Marine medium / sub-mesoscale dynamic process identification method based on space-based SAR (Synthetic Aperture Radar)

Abstract

The invention discloses a space-based synthetic aperture radar (SAR)-based ocean medium / sub-mesoscale dynamic process identification method, which belongs to the technical field of ocean remote sensing and artificial intelligence crossing, and comprises the following steps of: inputting data, preprocessing the data, constructing a data set, constructing a model, improving by adopting a YOLOv12 architecture, constructing an edge enhanced perceptron, and fusing low, medium and high-level features; the Neck layer uses edge enhanced ER-MLP and CBAM modules to improve the large target identification capability; the scale sensitivity problem is solved through multi-scale slicing, an overlapping ratio and a confidence coefficient threshold value are matched, and a hierarchical scanning strategy is used for anchor frame combination; and coordinate transformation is realized through geographic affine transformation, geographic information data are output, and the detection precision and practicability are improved. According to the method, the detection precision of the ocean dynamic target is improved, the problems of scale sensitivity and redundancy are solved, accurate geographic positioning and area calculation are realized, and the practicability and reliability of the model are enhanced.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of marine remote sensing and artificial intelligence, and specifically relates to a method for identifying meso / sub-mesoscale dynamic processes in the ocean based on space-based SAR. Background Technology

[0002] Currently, eddies and internal waves, as ubiquitous dynamic processes in the ocean, play important roles in energy transfer, material transport, and biogeochemical cycles. Eddies typically manifest as rotating water flow structures with spatial scales ranging from a few kilometers to hundreds of kilometers, exhibiting significant kinetic and momentum transport characteristics; internal waves, on the other hand, are waves propagating within the density stratification of the ocean, influencing ocean mixing, energy transfer, and material transport. Traditionally, the identification of these phenomena relies primarily on threshold segmentation methods using in-situ observations, numerical simulations, or satellite remote sensing data (such as detecting eddies from sea surface height anomalies and identifying internal waves using band features of synthetic aperture radar (SAR)). These methods suffer from two core problems: (1) difficulty in handling complex interferences in the marine environment, such as wind, waves, ship tracks, and oil spills; and (2) a lack of effective adaptability to multi-scale, weak feature signals, especially when the target scale varies greatly or the target shape is complex, resulting in a significant decrease in identification accuracy. However, these methods usually require manual parameter setting and have limited adaptability to complex marine environments, especially when multi-scale features coexist or when noise interference occurs, leading to a significant decrease in extraction accuracy.

[0003] In recent years, deep learning technology, especially target detection technology, has made significant progress in the field of image recognition. Its high automation, high precision, and strong generalization capabilities have shown great potential in marine dynamic feature extraction. Applying deep learning technology to automatically identify marine dynamic processes can not only improve recognition efficiency and accuracy but also provide richer data support for marine scientific research, revealing the spatiotemporal distribution characteristics and evolution patterns of eddies and internal waves. This will help to better understand marine dynamic processes, predict changes in the marine environment, and provide scientific basis for marine resource development, marine environmental protection, and marine disaster early warning. However, applying deep learning technology to marine dynamic feature extraction still faces many key challenges. These challenges stem from both the complexity of marine dynamic processes themselves and the limitations of existing technological frameworks. Specifically, marine dynamic data (such as SAR, infrared, or altimeter data) have significant high-dimensional nonlinear characteristics, and their signal-to-noise ratio is generally lower than that of land remote sensing data. For example, mesoscale eddies in altimeter data only manifest as sea surface height anomalies of -7 to 15 cm, while the backscattering intensity variation of internal waves in SAR images typically does not exceed 3 dB. Such weak signal characteristics require models with stronger feature decoupling capabilities, but existing CNN architectures are easily affected by marine environmental noise (such as sea surface wind fields and ship wakes) when dealing with these weak features. Furthermore, the amplification of thermal noise and its simultaneous additive and multiplicative noise characteristics significantly increase the complexity of target detection. Secondly, vortices and internal waves often exhibit coupling phenomena in the spatiotemporal domain. Observations in the South China Sea show that internal wave packets and vortex boundary layers frequently interact, leading to an increased false detection rate for features. Although attention mechanisms (such as the CBAM module) have been attempted for feature decoupling, their processing efficiency across cross-scale dynamic ranges (such as simultaneously capturing kilometer-scale vortices and ten-meter-scale internal wave ripples) still needs improvement.

[0004] In summary, traditional methods rely on manual parameter setting, resulting in poor adaptability; existing deep learning models lack sufficient feature decoupling capability for weak ocean signals (such as internal wave backscattering intensity variations ≤3dB); when multiple scale targets (kilometer-scale vortices and ten-meter-scale internal waves) coexist, it is difficult to balance model efficiency and accuracy; and the high computational complexity makes it difficult to embed into spaceborne platforms for real-time processing. It is evident that vortices and internal waves exhibit high complexity and diversity in their spatiotemporal distribution, morphological characteristics, and dynamic behavior, making it difficult for traditional identification methods to achieve efficient and accurate unified detection. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a space-based SAR-based method for identifying oceanic meso / sub-mesoscale dynamic processes. After inputting data, preprocessing data, and labeling datasets, a YOLOv12 architecture is employed. The backbone network uses 7×7 convolutions and a region attention mechanism, and an edge-enhanced perceptron is designed to fuse low, medium, and high-level features. The Neck layer uses edge-enhanced ER-MLP and CBAM modules to improve large target recognition. A multi-scale slicing strategy addresses scale sensitivity by setting three slicing sizes and assigning overlap rate and confidence thresholds. Anchor frame merging uses a hierarchical scanning strategy to dynamically control the overlap threshold. Geographic affine transformation achieves coordinate transformation through six parameters, and the area is calculated using the Karney algorithm and Gaussian integrals to output geographic information data, improving detection accuracy and practicality.

[0006] The technical solution adopted in this invention:

[0007] A method for identifying meso / sub-mesoscale dynamic processes in the ocean based on space-based SAR includes the following steps:

[0008] Step 1, Preprocessing data: Input SAR data containing internal waves or vortices for preprocessing. Preprocessing includes radiometric correction, thermal noise removal, land masking, and geocoding.

[0009] Step 2, construct the dataset: Divide the preprocessed SAR data into training and validation sets in a 9:1 ratio, and augment the data by flipping and rotating.

[0010] Step 3, Model Construction: A model is constructed based on a lightweight neural network architecture, including improvements to the backbone network layers and optimizations to the Neck layer. The specific process is as follows:

[0011] a. Improve the backbone network layer: Improve the YOLO architecture and build an edge enhancement perceptron to obtain the transformed feature maps;

[0012] b. Optimize the Neck layer: Replace the traditional FPN with the ER-MLP layer to improve the ability to recognize large targets;

[0013] Step 4, Train the model: Use the training set as data samples to train the model;

[0014] Step 5, Multi-scale slice detection: Pyramid slices are made on the images of data samples in the trained model. The overlap rate η_k and confidence threshold conf_k are set. After multi-scale detection, the anchor boxes are merged to obtain the detection results, including the pixel coordinate system.

[0015] Step 6, output geographic information data through geographic affine transformation: map the pixel coordinate system to the geographic coordinate system through geographic affine transformation, and output the center position and area of ​​the vortex / inner wave.

[0016] Preferably, in step S3a, the specific process of improving the YOLO architecture is as follows:

[0017] Based on the YOLO architecture, a new target detection network is constructed to achieve effective detection of multi-morphological targets in complex marine environments: In the backbone network layer, the YOLOv12 structure is adopted, and the initial layer uses 7×7 convolutions to extract the initial feature map; then, combined with the region attention mechanism module unique to the YOLOv12 structure, the attention to multi-morphological targets in marine dynamic processes is enhanced.

[0018] Preferably, in step S3a, the specific process of constructing the edge enhancement perceptron to obtain the changed feature map is as follows:

[0019] A learnable offset is introduced into the receptive field of deformable convolution, allowing the convolution kernel to dynamically adjust the position of sampling points to adapt to the shape of multi-morphological targets. The initial feature map is further processed by deformable convolution to obtain feature map x. i Subsequently, the feature map x was processed using MLP. i Cross-layer interaction and fusion are performed, organically combining high-resolution detail information from the low layer, local structural information from the middle layer, and global semantic information from the high layer into the MLP to generate multi-scale, multi-level feature representations, namely the multi-scale fusion layer; finally, spatial information fusion and channel transformation are performed on the feature representations through separable convolution and 1×1 convolution respectively to obtain the transformed feature map.

[0020] Feature map x i The transformation is expressed by the following formula:

[0021]

[0022] F(x) = Conv(l(Down(f(x)) i )))),i∈1,2,3,4

[0023] in, Let x represent a multilayer perceptron, ω represent deformable convolution, and x represent a multilayer perceptron. i Let f(x) represent the feature map of the i-th layer, f(x) represent the fused feature obtained after deformable convolution and multilayer perceptron operations; Down represents the downsampling operation, l represents the Concat connection, Conv represents the fusion convolution operation, and F(x) represents the transformed feature map.

[0024] Preferably, in step S3b, the specific process of optimizing the Neck layer is as follows:

[0025] (1) Cancel the original last layer's conv+concat+C3k2 operation; (2) Through the multi-scale fusion layer in the backbone network layer, complete cross-layer interaction and edge perception, and connect a CBAM module to form a dual attention mechanism multi-scale layer based on channel attention mechanism and spatial attention mechanism. Utilize the cascaded collaboration of channel attention mechanism and spatial attention mechanism to achieve adaptive recalibration of multi-scale features; (3) Directly use the dual attention mechanism multi-scale layer as the input layer of the third-stage detection head to perform large target stage detection and improve the recognition ability of large-size internal waves and vortices.

[0026] Preferably, in step S5, the multi-scale slice detection includes a slicing process and a detection process, as detailed below:

[0027] Slicing process: Based on a scale-adaptive strategy for pyramid slicing, assuming the original image is... The slice window size is S k =(w k ,h k ), where k∈{1,2,3} represents different scale levels, and the slicing process is described as follows:

[0028] I k,i,j =∪(x i +w k ,y j +h k )

[0029] x i+1 =x i +w k ·(1-η k )

[0030] y j+1 =y j +h k ·(1-η k )

[0031] Among them, I k,i,j This represents the slice image at position (i,j) at the k-th scale level extracted from the original image, y j h represents the starting coordinates of the slice in the vertical direction of the original image. k x represents the height of the slice at the k-th scale level. i w represents the starting coordinates of the slice in the horizontal direction of the original image. k x represents the width of the slice at the k-th scale level. i+1 Indicates the starting column coordinate of the next slice, y j+1 Indicates the starting row coordinates of the next slice, η k This represents the overlap rate of slices at the k-th scale level;

[0032] Detection process: Described using the following code:

[0033] pseudocode:

[0034] Image input and ndarray conversion

[0035]

[0036] Save the array and pass it to the model for testing.

[0037] Three pyramid slice size parameters are set: small slices are used to capture subscale targets, while medium and large slices are used for sub-mesoscale and large-scale target recognition. When using small slices, a higher overlap rate is set for coverage, and a higher confidence threshold is set to ensure the integrity of small-scale information and the authenticity of targets, as detailed below:

[0038] Cfg = {win:(600,850,1200),

[0039] scale:(0.1,0.2,0.5),

[0040] OL:(0.4,0.3,0.2),

[0041] conf:(0.5,0.45,0.45)}

[0042] Wherein, Cfg is the configuration dictionary used to store parameters related to multi-scale tiling and target detection, win is the window tiling size used to process C and L band images, scale is the scaled tiling size used to process X band images, OL is the tiling overlap rate, and conf is the threshold for determining target confidence.

[0043] Preferably, in step S5, the specific process of merging the anchor frames is as follows:

[0044] A hierarchical scanning merging strategy is designed. The first scan performs coarse merging, merging anchor boxes larger than a threshold using the minimum bounding matrix to optimize most candidate anchor boxes. The second scan performs fine merging, further optimizing the first coarse merging to obtain the detection result, i.e., the pixel coordinate system. The threshold is set using the following formula:

[0045]

[0046] Where S1 is the area of ​​anchor frame 1, S2 is the area of ​​anchor frame 2, Min(S1,S2) is the smaller of the two anchor frame areas, S1∩S2 is the area of ​​the overlapping part of the two anchor frames, and θ is the ratio value, i.e. the threshold.

[0047] The hierarchical scan merging strategy is described by the following code procedure:

[0048] pseudocode:

[0049] Step 1:

[0050]

[0051]

[0052] Step 2:

[0053]

[0054] Step 3:

[0055]

[0056] Step 4: #Secondary Merge

[0057] Repeat Step 3

[0058] Preferably, in step S6, geographic information data is output using geographic affine transformation, as follows:

[0059] Geographic affine transformation uses six parameters to perform various geometric transformations while preserving the parallelism of lines after the transformation. It linearly maps pixel coordinates to geographic coordinates through matrix operations. The affine transformation formula can be expressed as follows:

[0060]

[0061] Where (x,y) are the original pixel coordinates, (x',y') are the transformed geographic coordinates; a and e are the scaling factors in the x and y directions, respectively; c and f are the geographic coordinates of the top left corner of the image, respectively; and b and d are the shearing or rotation parameters, usually b = d = 0.

[0062] The formula for calculating the latitude and longitude of the center point is as follows:

[0063]

[0064] Among them, lon center The longitude of the image center point is W, and the image width is lat. center H represents the latitude of the image center point, and H represents the image height.

[0065] The Karney geodesic algorithm is used to transform the coordinates of the anchor frame vertices (x1, y1) and (x2, y2), and the area is accurately calculated based on the Gaussian integral formula for the WGS84 ellipsoid.

[0066]

[0067] Δλ i =λ i+1 -λ i

[0068]

[0069] Where, φ i , λ i φ represents the latitude and longitude of the polygon's vertices. i+1 For φ i The longitude of the adjacent vertices, Δλ i φ represents the difference in longitude between adjacent vertices. avg,i e is the average latitude of adjacent vertices 2 Let be the square of the first eccentricity, 'a' be the semi-major axis of the ellipsoid (6378137 m in WGS84), and 'e' be the eccentricity. 2 Let S be the square of the first eccentricity, n be the number of vertices of the polygon, and S be the eccentricity of the polygon. i This is the calculated area value for the i-th subregion.

[0070] Compared with existing technologies, this invention proposes a method for identifying meso / sub-mesoscale dynamic processes in the ocean based on space-based SAR. The advantages of this method are:

[0071] 1. In terms of lightweight design: Compared with the higher-precision YOLO12-s in the YOLO series, the YOLOv12 structure design is adopted, replacing the traditional FPN with an ER-MLP layer. While maintaining higher detection accuracy, the number of parameters is reduced by about 70%, and the computational complexity is reduced by 60%. The introduction of YOLOv12's AreAtten mechanism and Edge Enhancement Perceptron (ER-MLP) enhances the focus on marine dynamic targets. This breakthrough in "high precision and low overhead" characteristics provides a practical solution for real-time target detection in resource-constrained scenarios (such as airborne equipment), achieving a better balance between computational efficiency and detection performance.

[0072] 2. In terms of multi-scale feature extraction: Deformable convolution is introduced to dynamically adjust the shape of the convolution kernel by learning the offset, which improves the edge enhancement module and enhances the target contour recognition capability. The multi-scale inference strategy solves the bottleneck of target recognition accuracy at different scales and realizes the unified recognition of mesoscale / sub-mesoscale ocean dynamic processes (internal waves, eddies).

[0073] 3. In terms of robustness: The integration of region attention mechanism and sample training significantly reduces the false detection rate in complex backgrounds;

[0074] 4. In terms of geographic information extraction: Integrating affine transformation models enables accurate target positioning and measurement, improving application feasibility. Attached Figure Description

[0075] Figure 1 A structural diagram is constructed for the model of this invention;

[0076] Figure 2 This is a schematic diagram of the edge enhancement sensor framework of the present invention;

[0077] Figure 3 This is a block diagram of the optimized neck layer of the present invention, wherein (a) is the neck layer diagram of YOLO, and (b) is the adjusted neck layer diagram;

[0078] Figure 4 This is a schematic diagram of the multi-scale slice detection process of the present invention;

[0079] Figure 5 This is a schematic diagram of the hierarchical scanning merging strategy for anchor frame merging in this invention;

[0080] Figure 6 This is a schematic diagram of anchor frame merging failure and optimization according to the present invention, wherein (a) is a merging failure scenario diagram using the traditional IOU threshold, and (b) is a merging diagram using the improved threshold.

[0081] Figure 7 This is a map showing the acquisition of geographic information data according to the present invention. Detailed Implementation

[0082] The technical solutions of the embodiments of this application will be further described clearly and completely below with reference to the accompanying drawings. It should be noted that the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0083] To make the inventive objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be further described in detail below with reference to the accompanying drawings: In order to better understand the above-mentioned objectives, features, and advantages of this invention, the advantages of this invention will be further illustrated below by comparing the embodiments with the accompanying drawings and specific implementation methods.

[0084] This invention proposes a method for identifying meso / sub-mesoscale dynamic processes in the ocean based on space-based SAR, and the steps of this method are described in detail:

[0085] Step 1, Preprocessing data: Input SAR data containing internal waves or vortices for preprocessing. Preprocessing includes radiometric correction, thermal noise removal, land masking, and geocoding.

[0086] Step 2, construct the dataset: Divide the preprocessed SAR data into training and validation sets in a 9:1 ratio, and augment the data by flipping and rotating.

[0087] Step 3, Model Construction: A model is constructed based on a lightweight neural network architecture, including improvements to the backbone network layers and optimizations to the Neck layer. The specific process is as follows:

[0088] a. Improve the backbone network layer: Improve the YOLO architecture and build an edge enhancement perceptron to obtain the transformed feature maps;

[0089] b. Optimize the Neck layer: Replace the traditional FPN with the ER-MLP layer to improve the ability to recognize large targets;

[0090] Specifically, in step S3a, such as Figure 1 As shown, the specific process of improving the YOLO architecture is as follows:

[0091] YOLO is a target detection algorithm based on convolutional neural networks. With its high inference speed and good detection accuracy, it has strong applicability. Through lightweight design, multi-scale feature fusion and decoupled detection head, it has initially achieved effective suppression of background information and can be used for target detection in complex marine environments.

[0092] Based on the YOLO architecture, a new target detection network is constructed to achieve effective target detection in complex marine environments: In the backbone network layer, the YOLOv12 structure is adopted, and the initial layer uses 7×7 convolutions. By using the convolutional layer to extract the initial feature map, a wider range of information can be captured, ensuring the feature extraction capability; and combined with the region attention mechanism module (AreAtten) unique to the YOLOv12 structure, the attention to multi-morphological targets in marine dynamic processes is enhanced.

[0093] Specifically, in step S3a, an edge enhancement perceptron is constructed to obtain the changed feature map, such as... Figure 2 As shown, the details are as follows:

[0094] Vortices are circular or elliptical structures with large scale and clear edges; while internal waves are striped or wavy structures with smaller scale and varied shapes. Internal waves and vortices in the ocean have significant "striped / spiral banded" edge features, which are crucial for accurate target detection and recognition. Low-level features contain more edge and texture details, but due to their high resolution and weak semantic information, traditional convolution is difficult to effectively extract complex edges. In order to promote the interaction between low-level information and high-level global information and enhance the perception of multi-morphological targets, an edge enhancement perceptron for cross-layer interaction and fusion is designed.

[0095] A learnable offset is introduced into the receptive field of deformable convolution, allowing the convolution kernel to dynamically adjust the position of sampling points to adapt to the shape of multi-morphological targets. The initial feature map is further processed by deformable convolution to obtain feature map x.i This improves the model's adaptability to target geometric deformation, enabling it to more accurately capture the contours and detailed features of complex targets. Subsequently, the feature map x is processed using an MLP (Multilayer Perceptron, composed of multiple fully connected layers stacked together, where each fully connected layer's neurons are connected to all neurons in the previous layer). i Cross-layer interaction and fusion are performed, organically combining high-resolution detail information from the low layer, local structural information from the middle layer, and global semantic information from the high layer into the MLP to generate multi-scale, multi-level feature representations, i.e., multi-scale fusion layers. Finally, spatial information fusion and channel transformation are performed on the feature representations through separable convolution and 1×1 convolution respectively to obtain transformed feature maps, which helps in the localization of boundary features and improves the precision and robustness of feature fusion.

[0096] Feature map x i The transformation is expressed by the following formula:

[0097]

[0098] F(x) = Conv(l(Down(f(x)) i )))),i∈1,2,3,4

[0099] in, Let x represent a multilayer perceptron, ω represent deformable convolution, and x represent a multilayer perceptron. i Let f(x) represent the feature map of the i-th layer, f(x) represent the fused feature obtained after deformable convolution and multilayer perceptron operations; Down represents the downsampling operation, l represents the Concat connection, Conv represents the fusion convolution operation, and F(x) represents the transformed feature map.

[0100] The core advantage of ER-MLP (Edge Enhancement Perceptron) over 1x1 convolution lies in its powerful nonlinear global feature fusion capability, enabling it to more effectively integrate heterogeneous features from different levels in the backbone network. 1x1 convolution is essentially a linear channel transformation operation; although it can introduce some nonlinearity through subsequent activation functions, its single-layer structure can only achieve low-order feature combinations. ER-MLP, through multi-layer fully connected structures and stacked nonlinear activation functions, can achieve high-order feature mapping, thus completing deep interaction of cross-level features within a unified high-dimensional space. Furthermore, in terms of the spatial scope of feature integration, 1x1 convolution only mixes in the channel dimension, with feature processing at each spatial location being independent. ER-MLP's global fully connected nature allows direct interaction between features at any two spatial locations, which is crucial for establishing long-range dependencies (such as aligning distant local features with global category labels). Especially when combined with deformable convolution, ER-MLP can intelligently fuse geometrically deformed local features with global semantic information, retaining the spatial adaptive advantages of deformable convolution while improving feature discriminativity through cross-level semantic interaction.

[0101] Specifically, in step S3b, the Neck layer is optimized, such as... Figure 3 As shown, the specific process is as follows:

[0102] Adjusting the Neck layer improves the semantic information flow of hierarchical features. The Neck layer is a key module located between the backbone and the head, and its main function is to optimize and integrate multi-level features from the backbone, thereby improving the flow and transmission of semantic information. In the traditional YOLO structure, the Neck layer uses layer-by-layer upsampling and feature aggregation (such as FPN or PANet) to generate the input features of the Head layer. Although this method can achieve the fusion of multi-level features, it still has certain limitations in terms of computational efficiency and feature representation ability. Especially when dealing with complex scenes, layer-by-layer upsampling may lead to the loss of detailed information or the dilution of semantic information, resulting in blurring of large-sized targets in subsequent feature layers.

[0103] In order to further improve the performance of the model, the Neck layer of YOLO was adjusted. The specific operations are as follows: (1) The original conv+concat+C3k2 operation of the last layer was canceled. conv (convolution) is used for feature extraction, concat (concatenation) merges features of different scales, and C3k2 (C3 module, stride 2) is used for downsampling. These operations rely on passing features layer by layer, which will cause edge details to be lost in multiple convolutions and samplings, and the computation is redundant. After cancellation, the feature loss of the intermediate layer is reduced; (2) Cross-layer interaction and edge perception are completed through the multi-scale fusion layer in the backbone network layer, and a CBAM module is connected to form a dual attention mechanism multi-scale layer based on channel attention mechanism and spatial attention mechanism. The cascaded collaboration of channel attention mechanism and spatial attention mechanism is used to realize the adaptive recalibration of multi-scale features; (3) The multi-scale layer based on dual attention mechanism is directly used as the input layer of the third stage detection head to perform large target stage detection, thereby improving the recognition ability of large-size internal waves and vortices. This ER-MLP layer design provides a new feature fusion idea for YOLO series models and has broad application potential.

[0104] Step 4, Train the model: Use the training set as data samples to train the model;

[0105] Step 5, Multi-scale slice detection: Pyramid slices are made on the images of data samples in the trained model. The overlap rate η_k and confidence threshold conf_k are set. After multi-scale detection, the anchor boxes are merged to obtain the detection results, including the pixel coordinate system.

[0106] Specifically, in step S5, the multi-scale slice detection includes a slicing process and a detection process, such as... Figure 4 As shown, the specific process is as follows:

[0107] Slicing Process: Traditional fixed-size slicing methods are ill-suited to the multi-scale characteristics of ocean dynamic processes. When the slice size is too large, the representational ability of sub-scale ocean dynamic features (such as small and medium-sized eddies and short-wavelength internal waves) in the feature space is significantly reduced, making it difficult for neural networks to effectively extract their discriminative features. Conversely, when the slice size is too small, the global structural features of sub-mesoscale and larger targets (such as large internal waves) exhibit information fragmentation, resulting in the loss of key spatial context information. The essence of this scale sensitivity lies in the trade-off relationship between feature representation and spatial resolution, i.e., the trade-off relationship. To address this, a scale-adaptive strategy based on pyramid slicing is proposed, assuming the original image is... The slice window size is S k =(w k ,h k ), where k∈{1,2,3} represents different scale levels, and the slicing process is described as follows:

[0108] I k,i,j =∪(x i +w k ,y j +h k )

[0109] x i+1 =x i +w k ·(1-η k )

[0110] y j+1 =y j +h k ·(1-η k )

[0111] Among them, I k,i,j This represents the slice image at position (i,j) at the k-th scale level extracted from the original image, y j h represents the starting coordinates of the slice in the vertical direction (row direction) of the original image. k x represents the height of the slice at the k-th scale level. i w represents the starting coordinates of the slice in the horizontal (column) direction of the original image. k x represents the width of the slice at the k-th scale level. i+1 Indicates the starting column coordinate of the next slice, y j+1 Indicates the starting row coordinates of the next slice, η k This represents the overlap rate of slices at the k-th scale level;

[0112] The detection process can be described using the following code:

[0113] pseudocode:

[0114]

[0115] Save the array and pass it to the model for testing.

[0116] Three pyramid slice size parameters are set: small slices are used to capture subscale targets, while medium and large slices are used for sub-mesoscale and large-scale target recognition. When using small slices, which contain less spatial texture information, a higher overlap rate is set for coverage, and a higher confidence threshold is set to ensure the integrity of small-scale information and the authenticity of targets, as detailed below:

[0117] Cfg = {win:(600,850,1200),

[0118] scale:(0.1,0.2,0.5),

[0119] OL:(0.4,0.3,0.2),

[0120] conf:(0.5,0.45,0.45)}

[0121] Wherein, Cfg is the configuration dictionary used to store parameters related to multi-scale tiling and target detection, win is the window tiling size used to process C and L band images, scale is the scaled tiling size used to process X band images, OL is the tiling overlap rate, and conf is the threshold for determining target confidence.

[0122] Specifically, in step S5, the process of merging the anchor frames is as follows:

[0123] While multi-scale slicing detection strategies can effectively improve target recognition rates, they inevitably generate a large number of redundant detection boxes, severely impacting subsequent analysis efficiency. To address this critical issue, an optimization algorithm based on hierarchical merging is proposed. By dynamically controlling the anchor box overlap threshold and the minimum bounding box fusion strategy, the compactness and reliability of the detection results are significantly improved. The following uses experimental results from Sentinel-1 SAR imagery as an example to illustrate the algorithm implementation and optimization effects:

[0124] Design a hierarchical scan merging strategy, such as Figure 5 As shown; the first scan performs coarse merging, merging anchor boxes larger than the threshold using the minimum bounding matrix, optimizing most candidate boxes; the second scan performs fine merging, further optimizing the first coarse merging, obtaining the detection result, i.e., the pixel coordinate system; the study found that the traditional IoU threshold has inherent defects when the anchor box size difference is large, such as Figure 6 As shown, this is because when the areas of anchor frame 1 and anchor frame 2 are significantly different, the IoU value will tend to zero, causing the merging judgment to fail; in order to achieve controllability of anchor frames, the threshold setting formula is as follows:

[0125]

[0126] Where S1 is the area of ​​anchor frame 1, S2 is the area of ​​anchor frame 2, Min(S1,S2) is the smaller of the two anchor frame areas, S1∩S2 is the area of ​​the overlapping part of the two anchor frames, and θ is the ratio value.

[0127] The hierarchical scan merging strategy process can be described by the following code procedure:

[0128] pseudocode:

[0129] Step 1:

[0130]

[0131]

[0132] Step 2:

[0133]

[0134] Step 3:

[0135]

[0136] Step 4: #Secondary Merge

[0137] Repeat Step 3

[0138] Step 6, output geographic information data through geographic affine transformation: map the pixel coordinate system to the geographic coordinate system through geographic affine transformation, and output the center position and area of ​​the vortex / inner wave.

[0139] Specifically, in step S6, geographic information data is output using geographic affine transformation, such as... Figure 7 As shown, the details are as follows:

[0140] Geographic affine transformation is a two-dimensional linear coordinate transformation method used to describe the geometric correction of spatial data (such as remote sensing imagery and maps). Its core is to achieve various geometric transformations (including translation, rotation, scaling, and shearing) using six parameters (a, b, c, d, e, f) while maintaining the parallelism of the transformed lines. It is widely used in geometric correction of remote sensing images, map projection conversion, and multi-source data registration, such as aligning UAV imagery with WGS84 coordinates or eliminating projection distortion of satellite imagery. Through matrix operations, it linearly maps the pixel coordinate system to the geographic coordinate system. The affine transformation formula can be expressed as follows:

[0141]

[0142] Where (x,y) are the original pixel coordinates, (x',y') are the transformed geographic coordinates; a and e are the scaling factors in the x and y directions, respectively; c and f are the geographic coordinates of the top left corner of the image, respectively; and b and d are the shearing or rotation parameters, usually b = d = 0.

[0143] The formula for calculating the latitude and longitude of the center point (assuming image width W and height H) is as follows:

[0144]

[0145] Among them, lon center The longitude of the image center point is W, and the image width is lat. center H represents the latitude of the image center point, and H represents the image height.

[0146] The Karney geodesic algorithm is used to transform the coordinates of the anchor frame vertices (x1, y1) (top left) and (x2, y2) (bottom right), and the area is accurately calculated based on the Gaussian integral formula for the WGS84 ellipsoid.

[0147]

[0148] Δλ i =λ i+1 -λ i

[0149]

[0150] Where, φ i , λ i φ represents the latitude and longitude of the polygon's vertices. i+1 For φ i The longitude of the adjacent vertices, Δλ i φ represents the difference in longitude between adjacent vertices. avg,i e is the average latitude of adjacent vertices 2 Let be the square of the first eccentricity, 'a' be the semi-major axis of the ellipsoid (6378137 m in WGS84), and 'e' be the eccentricity. 2 Let S be the square of the first eccentricity, n be the number of vertices of the polygon, and S be the eccentricity of the polygon. i This is the calculated area value for the i-th subregion.

[0151] The specific experiments of this invention are described in detail below with reference to the technical solution and accompanying drawings:

[0152] To comprehensively verify the effectiveness and superiority of the model, numerous comparative experiments were conducted using the YOLO series models (including YOLOv8, YOLOv10-12, etc.) as benchmarks. The experiments covered target detection tasks in various complex scenarios, focusing on core metrics such as detection accuracy, computational cost, and parameter count. Experimental results show that the new YOLOv12 architecture exhibits significant advantages across multiple dimensions. Specifically, with only 2.79M parameters, the model not only meets the requirements for fast, real-time detection tasks but also surpasses the aforementioned models in internal wave-vortex detection accuracy.

[0153] In particular, due to the complex environment of SAR images, when detecting entire SAR images, it is impossible to avoid input from other objects with similar characteristics, such as oil spills, waves, and ship stern tracks, leading to false detections. Therefore, in addition to improving the model, the experiment was divided into two parts, with the presence or absence of negative sample data for training as the dividing line. Inference on a single image of size (640×640) was also tested, and the actual end-to-end detection FPS (Frames Per Second) was calculated. Extensive experiments show that internal wave features are more sensitive and are the main cause of false detections of ocean dynamic processes. The experimental process is shown in Table 1.

[0154] Table 1 Comparative Experiment

[0155]

[0156]

[0157] Current experimental results demonstrate that the model exhibits superior detection performance on the standard test dataset: in training mode including negative samples, the mAP50 reaches 87.8%; after excluding negative samples, the performance further improves to 90.5%. Notably, the identification of internal wave features is more significantly affected by the perturbation of negative samples, with accuracy fluctuations reaching approximately 5 points. Despite this, while learning from both positive and negative samples, the model achieves detection accuracies of 85.7% and 90% for internal waves and vortices, respectively. When training only on target samples, the accuracy on the test set reaches 90.7% and 90.2%, respectively, both superior to other models with similar parameters. In particular, compared to other models, the proposed method improves the model's sensitivity to internal waves, ensuring high vortex identification accuracy while significantly enhancing the model's ability to identify internal waves. Even with perturbation factors, the model maintains a high mAP, highlighting the advantage of cross-level fusion in balancing accuracy and robustness, resulting in greater reliability in handling complex marine environments.

[0158] The experimental results of this study demonstrate a significant breakthrough in model lightweighting: requiring only 2.79M parameters, a 90.5% mAP is achieved on the validation set, a performance that shows a clear advantage over the YOLO12-n baseline model. Particularly noteworthy is that, compared to similar YOLO12-s models, this model reduces the number of parameters by approximately 70% and computational complexity by 60% while maintaining higher detection accuracy. This breakthrough in "high accuracy-low overhead" characteristics provides a practical solution for real-time target detection in resource-constrained scenarios (such as airborne equipment), achieving a better balance between computational efficiency and detection performance.

[0159] ablation experiment

[0160] To demonstrate the effectiveness of the proposed network and modules, ablation experiments were conducted on the adjustments to each part. Area attention has already proven its effectiveness in YOLO12; therefore, using YOLO12-n as the base model, the effectiveness of adjustments to ER-MLP and the Neck was analyzed, and experimental analyses were performed on the respective datasets. In the ablation experiments, a, b, c, and d represent the adjustments to the ER-MLP (neck) module, the use of 7×7 convolutions, the ER-MLP (backbone) + Neck layer, and the addition of the CBAM module, respectively.

[0161] Table 2 Comparison of ablation experiments

[0162]

[0163] Four ablation experiments were conducted to verify the effectiveness of the proposed module and the modified structure. Table 3 shows that while adding the ER-MLP layer to the neck did not significantly improve accuracy, it introduced a global nonlinear expression to the model, which was confirmed in subsequent experiments. In Ablation-2, the first layer of the backbone neural network was replaced with a 7x7 convolutional kernel. This was done to capture a wider range of local features (such as edges and textures) in the early stages of the network. At this point, the model's detection performance improved somewhat, while the number of parameters remained essentially unchanged. In Ablation-3, ER-MLP (backbone) was added, and the neck layer was adjusted. At this point, mAP50 reached 89.4%, a 2% improvement compared to the base model. The recognition accuracy of inner waves and vortices was significantly improved, reaching 89.3% and 89.5%, respectively. Finally, in the Ablation-4 experiment, the CBAM module was added to achieve adaptive recalibration of the ER-MLP multi-scale features. At this point, the recognition accuracy of internal waves and vortices was relatively stable, and the accuracy of internal waves, vortices, and the overall model improved by 3.9%, 2.1%, and 3.1%, respectively, compared with the base model. This is thanks to the ER-MLP module, which effectively integrates multi-level features, optimizes the localization of target boundary regions, and improves the model's ability to capture ocean dynamic processes.

[0164] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0165] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for identifying oceanic meso / sub-mesoscale dynamic processes based on space-based SAR, characterized in that, include: Step 1, Preprocessing data: Input SAR data containing internal waves or vortices for preprocessing. Preprocessing includes radiometric correction, thermal noise removal, land masking, and geocoding. Step 2, construct the dataset: Divide the preprocessed SAR data into training and validation sets in a 9:1 ratio, and augment the data by flipping and rotating. Step 3, Model Construction: A model is constructed based on a lightweight neural network architecture, including improvements to the backbone network layers and optimizations to the Neck layer. The specific process is as follows: a. Improve the backbone network layer: Improve the YOLO architecture and build an edge enhancement perceptron to obtain the transformed feature maps; b. Optimize the Neck layer: Replace the traditional FPN with the ER-MLP layer to improve the ability to recognize large targets; Step 4, Train the model: Use the training set as data samples to train the model; Step 5, Multi-scale slice detection: Pyramid slices are made on the images of data samples in the trained model. The overlap rate η_k and confidence threshold conf_k are set. After multi-scale detection, the anchor boxes are merged to obtain the detection results, including the pixel coordinate system. Step 6, output geographic information data through geographic affine transformation: map the pixel coordinate system to the geographic coordinate system through geographic affine transformation, and output the center position and area of ​​the vortex / inner wave.

2. The method for identifying oceanic meso / sub-mesoscale dynamic processes based on space-based SAR according to claim 1, characterized in that, In step S3a, the specific process of improving the YOLO architecture is as follows: Based on the YOLO architecture, a new target detection network is constructed to achieve effective detection of multi-morphological targets in complex marine environments: In the backbone network layer, the YOLOv12 structure is adopted, and the initial layer uses 7×7 convolutions to extract the initial feature map; then, combined with the region attention mechanism module unique to the YOLOv12 structure, the attention to multi-morphological targets in marine dynamic processes is enhanced.

3. The method for identifying oceanic meso / sub-mesoscale dynamic processes based on space-based SAR according to claim 1, characterized in that, In step S3a, the specific process of constructing the edge enhancement perceptron to obtain the changed feature map is as follows: A learnable offset is introduced into the receptive field of deformable convolution, allowing the convolution kernel to dynamically adjust the position of sampling points to adapt to the shape of multi-morphological targets. The initial feature map is further processed by deformable convolution to obtain feature map x. i Subsequently, the feature map x was processed using MLP. i Cross-layer interaction and fusion are performed, organically combining high-resolution detail information from the low layer, local structural information from the middle layer, and global semantic information from the high layer into the MLP to generate multi-scale, multi-level feature representations, namely the multi-scale fusion layer; finally, spatial information fusion and channel transformation are performed on the feature representations through separable convolution and 1×1 convolution respectively to obtain the transformed feature map. Feature map x i The transformation is expressed by the following formula: F(x)=Conv(l(Down(f(x i )))),i∈1,2,3,4 in, Let x represent a multilayer perceptron, ω represent deformable convolution, and x represent a multilayer perceptron. i Let f(x) represent the feature map of the i-th layer, f(x) represent the fused feature obtained after deformable convolution and multilayer perceptron operations; Down represents the downsampling operation, l represents the Concat connection, Conv represents the fusion convolution operation, and F(x) represents the transformed feature map.

4. The method for identifying oceanic meso / sub-mesoscale dynamic processes based on space-based SAR according to claim 1, characterized in that, In step S3b, the specific process of optimizing the Neck layer is as follows: (1) Cancel the original last layer's conv+concat+C3k2 operation; (2) Through the multi-scale fusion layer in the backbone network layer, complete cross-layer interaction and edge perception, and connect a CBAM module to form a dual attention mechanism multi-scale layer based on channel attention mechanism and spatial attention mechanism. Utilize the cascaded collaboration of channel attention mechanism and spatial attention mechanism to achieve adaptive recalibration of multi-scale features; (3) Directly use the dual attention mechanism multi-scale layer as the input layer of the third-stage detection head to perform large target stage detection and improve the recognition ability of large-size internal waves and vortices.

5. The method for identifying oceanic meso / sub-mesoscale dynamic processes based on space-based SAR according to claim 1, characterized in that, In step S5, the multi-scale slice detection includes a slicing process and a detection process, as detailed below: Slicing process: Based on a scale-adaptive strategy for pyramid slicing, assuming the original image is... The slice window size is S k =(w k ,h k ), where k∈{1,2,3} represents different scale levels, and the slicing process is described as follows: I k,i,j =∪(x i +w k ,y j +h k ) x i+1 =x i +w k ·(1-th k ) and j+1 / and j +h k ·(1-η k ) Among them, I k,i,j This represents the slice image at position (i,j) at the k-th scale level extracted from the original image, y j h represents the starting coordinates of the slice in the vertical direction of the original image. k x represents the height of the slice at the k-th scale level. i w represents the starting coordinates of the slice in the horizontal direction of the original image. k x represents the width of the slice at the k-th scale level. i+1 Indicates the starting column coordinate of the next slice, y j+1 Indicates the starting row coordinates of the next slice, η k This represents the overlap rate of slices at the k-th scale level; Detection process: Described using the following code: pseudocode: Three pyramid slice size parameters are set: small slices are used to capture subscale targets, while medium and large slices are used for sub-mesoscale and large-scale target recognition. When using small slices, a higher overlap rate is set for coverage, and a higher confidence threshold is set to ensure the integrity of small-scale information and the authenticity of targets, as detailed below: Cfg = {win:(600,850,1200), scale:(0.1,0.2,0.5), OL:(0.4,0.3,0.2), conf:(0.5,0.45,0.45)} Wherein, Cfg is the configuration dictionary used to store parameters related to multi-scale tiling and target detection, win is the window tiling size used to process C and L band images, scale is the scaled tiling size used to process X band images, OL is the tiling overlap rate, and conf is the threshold for determining target confidence.

6. The method for identifying oceanic meso / sub-mesoscale dynamic processes based on space-based SAR according to claim 1, characterized in that, In step S5, the specific process of merging the anchor frames is as follows: A hierarchical scanning merging strategy is designed. The first scan performs coarse merging, merging anchor boxes larger than a threshold using the minimum bounding matrix to optimize most candidate anchor boxes. The second scan performs fine merging, further optimizing the first coarse merging to obtain the detection result, i.e., the pixel coordinate system. The threshold is set using the following formula: Where S1 is the area of ​​anchor frame 1, S2 is the area of ​​anchor frame 2, Min(S1,S2) is the smaller of the two anchor frame areas, S1∩S2 is the area of ​​the overlapping part of the two anchor frames, and θ is the ratio value, i.e. the threshold. The hierarchical scan merging strategy is described by the following code procedure: pseudocode: Step 1: Step 2: Step 3: Step 4: #Secondary Merge Repeat Step 3 7. The method for identifying oceanic meso / sub-mesoscale dynamic processes based on space-based SAR according to claim 1, characterized in that, In step S6, geographic information data is output using geographic affine transformation, as detailed below: Geographic affine transformation uses six parameters to perform various geometric transformations while preserving the parallelism of lines after the transformation. It linearly maps pixel coordinates to geographic coordinates through matrix operations. The affine transformation formula can be expressed as follows: Where (x,y) are the original pixel coordinates, (x',y') are the transformed geographic coordinates; a and e are the scaling factors in the x and y directions, respectively; c and f are the geographic coordinates of the top left corner of the image, respectively; and b and d are the shearing or rotation parameters, usually b = d = 0. The formula for calculating the latitude and longitude of the center point is as follows: Among them, lon center The longitude of the image center point is W, and the image width is lat. center H represents the latitude of the image center point, and H represents the image height. The Karney geodesic algorithm is used to transform the coordinates of the anchor frame vertices (x1, y1) and (x2, y2), and the area is accurately calculated based on the Gaussian integral formula for the WGS84 ellipsoid. Dl i =λ i+1 -l i Where, φ i , λ i φ represents the latitude and longitude of the polygon's vertices. i+1 For φ i The longitude of the adjacent vertices, Δλ i φ represents the difference in longitude between adjacent vertices. avg,i e is the average latitude of adjacent vertices 2 Let be the square of the first eccentricity, 'a' be the semi-major axis of the ellipsoid (6378137 m in WGS84), and 'e' be the eccentricity. 2 Let S be the square of the first eccentricity, n be the number of vertices of the polygon, and S be the eccentricity of the polygon. i This is the calculated area value for the i-th subregion.

Images