A method for reconstructing kuroshio front stereo centerline based on attention mechanism correction
By employing the Sobel algorithm and attention mechanism correction method, combined with the characteristics of the flow field and sea surface height anomaly field, a high-precision reconstruction of the three-dimensional centerline of the Kuroshio front was achieved. This solves the problem of the difficulty in characterizing the three-dimensional structure of the Kuroshio front in existing technologies and provides a deeper understanding of the dynamic mechanism and ecological effects of the Kuroshio Current.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE 715TH RES INST OF CHINA SHIPBUILDING IND CORP
- Filing Date
- 2026-01-22
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies that rely solely on sea surface information are insufficient to fully reveal the three-dimensional structure and vertical variations of the Kuroshio front, especially in accurately capturing the three-dimensional oscillation of its axis and the characteristics of the subsurface front, thus limiting a deeper understanding of the Kuroshio's dynamic mechanisms and its ecological effects.
The Sobel algorithm was used to extract the initial front from the three-dimensional temperature field, and the characteristics of the flow field and sea surface height anomaly field were combined for correction. An attention mechanism was introduced to optimize the weight distribution of each element, and finally a high-precision three-dimensional reconstruction of the Kuroshio front centerline was achieved.
It effectively depicts the three-dimensional reconstruction of the Kuroshio front centerline and accurately describes the vertical oscillation characteristics of the Kuroshio axis, providing an effective tool for in-depth research on the three-dimensional characteristics of the Kuroshio front.
Smart Images
Figure CN122116162A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of Kuroshio front extraction technology, and in particular, a method for reconstructing the three-dimensional centerline of Kuroshio front based on attention mechanism correction. Background Technology
[0002] The Kuroshio front, as the transition zone between the Kuroshio Current and surrounding water masses, is an area of exceptionally active energy exchange and biological activity. Accurate characterization of its front is crucial for understanding ocean dynamics, material transport, and ecosystem dynamics. The maturity of satellite remote sensing technology has made sea surface temperature (SST) data the primary basis for studying the Kuroshio front. SST-based front detection methods have been widely applied to the identification and tracking of Kuroshio paths at sea surface. These methods can clearly depict the spatial distribution and seasonal variations of the Kuroshio at the surface, resulting in a wealth of research findings on sea surface fronts. However, as a powerful current traversing the upper and middle layers of the ocean, the Kuroshio's influence extends far beyond the surface. Relying solely on sea surface information is insufficient to fully reveal the three-dimensional structure and vertical variations of the Kuroshio front, especially failing to accurately capture its three-dimensional axial oscillation and subsurface frontal characteristics. This "surface bias" severely restricts a deeper understanding of the Kuroshio's dynamic mechanisms and ecological effects.
[0003] With the fusion of multi-source satellite data and the development of artificial intelligence technology, three-dimensional reconstruction of fronts combining multi-layered information is gradually becoming possible. Sea surface height anomalies reflect ocean dynamic topography, while temperature-salinity current fields contain information about water mass boundaries and vertical variations of the front. In recent years, the attention mechanism in deep learning has demonstrated outstanding performance in computer vision and processing. Its core lies in its ability to adaptively focus on key information and suppress noise interference. This characteristic is highly suitable for multi-source data fusion and feature weighting problems in ocean front reconstruction. By introducing the attention mechanism, the contribution weights of different depths and data sources to front correction can be dynamically adjusted, thereby more accurately capturing the position and changes of the Kuroshio front in complex ocean environments. This provides a powerful technical tool for in-depth research on the multi-scale variability, vertical movement, and ecological effects of the Kuroshio Current. Summary of the Invention
[0004] The sea surface temperature obtained by relying on relatively mature satellite remote sensing data technology provides a wealth of methods for studying the Kuroshio front, and its characteristics are relatively clear. However, as a powerful western boundary current, the Kuroshio's influence is not limited to the surface layer. Existing technologies that only focus on the surface layer cannot adequately characterize its three-dimensional frontal features, thus limiting a deeper understanding of its dynamic processes and ecological impacts.
[0005] Based on this, the present invention proposes a method for reconstructing the three-dimensional centerline of the Kuroshio Front based on attention mechanism correction.
[0006] The technical solution of this invention is as follows:
[0007] A method for reconstructing the three-dimensional centerline of the Kuroshio front based on attention mechanism correction is proposed. First, the Sobel algorithm is used to extract the initial frontline from the three-dimensional temperature field. Then, the three-dimensional centerline is corrected using the characteristics of the flow field and sea surface height anomaly field. Finally, an attention mechanism is introduced to optimize the weight allocation of each element, ultimately achieving high-precision three-dimensional reconstruction of the Kuroshio front centerline. The specific steps are as follows:
[0008] (1) Obtaining the three-dimensional Kuroshio frontal zone based on the Sobel algorithm:
[0009] Using global ocean eddy-resolved analytical data, with horizontal resolution and vertical layer number set, the temperature field was selected as the baseline field, and the current field and sea surface height anomaly field were used as correction fields. The Sobel method was employed to calculate the gradient. Size;
[0010] Based on the ocean characteristics near the frontal zone, the ocean is divided into shallow water, intermediate water and deep water. Gradient critical values for different depths are selected, and the threshold of the frontal zone is selected based on dynamic characteristics, thereby extracting the three-dimensional Kuroshio frontal zone.
[0011] (2) Extracting the initial three-dimensional centerline of the Kuroshio Current:
[0012] For a mesh plane at the same depth, take a certain point within it. Construct an n×n matrix with the center point, where n≥2. If the point is a maximum value in any direction, retain it; otherwise, set it to 0.
[0013] Connect the processed points and repeat the same operation on meshes of different depths to obtain the initial three-dimensional center line of the Kuroshio Front.
[0014] (3) Outlier removal is achieved by using the characteristics of the flow field and sea surface height anomaly field:
[0015] First, the flow field and sea surface height anomaly field characteristics associated with the frontal zone temperature field are analyzed. In the same depth grid plane, only points where the difference between the frontal line and streamline direction and the difference between the frontal line and contour line trend direction do not exceed the set temperature are retained. This completes the outlier removal based on the frontal zone flow field characteristics, thereby constraining the frontal line.
[0016] (4) The attention mechanism is used to adjust the weights of the correction field:
[0017] Using a weighted processing method, in shallow, middle and deeper waters, only points where the difference in trend direction between the front and the contour lines does not exceed the set temperature are retained. As the depth increases, the azimuth value is increased to reduce the impact of sea surface height anomaly field correction.
[0018] Preferably, in step (1), the monthly average product data of global ocean eddy resolution reanalysis launched by the Copernicus Marine Service of the European Union is used, with the horizontal resolution set to 1 / 12° and the vertical direction divided into 50 layers.
[0019] Preferably, in step (1), the gradient The calculation formula is:
[0020] ,
[0021] Where T represents temperature.
[0022] Preferably, in step (1), the shallow water body is set to less than 30 meters, the middle water body to 30 meters to 650 meters, and the deep water body to more than 650 meters.
[0023] Preferably, in step (1), the frontal threshold of the shallow and deep water bodies is set to 0.045℃ / 0.02°, and the frontal threshold of the middle water body is set to 0.09℃ / 0.02°.
[0024] Preferably, in step (3), only points whose difference in direction between the front line and streamline, and whose difference in trend direction between the front line and contour line does not exceed 45° are retained in the grid plane at the same depth.
[0025] Preferably, in step (4), the angle is set to no more than 45° for shallow water within 30 meters, no more than 55° for water between 30 and 150 meters in the middle layer, no more than 70° for water between 150 and 300 meters, and no more than 90° for deeper water.
[0026] The beneficial effects of this invention are as follows: First, the Sobel algorithm is used to extract the initial front line from the three-dimensional temperature field. Then, the characteristics of the flow field and sea surface height anomaly field are used to correct the three-dimensional center line. An attention mechanism is introduced to adjust the weight of the correction field, thereby realizing the three-dimensional reconstruction of the Kuroshio front center line and effectively characterizing the vertical swing characteristics of the Kuroshio axis, providing an effective tool for describing the three-dimensional characteristics of the Kuroshio front. Attached Figure Description
[0027] Figure 1 A schematic diagram of the frontal surface of an oceanic front; Figure 2a A schematic diagram of the Kuroshio Current region and surrounding water masses in the East China Sea (winter). Figure 2b A schematic diagram of the Kuroshio Current region and surrounding water masses in the East China Sea (summer). Figure 3 This is a framework diagram of the three-dimensional frontal line extraction method of the present invention.
[0028] Figure 4a A three-dimensional feature map of the frontal zone (April); Figure 4b A three-dimensional feature map of the frontal zone (July); Figure 4c A three-dimensional feature map of the frontal zone (October); Figure 4d This is a three-dimensional feature map of the frontal zone (January).
[0029] Figure 5-1 An artist's rendering of the initial three-dimensional centerline of the Kuroshio front extracted from shallow water (6m); Figure 5-2 To Figure 5-1 The effect of using flow field characteristics to remove outliers is shown in the diagram. Figure 5-3 To Figure 5-2 The effect of removing outliers using the weighted sea surface height anomaly field characteristics is shown in the figure, where the purple line represents the fitted three-dimensional center line of the Kuroshio front in the shallow water (6m).
[0030] Figure 5-4 An artist's rendering of the initial three-dimensional centerline of the Kuroshio front extracted from the middle water layer (130m); Figure 5-5 To Figure 5-4 The effect of using flow field characteristics to remove outliers is shown in the diagram. Figure 5-6 To Figure 5-5 The effect of removing outliers using the weighted sea surface height anomaly field characteristics is shown in the figure. The purple line represents the fitted three-dimensional center line of the Kuroshio front in the middle water layer (130m).
[0031] Figure 5-7 An artist's rendering of the initial three-dimensional centerline of the Kuroshio front extracted from deep water (650m); Figure 5-8 To Figure 5-7 The effect of using flow field characteristics to remove outliers is shown in the diagram. Figure 5-9 To Figure 5-8 The effect of removing outliers using the weighted sea surface height anomaly field characteristics is shown in the figure, where the purple line represents the fitted three-dimensional center line of the Kuroshio front in the deep water body (650m). Detailed Implementation
[0032] The present invention will be further described below with reference to specific embodiments and accompanying drawings:
[0033] Figure 1 A schematic diagram of an oceanic front is provided. An oceanic front generally refers to a narrow transition zone between two or more adjacent water bodies with significantly different properties in a horizontal direction, and is an important mesoscale phenomenon. However, in reality, the ocean is a vast three-dimensional water body, and the transition zone cannot exist only on a single horizontal plane; therefore, its front also possesses a three-dimensional structure. The Kuroshio oceanic front is formed by the interaction between the high-temperature, high-salinity, high-color, and highly transparent Kuroshio water and the low-temperature, low-salinity, high-turbidity, and low-transparency East China Sea shelf water. Research on the Kuroshio front has wide applications in marine fisheries, environmental protection, and underwater acoustics technology; therefore, the accurate extraction of the three-dimensional front line is particularly important.
[0034] Figure 2a , Figure 2b Schematic diagrams of the Kuroshio Current region and surrounding water masses in the East China Sea during winter and summer are provided. This sea area has a vast continental shelf, complex and varied topography, and significant monsoon characteristics, resulting in complex hydrodynamic phenomena and a rich variety of oceanic front types and their changing characteristics. In terms of type, the oceanic fronts in this sea area mainly include tidal shelf fronts, estuarine fronts, coastal current fronts, upwelling fronts, and strong western boundary current fronts. The oceanic fronts also differ between summer and winter; in summer, the upwelling front is more prominent, the marginal fronts of the lower layers of the Taiwan Warm Current and the cold water mass at the bottom of the Yellow Sea are more obvious, and the Kuroshio front is also located more eastward.
[0035] This invention provides a method for reconstructing the three-dimensional centerline of the Kuroshio front based on attention mechanism correction, and its three-dimensional front extraction framework is as follows: Figure 3 As shown, based on the initial frontal line extracted from the three-dimensional temperature field using the Sobel algorithm, the three-dimensional centerline is corrected by comprehensively utilizing the characteristics of the flow field and the sea surface height anomaly field. Furthermore, an attention mechanism is introduced to optimize the weight allocation of each element, ultimately achieving a high-precision three-dimensional reconstruction of the Kuroshio front centerline. The specific steps are as follows:
[0036] (1) Obtaining the three-dimensional Kuroshio frontal zone based on the Sobel algorithm:
[0037] Using the monthly average product data of global ocean eddy reanalysis from the EU Copernicus Marine Service Information, the horizontal resolution was set to 1 / 12°, and the vertical direction was divided into 50 layers. The temperature field was selected as the reference field, and the flow field and sea surface height anomaly field were selected as the correction fields.
[0038] Since gradients can reflect the degree of change in a corresponding variable, sea surface temperature gradients can be used to detect sea surface temperature fronts. Gradient-based methods include different types, such as the maximum gradient method and various edge detection algorithms that use gradient operators as templates.
[0039] In this invention, the Sobel method is used to calculate the gradient. The size is calculated using the following formula:
[0040] ,
[0041] Where T represents temperature.
[0042] To address the issue of selecting critical values for different depth gradients, this invention divides the ocean into three layers based on the characteristics of the ocean near frontal zones: shallow water (less than 30 meters), intermediate water (30 to 650 meters), and deep water (greater than 650 meters).
[0043] In selecting the threshold for the frontal zone, the thresholds for shallow and deep water are set to 0.045℃ / 0.02° based on dynamic characteristics, and the thresholds for mid-water are set to 0.09℃ / 0.02°. Here, "℃ / 0.02°" represents degrees Celsius / horizontal resolution, specifically the temperature change rate of each grid (0.02° longitude / latitude) based on the grid division of the Earth's surface, thereby extracting the three-dimensional Kuroshio frontal zone.
[0044] (2) Extracting the initial three-dimensional centerline of the Kuroshio Current:
[0045] For a mesh plane at the same depth, at a certain point A 3×3 matrix is constructed around the center point. If the point is a maximum value in any direction, it is retained; otherwise, it is set to 0. The processed points are then connected. The same operation is performed on meshes of different depths to obtain the initial 3D centerline of the Kuroshio Front. The processed centerline contains many outliers, which previous researchers have mostly removed using purely mathematical methods.
[0046] (3) Outlier removal is achieved by using the characteristics of the flow field and sea surface height anomaly field:
[0047] First, we analyze the characteristics of the flow field and sea surface height anomaly field associated with the temperature field in the frontal zone. On the same depth grid plane, we retain only the points where the difference between the direction of the frontal line and the streamline, and the difference between the trend direction of the frontal line and the contour line, do not exceed the set temperature, such as 45°. This completes the outlier removal based on the characteristics of the frontal zone flow field, thereby constraining the frontal line and making the direction of the retained centerline more consistent with the characteristics of the Kuroshio Current.
[0048] (4) The attention mechanism is used to adjust the weights of the correction field:
[0049] It is worth noting that although the sea surface height anomaly field can reflect the state of the water body below the surface to a certain extent, it mainly serves as a manifestation of sea surface characteristics. Therefore, a weighted processing method is adopted. In the shallow layer, only points where the trend direction difference between the front and the contour line does not exceed 45° are retained. As the depth increases, the azimuth value is increased to reduce the impact of the sea surface height anomaly field correction. In the middle layer of water, the azimuth value is set to no more than 55° for 30-150 meters, no more than 70° for 150-300 meters, and no more than 90° for deeper water.
[0050] In this invention, April, July, October, and January are used as representative months for spring, summer, autumn, and winter, and the following is given: Figures 4a to 4dThe image shows the three-dimensional characteristics of the Kuroshio front. As can be seen from the image: In spring, the Kuroshio front is clearly visible from the surface to the depths, with significant bifurcation in the shallow and mid-level waters near Japan, and a clear path in the deep layers; In summer, the shallow front significantly diminishes or even disappears, while the mid-level waters remain visible, and the deep front diminishes less than the surface front; In autumn, the shallow front gradually increases as winter approaches, the mid-level waters are visible but the branching weakens, and the deep front diminishes to almost nothing; In winter, the shallow front continues to strengthen, but the path becomes slightly more complex, the mid-level waters are visible and the branching is stronger compared to autumn, and the deep front gradually becomes clearer.
[0051] Furthermore, in this invention, the three-dimensional centerline of the Kuroshio Front is corrected using the characteristics of the flow field and sea surface height anomaly field. Figures 5-1 to 5-9 The extraction results for shallow (6m), mid-level (130m), and deep (650m) water bodies are shown respectively. As can be seen from the figures, the front after weighted denoising using two characteristics better characterizes the front, achieving a three-dimensional reconstruction of the Kuroshio front centerline and effectively reducing interference points. Simultaneously, the expression of the front's tail direction is more accurate, effectively depicting the vertical oscillation characteristics of the Kuroshio axis. For example, the oscillation characteristics of the latter half of the Kuroshio front change at different water depths; compared to the surface layer, the mid-level water body turns earlier and more westward, while the deep water body turns later and more northward.
[0052] It should be noted that the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Equivalent modifications made based on the above embodiments are all within the scope of protection of the present invention.
Claims
1. A method for reconstructing the three-dimensional centerline of the Kuroshio Front based on attention mechanism correction, characterized in that, Includes the following steps: (1) Obtaining the three-dimensional Kuroshio frontal zone based on the Sobel algorithm: Using global ocean eddy-resolved analytical data, with horizontal resolution and vertical layer number set, the temperature field was selected as the baseline field, and the current field and sea surface height anomaly field were used as correction fields. The Sobel method was employed to calculate the gradient. Size; Based on the ocean characteristics near the frontal zone, the ocean is divided into shallow water, intermediate water and deep water. Gradient critical values for different depths are selected, and the threshold of the frontal zone is selected based on dynamic characteristics, thereby extracting the three-dimensional Kuroshio frontal zone. (2) Extracting the initial three-dimensional centerline of the Kuroshio Current: For a grid plane at the same depth, take one of the points Construct an n×n matrix with the center point, where n≥2. If the point is a maximum value in any direction, retain it; otherwise, set it to 0. Connect the processed points and repeat the same operation on meshes of different depths to obtain the initial three-dimensional center line of the Kuroshio Front. (3) Outlier removal is achieved by using the characteristics of the flow field and sea surface height anomaly field: First, the flow field and sea surface height anomaly field characteristics associated with the frontal zone temperature field are analyzed. In the same depth grid plane, only points where the difference between the frontal line and streamline direction and the difference between the frontal line and contour line trend direction do not exceed the set temperature are retained. This completes the outlier removal based on the frontal zone flow field characteristics, thereby constraining the frontal line. (4) The attention mechanism is used to adjust the weights of the correction field: Using a weighted processing method, in shallow, middle and deeper waters, only points where the difference in trend direction between the front and the contour lines does not exceed the set temperature are retained. As the depth increases, the azimuth value is increased to reduce the impact of sea surface height anomaly field correction.
2. The method for reconstructing the three-dimensional centerline of the Kuroshio Front based on attention mechanism correction according to claim 1, characterized in that: In step (1), the monthly average product data of global ocean eddy resolution reanalysis launched by the Copernicus Marine Service of the European Union is used, with the horizontal resolution set to 1 / 12° and the vertical direction divided into 50 layers.
3. The method for reconstructing the three-dimensional centerline of the Kuroshio Front based on attention mechanism correction according to claim 1, characterized in that: In step (1), the gradient The calculation formula is: ， Where T represents temperature.
4. The method for reconstructing the three-dimensional centerline of the Kuroshio Front based on attention mechanism correction according to claim 1, characterized in that: In step (1), the shallow water body is set to less than 30 meters, the middle water body to 30 to 650 meters, and the deep water body to more than 650 meters.
5. The method for reconstructing the three-dimensional centerline of the Kuroshio Front based on attention mechanism correction according to claim 1, characterized in that: In step (1), the frontal threshold for shallow and deep water is set to 0.045℃ / 0.02°, and the frontal threshold for mid-water is set to 0.09℃ / 0.02°.
6. The method for reconstructing the three-dimensional centerline of the Kuroshio Front based on attention mechanism correction according to claim 1, characterized in that: In step (3), on the grid plane at the same depth, only points whose difference in direction between the front line and streamline, and whose difference in trend direction between the front line and contour line does not exceed 45° are retained.
7. The method for reconstructing the three-dimensional centerline of the Kuroshio Front based on attention mechanism correction according to claim 1, characterized in that: In step (4), the angle is set to no more than 45° for shallow water within 30 meters, no more than 55° for water between 30 and 150 meters in the middle layer, no more than 70° for water between 150 and 300 meters, and no more than 90° for deeper water.