Detecting the Lake Area Seasonal Variations in the Tibetan Plateau from Multi-Sensor Satellite Data Using Deep Learning

Abstract

Monitoring lake area changes with a higher spatial and temporal resolution can facilitate a more detailed analysis of climate-related changes in the Tibetan Plateau. In the Landsat era, optical remote sensing observation with water body index-based methods mainly contributed to alpine lake investigation. However, monitoring the seasonal or monthly change of a lake area is challenging since optical data are easily contaminated by the high cloud cover in the Tibetan Plateau. To cope with this, we generated new time series datasets including Sentinel-1 Synthetic Aperture Radar (SAR) and the Landsat-8 Operational Land Imager (OLI) observations. Meanwhile, we presented an improved deep learning model with spatial and channel attention mechanisms. Based on these datasets, we compared several deep learning models and found that the CloudNet+ had better performance. Taking this architecture as a baseline, we added spatial and channel attention mechanisms to generate our AttCloudNet+ for extracting the lake area. The results revealed that AttCloudNet+ had a better performance compared with the CloudNet+ and other CNNs (e.g., DeepLabv3+, UNet). For the accuracy of the lakeshore prediction, results from AttCloudNet+ demonstrated closer distance to the truth-value than other models. The obtained mean RMSE and MAE were 21.6 and 16.6 m, respectively. In contrast, the mean RMSE and MAE of the DeepLabv3+ were 99.5 and 76.0 m, while the corresponding RMSE and MAE for UNet were 91.1 and 64.9 m. In addition, we found our AttCloudNet+ was more robust than UNet and DeepLabv3+ because AttCloudNet+ is less influenced by the input optical images compared with DeepLabv3+ and UNet. By combining the results from different seasons and satellite sensors, we are capable of generating the complete lake area seasonal dynamics of the 15 largest lakes. The mean correlation coefficient (R²) between our seasonal lake area time series and the water level of LEGOS is 0.81, which is much better than the previous study (0.25). This indicates that our method can be used to monitor lake area seasonal variation, which is important for understanding regional climate change in the Tibetan Plateau and other similar areas.

1. Introduction

The Tibetan Plateau and surroundings, known as ‘The Third Pole’ and ‘Water Tower of Asia’, are sensitive to regional and global climate change [1,2,3,4]. The air temperature in the Tibetan Plateau has increased by 0.16–0.67 °C per decade since the 1950s [5,6]. The changes in lakes (e.g., lake area, level, and volume) of the Tibetan Plateau show a slight decrease from 1976 to the mid-1990s [7]. Following this period, the increase in precipitation was around 5.1 mm per decade [8,9], which has significantly affected alpine lake expansion since the 1990s [5,10,11]. The area changes occurring in the alpine lakes provide an intuitive reflection of the ecological and hydrological changes in the Tibetan Plateau [10]. Consequently, monitoring lake area changes with a higher spatial and temporal resolution can facilitate a more detailed analysis of climate-related changes in this region [12,13].

The lake area variations are extensively investigated in previous studies based on optical remote sensing images, mainly focusing on the annual variation of the lake area [14,15,16,17,18]. Higher resolution lake monitoring (e.g., intra-annual lake area monitoring) is less investigated, which is attributed to the optical remote sensing images that are easily contaminated by cloud [19]. Especially in the Tibetan Plateau, the daily cloud coverage is greater than 45% [20]. Therefore, Yao et al. (2019) [21] developed a method to use Landsat images contaminated by cloud to map lakes and reservoirs. However, severely contaminated images can still cause a decrease in mapping accuracy. To cope with this, observations of Synthetic Aperture Radar (SAR) can be used. SAR can penetrate the cloud and provide an alternative way to reveal the lakes’ seasonal variations. Zhang et al. (2020b) [13] and Dai et al. (2022) [22] utilized the Sentinel-1 SAR data to study the seasonal patterns of lakes and preliminarily indicated the seasonal cycle of lakes or glacier lakes. However, the former shows less consistency with other independent data (e.g., water level), and the latter has worse detection results of glacier lakes in winter. Specifically, it is interruptive to distinguish the lake area from the land by using SAR images acquired in the winter period [23]. By taking the largest lake, i.e., Selin Co, as an example, one can hardly find the boundary between ice and land from SAR images collected in winter (Figure S1 shown in the Supplementary Materials). In contrast, the lake shorelines can be clearly distinguished by optical images during this period. Overall, the quality of SAR images have been found to be bad in winter. On the contrary, the optical images are mainly affected by clouds in summer, and less influenced in winter. Hence, the less noise-affected optical images in winter and spring, and all-weather SAR observation in summer and autumn (few clean optical images from any period are added to supplement the information), have generated a new dataset. In this study, we derived seasonal variations of lakes in the Tibetan Plateau with reasonable accuracy from this new dataset.

In terms of lake area extraction, a vast number of approaches have been developed in recent years [24,25,26,27,28,29]. Researchers have utilized the threshold derived from either single-band observation or the water index developed (e.g., Normalized Difference Water Index) to distinguish water from land. These methods can be easily implemented, however, their robustness is problematic since various thresholds should be applied for the complex capture conditions of images. To overcome this problem, data-driven based machine learning has provided an alternative method. For instance, researchers have applied machine learning algorithms, such as Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Networks (ANNs) to extract lake areas from remote sensing images [30,31,32]. The extraction of features derived from machine learning algorithms frequently hinges on the effectiveness of manual feature selection [33,34]. The simple structure of these algorithms often impedes the capture of complicated features [35]. Instead, deep learning models are able to automatically learn features from complex datasets. For example, convolutional neural networks can automatically learn deep semantic features in remote sensing images through a series of convolutional and pooling operations [36,37]. Specifically, it has been widely used in computer vision, natural language processing, and other applications [36,37,38]. For instance, deep learning is utilized to detect and count plants from low-cost drones images, and excellent research performance has been achieved in transportation and other fields [39,40]. In addition, it has also been adopted by the remote sensing community for similar tasks in climate change research (e.g., atmospheric rivers detecting) [41]. Motivated by the fruitful achievements obtained in segmentation tasks by using deep learning [42,43,44,45,46] in computer vision tasks, we would like to utilize similar principles for our lake area detection by segmenting the remote sensing observations to water and land as a binary segmentation problem. In order to make full use of spectral and spatial features of remote sensing images, we introduced channel and spatial attention mechanisms [47]. Channel attention can evaluate and assign weights to the input channels, spatial attention can notice location-based and the most drastic changing area, such as the boundary between the lake area and land. We evaluated several efficient deep learning models and found the best one for processing the SAR images is CloudNet+ [48]. Furthermore, in order to make the model more suitable for lake recognition, we developed an architecture named AttCloudNet+, which added the spatial and channel attention mechanisms, and utilized multi-sensor data as input on the original CloudNet+ for lake area extraction.

The paper is organized as follows. Section 2 introduces the study region and datasets. Section 3 describes the architecture of our model, the data processing procedure, and the lake area extraction processes. Section 4 presents the model performance, results validation, and comparisons with datasets from other institutions. The performance of our model and other CNN models or conventional methods are compared in Section 5. Furthermore, Section 5 also discusses the effect factors contribute to misclassification. The conclusion is described in Section 6 at the end. Overall, this study is noteworthy for developing a modified deep learning architecture to study seasonal lake area variations by the synergic use of both SAR and optical images.

2. Study Area and Datasets

2.1. Study Area

The Tibetan Plateau has an average elevation of 4000 m above sea level and covers an area of $3\times {10}^6$ km² [13]. A significant interaction exists between the Tibetan Plateau and the global climate [7]. The climate on the Tibetan Plateau is dominated by westerlies in winter and the Indian monsoon in summer [18]. Instead, the heating of Tibetan Plateau affects Rossby waves propagating east of Japan, enhancing the westerlies and the easterlies [49,50]. The maximum and minimum monthly air temperatures in the Tibetan Plateau are about 10 °C and −10 °C [51]. Temperature records also show a warming rate of 0.044 °C per year, which is 0.014 °C higher than the global warming rate [51]. In addition, the precipitation mostly (60–90%) declines from June to September, and it shows a decreasing trend from southeast to northwest [52]. In this study, we selected 11 study regions, which cover the 15 largest lakes in the Tibetan Plateau with the corresponding locations shown in Figure 1. These lakes are distributed in different climate zones, including the westerlies, Indian summer monsoon, and East Asian summer monsoon [7]. Among them, Qinghai Lake, Selin Co, Nam Co, Chibzhang Co, Ayakkum Lake, and Zhari Namco are larger than 1000 km² [17].

2.2. Landsat-8 and Sentinel-1 Datasets

Landsat-8 revisits the earth with a sixteen-day cycle and operates at several wavelengths between 0.43 and 12.50 μm, with a spatial resolution of 15–80 m. In this study, band 6 shortwave infrared (B6-SWIR) of Landsat-8 OLI is selected for later usage. The characteristics of the SWIR is strongly absorbed by water and reflected by vegetation as well as the dry soil [24]. Subsequently, we considered Landsat-8 images from 2015 to 2020 with less than 10% cloud cover from the United States Geological Survey (USGS) website (https://glovis.usgs.gov) accessed on 1 December 2023 and Google Earth Engine (https://developers.google.cn/earth-engine) accessed on 1 December 2023 to compose a high quality dataset. Overall, 407 mosaic images were applied. The detailed operation time, spatial/temporal resolution, and channels of these sensors are illustrated in

Table 1. Characteristics of Landsat-8 and Sentinel-1 datasets used in this study.

	Landsat-8 OLI	Sentinel-1 SAR
Time range	2013–present	2014–present
Spatial resolution	30 m	10 m
Temporal resolution	16 days	6 days
Channels	11 bands	HH/VV/HV
Spectral range	0.43–12.50 μm	3.75–7.50 cm

[53].

Sentinel-1 loaded a SAR operating at the C-band, providing radar images with a spatial resolution of 10 to 40 m [54]. Sentinel-1 has a six-day repeat cycle at the equator, and its revisit rate is significantly greater at higher latitudes, which is three days for Europe/Canada and less than a day for the Arctic region [55].

Sentinel-1 has four different data acquisition modes: wave (WV), strip map (SM), extra-wide swath (EW), and interferometric wide swath (IW) [56]. Information regarding the load parameters of SAR images, the polarizations, and their combination methods can be found on the European Space Agency (ESA) website (https://sentinel.esa.int/web/sentinel/missions) accessed on 15 December 2023. In our study, 770 Sentinel-1 SAR Level-1 Ground Range Detected mosaic scenes which were integrated from Sentinel-1 ascending and descending orbits with a single polarization (VV) in interferometric wide swath (IW) mode (10 m) between 2015 and 2020 over Tibetan Plateau are utilized.

2.3. Remote Sensing Database Generation from Various Sensor Observations

To analyze the seasonal variation of lakes, a database consisting of a monthly remote sensing dataset is required.

Figure 2 shows the available clear images of Landsat and Sentinel-1 over Selin Co lake in 2019. It can be seen that Landsat images are not available between May and October due to cloud contamination. Thus, Sentinel-1 images in these months will be used for the database together with the optical images. In general, lake ice and land have similar backscatter in Sentinel-1 images, while they are more distinguishable in Landsat images, as we can see in Figure 2a–c, etc. Figure 3 shows two examples of Sentinel-1 (Figure 3a,b) and Landsat OLI (Figure 3c,d) images in the Selin Co region. Lake ice and land can exhibit similar texture and gray scale values in Sentinel-1 images (Figure 3a,b). Landsat images show that the land appears bright while the lake is dark. Figure 3e shows the gray values of a horizontal profile in Figure 3b,d, which indicates that the grayscale of the lake ice is 200 gray values higher than the land in the Landsat image. In contrast, the lake ice and land have similar grayscale values in the Sentinel-1 image. This further confirms that the lake ice and land are more distinguishable in the Landsat image than in the Sentinel-1 image. Hereby, to create the monthly database, we supplemented the high-quality Landsat 8 image when it was available.

2.4. Data for Comparison

We used several water level and lake area datasets to be compared with our estimated results. Firstly, the Hydroweb database (https://www.theia-land.fr/en/hydroweb/) accessed on 1 December 2023 created by Laboratoire d’Etudes en Géophysique et Océanographie Spatiales (LEGOS), France, in 2003 was used. It delivers the water level of 230 lakes and is completely based on satellite altimetry and imagery. We found 12 matchups out of 15 lakes in the Hydroweb database except the Bangongco, Xijir Ulan lake, and Gyaring lake for this study. Additionally, we used the seasonal water level from the Institute of Tibetan Plateau, Chinese Academy of Sciences (CAS) [12] to compare with our result of 15 lakes. These comparative water-level data were created by multiple altimetric missions and verified by in situ measurement data. We also compared all of the lakes with corresponding seasonal area series from Wuhan University (WHU) [13], which utilized the Sentinel-1 SAR data to study the seasonal patterns of lakes. The comparison dataset also includes annual lake area values using an index-based method from the Landsat series data calculated by the CAS [11].

Table 2. Hydrological data information from different institutions for comparison.

Data Source	Data Type	Coverage Region	Coverage Period	Number of Matches
LEGOS	Lake level	230 lakes of the world	2015–2017	12
CAS	Lake level	70 lakes in the Tibetan Plateau	2003–2020	15
	Lake area	Lakes (>1 km2) in the Tibetan Plateau	1970s–2018	15
WHU	Lake area	Lakes (>50 km2) in the Tibetan Plateau	2015–2017	15

summarizes the specific information of different hydrological data of the three institutions for comparison.

3. Method

3.1. Improving the CloudNet+

The CloudNet+ is a deep learning-based feed-forward CNN to address the problem of cloud detection in Landsat-8 imagery. By packaging multiple modules, CloudNet+ is trained with a novel loss function named Filtered Jaccard Loss [48]. The powerful potency of cloud detection in the remote sensing images of the CloudNet+ motivated us to transfer it to detect the lake area. CloudNet+ is an encoder–decoder based fully on a convolutional neural network. Firstly, the encoder of the CloudNet+ conducts a semantic extraction on remote sensing images. In this processing, images are sent to convolution layers and max pooling layers to extract the high-level features of lakes and non-lake regions. Then, CloudNet+ used the extracted features to retrieve lake attributes and generate lake images for the decoder.

To improve the capability of the CloudNet+ to identify lakes and non-lake regions, we added the attention mechanisms, including channel and spatial attention [47], to feed into the encoding processing. The attention mechanism is to learn or extract the weight distribution from the image features, and then apply the weight distribution on the original features to change the feature distribution, thereby enhancing the effective features and suppressing the ineffective features or noise [47]. The channel attention can evaluate and assign weights to the three input channels, two Sentinel-1 channels, and one Landsat OLI B6-SWIR channel in this study. We selected the VV polarization derived from the least noise-influenced Sentinel-1 image in the month and replicated it to generate two Sentinel-1 channels. The channel attention is computed as:

$$M_c\left(F_{\mathrm{prev}}\right)=\delta \left(\mathit{MLP}\left(\mathit{AvgPool}\left(F_{\mathrm{prev}}\right)\right)\mathit{MLP}\left(\mathit{MaxPool}\left(F_{\mathrm{prev}}\right)\right)\right)$$

(1)

where the $M_c$ represents the channel attention mechanism, the $F_{\mathrm{prev}}$ is the result of the encoder, $\delta$ represents the activation function, MLP is a multi-layer perceptron with weight sharing, AvgPool is the average pooling operations, and MaxPool is the maximum pooling operations.

First, we set the input $F_{\mathrm{prev}}$ with a shape: $H\times W\times C$, where H, W and C are height, width, and number of channels, respectively. Then, we carried out the global average pooling and max pooling to obtain two-channel descriptions with a shape: $1\times 1\times C$. Then, we sent these descriptions separately into a shared two-layer neural network and used ReLU as the activation function to increase the non-linear relationship between the layers. With the two features, we obtained the output weight coefficient $M_c$ using a sigmoid activation function. Finally, we multiplied the $M_c$ by the original feature F to get the channel-based feature.

The spatial attention can notice location-based and the most drastic changing area, such as the boundary between the lake area and land. Compared with channel attention, spatial attention mainly focuses on location information. The computing process can be derived as:

$$M_s\left(F_{\mathrm{prev}}\right)=\delta \left(f^{n\times n}\left(\left[\mathit{AvgPool}\right(F_{\mathrm{prev}}\left)\right]; \left[\mathit{MaxPool}\left(F_{\mathrm{prev}}\right)\right]\right)\right)$$

(2)

where $f^{n\ast n}$ is a convolutional computing of n $\times$ n window.

The procedure of spatial attention is similar to that of channel attention. Given a feature $F$ with a size of $H\times W\times C$, we first carried out average pooling and max pooling along the channel axis, respectively, to obtain spatial feature descriptions ($H\times W\times 2$). Then, we used a $7\times 7$ kernel to convolve the previous layers, and we input the convolution output into the Sigmoid activation function to obtain the weight coefficient $M_s$. Finally, we multiplied $M_s$ by the feature $F$ to obtain the location-based feature. Last, we concatenated the new location-based and channel-based features into an integrated feature and sent it to the decoder. Subsequently, deconvolution is applied for the integrated feature to help segment the interested area, which is a lake in this study, and restored the size of patches.

3.2. Preprocessing Remote Sensing Data

The preprocessing in this study follows the routine of radiometric calibration, bit conversion, projection, image resampling, and image subset. First of all, we used ENVI 5.2 (https://envi.geoscene.cn) accessed on 15 December 2023 to complete radiometric calibration on Landsat images. Then, 16-bit OLI images and SAR images were converted to 8-bit using the Geospatial Data Abstraction Library (GDAL, https://www.gdal.org) accessed on 15 December 2023 to enable it to have a better visualization [57]. Next, we transformed the co-ordinates of all images to Universal Transverse Mercator (UTM) using GDAL and resampled the spatial resolution of the Landsat images (30 m) to 10 m. The resolution after resample can be consistent with that of Sentinel-1 images. In the end, each Landsat or Sentinel-1 image was split to patches with a size of 512 $\times$ 512 pixels to be fed further into the deep learning models. Each patch has a 150-pixel overlap with its adjacent patches to fully use the images’ edge information. It should be noted that the aforementioned preprocessing operations are for reference only, and that the preprocessing steps have different limitations in different experimental environments. For instance, in the context of bit conversion, there is a possibility of losing image information. Furthermore, resampling changes the resolution, which can result in jagged pixels, which consequently changes the file size. The set of overlap degree in the subset operation should be established according to the actual experimental situation. If the value is too high, the dataset will be large and difficult to store. If too low, the edge information will be disregarded.

3.3. Extracting Lake Area Using Deep Learning

Figure 4 shows the flowchart of this study. We applied the LaeNet [58] to process the Landsat-8 OLI images in winter and spring, and the AttCloudNet+ (in Section 3.1) to process the combined Landsat-Sentinel dataset in summer and autumn. The images around Selin Co are selected as training and validation samples. In the prediction phase, we applied our models to the test sample, which were Landsat and Sentinel-1 patches across the remaining 14 lakes (Figure 1) between 2015 and 2020.

3.3.1. Extracting Lake Area Using Optical Datasets

We used the LaeNet, a novel end-to-end lightweight multi-task CNN with no down-sampling [58], to extract lake regions from Landsat images. The LaeNet has been proven to have advantages in lake detection compared with the DeepLabv3+, UNet, and other UNet-based neural networks. Firstly, we prepared 1832 training samples using the B6-SWIR data of Landsat-8 OLI images and selected 1466 patches for training the LaeNet and 366 patches for the validation phase. We built the lake/non-lake label images using the Band 6 with a specific threshold corresponding to the image. Furthermore, we consulted the domain expert to inspect and fine-tune the label. The configuration for the training process was as follows, the learning rate was 0.0001, the batch size was 4, and the max iteration number was set at 2000. The test data contained 34,184 Landsat patches and we predicted the lake area via test data immediately after the training.

3.3.2. Extracting Lake Area Using Multi-Sensor Datasets

To overcome the extraction difficulties, we utilized AttCloudNet+ to grasp lakes’ spectral and spatial features based on Landsat-Sentinel datasets. We applied monthly SAR images together with the solo optical image in the testing process. The training sample included 3664 Sentinel-1 and 1832 Landsat-8 OLI B6-SWIR patches from April to November 2017, and training and validation were separated with percentages of 80% and 20%, respectively. We manually delineated the lake region in Sentinel-1 images and applied it as the truth label in the ArcGIS platform. The training configuration for the procedure is the same as LaeNet indicated in Section 3.3.1. Then, we set the optimal weight parameters to predict the test data of 214,080 patches, which are derived from subdividing the combined image dataset between 2015 and 2020. In the end, the seasonal binary results corresponding to the SAR images are obtained in the prediction phase.

Figure 5 is an example of our AttCloudNet+’s output in the Selin Co region in July 2020. When using Sentinel-1 images exclusively, the estimated lake area shows many false positives (red boxes in Figure 5a). The estimation was more reasonable when combined Sentinel-1 and Landsat images (Figure 5b).

3.4. Postprocessing

The postprocessing includes geocoding, patches mosaic, lake boundary validation, lake area calculation, and uncertainty calculation. We geocoded the image patches with the UTM co-ordinate system using GDAL and mosaic patches using the Python API of ArcGIS. The boundaries of the lakes predicted by our model were compared with the delineated boundaries through visual interpretation. The Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) were calculated as:

$$d_i=\sqrt{{(x_i^{dt}-x_i^{\mathit{pt}})}^2+{(y_i^{\mathit{dt}}-y_i^{\mathit{pt}})}^2}$$

(3)

$$R\mathit{MSE}=\sqrt{{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{d}_i^2}$$

(4)

$$\mathit{MAE}=\stackrel{-}{d}={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{d}_i$$

(5)

where ($x_i, y_i$) is the co-ordinate of one lake boundaries pixel, the superscript pt indicates that this pixel is estimated by our model, and dt indicates that it is delineated by visual interpretation. $d_i$ is the distance between ($x_i^{pt}$, $y_i^{pt}$) and ($x_i^{dt}$, $y_i^{dt}$).

We calculated the seasonal variation based on the number of pixels predicted as water by our model (water pixels). As variations of the lake area cause the retreat or advance of the shoreline, we created a 1000-m buffer around the predicted lake shoreline as the changing region [59]. The lake area variation depends on the water pixels within the changing region. Then, we summed the water pixels in and outside the changing region to calculate the area of the whole lake ($S$):

$$S={\displaystyle \sum _{i=1}^N}{r}_i^2$$

(6)

where r is the pixel size of the image, and N is the number of water pixels. We also calculated the uncertainty of the predicted lake areas as the perimeter of the lake multiplies the mean RMSE. The uncertainty of the predicted lake area (E) is calculated as:

$$E=\pm C\times {\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{\mathit{RMSE}}_i$$

(7)

where C is the perimeter of a lake, n is the number of the lakes that have been delineated by visual interpretation (n = 3), and ${RMSE}_i$ is calculated as Equation (4).

3.5. Error Metrics

The indicators to assess performance include Accuracy, F1_Score, Precision, Recall, and mIoU (mean Intersection over Union) (Equations (8)–(12)).

$$\mathit{Accuracy}={\displaystyle \frac{\mathit{TP}+\mathit{TN}}{\mathit{TP}+\mathit{TN}+\mathit{FP}+\mathit{FN}}}$$

(8)

$$\mathit{Precision}={\displaystyle \frac{\mathit{TP}}{\mathit{TP}+\mathit{FP}}}$$

(9)

$$\mathit{Recall}={\displaystyle \frac{\mathit{TP}}{\mathit{TP}+\mathit{FN}}}$$

(10)

$$\mathit{F}1\_\mathit{Score}={\displaystyle \frac{2\times \mathit{Precision}\times \mathit{Recall}}{\mathit{Precision}+\mathit{Recall}}}$$

(11)

$$\mathit{mIoU}={\displaystyle \frac{1}{2}}\times ({\displaystyle \frac{\mathit{TP}}{\mathit{TP}+\mathit{FP}+\mathit{FN}}}+{\displaystyle \frac{\mathit{TN}}{\mathit{TN}+\mathit{FN}+\mathit{FP}}})$$

(12)

The TP, TN, FP, and FN are the number of true positives, true negatives, false positives, and false negatives, respectively. In this study, TP is the true classification for the lake pixels and TN for non-lake pixels. The FP are the land pixels that are classified as lake water. The FN are the lake pixels that our method misidentified as lake water. The mIoU is a metric showing how close the model’s prediction is to the truth label.

4. Results

4.1. Performance of the AttCloudNet+

We compared the performance of AttCloudNet+ with other CNN models: (1) UNet, (2) DeepUNet, (3) DeepLabv3+, (4) AttResUNet, (5) SegNet, (6) original CloudNet+, and (7) CloudNet+ with sequential Channel and Spatial Attention.

Table 3. The performance of different deep learning models on lake area extraction.

Model Structure	Accuracy	Precision	Recall	F1_Score	mIoU
UNet	0.619	0.619	0.988	0.716	0.309
DeepUNet	0.986	0.969	0.964	0.964	0.910
DeepLabv3+	0.971	0.974	0.975	0.975	0.939
AttResUNet	0.986	0.968	0.965	0.964	0.910
SegNet	0.986	0.970	0.965	0.965	0.913
Original CloudNet+	0.983	0.981	0.990	0.985	0.938
CloudNet+ -Sequential Channel and Spatial Attention	0.979	0.974	0.989	0.981	0.930
AttCloudNet+	0.985	0.982	0.992	0.986	0.945

shows the indicators of all these models and the best model for each metric is bold. It can be seen that our AttCloudNet+ has achieved the best performance in terms of F1_Score, Precision, Recall, and mIoU. Other models (e.g., DeepUNet, DeepLabv3+, AttResUNet, and SegNet) exhibited a lower F1_Score, Precision, Recall, or mIoU, but they have a small gap with AttCloudNet+. The UNet obtained the lowest score probably because its simple structure failed to handle the complex textures of Sentinel-1 images.

4.2. Evaluating Detected Lake Boundaries

Figure 6 shows an example of the predicted lake boundary and the visual interpretation boundary in eastern Selin Co. For the boundary estimated by AttCloudNet+ using combined Landsat/Sentinel-1 dataset, the RMSE is 30.0 m, and the MAE is 21.8 m, indicating that the error is in the order of three Sentinel-1 pixels (Figure 6a). According to the results of LaeNet using optical images, the RMSE is 30.8 m and the MAE is 22.5 m (Figure 6b). This indicated AttCloudNet+ has a better performance than LaeNet. However, the performance difference between the two models was small, so we still used the LaeNet during the winter and spring because the model size and computation cost of LaeNet is much smaller than those of AttCloudNet+. We also conducted validations in Nam Co and Yamzho Yumco and the analysis of RMSE and MAE are revealed in

Table 4. Errors of lake boundaries predicted by DeepLabv3+, UNet, LaeNet, and our deep learning model compared with boundaries delineated by visual interpretation in Selin Co, Nam Co, and Yamzho Yumco.

Model	Comparison Sites	Combined RMSE (m)	Combined MAE (m)	Optical RMSE (m)	Optical MAE (m)
DeepLabv3+	Selin Co	229.1	173.5	58.2	42.5
	Nam Co	10.6	11.0	30.1	24.6
	Yamzho Yumco	43.9	17.5	32.8	22.2
	Mean	99.5	76.0	40.3	29.8
UNet	Selin Co	232.9	162.2	29.1	21.9
	Nam Co	24.6	19.4	24.1	20.2
	Yamzho Yumco	15.8	13.0	28.6	22.9
	Mean	91.1	64.9	27.3	21.7
AttCloudNet+	Selin Co	30.0	21.8
	Nam Co	16.7	12.7
	Yamzho Yumco	18.2	15.2
	Mean	21.6	16.6
LaeNet	Selin Co			30.8	22.5
	Nam Co			20.1	16.0
	Yamzho Yumco			23.7	18.8
	Mean			24.9	19.1

, which are RMSE = 21.6, MAE = 16.6 for the AttCloudNet+-predicted boundaries, and RMSE = 24.9, MAE = 19.1 for LaeNet-predicted boundaries.

In addition, to explore the performance of AttCloudNet+ proposed in this study and other deep learning models (i.e., DeepLabv3+ and UNet) in predictable results, the predicted lakeshores are compared with the manually delineated boundary of corresponding locations. The results show that the accuracy of prediction results of the AttCloudNet+ is higher than DeepLabv3+ and UNet models. Detailed error results are shown in Table 4. The RMSE of DeepLabv3+ for combined data reached 99.5 m, compared with 21.6 m for the AttCloudNet+, which improved the lake identification accuracy by 78.2%. Similarly, for UNet, its RMSE for combined data reaches 91.1 m, and compared with AttCloudNet+, the accuracy of lake identification is improved by 76.2%.

The comparison results demonstrate that our model has significantly enhanced the accuracy of the extraction process. However, there is still a slight differential of approximately one pixel when compared to the manual extraction results. We believe there are a number of potential reasons for this discrepancy. Firstly, the production of training data utilizes a threshold method in conjunction with a manual adjustment approach. While this method can achieve near-real-value precision, it is not an exact reflection of the true value. Consequently, it may result in minor errors in the generated results. Secondly, during the process of data prediction, complex terrain areas are often accompanied by mountain shadows. In some cases, the prediction results may misidentify the shadows as water bodies, which could lead to errors. Thirdly, since the initial weight parameter values of the neural network are randomly assigned, there are slight differences in the weight parameter results saved after each training. This could also contribute to differences in the final prediction results.

The previous results indicate that AttCloudNet+ performs better than UNet and DeepLabv3+. The possible reason can be that the optical image has different influences over the three models. To verify this difference, we designed an experiment to explore the predicted results effects of AttCloudNet+, DeepLabv3+, and UNet on SAR images when using optical images acquired at different times, and these were added for the combined data. In this experiment, SAR image data in August 2020, similar to the visual interpretation data, were still selected. Optical image data in 2015, 2017, and 2020 were used as supplementary information, respectively, to assist the lake recognition process of SAR images. Figure 7 illustrated that the AttCloudNet+ model can use a few high-quality optical images with any acquisition time to supplement the SAR image data to assist in extracting lake areas. It solves the challenge of identifying the targets from SAR images, as shown in Figure 7. However, when using DeepLabv3+ to process the combined images, the extraction results are lake regions in the supplementary optical images rather than the target region in the SAR image, as shown in Figure 7e–h.

Table S1 in the Supplementary Materials shows the comparison between the predicted results of AttCloudNet+, DeepLabv3+, and UNet. The results of these models, obtained from datasets supplementing different years of optical images, were all compared with the manually delineated shorelines. The reasons for differences between these models may be as follows: (1) when different models use combined data for prediction, they need to consider the channel allocation weight most suitable for the model; (2) the structural characteristics of AttCloudNet+ re complex, indicated by integrating multiple blocks; this enables AttCloudNet+ to reveal spatial details information of input images. In addition, the structure of each model will also affect the processing of the combined data. For example, the dilated convolution in the structure of DeepLabv3+ will enable it to have a larger perceptive field in the convolution process. This resulted in extracting more spatial feature information from optical images, making the model more susceptible to the influence of optical image features.

Taking the results from the DeepLabv3+ model as an example, when we used the B6 band in 2015 as a supplement, the results were more affected by the B6 band image. This condition resulted in an RMSE of 199.8 m. In addition, if the optical images in 2020 were used as supplementary data, the RMSE between the predicted result and the visual interpretation boundary was reduced to 67.6 m. This variation arises because the acquisition time of the optical images is similar to that of the SAR images. The same principles apply to UNet. Instead, in the prediction process of AttCloudNet+, regardless of the acquisition time of supplementing optical images fed into the combined data, AttCloudNet+ can precisely extract the lake area from the combined images. The RMSE of the predicted results was stable at around 2 pixels, a specific value as shown in Table S1 in the Supplementary Materials. In conclusion, when predicted combined images use DeepLabv3+ and UNet, their results mainly extract lakes from complementary optical images. In contrast, our AttCloudNet+ can use optical information to compensate for the defects of SAR images. At the same time, the results were not affected. This effectively solved the recognition difficulties caused by excessive noises and shadows in SAR images.

4.3. Seasonal Variations of Lake Area from 2015 to 2020

We compared our estimated lake area with the water level dataset and other datasets for the lake area. Figure 8a shows the linear regression between our estimated lake area and the LEGOS water level in Selin Co from 2015 to 2020, which has an R² of 0.87. This indicates that water level data can be used as a surrogate for the lake area. We also analyzed the correlation in the same lake between the seasonal lake area variation of WHU and the water level of LEGOS at the same time, and the results showed that R² was 0.46, as shown in Figure 8b.

In addition, we also calculated the correlation between the seasonal area series and the water level of the other eleven lakes, as shown in Table S2 in the Supplementary Materials. The results show that the average correlation coefficient between the seasonal variation time series of the WHU lake area and the water level is 0.25, while the average correlation coefficient of our method is 0.81. Apart from LEGOS, we compared our estimated lake area with (1) the lake area calculated by Wuhan University (WHU), (2) the annual lake area calculated by the Institute of the Tibetan Plateau, Chinese Academy of Sciences (CAS), and (3) the water level data derived from CAS for the period between 2015 and 2020, shown in Figure 9. This shows that our AttCloudNet+ predicted the lake area reached its peak value of 2385.96 km² in July 2016, then it decreased to 2352.93 km² from October 2016 to February 2017. The LEGOS water level shows a similar increasing pattern from March to August, and reached 4545.21 m in September then decreased from September to January and reached 4544.50 m in January. However, the water level decreased from January to August 2019, while our estimated lake area increased. CAS has a four-year (2015–2018) lake area record in Selin Co. Our estimated lake areas are consistent with the area of CAS in 2015, 2017, 2018, and 10 km² larger than the CAS in 2016. The average trend of our lake seasonal area series is basically the same as LEGOS, but there are still discrepancies in some months. Although we have made a breakthrough in the seasonal variation of area, it is still limited by image quality, identification errors and other factors. We will dedicate the discussion section to this subject in detail.

In the process of calculating the lake area, we also unexpectedly found that, in the estuary area, there is a significant water accumulation or reduction with the change in seasonal water quantity. We chose the Selin Co region as an example to seriously compare the changes of estuary areas in different seasons, especially in the wet season. The results showed that in the summer of 2015, the area of the river entering the lake estuary expanded by 0.21 km² compared with that in spring. In the winter of 2015, the area of the same region was reduced by 0.61 km² compared with the summer wet season (Figure S16 shown in the Supplementary Materials). Other years showed similar trends. This change is consistent with the overall seasonal variation trend of the lake, and helps us to realize that the shrinking rate in the dry season is higher than the expanding rate in the wet season, which is conducive to further understanding the differences of seasonal climate change in the Tibetan Plateau region.

Figures S2–S15 show the seasonal variations of the other 14 lakes. In Qinghai Lake (Figure S2, shown in the Supplementary Materials), the largest lake in Tibetan Plateau, our detected lake area was consistent with the LEGOS water level and CAS lake area. However, the seasonal lake area series of WHU between 2015 and 2017 is underestimated compared with CAS. For instance, the peak area value of WHU was 4398.99 km² in 2017, but 4579.44 km² for ours, and 4567.25 km² for CAS. Area variations in other lakes also show similar trends with the LEGOS water level except for the Bangongco and Gyaring Lakes.

The seasonal series shown in Figures S2–S15 illustrated that most lakes, especially in the Central Tibetan Plateau, show summer expansion and a winter shrinkage trend. In contrast, some lakes, such as the Aqqikkol lake in the Northern Tibetan Plateau, do not exhibit noticeable seasonal variation trends. Qinghai Lake, Selin Co, Chibzhang Co, Ayakkum Lake, Zhari Namco, Tangra Yumco, Ulan Ul Lake, Har Lake, and Xijir Ulan Lake, showed an abrupt increase in 2017 and 2018. The lake area of Gyaring Lake shows a decreasing trend from 2015 to 2017 and an increasing trend from 2017 to s2019, which is consistent with the results from [60].

5. Discussion

5.1. Effect of Snow Cover

The AttCloudNet+ can extract lake areas in most months throughout a year, except the months with snow coverage over lakes and land. However, this is not severe for seasonal lake area monitoring since the snow season is short in the local area. Figure 10 shows the Landsat OLI true color images and B6-SWIR images in Selin Co from December 2019 to April 2020. It is difficult to distinguish the land and lake when most regions were covered by snow in January 2020 (Figure 10b). However, snow cover disappeared in most land regions within one month, and the lake boundary became clear in February (Figure 10c). Thus, the AttCloudNet+ failed to detect the lake area in one month (January) out of these five months (December–April). Because of this ephemeral snow cover, the effect of snow makes little impact on monitoring the seasonal lake area.

5.2. Effect of Radar Shadow

The Bangong Co is located in the mountainous area in northwestern part of the Tibetan Plateau, where radar shadow is severe. The radar shadow and lake have similar gray scale values and texture in Sentinel-1 images, which may cause misclassification. We calculated the misclassification area of the mountain shadow on the lake extraction process, comparing the results with the inter-annual data from the CAS. Our analysis revealed that the average misidentified area affected by shadows in Bangong Co from 2015 to 2018 was 6.84 km², representing 0.98% of the total area. Constantly, we estimated the average uncertainty in Bangong Co based on the shadow effect, with a result of 20.11 km². Figure 11 shows examples of radar shadow, which our deep learning models cannot differentiate from the lake water in the Sentinel-1 images. A possible solution to remove the radar shadow is to use the dual-polarization radar images and Digital Elevation Model (DEM), as well as more powerful deep learning network models, to avoid the influence of shadows on recognition results. Also, applying linear image stretching to the original Sentinel-1 images can alleviate the effect of the radar shadow [61].

5.3. Lake Area Extracted Using Conventional Methods

Figure 12 shows the Selin Co lake area extracted using the NDWI, LaeNet, and Otsu thresholds [62], and the AttCloudNet+. Figure 12b shows that the NDWI enhanced the presence of water bodies in the Landsat image. However, it is difficult to set a threshold that can extract water pixels precisely in such a large region. Figure 12e shows the lake area detected using the Otsu threshold algorithm from the Sentinel-1 images. Many false positives outside Selin Co can be found, which is indicated by arrows. In contrast, LaeNet and the AttCloudNet+ can extract features from remote sensing images and automatically learn high-level semantic information (Figure 12c,f). The conventional methods remain prevalent approaches in water extraction. These traditional methods are straightforward to use and ideal for small-scale extraction scenarios. However, for higher accuracy and large volumes of data, we advocate the use of deep learning algorithms.

5.4. Future Improvement

The proposed method enhanced the temporal resolution of water body extraction, thereby facilitating higher time-frequency monitoring, which is crucial for discerning seasonal alterations in lakes. Furthermore, the method is beneficial for a comprehensive examination of the influence of seasonal and even monthly variations on the hydrological processes of lakes and glaciers on the Tibetan Plateau. Despite the method’s advancements in temporal resolution for lake extraction, it still exhibits certain limitations. In future studies, we will try to integrate additional multi-source remote sensing data (e.g., thermal infrared data, dual-polarization radar images, digital elevation models, etc.). For instance, water bodies exhibit lower surface temperatures than rocks or soils during the daytime. Hereby, thermal infrared remote sensing data can be leveraged to incorporate and enhance the precision and reliability of the lake extraction process. Furthermore, the incorporation of DEM data can facilitate more accurate locations susceptible to mountain shadows, thereby mitigating the impact of these shadows on feature extraction. Additionally, the integration of cutting-edge algorithms, such as transformers, within the field of computer science can enable the development of lake extraction. It will optimize the structure of deep learning models, and enhance the precision and reliability of prediction outcomes. In subsequent improvements, the number of lake pixels can be classified and calculated based on the color and geometric features of the target features, thus enabling the lake area to be obtained. The geometric relationship between fixed features in the image field of view and the target lake is used to locate the lake edges. Hence, obtaining the boundary changes of the lake over time based on different time-phase images will become more considerable.

6. Conclusions

In this study, AttCloudNet+ is being proposed to analyze seasonal area variations of the 15 largest lakes in the Tibetan Plateau by using combined remote sensing images from Landsat-8 and Sentinel-1. Landsat images were used to generate the lake area in spring and winter, while the combined images were applied in summer and autumn. In this way, a complete seasonal cycle dataset is generated. In addition, the predicted metrics show our AttCloudNet+ outperforms mainstream deep learning approaches (i.e., DeepLabv3+, UNet, etc.). Moreover, validation revealed that our results have a closer distance compared with other models. The results of the robustness experiment revealed that by adding optical images with different acquisition time as supplementary data, the AttCloudNet+ could extract the lake area from SAR images stably. Overall, the issue of cloud cover interfering with optical images was overcome and the accuracy of SAR image segmentation was enhanced. Our method partly addresses the challenge of seasonal variations in lake areas in regions with cloudy climates, significantly improving the temporal resolution of dynamic monitoring of the lake area.

The correlation analysis showed that the mean correlation coefficient R² between seasonal area time series derived in our paper and the water level of LEGOS is 0.81, which is much higher than the previous study’s performance (0.25). Furthermore, the area of 15 Tibetan Plateau lakes from 2015 to 2020 showed seasonal variation cycles with the lake area increases from July to September and decreases from October to February each year. A comparison between our detected lake area variations and other lake-level or area datasets (LEGOS, THU, WHU, and CAS) showed consistency in 13 out of the 15 lakes. This method can be further implemented in other areas to further understand the lake changes in the continental and global scale. This will be helpful to understand the flux between land and atmosphere. If lake extraction is carried out in areas outside the Tibetan Plateau, we recommend introducing images of the target region and retraining the network structure to achieve the corresponding weight parameters for migration. Furthermore, it will also be important for better water resource distribution, as well as a benefit to modelers, to understand the climate model. Finally, it will be essential to come up with remedies to prevent global climate change.