A method and system for detecting internal ocean waves using synthetic aperture radar imagery

Abstract

The application discloses a kind of using synthetic aperture radar image detection ocean internal wave method and system, method includes: obtaining preprocessed synthetic aperture radar image, send into structure sampling module, and rearrange image spatial dimension information to channel dimension by slice, output initial feature map.Initial feature map is input into the structure continuity perception backbone network of multilevel cascade feature extraction stage, relies on window multi-head self-attention and cross-stage local connection structure, fuses deep and shallow features to obtain deep features.Again send into the multiscale context aggregation module containing multiple parallel maximum pooling branch, after channel splicing convolution is operated with different pooling kernel, obtain multiscale fusion features.Finally, input high-fidelity decoder of feature pyramid architecture, after depth separable convolution decoupling processing, generate ocean internal wave detection result.Provide reliable technical support for ocean internal wave detection.

Description

Technical Field

[0001] This invention belongs to the field of marine internal wave detection technology, and more specifically, relates to a method and system for detecting marine internal waves using synthetic aperture radar imagery. Background Technology

[0002] Synthetic Aperture Radar (SAR) has gained widespread attention for internal wave detection due to its advantages of all-weather, all-day imaging and high resolution. However, internal wave detection based on SAR images still faces many challenges. In SAR images, internal wave characteristics typically appear as narrow, low-contrast stripe patterns embedded in complex sea clutter, with wavelengths ranging from hundreds of meters to tens of kilometers, exhibiting significant cross-scale features. This elongated spatial morphology and non-uniform visibility are significantly affected by complex marine environmental conditions such as wind speed, sea state, and imaging geometry, further exacerbating the difficulty of internal wave detection and posing a significant challenge to accurate and robust internal wave detection.

[0003] Currently, research on monitoring internal waves over large ocean areas mainly falls into three categories: in-situ observation, traditional image processing methods such as wavelet transform and edge detection, and deep learning-based techniques such as target detection. In-situ observation using moored buoys or shipborne acoustic instruments can provide accurate physical observation data, but it is costly and not suitable for large-scale or real-time monitoring. Traditional image processing methods rely on manual design and are effective under controlled conditions, but they are highly sensitive to parameter adjustments and lack generalization ability. Deep learning, such as detectors based on Faster R-CNN and attention-enhanced structural networks such as IWE-Net, significantly improves the automation and efficiency of detection, but standard convolution operations and hierarchical feature aggregation often fail to preserve the elongated and coherent spatial structure of internal waves, leading to missed detections or inaccurate localization, especially when detecting weak and discontinuous wavebands. Summary of the Invention

[0004] This invention aims to address the challenges of weak, scale-variable, and spatially continuous ocean internal wave features in synthetic aperture radar (SAR) imagery, as well as the low efficiency of manual detection. It proposes an intelligent method for detecting ocean internal waves. Based on structure sampling, attention backbone networks, multi-scale context aggregation, and a high-fidelity decoding architecture, it integrates local and global features to achieve automated detection and identification of ocean internal waves.

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, as a first aspect of this invention, the present invention provides a method for detecting internal ocean waves using synthetic aperture radar imagery, comprising: S1. Acquire preprocessed synthetic aperture radar image, input the image into the structure sampling module, the structure sampling module rearranges the spatial dimension information of the image to the channel dimension through slicing operation, and outputs the rearranged data as the initial feature map; S2. Input the initial feature map into the structural continuity perception backbone network. The backbone network includes multiple cascaded feature extraction stages, wherein at least one feature extraction stage embeds a window-based multi-head self-attention module and a cross-stage local connection structure. The window-based multi-head self-attention module is used to perform local window partitioning and window shifting calculation on the input features. The cross-stage local connection structure is used to concatenate and fuse the calculated features with the shallow features of the identity mapping to output deep features. S3. Input the deep features into the multi-scale context aggregation module. The multi-scale context aggregation module includes multiple parallel max pooling branches. Each branch uses a pooling kernel of a different size to perform pooling operations on the deep features. The pooling results of each branch are concatenated and convolved with the input deep features in the channel dimension to output multi-scale fused features. S4. Input the multi-scale fused features into a multi-scale high-fidelity decoder. The decoder adopts a top-down feature pyramid structure and has a depth-separable convolutional layer at the feature fusion node. The depth-separable convolutional layer is used to decouple the fused features from spatial filtering and channel mixing. The detection result of ocean internal waves is generated based on the processed feature map.

[0006] Furthermore, the structure sampling module in S1 uses a focus-based operator for structure sampling; the slicing operation specifically involves dividing the input image X into four spatially offset sub-samples, connecting the four sub-samples along the channel dimension, and then generating a feature map of size H / 2×W / 2 through a convolution.

[0007] Furthermore, the deepest layer of the backbone network in S2 is equipped with a C3STR module to replace the standard C3 module; the C3STR module follows the CSP design concept and replaces the traditional convolution bottleneck in the main branch with the SwinTransformer module.

[0008] Furthermore, the processing procedure of the C3STR module specifically includes: dividing the input feature X into two parts, wherein the first part X1 is processed based on the bottleneck of SwinTransformer, the second part X2 is retained through identity mapping, and finally the processed first part and the second part are concatenated and fused.

[0009] Furthermore, the window-based multi-head self-attention module in S2 is specifically the SwinTransformer module embedded in the main branch of the C3STR module, and its processing includes: The input feature map is divided into multiple non-overlapping local windows, and multi-head self-attention features are computed independently within each local window; A window shift calculation is performed on the feature map to establish cross-window information interaction; wherein, the window shift calculation process specifically includes: Set window size to The displacement size is ; Translate left and up along the height and width dimensions of the feature map, respectively. Each pixel unit is used to cyclically fill the overflowing area of ​​the original feature map to the opposite side, forming a continuous feature map; Perform routine processing on the feature map after cyclic shifting. The window is divided to obtain several complete windows containing different original spatial regions; Construct an indexing matrix of the same size as the feature map, and assign unique sub-window numbers to different regions based on the continuity of the original spatial location; within each complete window after division, if the sub-window numbers corresponding to two pixels are different, set the value of their corresponding positions to negative infinity in the similarity matrix calculated by self-attention; if the numbers are the same, keep the original value. Multi-head self-attention features are calculated based on the masked similarity matrix, and the output feature map is then shifted to the right and downward along the height and width dimensions, respectively. Each pixel unit restores the original spatial arrangement structure.

[0010] Furthermore, the cross-stage local connection structure in S2 is specifically as follows: The input feature map is divided into a first sub-feature map and a second sub-feature map along the channel dimension; The first sub-feature map is input to the feature processing branch, which includes at least one cascaded window-based multi-head self-attention module for performing spatial feature transformation on the first sub-feature map to obtain the transformed first sub-feature map. The second sub-feature map is directly transmitted through the identity mapping branch; the transformed first sub-feature map and the second sub-feature map are concatenated in the channel dimension and then fused through a convolutional layer to obtain fused features; The fused features are residually concatenated with the input feature map to output the deep features.

[0011] Furthermore, the multi-scale context aggregation module in S3 employs the SPPCSPC module; the SPPCSPC module utilizes parallel pooling operations with kernels of various sizes to aggregate context cues within different spatial ranges, wherein the kernel sizes are... .

[0012] Furthermore, the depth-separable convolutional layer in S4 sequentially includes depthwise convolution (DWConv) and pointwise convolution (PWConv). First, spatial filtering is performed through depthwise convolution, and then channel mixing is performed through pointwise convolution.

[0013] As a second aspect of the invention, a system for detecting internal ocean waves using synthetic aperture radar imagery is also provided, comprising: The image slicing and spatial rearrangement unit is used to acquire preprocessed synthetic aperture radar images, input the images into the structure sampling module, and the structure sampling module rearranges the spatial dimension information of the images to the channel dimension through slicing operations, and outputs the rearranged data as the initial feature map; The backbone network feature extraction unit is used to input the initial feature map into the structural continuity perception backbone network. The backbone network includes multiple cascaded feature extraction stages, wherein at least one feature extraction stage embeds a window-based multi-head self-attention module and a cross-stage local connection structure. The window-based multi-head self-attention module is used to perform local window partitioning and window shifting calculation on the input features. The cross-stage local connection structure is used to concatenate and fuse the calculated features with the shallow features of the identity mapping to output deep features. A multi-scale context aggregation unit is used to input the deep features into a multi-scale context aggregation module. The multi-scale context aggregation module includes multiple parallel max pooling branches. Each branch uses a pooling kernel of a different size to perform pooling operations on the deep features. The pooling results of each branch are concatenated and convolved with the input deep features in the channel dimension to output multi-scale fused features. The decoding and detection unit is used to input the multi-scale fused features into a multi-scale high-fidelity decoder. The decoder adopts a top-down feature pyramid structure and has a depth-separable convolutional layer at the feature fusion node. The depth-separable convolutional layer is used to decouple the fused features from spatial filtering and channel mixing, and generates the detection result of ocean internal waves based on the processed feature map.

[0014] As a third aspect of the invention, a computer-readable storage medium is also provided, on which a computer program is stored, which is executed by a processor as described in any one of the claims: a method for detecting internal ocean waves using synthetic aperture radar imagery.

[0015] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. The method for detecting internal ocean waves using synthetic aperture radar (SAR) imagery of the present invention employs a slicing operation through a structured sampling module. This rearranges the spatial dimensional information of the SAR imagery to the channel dimension, dividing the input image into four spatially offset subsamples and stitching them along the channel dimension. Then, convolution is used to generate an initial feature map with half the size. This processing method, while achieving image downsampling and reducing the computational load on subsequent networks, fully preserves the original spatial details and texture features of internal ocean waves in the SAR imagery. It avoids the loss of detail information caused by conventional downsampling operations, standardizes the feature arrangement of image data, and provides a uniform and complete basic feature input for the subsequent feature extraction process of the network, ensuring that subsequent feature mining can be carried out systematically based on complete original image information.

[0016] 2. The method for detecting ocean internal waves using synthetic aperture radar (SAR) imagery of this invention constructs a structural continuity perception backbone network comprising multiple cascaded feature extraction stages. Within the feature extraction stage, a window-based multi-head self-attention module and a cross-stage local connectivity structure are embedded. A C3STR module replaces the standard convolutional module. Self-attention computation is performed using window partitioning, window shifting, and masking mechanisms. Simultaneously, features are split into channels, and feature transformations and identity mappings are applied separately before concatenation and fusion. This processing method can extract local window features of ocean internal waves, establish long-distance information interaction across windows, and fuse deep features with shallow original features based on the cross-stage local connectivity structure. This effectively maintains the continuity of the spatial structure of ocean internal waves, alleviates the feature loss problem during deep network propagation, and extracts deep features of ocean internal waves with strong semantic relevance.

[0017] 3. The method for detecting ocean internal waves using synthetic aperture radar imagery of the present invention configures a multi-scale context aggregation module with multiple parallel max-pooling branches. It performs multi-scale pooling operations on deep features using pooling kernels of different sizes. The pooling results of each branch are then combined with the original deep features through channel concatenation and convolution. A top-down feature pyramid structure multi-scale high-fidelity decoder is then introduced. Depth-separable convolutional layers are set at the feature fusion nodes to complete spatial filtering and channel mixing decoupling. This processing method can aggregate contextual information from different spatial ranges, adapting to the actual distribution of ocean internal waves with significant differences in scale and morphology. It compensates for the insufficient expressive power of single-scale features, simplifies the computational logic through depth-separable convolutional decoupling, restores the contour structure features of ocean internal waves, and outputs the ocean internal wave detection results based on the optimized feature map. Attached Figure Description

[0018] Figure 1This is a flowchart of a method for detecting internal ocean waves using synthetic aperture radar imagery according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a structural continuity sensing backbone network according to an embodiment of the present invention; Figure 3 This diagram illustrates a comparison of the performance of the method in this embodiment with other models. SCP-YOLO is the model used in this embodiment; YOLOv12 is a detection model based on a traditional convolutional neural network; KAJNConv2D is a detection method model combining a Kolmogorov-Arnold network with convolution; FastKAN is a detection model based on a fast Kolmogorov-Arnold network; and YoLOv8 is a mainstream object detection model. Figure 4 This is a schematic diagram of the system units in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0020] Example 1 Please refer to Figure 1 This embodiment 1 provides a method for detecting internal ocean waves using synthetic aperture radar imagery, comprising: S1. Acquire preprocessed synthetic aperture radar image, input the image into the structure sampling module, the structure sampling module rearranges the spatial dimension information of the image to the channel dimension through slicing operation, and outputs the rearranged data as the initial feature map; S2. Input the initial feature map into the structural continuity perception backbone network. The backbone network includes multiple cascaded feature extraction stages, wherein at least one feature extraction stage embeds a window-based multi-head self-attention module and a cross-stage local connection structure. The window-based multi-head self-attention module is used to perform local window partitioning and window shifting calculation on the input features. The cross-stage local connection structure is used to concatenate and fuse the calculated features with the shallow features of the identity mapping to output deep features. S3. Input the deep features into the multi-scale context aggregation module. The multi-scale context aggregation module includes multiple parallel max pooling branches. Each branch uses a pooling kernel of a different size to perform pooling operations on the deep features. The pooling results of each branch are concatenated and convolved with the input deep features in the channel dimension to output multi-scale fused features. S4. Input the multi-scale fused features into a multi-scale high-fidelity decoder. The decoder adopts a top-down feature pyramid structure and has a depth-separable convolutional layer at the feature fusion node. The depth-separable convolutional layer is used to decouple the fused features from spatial filtering and channel mixing. The detection result of ocean internal waves is generated based on the processed feature map.

[0021] Please refer to Figure 2 This embodiment 1 further elaborates on the above steps.

[0022] (1) Image slicing and spatial rearrangement Before using synthetic aperture radar imagery for ocean internal wave detection, preprocessed radar imagery data must first be acquired. These images exhibit high texture complexity and speckle noise. To reduce computational load while preserving key texture information, this embodiment employs a structured sampling module based on the Focus operator to construct the initial feature extraction layer. The spatial dimensions of the image are rearranged to the channel dimensions through slicing operations, and the rearranged data is output as the initial feature map. The specific process is as follows: First, input image The image is segmented spatially into four non-overlapping subsamples with offset spatial positions. These four subsamples correspond to four different phase pixel sets sampled at a stride of 2 on the spatial grid of the original image. Through this slicing method, the original spatial structure information is rearranged, increasing the number of channels to four times the original without changing the spatial size. ,in These are four non-overlapping subsamples with offset spatial positions. This process essentially "transfers" spatial semantic information to the channel dimension, greatly enriching the receptive field basis for subsequent convolutional operations.

[0023] Following the rearrangement of spatial information to the channel dimension, this embodiment introduces a standard convolution operation, combined with batch normalization and activation functions. Through the linear combination and nonlinear transformation of this convolutional layer, the model can extract preliminary edge and texture features from the rearranged high-dimensional channel data. Finally, a feature map of size H / 2 × W / 2 is output.

[0024] By employing a process of "slicing and reconstructing - channel stitching - convolutional dimensionality reduction," not only is the subtle structural information about internal ocean waves in SAR imagery preserved, but the computational burden on subsequent deep networks is also reduced. The output initial feature map serves as the input to the subsequent structural continuity perception backbone network, carrying rich structured semantics transformed from the original pixel space, thus laying a data foundation for detecting internal ocean waves.

[0025] (2) Backbone network feature extraction After the initial feature maps are output, they are fed into a structural continuity-aware backbone network for deep feature extraction. The backbone network architecture is built upon multiple cascaded feature extraction stages, with at least one stage embedding a window-based multi-head self-attention module and cross-stage local connectivity structures. Furthermore, the deepest layer of the backbone network introduces a C3STR module to replace the traditional C3 module. This C3STR module follows the CSP design philosophy, aiming to optimize gradient flow and reduce computational redundancy through a split-transform-fusion strategy.

[0026] First, the cross-stage local connection structure splits the overall input feature map along the channel dimension into two equal sub-feature streams: the first part (denoted as X1) is fed into the main branch for bottleneck processing based on the Swin Transformer, while the second part (denoted as X2) is directly transmitted through the identity mapping branch of the bypass, thereby preserving the original shallow texture and structural information and preventing the loss of details due to the network being too deep.

[0027] Within the main branch, this embodiment uses the SwinTransformer module as the core feature transformation bottleneck layer, replacing the standard convolutional layer in traditional convolutional neural networks. This module first divides the input feature map into multiple non-overlapping local windows, and independently computes multi-head self-attention features within each local window; Subsequently, window shift calculations are performed on the feature map to establish cross-window information interaction; wherein, the window shift calculation process specifically includes: First, set the window size to... The displacement size is ; Secondly, shift left and up along the height and width dimensions of the feature map, respectively. Each pixel unit is used to cyclically fill the overflowing area of ​​the original feature map to the opposite side, forming a continuous feature map; Next, the feature map after cyclic shifting is subjected to conventional processing. The window is divided to obtain several complete windows containing different original spatial regions; Then, an identification matrix of the same size as the feature map is constructed, and a unique sub-window number is assigned to different regions based on the continuity of the original spatial location. Within each complete window after division, if the sub-window numbers corresponding to two pixels are different, the value of their corresponding positions in the similarity matrix calculated by self-attention is set to negative infinity; if the numbers are the same, the original value is kept. Finally, multi-head self-attention features are calculated based on the masked similarity matrix, and the output feature map is shifted to the right and down along the height and width dimensions, respectively. Each pixel unit is used to restore the original spatial arrangement structure, resulting in the first part of the feature after window shifting. .

[0028] Subsequently, the first part of the features processed by the SwinTransformer and the second part of the features preserved through the identity mapping are concatenated along the channel dimension and then input into the subsequent convolutional layer for feature fusion. ,in This is the fused feature map. Finally, the fused feature map is residually connected to the initial input feature map of this module, outputting deep features with rich deep semantics and shallow details.

[0029] (3) Multi-scale context aggregation The deep features extracted by the backbone network are then fed into the multi-scale context aggregation module for feature enhancement and information aggregation. This module employs the SPPCSPC module, which is based on the design concept of cross-stage local networks. First, it reduces the dimensionality of the input deep feature map in the channel dimension using convolutional layers. Then, it splits the map into two independent feature branches: a main branch and a shortening branch. The shortening branch directly retains the original feature information for subsequent feature reconstruction, while the main branch is responsible for deep multi-scale context extraction.

[0030] In the main branch processing, this embodiment first preprocesses the features using a standard convolutional layer, and then constructs three parallel max pooling paths. These three paths are configured with square pooling kernels of different sizes, specifically set as follows: Specifically, the first parallel branch uses a 5×5 max-pooling kernel to extract small-scale local texture features; the second parallel branch uses a 9×9 max-pooling kernel to capture medium-scale morphological features; and the third parallel branch uses a 13×13 max-pooling kernel to aggregate large-scale background context information. All pooling operations employ a "same-size" padding strategy to ensure that the spatial resolution of the output feature map remains consistent with the input, only changing the semantic representation of the channel dimensions. ,in This is the output pooled feature map.

[0031] Next, the pooled feature maps output from the three parallel branches are concatenated with the original preprocessed feature map from the main branch (which has not undergone pooling) along the channel dimension. At this point, the number of channels in the feature map significantly increases, encompassing multi-scale information from local details to the global background. To reduce computational complexity and fuse these heterogeneous features, a 1×1 convolution operation is performed on the concatenated high-dimensional feature map to compress its channel count to a preset dimension, thus obtaining the output features of the main branch.

[0032] Finally, the multi-scale fused features output from the main branch are concatenated again with the previously retained shortened branch features along the channel dimension. To further smooth the features and eliminate the aliasing effect caused by concatenation, this embodiment uses a standard convolutional layer to perform final fusion processing on the recombined features, outputting multi-scale fused features with rich context-aware capabilities. The specific process is as follows: First, the standard convolutional layer is configured as a 3×3 convolutional kernel with a stride of 1 and padding of 1. This convolutional layer slides in the spatial domain, performing a weighted summation of the feature responses of each pixel and its 8-neighborhood in the reconstructed feature map. Utilizing the local receptive field of the convolutional kernel, the structural information preserved by the shortened branch is non-linearly mixed with the coarse-grained global context information extracted by the main branch at the pixel level.

[0033] After the convolution operation is completed, batch normalization and activation function operations are performed. Batch normalization is used to correct internal covariate offsets, ensuring the stability of the fused feature distribution and accelerating network convergence; the activation function introduces nonlinear factors into the fusion process, enabling the network to fit complex changes in the marine environment. After this series of processing steps, the number of channels in the feature map is compressed back to the preset baseline dimension, ultimately outputting multi-scale fused features with rich context awareness.

[0034] (4) Decoding detection The fused features obtained through multi-scale aggregation need to be restored in resolution and reconstructed in structure by a multi-scale high-fidelity decoder to complete target recognition. The multi-scale high-fidelity decoder employs a top-down feature pyramid structure, aiming to progressively transfer and fuse deep, strong semantic features into shallow, high-resolution features. The multi-scale high-fidelity decoder first upsamples the deepest feature map (e.g., by using nearest-neighbor interpolation to double the feature map size) to ensure its spatial resolution matches that of the corresponding shallow feature map in the backbone network. Then, the upsampled deep features are concatenated with the shallow features along the channel dimension or added element-wise to construct a fused feature map containing both semantic and localization information.

[0035] To address the fusion features generated at the aforementioned fusion nodes, this embodiment introduces a depthwise separable convolutional layer to replace the traditional computationally expensive standard convolution. This depthwise separable convolutional layer is used to decouple spatial filtering and channel mixing of the fusion features. The specific process is as follows: First, spatial filtering is performed using depthwise convolution (DWConv). A spatial dimension decoupling strategy is employed, configuring a set of single-channel convolutional kernels (e.g., 3×3 kernels) with a number of channels strictly equal to the input feature map. Each convolutional kernel operates on only one specific channel of the input feature map, i.e., the [missing channel]. The convolutional kernel only applies to the first... The feature maps of each channel are used for sliding window calculation. This operation independently extracts spatial correlations within each channel, focusing on capturing the local geometry, edge orientation, and texture details of ocean internal waves in SAR imagery, without involving information interaction between different channels. Through this spatial filtering method, this embodiment significantly reduces computational redundancy while preserving high-resolution spatial information.

[0036] Next, channel blending is performed using pointwise convolution PWConv. A 1×1 convolutional kernel is used, its core function being to perform channel-dimensional decoupling, i.e., channel blending. In this step, the 1×1 convolutional kernel performs a weighted summation and linear combination of all channels at each spatial pixel location in the feature map. This process integrates and reorganizes the spatial features of each independent channel extracted from the preceding depthwise convolutional layers across channels, non-linearly fusing semantic information scattered across different channels. ,in This refers to the fusion feature after decoupling.

[0037] Through the cascaded operations of depthwise convolution and pointwise convolution described above, this embodiment achieves decoupling of the fused features, preserving the high fidelity of the spatial dimension structure while completing the deep fusion of features in the channel dimension. Finally, the feature map after multi-level processing by the decoder is fed into the detection head to generate the bounding box coordinates and class confidence of the internal ocean waves, thus completing the internal ocean wave detection task.

[0038] Figure 3 This paper presents a visualization comparison of the proposed method model SCP-YOLO with existing mainstream deep learning models in the task of detecting internal ocean waves in SAR imagery. As can be seen from the figure, SCP-YOLO (… Figure 3a) It demonstrates significant advantages in various complex marine scenarios: In low-contrast regions (such as sample A), SCP-YOLO can accurately identify and completely delineate the slender internal wave structure, while other models (YOLOv12, KAJNConv2D, FastKAN, and YOLOv8) generally suffer from missed detections or incomplete bounding boxes; In scenarios where internal waves coexist at multiple scales (such as samples B and C), SCP-YOLO can simultaneously detect internal wave features at different scales with accurate bounding box positioning, while the comparison models often only detect internal waves at some scales or produce positioning offsets; Under complex sea state interference (such as samples D and E), SCP-YOLO effectively suppresses environmental noise interference through its structural continuity sensing backbone network, maintaining the continuity of internal wave fringes, while other models are prone to false detections or discontinuous detection results. This performance advantage directly verifies the effectiveness of the "structural sampling", "structural continuity sensing backbone network" and "multi-scale context aggregation" described in this patent, and fully demonstrates the significant advantages of SCP-YOLO in maintaining the integrity of the slender internal wave structure, adapting to multi-scale changes and resisting environmental interference.

[0039] Example 2 Please refer to Figure 4 This embodiment 2 provides a system for detecting internal ocean waves using synthetic aperture radar imagery, comprising: The image slicing and spatial rearrangement unit is used to acquire preprocessed synthetic aperture radar images, input the images into the structure sampling module, and the structure sampling module rearranges the spatial dimension information of the images to the channel dimension through slicing operations, and outputs the rearranged data as the initial feature map; The backbone network feature extraction unit is used to input the initial feature map into the structural continuity perception backbone network. The backbone network includes multiple cascaded feature extraction stages, wherein at least one feature extraction stage embeds a window-based multi-head self-attention module and a cross-stage local connection structure. The window-based multi-head self-attention module is used to perform local window partitioning and window shifting calculation on the input features. The cross-stage local connection structure is used to concatenate and fuse the calculated features with the shallow features of the identity mapping to output deep features. A multi-scale context aggregation unit is used to input the deep features into a multi-scale context aggregation module. The multi-scale context aggregation module includes multiple parallel max pooling branches. Each branch uses a pooling kernel of a different size to perform pooling operations on the deep features. The pooling results of each branch are concatenated and convolved with the input deep features in the channel dimension to output multi-scale fused features. The decoding and detection unit is used to input the multi-scale fused features into a multi-scale high-fidelity decoder. The decoder adopts a top-down feature pyramid structure and has a depth-separable convolutional layer at the feature fusion node. The depth-separable convolutional layer is used to decouple the fused features from spatial filtering and channel mixing, and generates the detection result of ocean internal waves based on the processed feature map.

[0040] Example 3 This embodiment 3 also provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement any step of a method for detecting internal ocean waves using synthetic aperture radar imagery.

[0041] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0042] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.

[0043] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting internal ocean waves using synthetic aperture radar imagery, characterized in that, include: S1. Acquire preprocessed synthetic aperture radar image, input the image into the structure sampling module, the structure sampling module rearranges the spatial dimension information of the image to the channel dimension through slicing operation, and outputs the rearranged data as the initial feature map; S2. Input the initial feature map into the structural continuity perception backbone network. The backbone network includes multiple cascaded feature extraction stages, wherein at least one feature extraction stage embeds a window-based multi-head self-attention module and a cross-stage local connection structure. The window-based multi-head self-attention module is used to perform local window partitioning and window shifting calculation on the input features. The cross-stage local connection structure is used to concatenate and fuse the calculated features with the shallow features of the identity mapping to output deep features. S3. Input the deep features into the multi-scale context aggregation module. The multi-scale context aggregation module includes multiple parallel max pooling branches. Each branch uses a pooling kernel of a different size to perform pooling operations on the deep features. The pooling results of each branch are concatenated and convolved with the input deep features in the channel dimension to output multi-scale fused features. S4. Input the multi-scale fused features into a multi-scale high-fidelity decoder. The decoder adopts a top-down feature pyramid structure and has a depth-separable convolutional layer at the feature fusion node. The depth-separable convolutional layer is used to decouple the fused features from spatial filtering and channel mixing. The detection result of ocean internal waves is generated based on the processed feature map.

2. The method for detecting internal ocean waves using synthetic aperture radar imagery according to claim 1, characterized in that, The structure sampling module in S1 uses a focus-based operator for structure sampling; the slicing operation specifically involves dividing the input image X into four spatially offset sub-samples, connecting the four sub-samples along the channel dimension, and then performing a convolution to generate a feature map of size H / 2×W / 2.

3. The method for detecting internal ocean waves using synthetic aperture radar imagery according to claim 1, characterized in that, The deepest layer of the backbone network in S2 is equipped with a C3STR module to replace the standard C3 module. The C3STR module follows the CSP design concept and replaces the traditional convolution bottleneck in the main branch with the Swing Transformer module.

4. A method for detecting internal ocean waves using synthetic aperture radar imagery according to claim 3, characterized in that, The C3STR module's processing procedure specifically includes: dividing the input feature X into two parts, where the first part X1 is processed based on the bottleneck of the Swing Transformer, and the second part X2 is preserved through identity mapping. Finally, the processed first part and the second part are concatenated and fused.

5. A method for detecting internal ocean waves using synthetic aperture radar imagery according to claim 1, characterized in that, The window-based multi-head self-attention module in S2 is specifically the SwinTransformer module embedded in the main branch of the C3STR module, and its processing includes: The input feature map is divided into multiple non-overlapping local windows, and multi-head self-attention features are computed independently within each local window; A window shift calculation is performed on the feature map to establish cross-window information interaction; wherein, the window shift calculation process specifically includes: Set window size to The displacement dimension is ; Translate left and up along the height and width dimensions of the feature map, respectively. Each pixel unit is used to cyclically fill the overflowing area of ​​the original feature map to the opposite side, forming a continuous feature map; Perform routine processing on the feature map after cyclic shifting. The window is divided to obtain several complete windows containing different original spatial regions; Construct an indexing matrix of the same size as the feature map, and assign unique sub-window numbers to different regions based on the continuity of the original spatial location; within each complete window after division, if the sub-window numbers corresponding to two pixels are different, set the value of their corresponding positions to negative infinity in the similarity matrix calculated by self-attention; if the numbers are the same, keep the original value. Multi-head self-attention features are calculated based on the masked similarity matrix, and the output feature map is then shifted to the right and downward along the height and width dimensions, respectively. Each pixel unit restores the original spatial arrangement structure.

6. A method for detecting internal ocean waves using synthetic aperture radar imagery according to claim 1, characterized in that, The cross-stage local connection structure in S2 is specifically as follows: The input feature map is divided into a first sub-feature map and a second sub-feature map along the channel dimension; The first sub-feature map is input to the feature processing branch, which includes at least one cascaded window-based multi-head self-attention module for performing spatial feature transformation on the first sub-feature map to obtain the transformed first sub-feature map. The second sub-feature map is directly transmitted through the identity mapping branch; the transformed first sub-feature map and the second sub-feature map are concatenated in the channel dimension and then fused through a convolutional layer to obtain fused features; The fused features are residually concatenated with the input feature map to output the deep features.

7. A method for detecting internal ocean waves using synthetic aperture radar imagery according to claim 1, characterized in that, The multi-scale context aggregation module in S3 employs the SPPCSPC module; the SPPCSPC module utilizes parallel pooling operations with kernels of various sizes to aggregate context clues within different spatial ranges, wherein the kernel sizes are... .

8. A method for detecting internal ocean waves using synthetic aperture radar imagery according to claim 1, characterized in that, The depth-separable convolutional layer in S4 includes depthwise convolution (DWConv) and pointwise convolution (PWConv) in sequence. First, spatial filtering is performed by depthwise convolution, and then channel mixing is performed by pointwise convolution.

9. A system for detecting internal ocean waves using synthetic aperture radar imagery, characterized in that, include: The image slicing and spatial rearrangement unit is used to acquire preprocessed synthetic aperture radar images, input the images into the structure sampling module, and the structure sampling module rearranges the spatial dimension information of the images to the channel dimension through slicing operations, and outputs the rearranged data as the initial feature map; The backbone network feature extraction unit is used to input the initial feature map into the structural continuity perception backbone network. The backbone network includes multiple cascaded feature extraction stages, wherein at least one feature extraction stage embeds a window-based multi-head self-attention module and a cross-stage local connection structure. The window-based multi-head self-attention module is used to perform local window partitioning and window shifting calculation on the input features. The cross-stage local connection structure is used to concatenate and fuse the calculated features with the shallow features of the identity mapping to output deep features. A multi-scale context aggregation unit is used to input the deep features into a multi-scale context aggregation module. The multi-scale context aggregation module includes multiple parallel max pooling branches. Each branch uses a pooling kernel of a different size to perform pooling operations on the deep features. The pooling results of each branch are concatenated and convolved with the input deep features in the channel dimension to output multi-scale fused features. The decoding and detection unit is used to input the multi-scale fused features into a multi-scale high-fidelity decoder. The decoder adopts a top-down feature pyramid structure and has a depth-separable convolutional layer at the feature fusion node. The depth-separable convolutional layer is used to decouple the fused features from spatial filtering and channel mixing, and generates the detection result of ocean internal waves based on the processed feature map.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor according to any one of claims 1-8, a method for detecting internal ocean waves using synthetic aperture radar imagery.

Images