SAR sea ice segmentation method, device and equipment based on multi-scale feature extraction fusion

Abstract

The invention provides an SAR sea ice segmentation method, device and equipment based on multi-scale feature extraction fusion. The SAR sea ice segmentation method comprises the following steps: in an SAR sea ice image data imaging process, performing incident angle correction on SAR sea ice image data to obtain an SAR image product of sea ice; performing multi-scale characteristic decomposition and reconstruction on the SAR image product by adopting a deep learning network embedded with a wavelet transformation module, and performing channel enhancement in the deep learning network by adopting a residual channel attention mechanism to obtain the SAR image product after speckle noise suppression; and based on a sea ice segmentation model integrated with a lightweight adaptive extraction module and a multi-scale feature extraction module, segmenting the SAR image product after speckle noise suppression to obtain a sea ice segmentation result.

Description

Technical Field

[0001] The present invention relates to fields related to artificial intelligence, SAR image processing, and SAR image segmentation, and in particular to a SAR sea ice segmentation method, device, and electronic equipment based on multi-scale feature extraction and fusion. Background Art

[0002] Synthetic Aperture Radar (SAR) is a key remote sensing tool for sea ice monitoring. Its ability to penetrate clouds allows it to acquire high-resolution imagery under a wide range of meteorological conditions. Changes in the distribution and types of sea ice have profound implications for Arctic shipping safety, global climate change, and the ecological environment. Accurately identifying and classifying sea ice types (e.g., new ice, old ice, open water, etc.) is not only crucial for marine resource development, shipping safety, and disaster prevention, but also provides important data support for climate change research. Sea ice classification requires segmentation, which can be used to determine the classification results. Traditional sea ice segmentation methods rely primarily on the echo intensity of SAR images or machine learning algorithms based on manual feature extraction, but these methods have limited performance in complex environments. In recent years, the introduction of deep learning technology has significantly improved the accuracy of sea ice segmentation, demonstrating strong adaptability and reliability, particularly under complex meteorological and oceanographic conditions. Summary of the Invention

[0003] In view of this, the present invention provides a SAR sea ice segmentation method, device and electronic equipment based on multi-scale feature extraction and fusion.

[0004] According to a first aspect of the present invention, a SAR sea ice segmentation method based on multi-scale feature extraction and fusion is provided, comprising: during the SAR sea ice image data imaging process, performing incident angle correction on the SAR sea ice image data to obtain a SAR image product of sea ice; using a deep learning network embedded with a wavelet transform module to perform multi-scale feature decomposition and reconstruction on the SAR image product, and using a residual channel attention mechanism in the deep learning network to perform channel enhancement to obtain a SAR image product after speckle noise suppression; based on a sea ice segmentation model integrated with a lightweight adaptive extraction module and a multi-scale feature extraction module, segmenting the SAR image product after speckle noise suppression to obtain a sea ice segmentation result.

[0005] According to an embodiment of the present invention, the step of performing incident angle correction on SAR sea ice image data includes: parsing radar parameters from a scene element file, determining a correction coefficient based on the radar parameters and the geometric relationship between the slant range and the incident angle; and uniformly correcting the backscatter coefficient of the SAR sea ice image data at different incident angles to a theoretical value at a reference incident angle based on the correction coefficient.

[0006] According to an embodiment of the present invention, the step of determining a correction coefficient based on radar parameters and a geometric relationship between slant range and incident angle includes: calculating the center slant range and range resolution based on the radar parameters; calculating the incident angle corresponding to each column based on the center slant range and range resolution; taking the beam center incident angle as a reference incident angle, and determining the correction coefficient based on the beam center incident angle and the incident angle corresponding to each column.

[0007] According to an embodiment of the present invention, the steps of performing multi-scale feature decomposition and reconstruction on a SAR image product using a deep learning network embedded with a wavelet transform module include: inputting the SAR image product into the deep learning network, extracting deep features of the SAR image product by alternately using a first convolutional layer and a discrete wavelet transform module; and alternately using an inverse wavelet transform module and a second convolutional layer to fuse and reconstruct the deep features obtained by decomposition, thereby obtaining original structural information that retains the deep features.

[0008] According to an embodiment of the present invention, the step of using the residual channel attention mechanism to perform channel enhancement in a deep learning network includes: using the residual channel attention mechanism to perform channel enhancement on each feature map extracted by the first convolutional layer, and transmitting the channel-enhanced feature map to the corresponding second convolutional layer.

[0009] According to an embodiment of the present invention, the residual channel attention mechanism includes a 1×1 convolutional layer and a depth convolutional layer, the 1×1 convolutional layer and the depth convolutional layer are connected, and the depth convolutional layer is used to perform cross-channel local interaction on the features after dimensionality reduction of the 1×1 convolutional layer.

[0010] According to an embodiment of the present invention, the step of segmenting the SAR image product after speckle noise suppression to obtain a sea ice segmentation result includes: performing feature extraction, feature fusion, multi-scale information processing and high-resolution reconstruction on the SAR image product after speckle noise suppression in sequence to determine the sea ice segmentation result.

[0011] According to an embodiment of the present invention, the steps of sequentially performing feature extraction, feature fusion, multi-scale information processing, and high-resolution reconstruction on a SAR image product after speckle noise suppression include: extracting an initial multi-scale feature map through a convolutional layer of a sea ice segmentation model; downsampling and feature enhancement of each initial multi-scale feature map using a lightweight adaptive extraction module, and fusing feature maps of different levels through jump connections; capturing detail information from the fused feature map using a multi-scale feature extraction module, and fusing it with shallow features to complete multi-scale information processing of the features; and upsampling the features after multi-scale information processing to the original resolution to obtain a segmentation result.

[0012] The second aspect of the present invention provides a SAR sea ice segmentation device based on multi-scale feature extraction and fusion, including: a correction module, used to correct the incident angle of SAR sea ice image data during the imaging process of SAR sea ice image data, so as to obtain a SAR image product of sea ice; a denoising module, used to use a deep learning network embedded with a wavelet transform module to perform multi-scale feature decomposition and reconstruction on the SAR image product, and use the residual channel attention mechanism in the deep learning network to perform channel enhancement, so as to obtain a SAR image product after speckle noise suppression; a segmentation module, used to segment the SAR image product after speckle noise suppression based on a sea ice segmentation model integrated with a lightweight adaptive extraction module and a multi-scale feature extraction module, so as to obtain a segmentation result of sea ice.

[0013] The third aspect of the present invention provides an electronic device comprising: one or more processors; a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors execute the method in any one of the above embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The above contents and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings, in which:

[0015] Figure 1 The flowchart of the SAR sea ice segmentation method based on multi-scale feature extraction and fusion according to one embodiment of the present invention is schematically shown;

[0016] Figure 2 The flowchart of the SAR sea ice segmentation method based on multi-scale feature extraction and fusion according to another embodiment of the present invention is schematically shown;

[0017] Figure 3 Schematically shows a flow chart of incident angle correction according to an embodiment of the present invention;

[0018] Figure 4 The overall structure of the deep learning network according to the embodiment of the present invention is schematically shown;

[0019] Figure 5 Schematically shows a structural diagram of a residual channel attention mechanism module of a deep learning network according to an embodiment of the present invention;

[0020] Figure 6 The overall structure of the sea ice segmentation model according to an embodiment of the present invention is schematically shown;

[0021] Figure 7 Schematically shows a structural diagram of a lightweight adaptive extraction module of a sea ice segmentation model according to an embodiment of the present invention;

[0022] Figure 8 A schematic diagram illustrating a structure of a multi-scale feature extraction module of a sea ice segmentation model according to an embodiment of the present invention is shown;

[0023] Figure 9 The following schematically shows a structural block diagram of a SAR sea ice segmentation device based on multi-scale feature extraction and fusion according to an embodiment of the present invention;

[0024] Figure 10 A block diagram of an electronic device suitable for implementing a SAR sea ice segmentation method based on multi-scale feature extraction and fusion according to an embodiment of the present invention is schematically shown. DETAILED DESCRIPTION

[0025] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present invention. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of embodiments of the present invention. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concept of the present invention.

[0026] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise," "include," etc. used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0027] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0028] When expressions such as "at least one of A, B, and C, etc." are used, they should generally be interpreted in accordance with the meaning commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0029] The applicant found in the research that the existing SAR sea ice segmentation methods still face some challenges.

[0030] (1) Changes in the incident angle lead to different scattering intensities in different areas of the image, which in turn affects the distinguishability of different sea ice types. For example, at low incident angles, multi-year ice exhibits higher scattering due to its rough surface, while at high incident angles, the volume scattering caused by its internal bubble layer is more significant. The incident angle correction method based on visual interpretation and then fitting a linear model is difficult to directly apply to an unlabeled SAR image.

[0031] (2) Coherent speckle noise in SAR images significantly affects the accuracy of sea ice segmentation. This granular noise, caused by radar signal interference, blurs the edges of sea ice and distorts texture features, causing abnormal fluctuations in radar backscatter values ​​for the same type of sea ice. Traditional denoising methods are prone to losing edge and texture details; adding a multi-look step to the product processing process will reduce resolution while suppressing speckle noise.

[0032] (3) Segmentation accuracy remains limited in complex scenarios, especially for differentiating sea ice types with similar echo characteristics. Because SAR images are affected by noise, imaging conditions, and the physical properties of sea ice, some sea ice types have very similar echo characteristics, making it difficult for the model to accurately distinguish them. Multi-scale targets in complex environments and the boundary between sea ice and open water further increase the difficulty of segmentation.

[0033] (4) The generalization ability of the model still needs to be improved. There are significant differences in the performance of SAR sea ice images in different regions or different categories. Due to the differences in geographical environment, climate conditions and imaging parameters, the performance of the model on cross-regional data is unstable. In addition, the characteristic distribution of different types of sea ice in SAR images also varies greatly, resulting in poor segmentation effect of the model for some categories.

[0034] In view of this, the present invention proposes a SAR sea ice segmentation method based on multi-scale feature extraction and fusion. The SAR sea ice segmentation method based on multi-scale feature extraction and fusion is further explained below with reference to the accompanying drawings.

[0035] Figure 1 The flowchart of the SAR sea ice segmentation method based on multi-scale feature extraction and fusion according to one embodiment of the present invention is schematically shown.

[0036] like Figure 1 As shown, the SAR sea ice segmentation method based on multi-scale feature extraction and fusion may include steps S110 to S130.

[0037] In step S110 , during the SAR sea ice image data imaging process, the SAR sea ice image data is subjected to incident angle correction to obtain a SAR image product of the sea ice.

[0038] In step S120, a deep learning network embedded with a wavelet transform module is used to perform multi-scale feature decomposition and reconstruction on the SAR image product, and a residual channel attention mechanism is used in the deep learning network to perform channel enhancement to obtain a SAR image product after speckle noise suppression.

[0039] In step S130 , based on the sea ice segmentation model integrated with the lightweight adaptive extraction module and the multi-scale feature extraction module, the SAR image product after speckle noise suppression is segmented to obtain a sea ice segmentation result.

[0040] The SAR sea ice segmentation method based on multi-scale feature extraction and fusion provided in this embodiment can be divided into an incident angle correction stage (i.e., step S110), a noise suppression stage (i.e., step S120), and a sea ice segmentation stage (i.e., step S130). In the incident angle correction stage, incident angle correction is performed during the SAR product imaging process, meaning that imaging and correction are completed simultaneously. This reduces error propagation and meets the requirements of automated processing, eliminating the need for manual intervention and improving processing efficiency and adaptability. In the noise suppression stage, the wavelet transform effectively suppresses SAR speckle noise while preserving details through multi-scale decomposition and adaptive thresholding. The introduction of the Residual Channel Attention Aggregation (RCAA) mechanism enables cross-scale information aggregation at multiple levels, making features at different depths more discriminative when fused. A deep learning network combined with the wavelet transform can automatically learn the optimal wavelet coefficient processing method, avoiding the limitations of traditional methods that rely on manually set thresholds or basis functions, resulting in more accurate feature extraction and reconstruction. In the sea ice segmentation stage, to address the problem of insufficient segmentation accuracy in complex scenarios, this embodiment introduces a Lightweight Adaptive Extraction (LAE) module and integrates it into the feature extraction network in the form of a residual network. By retaining more contextual information and high-resolution details on multi-scale feature maps, the impact of noise is reduced while retaining rich semantic information, thereby improving the ability to distinguish sea ice categories in complex scenarios. To address the problem of insufficient model generalization ability, a Multi-Scale Feature Extraction (MSFE) module is introduced. Through a dual aggregation mechanism of space and channels, it improves feature expression ability, replaces traditional skip connections, and better supports multi-level information fusion of the decoder, thereby enhancing the model's generalization ability in SAR sea ice images of different regions and categories.

[0041] In some embodiments, the incident angle correction stage (i.e., step S110) may include: parsing radar parameters from the scene element file, determining a correction coefficient based on the radar parameters and the geometric relationship between the slant range and the incident angle; and uniformly correcting the backscatter coefficients of the SAR sea ice image data at different incident angles to the theoretical value at the reference incident angle based on the correction coefficient. Figure 2 The flowchart of the SAR sea ice segmentation method based on multi-scale feature extraction and fusion according to another embodiment of the present invention is schematically shown. Figure 3 The flowchart of incident angle correction according to an embodiment of the present invention is schematically shown.

[0042] like Figure 2 As shown, in this embodiment, the SAR sea ice segmentation method based on multi-scale feature extraction and fusion can begin product processing with zero-level data (also known as L0-level raw data). Before generating a standard product, imaging algorithm processing is performed and the SAR sea ice image data is corrected for incident angle to obtain a SAR image product. This is followed by noise suppression and sea ice segmentation. These noise suppression and sea ice segmentation steps will be further explained below.

[0043] like Figure 3 As shown, the steps of determining the correction coefficient based on the radar parameters and the geometric relationship between the slant range and the incident angle may include: calculating the center slant range and the range resolution based on the radar parameters; calculating the incident angle corresponding to each column based on the center slant range and the range resolution; taking the beam center incident angle as a reference incident angle, and determining the correction coefficient based on the beam center incident angle and the incident angle corresponding to each column.

[0044] Specifically, the incident angle correction step may include the following:

[0045] (1) Read the scene element file and parse the radar parameters. For example, the data such as the beam center incident angle, bandwidth, sampling rate, pulse delay number, pulse repetition frequency, pulse width, etc. can be obtained.

[0046] (2) Calculate the center slope distance , where dpnum is the pulse delay number, PRF is the pulse repetition frequency, d is the pulse delay time, τ is the pulse width, and c is the speed of light.

[0047] (3) Calculation of distance resolution , where B w is the signal bandwidth.

[0048] (4) Calculate the incident angle corresponding to each column

[0049] ,

[0050] where R iis the slope distance corresponding to the i-th point, N is the number of distance points, , H is the platform height, which can be calculated from the GPS position information in the auxiliary data.

[0051] (5) Incident angle correction (the input is the image data after imaging). The attenuation of radar echo intensity in the range direction can be obtained from the geometric relationship:

[0052]

[0053] Where σ is the backscatter coefficient, which represents the scattering ability of the ground object to the radar signal, and θ is the incident angle, that is, the angle between the radar beam and the normal line of the target surface. rec Indicates received power, which refers to the target scattered signal power received by the radar system. eff It represents the effective receiving area, reflecting the radar receiving antenna's ability to collect echo signals. P0 represents the transmission power, that is, the initial power of the electromagnetic wave emitted by the radar. G(θ) represents the antenna gain. σ target represents the target scattering cross section, which characterizes the target's scattering efficiency against radar signals. According to the above formula, as the incident angle increases, the backscatter intensity decreases. Therefore, during the production of SAR products, the beam center incident angle can be selected as the reference angle for correction. The correction formula can be as follows:

[0054]

[0055] where θ ref is the reference angle, is the correction coefficient, σ 0 (θ i ) is the original backscattering coefficient (corresponding to the incident angle θ i ).

[0056] In the incident angle correction stage, this embodiment adopts an incident angle correction method for the SAR imaging process based on physical model embedding. By directly integrating radiation and geometric correction into the imaging algorithm, the correction coefficient is dynamically calculated using radar system parameters and real-time geometric relationships (such as slant range, viewing angle, and digital elevation model). This breaks away from the traditional method's reliance on manual visual interpretation and statistical fitting, and achieves universal modeling of the relationship between sea ice backscatter intensity and incident angle, significantly improving processing efficiency and adaptability.

[0057] Please continue reading Figure 1In some embodiments, the step of performing multi-scale feature decomposition and reconstruction on a SAR image product using a deep learning network embedded with a wavelet transform module may include: inputting the SAR image product into the deep learning network, extracting deep features of the SAR image product by alternately using a first convolutional layer and a discrete wavelet transform module; and alternately using an inverse wavelet transform module and a second convolutional layer to fuse and reconstruct the decomposed deep features to obtain original structural information that retains the deep features. Figure 4 The overall structure of a deep learning network according to an embodiment of the present invention is schematically shown. Figure 5 The structure diagram of the residual channel attention mechanism module of the deep learning network according to an embodiment of the present invention is schematically shown.

[0058] like Figure 4 As shown, the deep learning network of this embodiment can be composed of a backbone network and a residual channel attention mechanism module. By combining wavelet transform and channel attention mechanism, the deep learning network can reduce the coherent speckle noise of SAR images, enhance feature extraction capabilities, and improve the processing effect of remote sensing images.

[0059] First, the incident angle-corrected SAR image product can be fed into a CR module (Conv2d convolution + ReLU activation function) for preliminary feature extraction. Subsequently, a discrete wavelet transform (DWT) can be used to decompose the features at multiple scales to obtain information at different frequencies. Deep learning networks can alternate between the CR module and the DWT at different stages to continuously extract deep features while enhancing the ability to represent features at different scales. The CR module, used alternately with the DWT, can be referred to as the first convolutional layer.

[0060] During the feature fusion and reconstruction process, the deep learning network reconstructs the features decomposed from the DWT using an inverse wavelet transform (IWT), restoring the original structural information while preserving the enhanced depth features. By alternating between CR and IWT multiple times, the network maintains the advantages of decomposed features while retaining important information from the original image. The CR module, used alternately with IWT, can be called the second convolutional layer.

[0061] Continue reading Figure 4 , the residual channel attention mechanism (RCAA) can be used to perform channel enhancement on each feature map extracted by the first convolution layer, and the channel-enhanced feature map is transmitted to the corresponding second convolution layer. RCAA can include a 1×1 convolution layer and a depthwise convolution layer (DWConv). The 1×1 convolution layer and the depthwise convolution layer are connected, and the depthwise convolution layer is used to perform cross-channel local interaction on the features after dimensionality reduction in the 1×1 convolution layer. Figure 5As shown, in some embodiments, each RCAA module can consist of global average pooling, 1×1 convolution, depthwise convolution, and a sigmoid activation function. By calculating attention weights between channels, the RCAA module enhances key features while suppressing irrelevant information, thereby improving the network's expressiveness and robustness. Multiple RCAA modules operate in parallel on feature maps at different stages, further optimizing information flow and ensuring the network focuses on important feature areas.

[0062] Continue reading Figure 4 , further processed by a multi-layer perceptron (MLP) to produce the final output image. Overall, this deep learning network, by combining wavelet transforms, deep convolutions, and an attention mechanism, improves the ability to recover details from remote sensing images while maintaining feature extraction capabilities. This makes it suitable for tasks such as super-resolution, denoising, and object detection in remote sensing images.

[0063] Compared to traditional denoising methods, this embodiment combines wavelet transforms (including DWT and IWT) for multi-scale feature modeling. DWT decomposes the input image into subbands of different frequencies, preserving more high-frequency detail, while IWT gradually restores the global structure during reconstruction. This approach automatically learns the optimal wavelet coefficient processing method, avoiding the limitations of traditional methods that rely on manually set thresholds or basis functions, making the network more accurate in feature extraction and reconstruction.

[0064] This embodiment also improves channel-level feature expression by introducing RCAA. The deep learning network integrates RCAA modules in multiple stages, calculates attention weights through global average pooling, 1×1 convolution, depth convolution and Sigmoid function, and dynamically adjusts the channel importance of the feature map. Unlike the traditional channel attention mechanism, RCAA can aggregate cross-scale information at multiple levels, making features of different depths more discriminative when fused. This mechanism enhances the expressive power of key features while reducing the interference of irrelevant information, enabling the network to more accurately extract important areas in remote sensing images.

[0065] Figure 6 The overall structure of the sea ice segmentation model according to an embodiment of the present invention is schematically shown.

[0066] like Figure 6 As shown in Figure 1, the sea ice segmentation model can use a deep learning architecture based on multi-scale feature extraction and gradual reconstruction, which can be used for tasks such as super-resolution, denoising, and remote sensing image processing. The processing of the sea ice segmentation model can include four stages: feature extraction, feature fusion, multi-scale information processing, and high-resolution reconstruction.

[0067] During the feature extraction phase, for example, an a×a input image can be processed through multiple CRP (Conv2d convolution + ReLU activation function + Pooling) modules to gradually extract deeper features while reducing the resolution. For example, images can be gradually downsampled to a / 2×a / 2, a / 4×a / 4, and a / 8×a / 8. This gradual downsampling approach not only extracts higher-level semantic information but also reduces computational effort, improving network processing efficiency.

[0068] During the feature fusion and multi-scale information processing stages, to address the issue of insufficient sea ice classification differentiation in complex scenarios, a Lightweight Adaptive Extraction (LAE) module is integrated into the sea ice segmentation model network of this embodiment in the form of a residual network. This module retains more contextual information and high-resolution details in the multi-scale feature maps, reducing the impact of noise while preserving rich semantic information. To improve model generalization, this embodiment also introduces a Multi-Scale Feature Extraction (MSFE) module. This module enhances feature expression through dual spatial and channel aggregation, replacing traditional skip connections to better support multi-level information fusion in the decoder.

[0069] After completing multi-scale feature extraction, the sea ice segmentation model network enters the gradual reconstruction phase, utilizing the CRU (Conv2d convolution + ReLU activation function + upsampling) module. The CRU module combines convolutional layers for feature extraction, ReLU activation functions for nonlinear mapping, and upsampling modules for gradual resolution restoration, enabling the network to recover high-resolution images from low-dimensional features. Furthermore, the network utilizes multiple Feature Fusion Modules (FFMs) during the reconstruction process to fuse features from different stages. These modules can incorporate channel attention or feature weighting mechanisms, allowing information from different stages to complement each other more effectively and improve the final restoration quality. The CRU and FFM modules can be used alternately to ensure efficient information flow while features are gradually amplified.

[0070] In summary, if Figure 6As shown, the sea ice segmentation model can use a×a images as input to extract multi-scale feature maps, for example, it can include three resolutions of a / 2×a / 2, a / 4×a / 4, and a / 8×a / 8. Each feature map is downsampled and feature enhanced by the LAE module, and features at different levels are fused through jump connections. The fused features are first optimized by the MSFE module to capture finer detail information, and then fused with shallow features by the FFM module. Finally, the processed features are upsampled to the original resolution to generate the final segmentation results or other task outputs. It should be noted that the specific values ​​of the above resolutions are only examples and do not constitute a specific limitation of the present invention. The specific structures of the LAE and MSFE modules will be combined Figure 7 and Figure 8 Provide explanation.

[0071] Figure 7 The structure diagram of the lightweight adaptive extraction module of the sea ice segmentation model according to an embodiment of the present invention is schematically shown.

[0072] like Figure 7 As shown, the lightweight adaptive extraction (LAE) module may include a lightweight extraction branch and an adaptive extraction branch.

[0073] The lightweight branch can use grouped convolution for downsampling, reducing the model's computational overhead. Grouped convolution can reduce the number of parameters and computation required for the convolution operation by dividing the input channels into multiple groups for independent processing, while still effectively preserving the spatial information in the image, thereby reducing memory usage and accelerating the model's computational process. At the same time, the lightweight branch retains the key spatial information of the image, allowing the network to extract important local features while maintaining high efficiency.

[0074] The adaptive branch captures global context through average pooling and convolution operations, further enhancing the model's spatial understanding. Average pooling allows the module to extract global information from the image, while convolutions allow for deeper feature abstraction. The attention weights generated by these operations are applied to the features extracted by the lightweight branch and weightedly fused through element-by-element multiplication. This weighted fusion process effectively enhances the model's focus on important regions, allowing it to focus more on identifying key targets in segmentation tasks while ignoring less important background information. The fused output feature map is combined with the original input features via residual connections, further enhancing information flow and preventing information loss within the network. This ensures smoother gradient propagation during model training. This design not only helps prevent network overfitting but also improves model convergence speed and training stability.

[0075] Figure 8The structure of the multi-scale feature extraction module of the sea ice segmentation model according to an embodiment of the present invention is schematically shown.

[0076] like Figure 8 As shown in Figure 2, the Multi-Scale Attention Aggregation (MSFE) module can include spatial attention and channel attention mechanisms. Through the spatial and channel attention mechanisms, the feature extraction and context understanding capabilities in image segmentation tasks are significantly improved.

[0077] The spatial attention mechanism enables the model to focus on key areas, enhancing the extraction of details and important parts, while the channel attention mechanism optimizes feature selection, helping the model focus on more discriminative channel features. By aggregating multi-scale information, MSFE can handle objects of different scales, improving the segmentation of image details and complex backgrounds. Furthermore, MSFE can enhance the model's generalization capabilities. By fusing multi-scale features, the MSFE module enables the sea ice segmentation model to learn richer and more diverse feature representations, reducing overfitting and improving adaptability in different scenarios. This enables the MSFE module to not only improve segmentation accuracy but also enhance the model's robustness, strengthening its performance in new datasets and diverse environments, and greatly enhancing the model's generalization capabilities and applicability.

[0078] Based on the above embodiments, comparative experimental results of the sea ice segmentation method of the present invention and other common segmentation methods (e.g., segmentation methods based on U-Net, UKAN, DeeplabV3+, TransUnet, TransFuse, MobilenetV3, BisNet, DDRNet, and PIDNet models) can be shown in Table 1. In Table 1, PA represents pixel accuracy, mPA represents mean pixel accuracy, and mIoU represents mean intersection over union. As can be seen from Table 1, among the listed segmentation methods, the sea ice segmentation method of the present invention achieves the highest mIoU of 84.028 and mPA of 91.346. The results in Table 1 demonstrate that the sea ice segmentation method of the present invention effectively addresses the challenges of sea ice segmentation in SAR imagery, achieving significant improvements over existing methods.

[0079] Table 1

[0080]

[0081] Based on the above embodiments, the present invention has at least the following advantages compared with the prior art:

[0082] (1) Automatic correction of the incident angle. The incident angle correction method for the SAR imaging process proposed in this paper realizes full-process automation and can meet large-scale, near-real-time operational monitoring needs (such as polar sea ice monitoring) without human intervention. This method is driven by a physical model and effectively avoids the errors caused by traditional manual segmentation and statistical fitting. It has higher accuracy and robustness and can adapt to multi-platform and multi-mode SAR data. At the same time, by synchronizing the correction process with the imaging algorithm, the processing efficiency is significantly improved, providing standardized high-quality data input for subsequent AI analysis.

[0083] (2) Effectively suppress speckle noise while preserving texture details. This invention combines wavelet transform (DWT / IWT) for multi-scale feature modeling. DWT can decompose the input image into sub-bands of different frequencies, thereby preserving more high-frequency detail information, while IWT can gradually restore the global structure during the reconstruction process. This method automatically learns the optimal wavelet coefficient processing method, avoiding the limitations of traditional methods that rely on manually set thresholds or basis functions, making the network more accurate in extracting and reconstructing features.

[0084] (3) Higher segmentation accuracy in complex scenes. By introducing the LAE module and the MSFE module, the present invention can retain more contextual information and high-resolution details on multi-scale feature maps, significantly reducing the impact of noise while enhancing the ability to express semantic information. This makes the model perform better in complex scenes and significantly improves segmentation accuracy.

[0085] (4) The model has stronger generalization ability. This paper designs the MSFE module, which significantly improves the feature expression ability through a dual aggregation mechanism of space and channel. The MSFE module replaces the traditional skip connection and better supports the multi-level information fusion of the decoder, making the model more stable in SAR sea ice imagery of different regions and categories, and significantly enhancing the generalization ability.

[0086] Based on the above-mentioned SAR sea ice segmentation method based on multi-scale feature extraction and fusion, the present invention also provides a SAR sea ice segmentation device based on multi-scale feature extraction and fusion. Figure 9 The device is described in detail.

[0087] Figure 9 The structural block diagram of the SAR sea ice segmentation device based on multi-scale feature extraction and fusion according to an embodiment of the present invention is schematically shown.

[0088] like Figure 9 As shown, the SAR sea ice segmentation device 900 based on multi-scale feature extraction and fusion of this embodiment includes a correction module 910, a denoising module 920 and a segmentation module 930.

[0089] The correction module 910 can be used to perform incident angle correction on the SAR sea ice image data during the SAR sea ice image data imaging process to obtain a SAR image product of sea ice. In one embodiment, the correction module 910 can be used to perform the operation S110 described above, which will not be repeated here.

[0090] Denoising module 920 can be used to perform multi-scale feature decomposition and reconstruction on the SAR image product using a deep learning network embedded with a wavelet transform module. It can also perform channel enhancement using a residual channel attention mechanism within the deep learning network to obtain a SAR image product with speckle noise suppressed. In one embodiment, denoising module 920 can be used to perform operation S120 described above, which will not be further described here.

[0091] Segmentation module 930 can be used to segment the speckle noise suppressed SAR image product based on a sea ice segmentation model that integrates a lightweight adaptive extraction module and a multi-scale feature extraction module, thereby obtaining a sea ice segmentation result. In one embodiment, segmentation module 930 can be used to perform operation S130 described above, which will not be further described here.

[0092] Please refer to the previous article for details of the relevant content, which will not be repeated here.

[0093] According to embodiments of the present invention, any multiple of the above-mentioned modules may be combined into a single module, or any one of them may be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules may be combined with at least part of the functionality of other modules and implemented in a single module. According to embodiments of the present invention, at least one of the above-mentioned modules may be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or may be implemented in hardware or firmware through any other reasonable means of circuit integration or packaging, or may be implemented in any one of the three implementation methods of software, hardware, and firmware, or in any appropriate combination of any of these. Alternatively, at least one of the above-mentioned modules may be at least partially implemented as a computer program module that, when executed, performs the corresponding functionality.

[0094] Figure 10 A block diagram of an electronic device suitable for implementing a SAR sea ice segmentation method based on multi-scale feature extraction and fusion according to an embodiment of the present invention is schematically shown.

[0095] like Figure 10As shown, an electronic device 1000 according to an embodiment of the present invention includes a processor 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage unit 1008 into a random access memory (RAM) 1003. Processor 1001 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or a related chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. Processor 1001 may also include onboard memory for caching purposes. Processor 1001 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0096] RAM 1003 stores various programs and data required for the operation of electronic device 1000. Processor 1001, ROM 1002, and RAM 1003 are interconnected via bus 1004. Processor 1001 executes the programs in ROM 1002 and / or RAM 1003 to perform various operations according to the method flow of the embodiment of the present invention. It should be noted that the programs may also be stored in one or more memories other than ROM 1002 and RAM 1003. Processor 1001 may also execute the programs stored in one or more memories to perform various operations according to the method flow of the embodiment of the present invention.

[0097] According to an embodiment of the present invention, electronic device 1000 may further include an input / output (I / O) interface 1005, which is also connected to bus 1004. Electronic device 1000 may also include one or more of the following components connected to I / O interface 1005: an input section 1006 including a keyboard, mouse, etc.; an output section 1007 including devices such as a cathode ray tube (CRT), liquid crystal display (LCD), and speakers; a storage section 1008 including a hard disk; and a communication section 1009 including a network interface card such as a LAN card or modem. Communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to I / O interface 1005 as needed. Removable media 1011, such as a magnetic disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed in drive 1010 as needed, so that computer programs read from the removable media can be installed into storage section 1008 as needed.

[0098] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments, or may exist independently and not incorporated into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of the present invention.

[0099] According to an embodiment of the present invention, a computer-readable storage medium may be a non-volatile computer-readable storage medium, and may include, for example, but not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to an embodiment of the present invention, a computer-readable storage medium may include ROM 1002 and / or RAM 1003 described above, and / or one or more memories other than ROM 1002 and RAM 1003.

[0100] Embodiments of the present invention also include a computer program product, comprising a computer program containing program code for executing the method shown in the flowchart. When the computer program product is executed in a computer system, the program code causes the computer system to implement the SAR sea ice segmentation method based on multi-scale feature extraction and fusion provided in an embodiment of the present invention.

[0101] The computer program executes the above functions defined in the system / device of the embodiment of the present invention when executed by the processor 1001. According to the embodiment of the present invention, the system, device, module, unit, etc. described above can be implemented by a computer program module.

[0102] In one embodiment, the computer program may be stored on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may be transmitted and distributed in the form of a signal on a network medium, downloaded and installed via the communication portion 1009, and / or installed from the removable medium 1011. The program code contained in the computer program may be transmitted using any appropriate network medium, including but not limited to wireless, wired, or any suitable combination thereof.

[0103] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 1009 and / or installed from the removable medium 1011. When the computer program is executed by the processor 1001, the above-described functions defined in the system of the embodiment of the present invention are performed. According to the embodiment of the present invention, the systems, devices, means, modules, units, etc. described above can be implemented by computer program modules.

[0104] According to an embodiment of the present invention, the program code for executing the computer program provided by the embodiment of the present invention can be written in any combination of one or more programming languages. Specifically, these computer programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C" or similar programming languages. The program code can be executed entirely on the user computing device, partially on the user device, partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computing device (for example, using an Internet service provider to connect via the Internet).

[0105] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, program segment, or a part of code, and the above-mentioned module, program segment, or a part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0106] It will be understood by those skilled in the art that the features described in the various embodiments of the present invention may be combined and / or coupled in various ways, even if such combinations or couplings are not explicitly described in the present invention. In particular, the features described in the various embodiments of the present invention may be combined and / or coupled in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or couplings fall within the scope of the present invention.

[0107] The above describes embodiments of the present invention. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention. Although each embodiment has been described separately above, this does not mean that the measures in each embodiment cannot be advantageously used in combination. Without departing from the scope of the present invention, those skilled in the art may make various substitutions and modifications, which should all fall within the scope of the present invention.

Claims

1. A SAR sea ice segmentation method based on multi-scale feature extraction and fusion, characterized in that: include: During the SAR sea ice image data imaging process, the SAR sea ice image data is subjected to incident angle correction to obtain a SAR image product of the sea ice; A deep learning network embedded with a wavelet transform module is used to perform multi-scale feature decomposition and reconstruction on the SAR image product, and a residual channel attention mechanism is used in the deep learning network to perform channel enhancement to obtain a SAR image product after speckle noise suppression; Based on a sea ice segmentation model that integrates a lightweight adaptive extraction module and a multi-scale feature extraction module, the SAR image product after speckle noise suppression is segmented to obtain the sea ice segmentation result.

2. The SAR sea ice segmentation method according to claim 1, characterized in that: The step of performing incident angle correction on the SAR sea ice image data comprises: parsing radar parameters from the scene element file, and determining correction coefficients based on the radar parameters and the geometric relationship between slant range and incident angle; According to the correction coefficient, the backscattering coefficients of the SAR sea ice image data at different incident angles are uniformly corrected to the theoretical values ​​at the reference incident angle.

3. The SAR sea ice segmentation method according to claim 2, characterized in that: The step of determining the correction coefficient according to the radar parameters and the geometric relationship between the slant range and the incident angle includes: Calculating the center slant range and range resolution based on the radar parameters; Calculating the incident angle corresponding to each column according to the central slant distance and the range resolution; The beam center incident angle is used as the reference incident angle, and the correction coefficient is determined according to the beam center incident angle and the incident angle corresponding to each column.

4. The SAR sea ice segmentation method according to claim 1, characterized in that: The step of performing multi-scale feature decomposition and reconstruction on the SAR image product using a deep learning network embedded with a wavelet transform module includes: Inputting the SAR image product into the deep learning network, and extracting deep features of the SAR image product by alternately using the first convolutional layer and the discrete wavelet transform module; The inverse wavelet transform module and the second convolution layer are alternately used to fuse and reconstruct the deep features obtained by decomposition to obtain the original structural information that retains the deep features.

5. The SAR sea ice segmentation method according to claim 4, characterized in that: The step of using the residual channel attention mechanism to perform channel enhancement in the deep learning network includes: The residual channel attention mechanism is used to perform channel enhancement on each feature map extracted by the first convolutional layer, and the channel-enhanced feature map is transmitted to the corresponding second convolutional layer.

6. The SAR sea ice segmentation method according to claim 1 or 5, characterized in that: The residual channel attention mechanism includes a 1×1 convolutional layer and a depth convolutional layer, the 1×1 convolutional layer and the depth convolutional layer are connected, and the depth convolutional layer is used to perform cross-channel local interaction on the features after dimensionality reduction of the 1×1 convolutional layer.

7. The SAR sea ice segmentation method according to claim 1, characterized in that: The step of segmenting the SAR image product after speckle noise suppression to obtain the sea ice segmentation result includes: The SAR image product after speckle noise suppression is subjected to feature extraction, feature fusion, multi-scale information processing and high-resolution reconstruction in sequence to determine the segmentation result of the sea ice.

8. The SAR sea ice segmentation method according to claim 7, characterized in that: The steps of sequentially performing feature extraction, feature fusion, multi-scale information processing and high-resolution reconstruction on the SAR image product after speckle noise suppression include: Extracting an initial multi-scale feature map through the convolutional layer of the sea ice segmentation model; Using the lightweight adaptive extraction module to downsample and enhance the features of each of the initial multi-scale feature maps, and fusing feature maps of different levels through skip connections; The multi-scale feature extraction module is used to capture detail information from the fused feature map and fuse it with shallow features to complete multi-scale information processing of features; The features after multi-scale information processing are upsampled to the original resolution to obtain the segmentation result.

9. A SAR sea ice segmentation device based on multi-scale feature extraction and fusion, characterized in that: include: a correction module, configured to perform incident angle correction on the SAR sea ice image data during the imaging process of the SAR sea ice image data to obtain a SAR image product of the sea ice; a denoising module for performing multi-scale feature decomposition and reconstruction on the SAR image product using a deep learning network embedded with a wavelet transform module, and performing channel enhancement using a residual channel attention mechanism in the deep learning network to obtain a SAR image product after speckle noise suppression; A segmentation module is used to segment the SAR image product after speckle noise suppression based on a sea ice segmentation model integrated with a lightweight adaptive extraction module and a multi-scale feature extraction module to obtain a segmentation result of the sea ice.

10. An electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors are enabled to perform the method according to any one of claims 1 to 8.

Images