SAR Image Water Body Extraction Method Based on Multi-Scale Residual Attention Model

Abstract

The present invention discloses a method for extracting water bodies from SAR images based on a multi-scale residual attention model, and the steps are as follows: First, preprocess the original SAR image and make labels to obtain the training set and test set of the model; Secondly, in view of the problem that the water body shapes in remote sensing images are diverse, the encoder of the model is designed as a multi-scale residual module; In view of the problem of low recognition accuracy for narrow waters, the decoder of the model adopts a nested structure; Finally, a refinement attention module (SC-attention) is added behind each layer of the encoder to fuse the rich semantic information in the spatial features and the accurate position information in the channel features, so as to gradually refine the segmentation details to avoid the loss of water body boundary information. The present invention utilizes the multi-scale residual module to obtain the feature information of the water body at different scales, and the nested decoder can suppress the speckle noise and non-water body background in the image, retain more water body information, and thus obtain better segmentation results.

Description

Technical Field

[0001] The present invention relates to remote sensing image processing, and in particular to a method for extracting water bodies from SAR images based on a multi-scale residual attention model. Background Art

[0002] Lakes are an important part of the terrestrial ecosystem and play an important role in maintaining ecological balance, flood diversion, and domestic water use. As one of the most important parameters of lake morphology, the change of lake area reflects the changes in water resource management and the environment of flood control reservoirs, directly affecting human production and life. With the rapid development of remote sensing technology, a large amount of remote sensing satellite data can be used to monitor water bodies. However, optical satellites cannot penetrate clouds to provide more useful water body information, making it difficult to use optical satellite data for real-time monitoring of lake water surfaces. Currently, the emergence of SAR has solved the thorny problem of cloud interference because it has strong penetration ability and can overcome the obstacles of clouds or fog regardless of meteorological conditions. In addition, SAR can also make up for the lack of visible light data in providing night images.

[0003] In the past few years, a large number of SAR-based water body segmentation methods have emerged like mushrooms after a spring rain, mainly divided into two categories: unsupervised and supervised. Among them, the unsupervised method based on adaptive threshold segmentation is the most common water extraction algorithm, which segments water bodies based on the fact that the backscattering coefficient of water is much lower than that of other objects in SAR. Current automatic threshold segmentation methods include the OSTU method and the entropy threshold method, which find the optimal threshold by analyzing the bimodal histogram of SAR. However, the threshold is easily affected by human subjectivity and changes with time and space. When the proportion of water in the image is the smallest, the histogram cannot show a bimodal shape, indicating that the water feature will be greatly affected by the background, ultimately resulting in unsatisfactory water extraction results. Different from the threshold method, the convolutional neural model (CNN) based on deep learning can flexibly learn features from a large number of samples. Among them, the fully convolutional model FCN based on the CNN structure and the encoder-decoder structure have become the mainstream semantic segmentation model structures. In particular, U-Net has much better performance than traditional water body segmentation methods. However, if the general FCN or U-Net is directly used for water body extraction without modification, problems such as poor prediction accuracy, blurred water body boundaries, and visual degradation often occur. Therefore, how to improve the prediction accuracy and maintain the accurate segmentation of water body boundaries are two important problems faced by water body segmentation in remote sensing images. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to provide a method for extracting water bodies from SAR images based on a multi-scale residual attention model, which uses a multi-scale residual module to obtain feature information of water bodies at different scales, and can suppress speckle noise and non-water body background in the image through a nested decoder, retaining more water body information, thereby obtaining better segmentation results.

[0005] Technical solution: A method for extracting water bodies from SAR images based on a multi-scale residual attention model according to the present invention includes the following steps:

[0006] (1) To address the problem of low recognition accuracy for narrow water areas, a decoder structure with nested connections is introduced. This structure uses upsampling and several short skip connections instead of long skip connections. The nested decoder can suppress the influence of semantic gaps and retain more information, thus obtaining better segmentation results. To address the problem of diverse water body shapes in remote sensing images, a water body segmentation network based on multi-scale residuals and attention is proposed. The multi-scale residual module can analyze targets at different scales and extract deeper high-level feature information. To obtain accurate and continuous water body boundaries, a refinement attention module is added behind each encoder layer to fuse the rich semantic information in spatial features and the accurate position information in channel features, gradually refining the segmentation details to avoid the loss of water body boundary information.

[0007] (2) Obtain the SAR image, preprocess the source image to obtain the training dataset and test dataset of the model.

[0008] (3) Make labels based on the characteristics that the water body has a low scattering coefficient and appears darker in the SAR image.

[0009] (4) Establish a method for extracting water bodies from SAR images using a multi-scale residual attention model.

[0010] (5) Set the hyperparameters, total number of training times, learning rate, momentum parameter, and weight decay parameter of the model. Through repeated adjustment and comparative experiments, obtain the optimal parameters to obtain the optimal water body segmentation model.

[0011] (6) Use the test set to test the trained model, and finally verify the superiority of the model of the present invention in water body segmentation by comparing with traditional models.

[0012] The specific content of step (2) is as follows:

[0013] Perform radiometric calibration, terrain correction, logarithmic transformation, and speckle filtering preprocessing on the SAR image; superimpose and combine the SAR dual-polarization data and the derived SDWI data to form a new three-channel image, and use the new image as the input of the model. The SDWI formula is as follows:

[0014] K SDWI = ln(10 × VV × VH) (1)

[0015] where K SDWIRepresents the result value of band operation, where VV and VH represent SAR dual-polarization data; SDWI draws on the Normalized Difference Water Index (NDWI), and uses band operations between SAR dual-polarization data to enhance water body features, achieving good results in water body information extraction.

[0016] The specific steps of step (3) are as follows:

[0017] (3.1) Annotate each SAR image to obtain a binary classification label for water and non-water areas;

[0018] (3.2) Keep the input size of each model consistent;

[0019] (3.3) Crop the processed images and labels to finally obtain the training dataset, validation dataset, and test dataset of the model.

[0020] The specific steps of step (4) are as follows:

[0021] (4.1) To address the problem of diverse water body forms in remote sensing images, a water body segmentation model composed of a multi-scale residual module and a refinement attention module is proposed; the multi-scale residual module can analyze targets at different scales and extract deeper high-level feature information;

[0022] (4.2) To address the problem of low recognition accuracy for narrow water areas, a decoder structure with nested connections is introduced, which uses upsampling and several short skip connections instead of long skip connections;

[0023] (4.3) Add a refinement attention module after each layer of the encoder to fuse the rich semantic information in the spatial features and the accurate position information in the channel features, so as to gradually refine the segmentation details to avoid the loss of water body boundary information.

[0024] The specific design steps of the refinement attention module described in step (4.3) are as follows:

[0025] The refinement attention module introduces the ideas of spatial attention SA and channel attention CA; the SA and CA modules can learn the SA weight and CA weight respectively, and by weighting the two attention weights to the combined features, two new weighted features can be obtained, that is, the SA weight is assigned to the abstract elements, and the CA weight is assigned to the spatial elements; finally, the two weighted features are added to obtain a new fused feature; the refinement attention module utilizes channel features and spatial features, ultimately improving the accuracy of pixel-level classification in semantic segmentation. The refinement attention module is expressed as follows:

[0026] f CA =σ(MLP(AvgPool(X)+MaxPool(X)))

[0027] = σ(W1(W0(Favg)+W1(W0(Favg)))

[0028] f SA = σ(f 7×7 (AvgPool(X); MaxPool(X)))

[0029] = σ(f 7×7 (Fss))

[0030]

[0031] where W1 ∈ R c / r×c and W0 ∈ R c×c / r are the weights of the multi-layer perceptron, f 7×7 represents a convolutional layer with a filter size of 7×7, σ is the sigmoid function, represents the dot product operator, f CA is the CA module, f SA is the SA module; X is the input feature map;

[0032] When long skip connections are used in the model architecture, it is impossible to distinguish between waters with similar features and small waters, or to exclude the interference of radar noise and mountain shadows; for these problems, a nested connection method is proposed, which uses upsampling and several short skip connections instead of long skip connections; through nested connections, speckle noise and water shadows in the image can be suppressed, and more information can be retained, thus obtaining better segmentation results.

[0033] A computer storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the above-mentioned method for extracting water bodies from SAR images based on a multi-scale residual attention model.

[0034] A computer device, including a storage, a processor, and a computer program stored on the storage and executable on the processor, and when the processor executes the computer program, it implements the above-mentioned method for extracting water bodies from SAR images based on a multi-scale residual attention model.

[0035] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0036] 1. In view of the problem that water bodies in remote sensing images have various forms, the present invention proposes a water body segmentation network based on multi-scale residuals and attention. The multi-scale residual module can analyze targets at different scales and extract deeper-level high-level feature information;

[0037] 2. In view of the problem of incomplete extraction of water body boundaries, the present invention adds a refinement attention module (SC-attention) after each layer of the encoder to fuse the rich semantic information in the spatial features and the accurate position information in the channel features, so as to gradually refine the segmentation details to avoid the loss of water body boundary information, effectively solving the above problems;

[0038] 3. The present invention uses upsampling and several short skip connections instead of long skip connections. Through nested connections, it can suppress speckle noise and non-water body background in the image, retain more information, and thus obtain better segmentation results;

[0039] 4. The model trained by the present invention only takes one to two seconds to segment the water body in a single satellite image, and can quickly and accurately segment the target water body, having good application prospects;

[0040] 5. The present invention only needs to prepare corresponding data and adjust a small number of hyperparameters of the deep learning model according to the training situation of its own model during the experimental stage; the deep learning model automatically learns the features of the water body in the remote sensing image according to the input data and labels, and automatically saves the best model parameters. After the best model is saved, only by calling the model can the water body segmentation be realized, which is convenient for users to operate. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic flowchart of the method of the present invention.

[0042] Figure 2 It is a schematic diagram of the data set. Among them, Figure 2 (a) is the original data image, Figure 2 (b) is the data label.

[0043] Figure 3 It is a schematic diagram of the model Attention-MRNet of the present invention.

[0044] Figure 4 It is a schematic diagram of the multi-scale residual module in the model of the present invention.

[0045] Figure 5 It is a schematic diagram of the refinement attention module in the model of the present invention.

[0046] Figure 6 It is a schematic diagram for comparing the segmentation results of the model of the present invention and the traditional model. Among them, Figure 6 (a) is a schematic diagram for comparing the validation loss, Figure 6 (b) is a schematic diagram for comparing the Accuracy, <L Figure 6 (c) is a schematic diagram for comparing the Recall, Figure 6 (d) is a schematic diagram for comparing the Precision, Figure 6 (e) is a schematic diagram for comparing the IOU. Figure 6(f) is a schematic comparison diagram of F1.

[0047] Figure 7 This is a schematic comparison diagram of the extraction of typical water area information between the model of the present invention and the traditional model. Among them, Figure 7 (a) is the original data diagram, Figure 7 (b) is the data label, Figure 7 (c) is the extraction result of Attention-MRNet, Figure 7 (d) is the extraction result of FCN, Figure 7 (e) is the extraction result of Hrnet, Figure 7 (f) is the extraction result of Inception_v3, Figure 7 (g) is the extraction result of Unet, Figure 7 (h) is the extraction result of W-net, Figure 7 (i) is the extraction result of DenseNet121, Figure 7 (j) is the extraction result of ResNet34. Detailed implementation manner

[0048] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0049] As Figure 1 shown, the present invention designs a method for extracting water bodies from SAR images of a multi-scale residual attention model. The method specifically includes the following steps:

[0050] Step 1: In order to train the model, the present invention selects the middle and lower reaches of the Yangtze River and Poyang Lake as the training and testing of the model. Through the Sentinel-1 SAR satellite developed by the European Space Agency, the selected time of the data is 20 SAR images from May 27th to November 23rd, 2020 and 10 SAR images from July 22nd to November 7th, 2022. The image information is shown in Table 1:

[0051] Table 1 Data of the flood period in 2020 and the drought period in 2022 in Poyang Lake

[0052] Serial number Imaging time Imaging time Imaging time Product type Polarization mode 1 May 27, 2020 August 7, 2020 July 22, 2022 GRD VH, VV 2 June 2, 2020 August 19, 2020 August 3, 2022 GRD VH, VV 3 June 8, 2020 August 31, 2020 August 15, 2022 GRD VH, VV 4 June 14, 2020 September 12, 2020 August 27, 2022 GRD VH, VV 5 June 20, 2020 September 24, 2020 September 8, 2022 GRD VH, VV 6 June 26, 2020 October 6, 2020 September 20, 2022 GRD VH, VV 7 July 2, 2020 October 18, 2020 October 2, 2022 GRD VH, VV 8 July 8, 2020 October 30, 2020 October 14, 2022 GRD VH, VV 9 July 14, 2020 November 11, 2020 October 26, 2022 GRD VH, VV 10 July 20, 2020 November 23, 2020 November 7, 2022 GRD VH, VV

[0053] Step 2: After obtaining the sentinel-1 data set, first use SNAP to complete data preprocessing operations such as radiometric calibration, terrain correction, and logarithmic transformation.

[0054] Step 3: Considering the limited polarization data of Sentinel-1, the present invention combines the Sentinel-1 dual-polarization data and the derived SDWI data to form a new three-channel image, and uses the new image as the input of the model. Among them, SDWI is a water body information extraction method proposed by Jia Shichao et al. in 2018. The formula is as follows:

[0055] K SDWI = ln(10 × VV × VH) (1)

[0056] Where K SDWI represents the result value of the band operation, and VV and VH represent Sentinel-1 dual-polarization data. SDWI draws on the Normalized Difference Water Index (NDWI) and uses band operations between Sentinel-1 dual-polarization data to enhance water features, achieving good results in water information extraction.

[0057] Since the brightness in SAR images depends on the magnitude of the backscattering coefficient, the smaller the backscattering coefficient, the darker the SAR image. Therefore, water often appears darker than other ground features in SAR images. By combining the differences in backscattering coefficients and prior geographical knowledge, each SAR image is labeled to obtain binary classification labels for water and non-water areas. To fairly compare the performance of CNNs, the input size of each model is kept consistent, i.e., 512 × 512 pixels. Therefore, 30 three-channel original images and their corresponding labels are cropped, and finally 3286 training datasets, 340 validation datasets, and 253 test datasets are obtained. Figure 2 Shows elongated rivers, lake surfaces, mountain rivers, and urban waters. In the labeled image, the white areas represent water, while the black areas represent non-water areas.

[0058] Table 2 Dataset Size and Characteristics

[0059] Dataset Number of samples Date Training set 3286 2020-01—2022-10 Validation set 340 2020-01—2022-10 Test set 253 2020-01—2022-10

[0060] Step 4: The encoder of the model consists of a convolutional layer and four multi-scale residual modules. In each multi-scale residual module, the number of input channels is multiplied by 2, and the feature map is reduced by half through a max-pooling layer. The max-pooling layer is removed from the fourth multi-scale residual module, making the final feature map size 512 × 32 × 32. After a series of convolutions and downsamplings by the encoder, the number of channels in the middle layer of the model reaches 512. To enable the model to learn feature information at multiple abstract levels and prevent the problem of gradient disappearance or explosion in deep models, the output of the fourth multi-scale residual module is used as the input to the ASPP module, and the dilation rates of the dilated convolutions are set to 1, 2, and 5 respectively. This is to ensure that when the resolution of the feature map is small and the channel information is rich, the receptive field of each channel's feature information is further expanded to obtain deep-channel feature information. In addition, the result of the last dilated convolution is subjected to adaptive average pooling and bilinear interpolation operations to ensure that the number of input and output features remains unchanged.

[0061] The decoder of the model adopts an embedded structure. In addition to the multi-scale residual module and the upsampling convolutional layer, a refinement attention module (SC-attention) corresponds to each encoder layer. The output of the SC-attention module serves as the input to the decoder. Finally, after four layers of upsampling and a convolutional layer, the reduced-resolution image is upsampled to the size of the original image. Finally, the SE-attention module is combined to obtain the key feature information of each pixel channel, and the water body segmentation result is finally output. The detailed settings of the model are shown in Table 3, where Conv represents the convolutional layer, and En and Dnn represent the multi-scale residual module. The model structure is as Figure 3 shown.

[0062] Table 3 Model structures of the encoder and decoder

[0063]

[0064]

[0065] As a part of remote sensing images, water bodies often have irregular and different scales in most cases, and their sizes and shapes vary from region to region. This invention draws on the characteristics of the inception model. However, the feature information extracted by adjacent layers of the neural model is correlated. Using convolutional layers with kernel sizes of 5×5 and 7×7 in parallel in the model will greatly increase the memory requirement. Therefore, following its idea, two consecutive 3×3 and three 3×3 convolutional layers are used to replace the 5×5 and 7×7 convolutional operations, rather than directly using convolutional layers with kernel sizes of 5×5 and 7×7 with high memory requirements, and they are combined through concatenation operations. The results show that the multi-scale residual module enhances the nonlinearity while reducing the parameters, making the performance of the model better.

[0066] In addition, in order to prevent the model from crashing during the early stage of training, residual connectivity is introduced and 1×1 convolution is used to provide feature information of other scales, which is finally defined as the multi-scale residual module, as Figure 4 shown.

[0067] Since water bodies in remote sensing images are easily affected by non-water body local information such as the reflection of surrounding buildings, cloud shadows, and mountain shadows, these information often affect the overall segmentation accuracy of water bodies. To solve this problem, a branch is drawn from the fourth feature map obtained from the parallel fusion module through the ASPP module, and the dilation rates of the dilated convolutions are set to 1, 2, and 5 respectively. This is to ensure that when the resolution of the feature map is small and the channel information is rich, the receptive field of each channel's feature information is further expanded to obtain deep channel feature information. In addition, the results of the last dilated convolution are subjected to adaptive average pooling and bilinear interpolation operations to ensure that the number of input and output features remains unchanged.

[0068] The refined attention module (SC-attention) introduces the ideas of spatial attention SA (Spattial Attention) and channel attention CA (Channel Attention). The SA and CA modules can learn the SA weight and CA weight respectively. By weighting the two attention weights to the combined features, two new weighted features can be obtained, that is, the SA weight is assigned to the abstract elements, and the CA weight is assigned to the spatial elements. Finally, the two weighted features are added together to obtain a new fused feature. The refined attention module is as Figure 5 shown below:

[0069] The SC-attention module utilizes channel features and spatial features, ultimately improving the accuracy of pixel-level classification in semantic segmentation. The SC-attention module can be expressed as follows:

[0070] f CA = σ(MLP(AvgPool(X) + MaxPool(X)))

[0071] = σ(W1(W0(Favg) + W1(W0(Favg)))

[0072] f SA = σ(f 7×7 (AvgPool(X); MaxPool(X)))

[0073] = σ(f 7×7 (Fss))

[0074]

[0075] where W1 ∈ R c / r×c and W0 ∈ R c×c / r are the weights of the multi-layer perceptron (MLP), f 7×7 represents a convolutional layer with a filter size of 7×7, σ is the sigmoid function, represents the dot product operator, f CA is the CA module, and f SA is the SA module. X is the input feature map.

[0076] Step 5: Set the hyperparameters, total number of training times, adjustable learning rate, momentum parameter, weight decay parameter, and loss function of the model. Through repeated adjustment and comparative experiments, the best parameters are obtained to get the best precipitation intensity estimation model. To ensure the fairness of comparison, Adam is uniformly used as the optimizer for the model. Each model is set with 50 epochs, and one epoch contains 1643 iterations. The initial learning rate of the model is set to 0.00001. To better fit the model, the learning rate is multiplied by 0.1 every 30 epochs. Nesterov momentum and weight decay are set to 0.9 and 0.0003 respectively, aiming to accelerate the learning process of the model and prevent overfitting problems. All experiments are run on the GPU NVIDIA GeForce GTX1080 with the Pytorch 0.2.1 environment configured.

[0077] Step 6: The loss function is widely used to describe the gap between the predicted value and the label, and to evaluate the fitting ability of the model. For the binary classification of images, binary cross-entropy is selected as the loss function. Using binary cross-entropy, the loss curve is plotted to visually compare the efficiency of the models used. The lower and smoother the curve, the better the fitness and robustness of the corresponding model. Accuracy represents the correctness of sample prediction, whether positive or negative. Recall is the proportion of positive samples that are correctly predicted as positive. Precision represents the proportion of correct water samples among all samples predicted as true. Intersection over Union (IoU) represents the ratio of the intersection to the union of the true value and the predicted value. F1 score takes into account both precision and recall and is used to evaluate the binary classification results. The comparison between different models is as Figure 6 shown.

[0078] To further compare the deep convolutional neural models, a test set is used to test the models, and their accuracy, recall, precision, IoU, F1 score, and test time cost are calculated. Table 3 shows the metrics of all methods, and the best values are shown in bold. Figure 7 Show the water body extraction effects of each model.

[0079] Table 3 Accuracy, recall, precision, IOU, F1 score, and test time cost of Attention-MRNet, FCN, Hrnet, Inception_v3, Unet, W-net, DenseNet121, ResNet34, where the bold values are the best

[0080]

[0081] Therefore, through the above experimental comparison, it can be seen that the present invention captures the context features of water bodies at different scales based on the multi-scale residual module and extracts deeper high-level feature information. The SC-attention attention module is used to obtain the importance of each connection feature, fuse the rich semantic information in the spatial features and the accurate position information in the channel features, so as to gradually refine the segmentation details to avoid the loss of water body boundary information and reduce the number of false alarms caused by speckle noise and water shadows in SAR images. In order to combine feature maps of different scales and improve the accuracy of narrow water recognition, the decoder is designed as a nested structure to save a large amount of input data features from a multi-scale perspective, thereby obtaining better segmentation results.

Claims

1. A method for extracting water bodies from SAR images based on a multi-scale residual attention model, characterized in that, It includes the following steps: (1) Aiming at the problem of low recognition accuracy in narrow waters, a method for extracting water bodies from SAR images based on a multi-scale residual attention model is established. This model introduces a decoder structure with nested connections, which uses upsampling and several short skip connections instead of long skip connections; The decoder with nested connections can suppress the influence of semantic gaps and retain more information, thus obtaining better segmentation results; this model also includes a multi-scale residual and an attention module; the multi-scale residual module can analyze targets at different scales and extract deeper high-level feature information; in order to obtain accurate and continuous water body boundaries, a refinement attention module is added behind each encoder layer to fuse the rich semantic information in the spatial features and the accurate position information in the channel features, so as to gradually refine the segmentation details to avoid the loss of water body boundary information; The specific design steps of the refinement attention module are as follows: The refinement attention module introduces the ideas of spatial attention SA and channel attention CA; the SA and CA modules respectively learn the SA weight and the CA weight, and by weighting the two attention weights to the combined features, two new weighted features can be obtained, that is, the SA weight is assigned to the abstract elements and the CA weight is assigned to the spatial elements; finally, the two weighted features are added to obtain a new fused feature; the refinement attention module utilizes the channel features and the spatial features, and finally improves the accuracy of pixel-level classification in semantic segmentation. The refinement attention module is expressed as follows: f CA = σ(MLP(AvgPool(X) + MaxPool(X))) =σ(W1(W0(Favg)+W1(W0(Favg))) f SA = σ(f 7×7 (AvgPool(X); MaxPool(X)) = σ(f 7×7 (Fss)) where \(W1\in\mathbb{R}\) c / r×c and \(W0\in\mathbb{R}\) c×c / r are the weights of the multi-layer perceptron, \(f\) 7×7 represents a convolutional layer with a filter size of \(7\times7\), \(\sigma\) is the sigmoid function, represents the dot product operator, \(f\) CA is the CA module, \(f\) SA is the SA module; \(X\) is the input feature map; (2) Obtain SAR images, preprocess the source images to obtain the training dataset and the test dataset of the model; (3) Make labels based on the characteristics that the scattering coefficient of water bodies in SAR images is low and the presented color is relatively dark; (4) Set the hyperparameters, total number of training times, learning rate, momentum parameter, and weight decay parameter of the model, and obtain the best parameters through repeated adjustment and comparative experiments to obtain the best water body segmentation model; (5) Use the test set to test the trained model, and finally verify the superiority of the multi-scale residual and attention model in water body segmentation by comparing with traditional models.

2. The method for extracting water bodies from SAR images based on a multi-scale residual attention model according to claim 1, wherein The specific content of step (2) is as follows: Perform radiometric calibration, terrain correction, logarithmic transformation, and speckle filtering preprocessing on the SAR images; stack and combine the SAR dual-polarization data and the derived SDWI data into a new three-channel image, and use the new image as the input of the model. The SDWI formula is as follows: K SDWI = ln(10 × VV × VH) (1) Among them, K SDWI represents the result value of the band operation. VV and VH represent SAR dual-polarization data. SDWI draws on the normalized difference water index NDWI and uses the band operation between SAR dual-polarization data to enhance water body features, achieving a good effect in extracting water body information.

3. A method for extracting water bodies from SAR images based on a multi-scale residual attention model according to claim 1, characterized in that The specific content of step (3) is as follows: (3.1) Annotate each SAR image to obtain a binary classification label for the water area and the non-water area; (3.2) Keep the input size of each model consistent; (3.3) Crop the processed images and labels to finally obtain the training dataset, validation dataset, and test dataset of the model.

4. A computer storage medium, on which a computer program is stored, characterized in that, When the computer program is executed by a processor, it implements a method for extracting water bodies from SAR images based on a multi-scale residual attention model as described in any one of claims 1-3.

5. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a method for extracting water bodies from SAR images based on a multi-scale residual attention model as described in any one of claims 1-3.

Images