A Multi-Source SAR Water Body Extraction Method and System Based on MFAFNet

Abstract

This invention discloses a multi-source SAR water body extraction method and system based on MFAFNet. The method involves inputting an input SAR image into an MFAFNet network. The MFAFNet network extracts low-level features, high-level features, and multiple mid-level features from the SAR image through its backbone network. The low-level features are then fed into a decoder, the high-level features into a high-level multi-scale feature extraction module to extract high-level multi-scale features, and the multiple mid-level features into an intermediate-level feature fusion module to obtain intermediate-level features. Finally, the decoder decodes the low-level features, multi-scale features, and intermediate-level features to obtain the water body extraction result. This invention enables high-precision automatic detection of SAR water bodies from multiple sources across multiple frequency bands and resolutions, offering advantages such as high accuracy, low false alarm rates, and good detection performance.

Description

Technical Field

[0001] This invention relates to the field of image processing technology for Synthetic Aperture Radar (SAR) images, specifically to a multi-source SAR water body extraction method and system based on MFAFNet. Background Technology

[0002] Synthetic Aperture Radar (SAR) is unaffected by weather conditions, enabling all-day, all-weather imaging. It offers advantages such as wide monitoring range, low cost, and high efficiency. In SAR images, water bodies exhibit relatively low backscattering coefficients, appearing as dark areas. Utilizing SAR images for water body detection has significant application value in urban water body monitoring, coastline monitoring, and flood disaster monitoring, and has remained a research hotspot.

[0003] Many scholars have conducted research on water body detection, with traditional methods broadly categorized into two types: thresholding and classifier methods. Since Hinton proposed deep learning in 2006, increasingly more deep learning-based water body detection methods have emerged and become mainstream, achieving automated water body detection. Inspired by multi-scale feature fusion algorithms in computer vision, many researchers have explored various methods to improve classification accuracy using multi-scale features; others have used pixels as nodes and the relationships between pixels as edges, fully utilizing the spatial neighborhood information of labeled and observed images to effectively obtain spatial context information and refine the extraction boundaries. While these methods have addressed issues such as shadow interference and threshold setting in traditional methods to some extent, the existence of inter-class similarities and intra-class differences in water bodies means that automatic and high-precision detection of water bodies from SAR images of different frequency bands and resolutions still faces significant challenges. Current research on water body detection is primarily based on single-frequency SAR images, often failing to achieve satisfactory results when applied to other frequency bands, exhibiting weak generalization ability. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a multi-source SAR water body extraction method and system based on MFAFNet, which can realize high-precision automatic detection of SAR water bodies from multiple sources with multiple frequency bands and resolutions, and has the advantages of high accuracy, few false alarms and good detection effect.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A multi-source SAR water body extraction method based on MFAFNet includes:

[0007] S101, The input SAR image is input into the backbone network of the MFAFNet network. The MFAFNet network includes an encoder and a decoder. The encoder includes a backbone network, an intermediate-level feature fusion module, and an advanced multi-scale feature extraction module.

[0008] S102, the MFAFNet network extracts low-level features, high-level features and multiple mid-level features of SAR images through the backbone network, and sends the low-level features to the decoder, the high-level features to the advanced multi-scale feature extraction module to extract advanced multi-scale features, and the multiple mid-level features to the intermediate-level feature fusion module to fuse them to obtain intermediate-level features. Finally, the decoder decodes the low-level features, multi-scale features and intermediate-level features to obtain the water body extraction result.

[0009] Optionally, the backbone network is a ResNet_101 network with dilated convolutions, and its second and third layers output two intermediate features respectively, which are then fed into the intermediate feature fusion module to obtain intermediate features.

[0010] Optionally, the process of feeding the multiple intermediate-level features into the intermediate-level feature fusion module to obtain intermediate-level features includes: the intermediate-level feature fusion module upsampling the lower-level intermediate-level features to match their channel count with that of the higher-level intermediate-level features; then concatenating the upsampled feature map with the higher-level intermediate-level features after performing a convolution operation with a kernel size of 1×1; and finally concatenating the concatenated feature map with the feature map obtained after performing two convolution operations with a kernel size of 3×3 to obtain the intermediate-level features.

[0011] Optionally, the advanced multi-scale feature extraction module includes five parallel branches, which include four parallel attention pooling branches and one global average pooling branch. Each of the four parallel attention pooling branches contains a serially connected convolutional module and an effective attention mechanism (ECA) module. The convolutional modules of the four parallel attention pooling branches have different scales and dilatation rates. The feature maps output by the five parallel branches are stacked on the channels and then subjected to a convolution operation with a kernel size of 1×1 to obtain advanced multi-scale features.

[0012] Optionally, the Effective Attention (ECA) module processes the input feature map by performing global average pooling on the input feature map, followed by a 1D convolution operation with a kernel size of k, and then obtaining the weights of each channel through a Sigmoid activation function. The weights are then multiplied by the corresponding elements of the original input feature map to obtain the final output feature map.

[0013] Optionally, the calculation function expression for the convolution kernel size k is:

[0014]

[0015] In the above formula, ψ(C) represents a function of the channel dimension C of the input, γ and b are constant parameters, and |t| odd This represents the odd number closest to t.

[0016] Optionally, the step of decoding low-level features, multi-scale features, and intermediate-level features to obtain water body extraction results includes:

[0017] S201, attention modulation is performed on the low-level features through the attention modulation module (AMM) in the decoder to enhance the weight of the water target. Then, after reducing the channel redundancy features by using a convolution operation with a kernel size of 1×1, the first part of the decoded features is obtained. The multi-scale features and intermediate-level features are connected and upsampled to obtain the fused features as the second part of the decoded features.

[0018] S202, after concatenating the first part of the decoded features and the second part of the decoded features, refines the features through a 3×3 convolution, and then performs upsampling to obtain the final water extraction result.

[0019] Optionally, the upsampling in step S202 refers to bilinear interpolation upsampling by a factor of 4.

[0020] Furthermore, the present invention also provides a multi-source SAR water body extraction system based on MFAFNet, including a microprocessor and a memory interconnected thereto, wherein the microprocessor is programmed or configured to execute the multi-source SAR water body extraction method based on MFAFNet.

[0021] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program that is programmed or configured by a microprocessor to execute the MFAFNet-based multi-source SAR water body extraction method.

[0022] Compared with existing technologies, the present invention has the following main advantages: The method of the present invention includes inputting an input SAR image into an MFAFNet network. The MFAFNet network extracts low-level features, high-level features, and multiple mid-level features from the SAR image through a backbone network. The low-level features are then sent to a decoder, the high-level features are sent to a high-level multi-scale feature extraction module to extract high-level multi-scale features, and the multiple mid-level features are sent to an intermediate-level feature fusion module to fuse them into intermediate-level features. Finally, the decoder decodes the low-level features, multi-scale features, and intermediate-level features to obtain the water body extraction result. The present invention can achieve high-precision automatic detection of SAR water bodies from multiple sources with multiple frequency bands and resolutions, and has the advantages of high accuracy, low false alarm rate, and good detection effect. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the MFAFNet network structure in an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the intermediate-level feature fusion module in an embodiment of the present invention.

[0025] Figure 3 The data from Gaofen-3 are experimental results from an embodiment of this invention.

[0026] Figure 4 The data from Sentinel 1 in this embodiment of the invention are experimental results.

[0027] Figure 5 The data experimental results of TerraSAR data in the embodiments of the present invention are shown. Detailed Implementation

[0028] This embodiment provides a multi-source SAR water body extraction method based on MFAFNet, including:

[0029] S101, input the SAR image into the backbone network of the MFAFNet (Multi-level Feature Attention Fusion Network), such as... Figure 1 As shown, the MFAFNet network includes an encoder and a decoder. The encoder includes a backbone network, an intermediate-level feature fusion module (IFFM module), and an advanced multi-scale feature extraction module (ECASPP module).

[0030] S102, the MFAFNet network extracts low-level features, high-level features and multiple mid-level features of SAR images through the backbone network, and sends the low-level features to the decoder, the high-level features to the advanced multi-scale feature extraction module to extract advanced multi-scale features, and the multiple mid-level features to the intermediate-level feature fusion module to fuse them to obtain intermediate-level features. Finally, the decoder decodes the low-level features, multi-scale features and intermediate-level features to obtain the water body extraction result.

[0031] See Figure 1The MFAFNet network consists of an encoder and a decoder. The encoder finds patterns in the data and generates concise and useful representations. The decoder uses these representations to generate high-resolution data and new descriptive knowledge. Specifically, the backbone network performs initial feature extraction from the image; the intermediate feature fusion module fuses the intermediate features output from the backbone network to obtain intermediate features that enhance spatial detail representation; and the high-level multi-scale feature extraction module performs multi-scale feature extraction from the high-level features output from the backbone network. The decoding part contains three inputs: the backbone network extracts low-level features, high-level features, and multiple intermediate features from the SAR image, and the final water body extraction result is generated through the fusion and processing of these features.

[0032] In this embodiment, the backbone network is a ResNet_101 network with dilated convolutions (see K. He, X. Zhang, S. Ren and J. Sun, "Deep Residual Learning for Image Recognition," in IEEE Conf. Comp. Vis. Patt. Rec. (CVPR), Las Vegas, NV, USA, Jun. 2016, pp. 770-778.). Its second and third layers output two intermediate-level features, which are then fed into an intermediate-level feature fusion module to obtain intermediate-level features. Alternatively, other backbone networks can be selected as needed to achieve multi-level feature extraction. The ResNet_101 network, with its skip connections and optimized residuals, is well-suited for building semantic segmentation networks. First, the image resolution is reduced through a convolutional layer and a max-pooling layer. Then, preliminary semantic feature extraction is performed through four res-blocks composed of residual units. Among them, res4 uses a series of dilated convolutional layers with dilation rates of 2, 4 and 8 to replace the last convolutional layer, thereby increasing the number of network layers and enriching the semantic information of the feature maps.

[0033] In deep learning networks, fusing features at different scales is an important means to improve segmentation performance. Low-level features have higher resolution and contain more location and detail information, but due to fewer convolutions, their semantics are lower and they contain more noise. High-level features have stronger semantic information, but their resolution is very low, and their ability to perceive details is poor. How to efficiently fuse the two is the key to improving the segmentation model. In this embodiment, an intermediate-level feature fusion module (IFFM) is used to fully fuse the two intermediate-level features output by the ResNet-101 network to extract more detailed features of the water body. Figure 2As shown, multiple mid-level features are fed into the intermediate-level feature fusion module (IFFM module) to obtain intermediate-level features. The intermediate-level feature fusion module upsamples the lower-level mid-level features to match their channel count with the higher-level mid-level features. The upsampled feature map is then concatenated with the higher-level mid-level features after a 1×1 convolution operation. Finally, the concatenated feature map is concatenated with the feature map obtained after two 3×3 convolution operations to obtain the intermediate-level features. In this embodiment, the second-layer output of the ResNet-101 network is first upsampled by a factor of 2 to achieve 1024 channels to match the third-layer channel count. This is then followed by a 1×1 convolution and concatenated with the third-layer features after a 3×3 convolution, improving the network model's ability to represent SAR image targets. Finally, two 3×3 convolutions and branch feature fusion are performed to refine the target region boundaries before outputting to the decoder.

[0034] In this embodiment, the Advanced Multi-Scale Feature Extraction (ECASPP) module includes five parallel branches: four parallel attention pooling branches and one global average pooling (GAP) branch. Each of the four parallel attention pooling branches contains a serially connected convolutional module and an Efficient Channel Attention (ECA) module. The convolutional modules of the four parallel attention pooling branches have different scales and dilation rates. The feature maps output from the five parallel branches are stacked across channels and then subjected to a 1×1 convolution operation (1×1 Conv) to obtain advanced multi-scale features. In this embodiment, the input feature map of the ECASPP module contains 2048 channels and rich semantic information. The global average pooling branch downsamples the feature map to prevent overfitting. The four convolutions with different scales and dilation rates effectively capture the contextual information of the image from different receptive fields. ECA, combined with the channel dimension, performs quality selection on the multi-scale features, improving the multi-scale extraction effect and increasing the accuracy of water body detection. See also... Figure 1 The four convolutions with different scales and dilation rates are a 1×1 convolution operation (1X1Conv) and three 3×3 convolution operations with a dilation rate of 6 (3X3Conv rate=6).

[0035] In this embodiment, the Effective Attention Mechanism (ECA) module processes the input feature map by: performing global average pooling on the input feature map, then performing a 1D convolution operation with a kernel size k, and then passing it through a Sigmoid activation function to obtain the weights of each channel. The weights are then multiplied by the corresponding elements of the original input feature map to obtain the final output feature map. Attention mechanisms are widely used in deep learning tasks such as semantic segmentation and object detection. The ECA module is an existing convolutional neural network module, and its specific implementation can be found in: Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, Qinghua Hu; Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 11534-11542.

[0036] The Effective Attention (ECA) module first performs global average pooling on the input feature map, then performs a 1D convolution operation with a kernel size of k, and passes it through a Sigmoid activation function to obtain the weights for each channel. These weights are then multiplied by the corresponding elements of the original input feature map to obtain the final output feature map. One-dimensional convolution efficiently implements local cross-channel interactions to avoid dimensionality reduction and extract the dependencies between channels. The kernel size k determines the coverage of the interaction; different numbers of channels and different convolutional blocks may result in different interaction ranges, therefore k is considered to be related to the channel dimension C. Given the channel dimension C, the kernel size k is adaptively determined, and the calculation function expression for the kernel size k is:

[0037]

[0038] In the above formula, ψ(C) represents a function of the channel dimension C of the input, γ and b are constant parameters (which can be taken as needed; for example, in this embodiment, γ is 2 and b is 1), |t| odd This represents the odd number closest to t.

[0039] The decoder improves the continuity between pixels through step-by-step decoding; its internal structure is as follows: Figure 1 As shown. In this embodiment, the water extraction result obtained by decoding low-level features, multi-scale features, and intermediate-level features through a decoder includes:

[0040] S201, the low-level features are enhanced by attention modulation through the Attention Modulation Module (AMM) in the decoder to strengthen the weight of the water target. Then, the first part of the decoded features is obtained by reducing the channel redundancy features through a convolution operation with a kernel size of 1×1 (1X1Conv). The multi-scale features and intermediate-level features are concatenated and upsampled (in this embodiment, it is upsampled by 2) to obtain the fused features as the second part of the decoded features.

[0041] S202, after concatenating the first part of the decoded features and the second part of the decoded features, refines the features through a 3×3 convolution, and then performs upsampling to obtain the final water extraction result.

[0042] It should be noted that the Attention Modulation Module (AMM) is an existing convolutional neural network module. For its specific implementation, please refer to: Qin J, Wu J, Xiao X, et al. Activation modulation and recalibration scheme for weakly supervised semantic segmentation[C] / / Proceedings of the AAAIConference on Artificial Intelligence.2022,36(2):2117-2125. The decoder in this embodiment contains three inputs: first, the high-level multi-scale features output by the ECASPP module; second, the low-level features output by the backbone network; and third, the intermediate fusion features output by the intermediate feature fusion module. Since the low-level features contain a lot of redundant information, this paper introduces the Attention Modulation Module (AMM) to increase the boundary information weights in order to obtain more accurate water body edge information. The decoder first concatenates the two features output by the encoder and upsamples them by a factor of 2 to obtain a fused feature. The low-level features from ResNet-101 are then processed using an AMM attention mechanism to enhance the weight of the water target and reduced in dimensionality using a 1×1 convolution. This feature is then concatenated with the fused feature, and another 1×1 convolution is used to reduce a large amount of redundant features across channels. Finally, a 3×3 convolution is used to refine the features, followed by upsampling to obtain the final water extraction result.

[0043] As an optional implementation, the upsampling in step S202 of this embodiment refers to bilinear interpolation upsampling by 4 (denoted as Upsample by 4).

[0044] The following section will experimentally verify the multi-source SAR water extraction method based on MFAFNet in this embodiment. This embodiment uses over a dozen multi-band high-resolution SAR images, with water body labels manually added and confirmed by SAR experts. Data from three SAR systems were used in the experiment: Sentinel-1 data (C-band) with a resolution of 5m × 20m, Gaofen-3 data (C-band) with a resolution of 1m, and TerraSAR data (X-band) with a resolution of 3m. First, the large-scale SAR images were segmented using a sliding window method to generate 512 × 512 pixel images, ultimately generating 2895 samples, with a training set to validation set ratio of 4:1. In addition, one large-scale SAR image was reserved for independent testing for each SAR system. The experimental software environment in this embodiment is: PyTorch 1.20, CUDA 10.0, and Python 3.7; the hardware environment is: CPU Intel Xeon Gold 5120, GPU (single) NVIDIA RTX 2080Ti. During network training, the learning rate is set to 0.005, and the weight decay value is 0.0005. The batch size of the input images is 8, and the network is trained for 100 iterations. Finally, the epoch with the best training result is saved.

[0045] To better evaluate algorithm performance, this embodiment uses pixel accuracy (PA) and intersection over union (IoU) as evaluation metrics. PA is the percentage of correctly classified pixels divided by the total number of pixels in that class, i.e., the percentage of correctly classified pixels in the image. IoU is the intersection of the ground truth and predicted values ​​of the target divided by the union of the ground truth and predicted values ​​of the target. Its calculation function expression is shown in the following formula:

[0046]

[0047]

[0048] In the above formula, k represents the number of sample classes, with a total of k+1 classes, including one class as the background. P ii P represents the number of pixels that are correctly predicted for the target. ij and P jiThis represents the number of false positives and false negatives, i.e., samples that originally belonged to class i but were predicted as class j, and samples that originally belonged to class j but were predicted as class i. In this embodiment, the network model is trained simultaneously using datasets from three SAR systems, and then tested on images from different SAR systems. A large-scale SAR image is reserved for each system for independent testing. To verify the effectiveness of the method in this embodiment, the MFAFNet method is compared with several state-of-the-art classification networks, including Deeplabv3+, MF2AM, and SegFormer. The experimental results are as follows: Figure 3 (Data from Gaofen-3, image size 870×1384 pixels) Figure 4 (Sentinel 1 data, image size 867×1412 pixels) and Figure 5 (TerraSAR data, image size 839×1396 pixels) is shown. Figure 3 In the figure, (a) is a SAR image; (b) is a label; (c) to (f) are fusion diagrams of water body detection results and SAR images obtained by DeepLabV3+, MF2AM, SegFormer and MFAFNet, respectively. Figure 4 In the image, (a) is a SAR image; (b) is a label; (c) to (f) are fusion images of water body detection results and SAR images obtained by DeepLabV3+, MF2AM, SegFormer and MFAFNet, respectively. Figure 5 In the figure, (a) is a SAR image; (b) is a label; (c) to (f) are fusion diagrams of water body detection results and SAR images obtained by DeepLabV3+, MF2AM, SegFormer and MFAFNet, respectively.

[0049] Depend on Figure 3 The water detection results from the Gaofen-3 SAR image show that the SAR image mainly contains large, connected water areas with relatively clear water features, and the overall detection accuracy is high. Several algorithms detected most of the large water areas, while some false alarms and missed detections occurred in smaller water areas. DeepLabV3+ and MF2AM showed significant false alarms but relatively few missed detections; SegFormer had fewer false alarms but more missed detections; the MFAFNet network proposed in this embodiment significantly reduced false alarms compared to DeepLabV3+ and MF2AM, and also had significantly fewer missed detections than SegFormer, achieving very good detection results. Figure 4In the low-resolution Sentinel-1 SAR image of the mid-C band, the water body consists of a large area of ​​water adjacent to several smaller areas, as well as some scattered water regions. The water extraction results show that the water extraction results of the four networks have similar characteristics to those of the Gaofen-3 experiment. SegFormer, in particular, exhibited too many missed detections. The MFAFNet network proposed in this embodiment can detect water targets much better. Figure 5 The results are from a TerraSAR image experiment, where the water body consists of many regular small water bodies and several long, narrow rivers. Overall, the water extraction results of the four networks still retain the characteristics of Gaofen-3 and Sentinel-1. However, SegFormer not only had many missed detections but also a relatively large number of false alarms; the MFAFNet network proposed in this embodiment has the fewest missed detections and very few false alarms, indicating that the intermediate feature fusion module and the attention module proposed in this embodiment have a better grasp of the detailed features of the water body and have learned the water body features very well.

[0050] Depend on Figures 3-5 Experiments on water extraction from SAR images at different bands and resolutions show significant differences in image detail, as well as substantial variations in water body types and characteristics. Table 1 presents the water detection metrics for the four networks corresponding to the three experiments.

[0051] Table 1: Detailed comparison of the method in this embodiment with other methods.

[0052]

[0053]

[0054] As shown in Table 1, Gaofen-3 primarily detects large water areas, resulting in high overall detection accuracy, generally around 95%. Only the SegFormer network achieves around 90%, representing an improvement of nearly 3% compared to DeepLabV3+. Sentinel-1 data has lower resolution and contains non-connected water areas, leading to reduced detection accuracy. The SegFormer network achieves only around 25% accuracy, resulting in numerous missed detections. In contrast, the MFAFNet network in this embodiment achieves 84% ​​accuracy, significantly outperforming the other networks. TerraSAR images contain many small water areas, requiring high-level extraction of detailed water information, leading to a decrease in overall water detection accuracy. Other networks fail to reach 80%, while MFAFNet achieves 83.19% accuracy, a significant improvement over other networks and approximately 7% higher than DeepLabV3+. These metrics reveal that most networks can achieve good detection results for large-scale connected water bodies, but MFAFNet also demonstrates its advantages in detection performance. However, when dealing with many small, scattered water bodies or regular, narrow rivers, where stronger grasp of detailed and edge information is required, MFAFNet's advantages become fully apparent, and its detection accuracy is significantly improved compared to other networks.

[0055] To better verify the effectiveness of the MFAFNet network proposed in this embodiment, an ablation experiment was conducted. The average water detection accuracy of the four test models in three independent test areas obtained by each network is shown in Table 2.

[0056] Table 2: Ablation experiment.

[0057]

[0058] As shown in Table 2, using the original DeepLabV3+ network to detect SAR images of three different bands and resolutions, the average water body detection accuracy reaches 82.69%, and the IoU reaches 76.49%. In the original DeepLabV3+, replacing only the CAASP module improved accuracy by 2%; introducing only the AMM module improved accuracy by 2.5%; and adding only the IFFM module proposed in this embodiment improved accuracy by 3%. After adding all three modules (i.e., the network MFAFNet proposed in this embodiment), the average accuracy improved by 5.4%, with an average improvement of approximately 7% in the detection accuracy of complex water targets in Sentinel-1 and TerraSAR data, and an IoU improvement of nearly 4%. Therefore, the network proposed in this embodiment can effectively improve the water body detection accuracy of multi-band, multi-resolution SAR images, achieving satisfactory water body detection results and providing important guidance for the application of water body detection in multi-source SAR images.

[0059] In summary, the method presented in this embodiment demonstrates the best detection performance for water bodies, with the fewest false alarms and missed detections, thus verifying the superiority of this method in extracting water bodies. SegFormer exhibits the most missed detections and the most false alarms on TerraSAR data, while MF2AM shows the most false alarms on Gaofen-3 and Sentinel-1 satellites, indicating insufficient learning ability for detailed features. The network cannot effectively distinguish between different feature regions and their similar regions, resulting in low detection completeness and false alarms. In contrast, the intermediate-level feature fusion module proposed in this embodiment strengthens the transfer between features; the ECASPP module in the encoder discards redundant features while retaining useful features, improving the network's ability to learn water body features; and the introduction of AMM attention in the decoding module enhances the network's ability to grasp water body edge information. Therefore, MFAFNet achieves satisfactory water body extraction results.

[0060] In summary, SAR image water body detection has significant application value, and currently, single-band SAR images remain the primary data source. However, in applications, models trained or algorithms designed for SAR images of one frequency band often perform poorly in water body detection when applied to SAR images of different frequency bands and resolutions. To address this, this embodiment proposes a multi-level feature attention fusion network, MFAFNet, which enables water body extraction from multi-band, multi-resolution SAR images, significantly advancing the practical application value of deep learning networks in water body detection. The proposed intermediate-level feature fusion modules IFFM and ECASPP, along with the fused AMM attention, greatly enhance MFAFNet's ability to grasp detailed water body features and edge information, achieving significantly better water body detection performance than several other excellent classification networks.

[0061] Furthermore, this embodiment also provides a multi-source SAR water body extraction system based on MFAFNet, including a microprocessor and a memory interconnected thereon. The microprocessor is programmed or configured to execute the MFAFNet-based multi-source SAR water body extraction method. Additionally, this embodiment also provides a computer-readable storage medium storing a computer program for being programmed or configured by the microprocessor to execute the MFAFNet-based multi-source SAR water body extraction method.

[0062] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A multi-source SAR water body extraction method based on MFAFNet, characterized in that, include: S101, The input SAR image is input into the backbone network of the MFAFNet network. The MFAFNet network includes an encoder and a decoder. The encoder includes a backbone network, an intermediate-level feature fusion module, and an advanced multi-scale feature extraction module. S102, the MFAFNet network extracts low-level features, high-level features and multiple mid-level features of SAR images through the backbone network, and sends the low-level features to the decoder, the high-level features to the high-level multi-scale feature extraction module to extract high-level multi-scale features, and the multiple mid-level features to the intermediate-level feature fusion module to fuse them to obtain intermediate-level features. Finally, the decoder decodes the low-level features, multi-scale features and intermediate-level features to obtain the water body extraction result. The process of decoding low-level features, multi-scale features, and intermediate-level features to obtain water body extraction results includes: S201, attention modulation is performed on the low-level features through the attention modulation module (AMM) in the decoder to enhance the weight of the water target. Then, after reducing the channel redundancy features by using a convolution operation with a kernel size of 1×1, the first part of the decoded features is obtained. The multi-scale features and intermediate-level features are connected and upsampled to obtain the fused features as the second part of the decoded features. S202, after concatenating the first part of the decoded features and the second part of the decoded features, refines the features through a 3×3 convolution, and then performs upsampling to obtain the final water extraction result.

2. The multi-source SAR water body extraction method based on MFAFNet according to claim 1, characterized in that, The backbone network is a ResNet_101 network with dilated convolutions. Its second and third layers output two intermediate features, which are then fed into the intermediate feature fusion module to obtain intermediate features.

3. The multi-source SAR water body extraction method based on MFAFNet according to claim 1, characterized in that, The process of feeding multiple mid-level features into an intermediate-level feature fusion module to obtain intermediate-level features includes: the intermediate-level feature fusion module upsampling the lower-level mid-level features to match their channel count with that of the higher-level mid-level features; then concatenating the upsampled feature map with the higher-level mid-level features after performing a 1×1 convolution operation; and finally concatenating the concatenated feature map with the higher-level mid-level features after performing two 3×3 convolution operations.

4. The multi-source SAR water body extraction method based on MFAFNet according to claim 1, characterized in that, The advanced multi-scale feature extraction module includes five parallel branches, which include four parallel attention pooling branches and one global average pooling branch. Each of the four parallel attention pooling branches contains a serially connected convolutional module and an effective attention mechanism (ECA) module. The convolutional modules of the four parallel attention pooling branches have different scales and dilatation rates. The feature maps output by the five parallel branches are stacked on the channels and then subjected to a convolution operation with a kernel size of 1×1 to obtain advanced multi-scale features.

5. The multi-source SAR water body extraction method based on MFAFNet according to claim 4, characterized in that, The Effective Attention (ECA) module processes the input feature map by performing global average pooling, followed by a 1D convolution operation with a kernel size of k, and then using a Sigmoid activation function to obtain the weights of each channel. The weights are then multiplied by the corresponding elements of the original input feature map to obtain the final output feature map.

6. The multi-source SAR water body extraction method based on MFAFNet according to claim 5, characterized in that, The function expression for calculating the kernel size k is: In the above formula, A function representing the channel dimension C of the input. and b These are constant parameters, Indicates closest t Odd numbers.

7. The multi-source SAR water body extraction method based on MFAFNet according to claim 1, characterized in that, The upsampling in step S202 refers to bilinear interpolation upsampling by 4 times.

8. A multi-source SAR water body extraction system based on MFAFNet, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to perform the MFAFNet-based multi-source SAR water extraction method according to any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program, characterized in that, The computer program is used to be programmed or configured by a microprocessor to perform the MFAFNet-based multi-source SAR water extraction method according to any one of claims 1 to 7.

Images