Water-Body Detection in Sentinel-1 SAR Images with DK-CO Network

Abstract

Synthetic Aperture Radar (SAR) is an active microwave sensor with all-day/night and all-weather detection capability, which is crucial for detecting surface water resources. Surface water-body such as rivers, lakes, reservoirs, and ponds usually appear as dark areas in SAR images. Accurate and automated extraction of these water bodies can provide valuable data for the management and strategic planning of surface water resources and effectively help prevent and control drought and flood disasters. However, most deep learning-based methods rely on manually labeled samples for model training and testing, which is inefficient and may introduce errors. To address this problem, this paper proposes a novel water-body detection method that combines optimization algorithms and deep learning techniques to automate water-body label extraction and improve the accuracy of water-body detection. First, this paper uses a swarm intelligence optimization algorithm, Dung Beetle Optimizer (DBO), to optimize the initial cluster center of the K-means clustering algorithm, which is called the DBO-K-means (DK) method. The DK method divides the training images into four categories and extracts the water bodies in them to generate the water-body labels required for deep learning model training and testing, and the whole process does not require human intervention. Then, the labels generated by DK and training data set images are fed into the Classifier–Optimizer (CO) for training. The classifier performs a dense classification task at the pixel level, resulting in an initial result image with blurred boundaries of the water body. Then, the optimizer takes this preliminary result image and the original SAR image as input, performs fine-grained optimization on the preliminary result, and finally generates a result image with a clear water-body boundary. Finally, we evaluated the accuracy of water-body detection using multiple performance indicators including ACC, precision, F1-Score, recall, and Kappa coefficient. The results show that the values of all indicators exceed 93%, which demonstrates the high accuracy and reliability of our proposed water-body detection method. Overall, this paper presents a novel DK-based approach that improves the automation and accuracy of deep learning methods for detecting water bodies in SAR images by enabling automatic sample extraction and optimization of deep learning models.

1. Introduction

Surface water is a crucial component of Earth’s water resources, playing an indispensable role in various hydrological cycles [1,2,3,4,5]. Accurate detection of surface water is vital for sustainable water resource management, environmental protection, and socio-economic development [6,7,8,9,10]. Against this backdrop, the automatic and precise extraction of the water body using Synthetic Aperture Radar (SAR) data has begun to attract widespread scholarly attention [11,12,13,14,15].

As an active microwave sensor, SAR is capable of sensing targets during both day and night, irrespective of sunlight conditions. Its signal can penetrate cloud cover, rain, and fog, overcoming the limitations faced by optical remote sensing sensors under adverse weather conditions. Therefore, SAR can provide stable and reliable time-series data for surface water-body detection. However, the challenge of automatically and accurately extracting water-body information from vast amounts of SAR data lies in the need for a substantial amount of annotated water-body data. The acquisition of this data typically relies on manual annotation, which undoubtedly amplifies the workload involved in this process [16,17,18,19,20,21].

Deep learning, a powerful branch of artificial intelligence, excels in automatic learning and feature extraction. In recent years, water-body detection methods based on Convolutional Neural Network (CNN) models have shown promising results [22,23,24,25]. In [26], Li et al. proposed a high-precision water boundary extraction method for SAR images, combining super-resolution recovery based on lightweight residual CNNs with traditional extraction methods. In [27], Zhang et al. proposed a segmentation network architecture based on deep learning that extracts high-dimensional features from SAR images using dense depth-separable convolutions and dilated convolutions. It achieved exceptional accuracy, robustness, and speed, surpassing traditional methods. In [28], Hertel et al. compared the performance of two probabilistic CNNs for water segmentation based on SAR using Sentinel-1 data. However, the implementation of deep learning typically necessitates a substantial amount of labeled water-body data, the acquisition of which usually relies on manual annotation, undoubtedly amplifying the workload involved in this process [29,30,31,32,33,34,35]. Therefore, unsupervised learning methods, such as clustering algorithms, can be considered to reduce the huge workload and time required for manual labeling.

Clustering algorithms, as a type of unsupervised learning method, offer numerous advantages, leading to their wide application in early water-body detection research [35,36]. This confers an advantage in creating water-body labels, presumably reducing the workload of manual annotation and enhancing the efficiency of label generation. However, the segmentation results of clustering algorithms are strongly influenced by the choice of initial cluster centers. To address this issue, the application of swarm intelligence optimization algorithms becomes particularly important.

Swarm intelligence optimization algorithms can explore complex solution spaces, allowing them to converge to the optimal or near-optimal solution. Their global search capability also enables them to adapt to changing data distributions. This provides a new idea for the selection of initial cluster centers in clustering algorithms.

In this paper, we utilize the swarm intelligence optimization algorithm, Dung Beetle Optimizer (DBO), to optimize the initial cluster centers of the K-means clustering algorithm, naming it the DBO-K-means (DK) method. The DK method categorizes the training dataset into four different classes, generating water-body labels needed for deep learning without any manual intervention. This unsupervised method significantly reduces the workload associated with manual labeling in traditional supervised deep-learning methods. Subsequently, Classifier is used to perform pixel-level dense classification tasks, yielding an initial result image with vague water-body boundaries. Finally, the optimizer is used to perform fine-grained optimization on the initial result image, yielding a final result image with clear water-body boundaries. Experiments conducted in the study area demonstrate that the proposed method not only effectively improves the accuracy of water body detection but also significantly reduces the workload of manual labeling by shifting the detection process from supervised to unsupervised [36,37,38,39,40,41,42].

2. Materials and Methods

2.1. Study Area and SAR Data

The study area is East Dongting Lake, located in Yueyang City, Hunan Province, China, with coordinates ranging from 29°00′ N to 29°38′ N latitude and 112°43′ E to 113°14′ E longitude, as shown in Figure 1. The study area has a subtropical humid monsoon climate, characterized by abundant sunshine and significant annual variations in precipitation. The region is home to the Hunan East Dongting Lake National Nature Reserve, which joined the Ramsar Convention on Wetlands of International Importance in 1992 and was upgraded to a national nature reserve in 1994. It represents a typical wetland ecosystem in China.

In recent years, climate change has been severe, and drought and floods have occurred frequently in Hunan. East Dongting Lake serves as an essential flood storage lake in the Yangtze River system, and its seasonal characteristics are particularly striking. There is a great urgent need to have an insight into the situation of the surface water body for the ecological protection of this area.

Sentinel-1 SAR sensor works in the C-band and exhibits better performance for discriminating water bodies with the appropriate spatiotemporal resolution. Two SAR images from it are used in this study, and their information is shown in

Table 1. Detailed information on SAR data.

ID	Time (M/D/Y)	Range Spacing (m)	Azimuth Spacing (m)	Orbit Direction	Processing Level
1	24 July 2022	10	10	Ascending	L1-GRD (IW)
2	1 August 2022	10	10	Ascending	L1-GRD (IW)

. One is for training, and the other is for testing. The training image consists of 33,097 × 21,287 pixels. To avoid the computational burden, this image is cropped into 509 patches, with a size of 512 × 512 pixels.

The preprocessing of SAR images involves orbit correction, radiometric calibration, speckle filtering (Refined Lee), range-Doppler terrain correction, and converting from linear to dB.

2.2. Method

The proposed method consists of four steps, as shown in Figure 2. First, the DK method is employed to divide the training dataset into four categories. Then, the water body is extracted from these images to serve as the necessary labels for the Classifier-Optimizer (CO) training. Second, the classifier is used for pixel-level dense prediction, generating an initial result image with fuzzy water-body boundaries. Third, the optimizer takes the initial result image and the original SAR image as inputs for fine-grained optimization, producing a result image with clearer water-body boundaries.

2.2.1. DK Processing

To address the limitations of the K-means algorithm, such as initial centroid selection and convergence speed, the DBO is selected to optimize the results of K-means by mimicking the behavior of dung beetles. There are four categories of agents in DBO: ball-rolling dung beetles, brood balls, small dung beetles, and thieves [34]. During the rolling process, the position of the ball-rolling dung beetle is updated and can be expressed as:

$$\begin{gathered} x_i\left(t+1\right)=x_i\left(t\right)+\alpha \times k\times x_i\left(t-1\right)+b\times ∆x, \\ ∆x=\left|x_i\left(t\right)-X^{\omega }\right| \end{gathered}$$

(1)

where $t$ represents the current iteration number, $x_i\left(t\right)$ denotes the position information of the $i$th dung beetle, $k$ denotes a constant value which indicates the deflection coefficient, $\alpha$ is a natural coefficient that is assigned $-1$ or $1$, and $X^{\omega }$ indicates the global worst position.

When an obstacle is encountered, the dung beetle dances to reorient itself, and its position is updated as follows:

$$x_i\left(t+1\right)=x_i\left(t\right)+\mathit{tan}\left(\theta \right)\left|x_i\left(t\right)-x_i\left(t-1\right)\right|$$

(2)

where $\theta$ is the deflection angle.

To simulate the area where female dung beetles lay eggs a boundary selection strategy was used, which was defined as

$$\begin{gathered} L{b}^{\mathrm{*}}=max\left(X^{\mathrm{*}}\times \left(1-R\right),Lb\right), \\ U{b}^{\mathrm{*}}=min\left(X^{\mathrm{*}}\times \left(1+R\right),Ub\right) \end{gathered}$$

(3)

where $X^{*}$ denotes the best position, $Lb$ represents the lower bound, and $Ub$ represents the upper bound.

The position of the ovoid changes dynamically during the iterative process, as defined below.

$$B_i\left(t+1\right)=X^{\mathrm{*}}+b_1\times \left(B_i\left(t\right)-L{b}^{\mathrm{*}}\right)+b_2\times \left(B_i\left(t\right)-U{b}^{\mathrm{*}}\right)$$

(4)

where $B_i\left(t\right)$ is the position information of brood ball, and $b_1$ and $b_2$ are random vectors by size $1\times D$.

Additionally, it is necessary to identify the ideal foraging area to effectively guide the small dung beetle, thereby simulating its foraging behavior more simply. The optimal foraging area is defined as follows:

$$\begin{gathered} L{b}^b=max\left(X^b\times \left(1-R\right),Lb\right), \\ U{b}^b=min\left(X^b\times \left(1+R\right),Ub\right) \end{gathered}$$

(5)

where $X^b$ denotes the global best position. As a result, the position of the small dung beetle is updated as follows.

$$x_i\left(t+1\right)=x_i\left(t\right)+C_1\times \left(x_i\left(t\right)-L{b}^b\right)+C_2\times \left(x_i\left(t\right)-U{b}^b\right)$$

(6)

where $C_1$ represents a random number that follows normally distributed, and $C_2$ denotes a random vector belonging to $\left(\mathrm{0,1}\right)$.

Assuming that $X^b$ is the best place to compete for food, the location of dung beetles with stealing behavior is thus updated as described below:

$$x_i\left(t+1\right)=X^b+S\times g\times \left(|x_i\left(t\right)-X^{\mathrm{*}}|+|x_i\left(t\right)-X^b|\right)$$

(7)

where $g$ is a random vector thar follows normally distributed; $S$ indicates a constant value.

Though the iteration of DBO, the initial clustering centers for K-means are determined. Subsequently, each pixel within the dataset is assigned to a specific clustering category. The center of each cluster is updated and calculated as the mean of all points within that cluster. This procedure is repeated until every data point is closest to the center of its corresponding cluster, ensuring optimal clustering.

2.2.2. CO Processing

In our approach, the classifier employs the DeepLabV3+ network architecture specifically for surface water-body detection in SAR images using pixel-wise dense prediction tasks [35]. DeepLabV3+ features a complete encoder-decoder structure, where the encoder captures the semantic information of water in the SAR image, and the decoder recovers this information to pixel-level resolution, facilitating accurate detection of water boundaries.

To address the issue of reduced spatial resolution in traditional CNNs, DeepLabV3+ employs the atrous convolution, which expands its receptive field without increasing computational complexity or parameter count. This enhancement allows the network to efficiently capture multi-scale contextual information, essential for detecting water bodies of varying sizes and shapes in SAR images.

Furthermore, we leverage DeepLabV3+’s ability to handle objects of different scales by using multiple parallel atrous convolution layers with different sampling rates. The features extracted at each sampling rate are further processed and fused in different branches to generate the final result, utilizing the Atrous Spatial Pyramid Pooling (ASPP) module, as shown in Figure 3.

By applying the DeepLabV3+ network in our classifier, we can perform pixel-level classification tasks on SAR images to accurately identify water bodies. After training the classifier with the DK-generated labels, it can effectively classify pixels and distinguish water bodies with varying scales and shapes, crucial for water resource management and environmental protection. Finally, the output generated by the classifier is further refined by the optimizer to obtain a more accurate and clear representation of water-body boundaries, enabling more precise monitoring and analysis of the surface water body.

In the classifier, there has always been a trade-off between classification accuracy and boundary segmentation accuracy. Models with deeper layers and multiple max-pooling layers perform well in classification tasks. However, as the number of max-pooling layers increases, spatial information gradually disappears, resulting in coarse boundary segmentation outcomes. Therefore, the classifier can predict the presence and approximate location of objects, but it is not capable of accurately delineating their boundaries. Therefore, following the classifier segmentation, an initial result image with fuzzy water-body boundary is generated. FC-CRF is employed for fine-grained refinement of the initial segmentation image, yielding sharper water-body boundaries. The energy function used is presented in Equation (8) [14].

$$\begin{array}{c}E\left(x\right)={\displaystyle \sum _i}{\theta }_i\left(x_i\right)+{\displaystyle \sum _{ij}}{\theta }_{i}g\left(x_i,x_j\right)\\ {\theta }_i\left(x_i\right)=-logP\left(x_i\right)\end{array}$$

(8)

where $x$ represents the pixel’s label assignment, ${\theta }_i\left(x_i\right)$ is one-dimensional potential, and $P\left(x_i\right)$ denotes the label assignment probability of pixel $i$, as calculated by the DeeplabV3+.

The pairwise potential adopts a structure that enables efficient inference when utilizing a fully-connected graph, that is, connecting every pair of image pixels $i$ and $j$. Specifically, we employ the expression as described in Equation (9):

$$\theta \left(x_i,x_j\right)=\mu \left(x_i,x_j\right)\left[{\omega }_1\underset{apperancekernel}{\underbrace{exp\left(-\frac{{\left|\left|p_i-p_j\right|\right|}^2}{2{\sigma }_{\alpha }^2}-\frac{{\left|\left|I_i-I_j\right|\right|}^2}{2{\sigma }_{\beta }^2}\right)}}+{\omega }_2\underset{smoothnesskernel}{\underbrace{exp\left(-\frac{{\left|\left|p_i-p_j\right|\right|}^2}{2{\sigma }_{\gamma }^2}\right)}}\right]$$

(9)

where $\mu \left(x_i,x_j\right)$ equals $1$ if $x_i$ and $x_j$ have the same value, signifying that only nodes with distinct labels face penalties. The remaining expression incorporates two Gaussian kernels in distinct feature spaces. The appearance kernel relies on both pixel positions and grayscale value, while the smoothness kernel depends solely on pixel positions. The hyperparameters ${\sigma }_{\alpha }$, ${\sigma }_{\beta }$, and ${\sigma }_{\gamma }$ dictate the scale of the Gaussian kernels. The first kernel encourages pixels with similar grayscale value and position to have corresponding labels, whereas the second kernel takes into account only spatial closeness when promoting smoothness.

3. Experimental and Results

3.1. Qualitative Evaluation

In the comparative experiment, we select three unsupervised and two deep-learning methods of water-body detection in SAR images to evaluate the performance and effectiveness, namely SegNet [36], U-Net [37], Markov Random Field (MRF) [38], Fuzzy C-means (FCM) [39], and Active Contour Model (ACM) [40]. Their effectiveness and accuracy are demonstrated through both qualitative and quantitative assessments.

To provide a more accurate and intuitive comparison of the six methods’ results, we select two subregions within the study area for a detailed presentation. As shown in Figure 4, the green box represents Region I, while the red box represents Region II. The upper-right corner of Region I contains numerous vessels, which can cause significant interference in water-body detection; selecting this area allows for the assessment of the model’s ability to resist such interference. Region II primarily consists of small ponds and narrow water bodies, and selecting this area allows for the evaluation of the model’s ability to identify these smaller water bodies. As seen in the image, the scattering energy of the water body in Region I is higher than that in Region II; by choosing these two regions, it is possible to simultaneously determine whether the model training is sufficient and if it can accurately detect different types of water-body. The specific results are shown in Figure 5 and Figure 6.

Figure 5 shows the comparison results of various methods in Region I. As can be seen from the blue circles in Figure 5b,c, the clarity of the water-body edges within the blue circles is significantly improved after the optimization by the optimizer. At the same time, the water-body extraction appears more complete in Figure 5b,c compared to Figure 5a. In Figure 5d,e, although most of the water bodies are detected, compared with Figure 5a, it can be seen that the SegNet and U-Net have limited ability to extract small water body without an optimizer. In Figure 5f–h, the most noticeable areas are the red circles on the right. Unsupervised segmentation methods may struggle to accurately identify water bodies in the original SAR images, possibly due to the presence of many ships in the river. In Figure 5c, the water-body edge within the left red circle is quite clear. However, in Figure 5f,h, the water body is not completely extracted. In Figure 5g, the water-body boundary within the left red circle is very blurry.

Figure 6 shows the comparison results of various methods in Region II. In Figure 6b,c, after optimization by the optimizer, the small water body within the bellow blue circle is also detected. The water-body edges within the upper blue circle have become very clear and smooth. At the same time, the water-body extraction appears more complete in Figure 6b,c compared to Figure 6a. In Figure 6d,e, the results are similar to those in Figure 5; although most of the water-body are detected, compared to Figure 6a,c, it can be seen that the deep-learning method also has limited ability to extract small water body without an optimizer. In Figure 6f–h, the small pond within the right red circle is not detected well. In fact, the water body is not detected at all within the right red circles in Figure 6g,h. In Figure 6b,c, the water-body edges within the left red circle are very clear, especially in Figure 6c where the detection errors are corrected after optimization. However, in Figure 6f–h, the water-body boundaries within the left red circle are extremely blurred. Moreover, in Figure 6g,h, the water body within the left red circle is not completely detected.

3.2. Quantitative Evaluation

In quantitative evaluation, we compare the resulting images obtained by various methods with the ground truth labels to generate confusion matrices, and the results for Region I and Region II are shown in Figure 7 and Figure 8. As shown in

Table 2. Confusion matrix.

		Generated Label
		Water	Non-Water
Ground truth	Water	True Position (TP)	False Negative (FN)
	Non-water	False Positive (FP)	True Negative (TN)

, the definitions of the four indicators were given in the confusion matrix. ACC, Precision, F1-score, Recall, and Kaapa are chosen as evaluation indicators. The calculations of them are illustrated in Equations (10)–(14), respectively. Based on the confusion matrix, we perform a quantitative evaluation of the different methods. The averaged results are listed in

Table 3. The accuracy of water-body detection.

Method	ACC	Precision	F1-Score	Recall	Kappa (%)
	(%)	(%)	(%)	(%)
DK-CO	97.22	93.45	95.21	97.05	93.25
SegNet	95.68	93.15	94.35	95.58	92.65
U-Net	95.81	94.32	94.53	94.32	92.93
MRF	95.92	88.99	93.14	97.86	90.26
SubPixelACM	96.29	93.62	93.35	93.12	90.76
FCM	95.79	97.57	92.04	92.66	89.18

.

ACC:

$$\mathrm{A}\mathrm{C}\mathrm{C}=\frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}+\mathrm{F}\mathrm{P}+\mathrm{T}\mathrm{N}}$$

(10)

Precision:

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}$$

(11)

F1-Score:

$$\mathrm{F}1\text{-}\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=\frac{2\mathrm{T}\mathrm{P}}{2\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}}$$

(12)

Recall:

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}$$

(13)

Kappa:

$$\mathrm{K}\mathrm{a}\mathrm{p}\mathrm{p}\mathrm{a}=\frac{\mathrm{A}\mathrm{C}\mathrm{C}-\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}\right)\times \left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}\right)+\left(\mathrm{F}\mathrm{N}+\mathrm{T}\mathrm{N}\right)\times \left(\mathrm{F}\mathrm{P}+\mathrm{T}\mathrm{N}\right)}{{\left(\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}\right)}^2}$$

(14)

Among them, the proposed DK-CO achieves the best accuracy. The ACC score is 97.22%, the precision score is 93.45%, the F1-Score is 95.21%, the recall score is 97.05%, and the Kappa coefficient score is 93.25%. Unsupervised classification methods cannot effectively establish mapping relationships for classifying each pixel, resulting in lower metric values. Although the deep-learning method without an optimizer can detect major water-body, the ability to detect small water bodies needs to be improved. Compared to machine learning methods and deep learning without an optimizer method, the metrics are improved for the proposed method in this paper.

In order to objectively evaluate the performance of each method, a visual comparison of the accuracy assessment metrics, including accuracy (ACC), precision, F1-Score, recall, and Kappa, was presented in Figure 9. The methods can be ranked in terms of ACC as follows: DK-CO > SubPixelACM > MRF >FCM. This indicates that the DK-CO method surpasses the other approaches in the task of water body detection, showing its superior performance.

4. Discussion

In the community of water-body detection using SAR images, the automatic generation of sample labels significantly optimizes the training process based on deep-learning network models. Traditional deep-learning methods usually rely on a large amount of manually labeled data, and the process is time-consuming and cumbersome. However, the DBO-K-means-based method proposed in this paper achieves the automatic extraction of water-body labels by optimizing the initial cluster centers and segmenting the training images into four categories. This method without human intervention greatly reduces the workload of manual labeling and improves the efficiency of label generation. On the other hand, this method also improves the accuracy and consistency of label generation, since machine learning algorithms are able to maintain consistent data better than humans. Therefore, this advancement has far-reaching implications for large-scale or real-time water-body detecting projects, especially when large amounts of SAR image data need to be processed.

The method proposed in this paper combines the advantages of optimization algorithms and deep-learning techniques to make it perform well in the task of water-body detection. First, through a refined pixel-level classification task, the classifier generates initial result images with blurred water-body boundaries. The optimizer then refines this initial result to produce a final result image with sharp water-body boundaries. This combined method takes full advantage of the capabilities of optimization algorithms in exploring complex solution spaces, adapting to different data distributions, and finding global optimal or near-optimal solutions, as well as the advantages of deep learning in automatic learning and feature extraction. In addition, the accuracy of water-body detection is evaluated by several performance metrics, including ACC, precision, F1-Score, recall, and Kappa coefficient. The results show that all indicators exceeded 93%, which further indicated the accuracy and reliability of this method in water-body detection. Therefore, in future work, we plan to expand the dataset and test the model in more complex scenarios to further improve the performance and applicability of the model. At the same time, we will also look for more effective optimization algorithms and deep-learning network models to improve the accuracy and speed of water-body detection.

Although the method proposed in this paper performs well in experiments, there are still some limitations and open issues. First, our method is mainly tested in simple scenarios and may require further validation in more complex scenarios. Second, our dataset is relatively small, which may affect the generalization ability of the model. In addition, although the method in this paper reduces the workload of manual labeling, it still requires certain computing resources to perform the DBO-K-means and training of deep-learning models.

Additionally, this paper explored the time consumption of DK-CO, SegNet, U-Net, MRF, ACM, and FCM. Through ten experiments, it was concluded that the time consumption of traditional machine learning methods was relatively less, and the time consumption of deep learning is similar. The result is shown in Figure 10. For research purposes, it is worthwhile to use an optimizer to improve accuracy compared with traditional machine learning and deep learning without an optimizer.

5. Conclusions

A water-body detection method based on SAR images is proposed. The labels required for CO training are obtained by using the DK method. Then, the labels and training dataset are brought into CO. Finally, CO is used to perform a dense pixel-by-pixel classification task and fine-grained optimization of the water-body boundary to obtain a result map with a clear water-body boundary. The ACC, precision, F1-score, recall, and Kappa coefficients reach 97.22%, 93.45%, 95.21%, 97.05%, and 93.25%, respectively. The results show the effectiveness of the proposed method. Moreover, the method is applicable to other areas, where fine water-body and complex terrain can be detected. However, our study also has some limitations. First, our method is mainly tested in simple scenarios and may require further validation in more complex scenarios. In addition, our dataset is relatively small, which may have affected the generalization ability of the model. In future work, we plan to expand the dataset and test the model in more complex scenarios to further improve the performance and applicability of the model. This study can provide a new idea and method for remote sensing detection of water bodies.