sentinel-2 images from 2022 to map glacial lakes in the HMA region using the improved machine learning method. Overall, 23,108 glacial lakes were identified with a total area of 1847.5 km² distributed at elevations of 1700–6400 m. The P, R, and F1 score values were 0.89, 0.87, and 0.86, respectively, when compared to the manual interpretation of lake boundary extraction in test sites of the HMA region. The uncertainty in glacial lake area for the entire HMA region resulting from the use of source images from different months was determined as ±2–4%.