A cloud and local collaborative global ocean internal solitary wave detection method

Abstract

The application relates to the technical field of internal solitary wave detection and remote sensing image generation, and discloses a global internal solitary wave detection method based on cloud and local cooperation, which comprises the following steps: acquiring a GEE cloud platform COPERNICUS / S1_GRD_FLOAT dataset, and screening linear floating-point remote sensing data of a VV polarization channel in a Sentinel-1 SAR interference wide mode; performing land mask, radiation normalization and resampling pretreatment on the data in the GEE cloud platform to generate a standardized equivalent sea surface roughness image; extracting a fringe feature by using an edge detection operator to generate an edge feature binary image, screening out candidate image blocks containing internal solitary waves by grid blocking and edge density threshold judgment, and downloading the candidate image blocks to a local terminal; inputting the pre-trained EVA02-XGBoost hybrid model to remove false samples and misdetected samples, inputting the improved UTNet model to complete peak line detection, generating a pixel-level peak line mask image and outputting position information. The application can effectively improve the efficiency and recognition accuracy of internal solitary wave detection.

Description

Technical Field

[0001] This invention relates to the field of ocean solitary wave detection and remote sensing image generation technology, and more specifically, to a global ocean solitary wave detection method that combines cloud and local collaboration. Background Technology

[0002] Intra-oceanic solitary waves are important mesoscale dynamic processes occurring within the ocean's density stratification, significantly impacting ocean acoustic propagation, underwater navigation safety, marine ecology, and material transport. Synthetic Aperture Radar (SAR) is currently the core technology for large-scale global observation and monitoring of intra-oceanic solitary waves. Existing technologies for SAR image detection of intra-oceanic solitary waves mainly fall into two categories: traditional feature extraction methods and deep learning detection methods. Traditional methods often extract intra-oceanic solitary wave fringe features through edge detection operators and threshold segmentation to achieve preliminary identification. Deep learning methods improve identification accuracy through various segmentation networks and detection models. Some solutions even build single-end processing architectures to complete the entire process from data reading to parameter inversion. The output results mainly consist of numerical information such as the latitude and longitude coordinates and feature parameters of the intra-oceanic solitary waves, with only simple labeled visualization results. A standardized, high-precision image generation and processing chain has not yet been formed.

[0003] However, existing technologies for detecting isolated waves in the ocean still have several shortcomings in practical applications: At the processing architecture level, local processing of all data is subject to the computational pressure of massive SAR data, resulting in low processing efficiency and high deployment costs. Pure cloud processing is limited by computing resources and deployment conditions, making it difficult to run high-precision, large-scale segmentation models and failing to balance detection efficiency and segmentation accuracy. At the cloud preprocessing and initial screening level, existing solutions generally lack a refined radiation normalization correction step. Differences in backscattering coefficients under different incident angles can severely interfere with edge feature extraction, leading to high false detection and false negative rates in the initial screening stage. The transmission of invalid data increases bandwidth and local processing burden. At the visualization output level, existing technologies mostly focus on the numerical output of detection parameters and do not perform pixel-level image generation optimization for the wave crest line of isolated waves. The output labeled images have poor edge continuity and insufficient spatial registration accuracy, which cannot directly support the needs of refined analysis and operational monitoring and display of marine dynamic processes.

[0004] This reveals several technical problems in the existing technologies, including an inability to balance processing efficiency and accuracy, insufficient stability of initial screening features, and a lack of ability to generate high-precision pixel-level wave crest line images. Therefore, there is an urgent need for a cloud-based and local collaborative method for detecting solitary waves in the global ocean. This method would perform lightweight preprocessing and initial screening of candidate regions in the cloud, while achieving high-precision authenticity determination and pixel-level wave crest line segmentation locally. Simultaneously, it would generate spatially registered, visually accurate mask images with continuous stripe edges, effectively improving the accuracy and visualization output quality of solitary waves while ensuring efficient global data processing. Summary of the Invention

[0005] In view of this, the present invention proposes a cloud-based and local collaborative method for detecting isolated waves in the global ocean, aiming to solve the problems in the current technology where the processing architecture is difficult to balance detection efficiency and segmentation accuracy, the initial screening features are not stable enough, and there is a lack of high-precision pixel-level peak line visualization image generation capability.

[0006] This invention proposes a cloud-based and local collaborative method for detecting isolated waves in the global ocean, comprising the following steps: Obtain the COPERNICUS / S1_GRD_FLOAT dataset from the GEE cloud platform, and filter to obtain linear floating-point remote sensing data of the VV polarization channel in Sentinel-1 SAR interferometric wide swath mode; The linear floating-point remote sensing data is preprocessed using land masking, radiometric normalization, and resampling on the GEE cloud platform to generate a standardized equivalent sea surface roughness image. On the GEE cloud platform, an edge detection operator is used to extract stripe features from the standardized equivalent sea surface roughness image to generate a binary image of edge features. Based on grid segmentation and edge density threshold determination, candidate image blocks containing internal solitary waves are selected and downloaded to the local terminal. The candidate image blocks are input into the pre-trained EVA02-XGBoost hybrid model. After multi-scale feature extraction and classification decision, artifacts and false detection samples are removed to obtain the refined internal solitary wave image data. The refined internal solitary wave image data is input into a pre-trained improved UTNet model to complete the internal solitary wave crest line feature detection, and simultaneously generate a pixel-level internal solitary wave crest line mask image and output the crest line position information.

[0007] Furthermore, the specific execution process of the screening is as follows: Sentinel-1 SAR image data with interferometric wide-swath imaging mode were obtained by filtering from the COPERNICUS / S1_GRD_FLOAT dataset; The backscattering coefficient data of the VV polarization channel is extracted from the Sentinel-1 SAR image data of the interferometric wide swath mode; The original linear floating-point numerical format and geographic coordinate information of the backscattering coefficient data are retained to obtain linear floating-point remote sensing data for subsequent preprocessing.

[0008] Furthermore, the specific execution steps of the land mask are as follows: In the GEE cloud platform, the Hansen Global Surface Water dataset is called to extract the binary distribution data of water bodies in the target detection area, and the initial land-sea boundary line is generated based on the binary distribution data of water bodies. A buffer zone of a predetermined width is extended along the initial land-sea boundary line toward the sea area to cover the interference range of nearshore land-sea mixed pixels; Morphological dilation is performed on the boundary data after the extended buffer area to fill the scattered holes in the boundary and smooth the boundary contour, resulting in a continuous and closed effective sea area binary mask. The effective sea area binary mask is multiplied with the linear floating-point remote sensing data at the pixel level to remove interfering pixels in the land area, retaining only the effective remote sensing data within the sea area, thus obtaining the linear floating-point remote sensing data after land masking, completing the initial filtering of land-source noise.

[0009] Furthermore, the specific steps for implementing the radiation normalization are as follows: The local observation incident angle parameters corresponding to each effective sea area pixel in the linear floating-point remote sensing data after land masking are extracted pixel by pixel to obtain the actual observation angle information of each pixel. Using the CMOD5N geophysical model function, with fixed standard sea surface wind speed parameters as input conditions, first input the preset reference incident angle, and obtain the reference simulated value of the sea surface backscattering coefficient under the reference incident angle condition; Then, input the corresponding actual observed incident angle pixel by pixel to obtain the measured simulated value of the sea surface backscattering coefficient at each pixel observation angle; Obtain the ratio of the baseline simulated value to the measured simulated value of each pixel, and use the ratio as the radiometric correction coefficient for the corresponding pixel; The original backscattering intensity of the corresponding pixel is corrected pixel by pixel by using the radiation correction coefficient, and the backscattering data under different observation angles are unified to the same radiation reference to generate an equivalent intermediate image of sea surface roughness with radiation consistency.

[0010] Furthermore, the resampling is performed after the radiation normalization process is completed, and the specific steps are as follows: The equivalent sea surface roughness intermediate image is projected onto a unified geographic coordinate system to match the spatial reference benchmark of the target detection area. Bilinear resampling is performed on the projected image at a fixed spatial resolution, and speckle noise is suppressed by weighted averaging of neighboring pixels. By preserving the continuous morphological features of isolated inner wavy fringes, a standardized equivalent sea surface roughness image with consistent spatial resolution is obtained, which serves as a unified input benchmark for subsequent wavy feature extraction.

[0011] Furthermore, the edge detection operator is the Canny edge detection operator, and the specific steps for stripe feature extraction are as follows: Gaussian smoothing filtering is applied to the standardized equivalent sea surface roughness image to suppress random noise in the image, resulting in a smoothed sea surface roughness image. The gray-level gradient information is extracted along the horizontal and vertical directions of the smoothed image to obtain the gray-level gradient magnitude and gradient direction information of the corresponding pixels. Based on the gray-level gradient magnitude and gradient direction information, the continuous stripe boundaries corresponding to the inner solitary wave are screened out through non-maximum suppression and double threshold segmentation, generating a structured binary image of edge features.

[0012] Furthermore, the specific steps for determining the grid segmentation and edge density threshold are as follows: The binary image of the edge features is divided into multiple non-overlapping grid sub-blocks according to a fixed spatial size. Sub-blocks with insufficient size at the image edge are filled with zero values ​​to unify the spatial size of all sub-blocks. Calculate the percentage of edge pixels within each grid sub-block and obtain the edge density value of the corresponding sub-block; When the edge density of a sub-block is higher than the preset edge density judgment threshold, the sub-block is judged as a candidate image block containing internal solitary waves and included in the subsequent processing queue; when the edge density of a sub-block is lower than or equal to the edge density judgment threshold, the sub-block is judged as a background block without a target and is removed, thus completing the cloud-based lightweight target screening.

[0013] Furthermore, the EVA02-XGBoost hybrid model includes a feature preprocessing fusion module and an XGBoost decision module. The specific processing flow of this hybrid model is as follows: The candidate image block is subjected to multi-scale feature extraction by the EVA02 visual encoding module, and shallow texture features, mid-level structural features and high-level semantic features are obtained in sequence to generate a multi-channel depth feature image. The feature preprocessing fusion module performs statistical dimensionality reduction on the depth feature image, and performs numerical normalization and scaling on the dimensionality-reduced depth features and the texture statistical features pre-extracted for the candidate image blocks to unify the feature scale of the two; the normalized depth features and texture statistical features are concatenated by channels to obtain a multi-source complementary fusion feature set. The fused feature set is input into the XGBoost decision module in the EVA02-XGBoost hybrid model for true / false detection, artifacts and false detection samples are removed, and the refined internal solitary wave image data is output.

[0014] Furthermore, the improved UTNet model is configured with an input layer, an initial convolutional module, an encoder, a bottleneck layer, a semantic graph fusion module, a decoder, and an output layer in sequence; wherein the initial convolutional module is located at the input front of the encoder, and the encoder contains a neighborhood convolutional construction module and a bidirectional attention module, and the specific processing procedure is as follows: The input layer receives the finely examined internal solitary wave image data, completes the dimension alignment and size standardization of the input data, and outputs it to the initial convolution module. The initial convolutional module performs preliminary feature extraction on the image data output from the input layer to obtain a basic feature map. The neighborhood convolution construction module performs pixel-by-pixel neighborhood feature aggregation on the basic feature map using a 3×3 convolution kernel, captures the neighborhood association information of local stripes, and generates an initial global semantic feature map. The bidirectional attention module uses the initial global semantic feature map as the key and value, and calculates attention weights to achieve bidirectional information interaction and complementarity between the basic feature map and the initial global semantic feature map, thereby strengthening the long-range spatial continuity of the stripes and outputting a multi-scale semantic feature map.

[0015] Furthermore, the multi-scale semantic feature map output by the encoder is input into the bottleneck layer, and a global semantic feature map is generated through global long-range dependency modeling; The semantic graph fusion module fuses the multi-scale semantic feature map and the global semantic feature map based on learnable weights to generate a fused semantic graph. The decoder combines the fused semantic graph with the skip connection features retained by the encoder to upsample and restore the resolution step by step. After being output by the output layer, it generates a pixel-level peak line mask image that is spatially registered with the original image and outputs the peak line position information simultaneously.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: Firstly, it adopts a hierarchical processing architecture that combines cloud and local collaboration, which effectively solves the core contradiction that a single-end processing architecture cannot balance detection efficiency and segmentation accuracy. The cloud relies on the GEE platform to complete lightweight preprocessing of massive SAR data and initial screening and reduction of candidate regions, which greatly reduces the transmission bandwidth of invalid data and local computing power consumption, significantly improving the monitoring and processing timeliness of large-scale sea areas around the world. The local deployment of high-precision detection and segmentation models fully ensures the accuracy of internal solitary wave identification and wave crest line extraction. Secondly, a refined radiation normalization correction based on the CMOD5N model is introduced in the cloud preprocessing stage to eliminate the difference in backscattering coefficient caused by different observation incident angles, effectively improve the stability of subsequent edge feature extraction, reduce the probability of false detection and false negative detection in the initial screening stage, and further ensure the screening quality of candidate regions. Third, the improved UTNet model achieves pixel-level precise segmentation of the crest lines of internal isolated waves, and simultaneously generates a visual mask image with high spatial registration accuracy and smooth, continuous stripe edges. This overcomes the shortcomings of existing technologies that only output numerical parameters and have insufficient annotation image accuracy. The generated image results can directly support the refined analysis and operational monitoring and display of ocean dynamic processes. Overall, it takes into account the processing efficiency, detection accuracy, and visualization output quality of global large-scale monitoring, and provides reliable technical support for the routine observation and engineering application of internal isolated waves in the ocean. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of a cloud-based and local collaborative method for detecting isolated waves in the global ocean, provided as an embodiment of the present invention. Figure 2 The Sentinel-1 SAR remote sensing image before preprocessing provided in this embodiment of the invention; Figure 3 This is a preprocessed Sentinel-1 SAR remote sensing image provided in an embodiment of the present invention; Figure 4 This is a result image of the internal solitary wave generation region obtained using the edge density detection algorithm on the GEE cloud platform, provided by an embodiment of the present invention. Figure 5 The structure diagram of the EVA02-XGBoost model constructed for an embodiment of the present invention; Figure 6 The result diagram of the internal solitary wave generation region obtained by local fine-tuning using the EVA02-XGBoost model is provided for the embodiments of the present invention; Figure 7 A diagram of the improved UTNet model structure constructed for an embodiment of the present invention; Figure 8 The image shows the results of the internal isolated wave peak line detected locally using the improved UTNet model, as provided in this embodiment of the invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] In some embodiments of this application, see Figure 1 - Figure 8 This embodiment provides a cloud-based and local collaborative method for detecting solitary waves in the global ocean, including the following steps: S100: Obtain the COPERNICUS / S1_GRD_FLOAT dataset from the GEE (Google Earth Engine) cloud platform, and filter to obtain linear floating-point remote sensing data of VV polarization channels in Sentinel-1 SAR interferometric wide swath mode; S200: Perform land masking, radiometric normalization and resampling preprocessing on the linear floating-point remote sensing data on the GEE cloud platform to generate a standardized equivalent sea surface roughness image; S300: On the GEE cloud platform, the edge detection operator is used to extract stripe features from the standardized equivalent sea surface roughness image to generate a binary image of edge features; based on grid segmentation and edge density threshold determination, candidate image blocks containing internal solitary waves are selected and downloaded to the local terminal; S400: Input the candidate image patch into the pre-trained EVA02-XGBoost hybrid model, and after multi-scale feature extraction and classification decision, remove artifacts and false detection samples to obtain the refined internal solitary wave image data. S500: Input the refined internal solitary wave image data into the pre-trained improved UTNet model to complete the internal solitary wave peak line feature detection, and simultaneously generate a pixel-level internal solitary wave peak line mask image and output the peak line position information.

[0020] Specifically, see Figure 2In this embodiment, the data filtering process is completed within the GEE cloud platform, using the COPERNICUS / S1_GRD_FLOAT dataset as the data source and performing a three-level hierarchical filtering. First, initial filtering is performed based on the imaging mode metadata fields of the dataset, filtering out Sentinel-1 SAR image data with the interferometric wide-swath mode from the full dataset. This mode has a swath width of up to 250 km, enabling continuous coverage of global ocean areas. Its spatial resolution and swath width ratio can completely capture the characteristic information of isolated waves within the ocean at different scales, meeting the observation needs of isolated wave surveys on a global scale. Second, backscattering coefficient data of the VV polarization channel is extracted from the filtered interferometric wide-swath image data. The VV polarization mode is more sensitive to changes in sea surface gravity capillary waves. Under the micro-scale modulation of the sea surface induced by internal isolated waves, it can exhibit more significant backscattering differences, possessing a higher signal-to-noise ratio compared to HH horizontal polarization, effectively reducing background noise interference under complex sea conditions. Finally, the original linear floating-point numerical format of the backscattering coefficient data and the corresponding geographic coordinate information are retained without additional logarithmic transformation or numerical compression. The original linear physical quantities of the SAR data are fully preserved, the true scattering characteristics of isolated waves in the ocean are restored to the maximum extent, and the quantization error introduced by numerical transformation is avoided. The final result is linear floating-point remote sensing data for subsequent preprocessing.

[0021] Understandably, the standardized input data obtained through hierarchical screening not only ensures the breadth of global ocean coverage, but also effectively improves the identification of internal solitary wave target features through polarization selection. The linear floating-point format fully preserves the accuracy of the original physical quantities, providing a unified and high-quality data benchmark for subsequent full-process preprocessing and feature extraction.

[0022] Specifically, see Figure 2In this embodiment, the land masking process relies on the parallel computing capabilities of the GEE cloud platform to remove strong scattering interference from land areas and purify the effective marine data. First, the Hansen Global Surface Water dataset is accessed on the GEE cloud platform. This dataset, constructed based on Landsat satellite imagery from 1984-2021, has a spatial resolution of 30m, accurately depicting the land-sea boundary and the extent of permanent water bodies. Binary water distribution data of the target detection area is extracted, and the land-sea boundary contour is extracted based on this data to generate an initial land-sea boundary line. Second, a buffer zone of a preset width is extended along the initial land-sea boundary line towards the marine area. In this embodiment, the buffer width is set to 1km to completely cover the interference range of nearshore mixed land-sea pixels, preventing coastal mixed pixels from being misjudged as sea surface stripe features. Third, morphological dilation processing is performed on the boundary data after the buffer zone is extended to fill scattered holes at the boundary and smooth the boundary contour, eliminating interference from fragmented edges and obtaining a continuous and closed effective marine binary mask. Finally, the effective sea area binary mask is multiplied with the linear floating-point remote sensing data at the pixel level, the pixel values ​​of the land area are set to zero, and only the effective remote sensing data within the sea area is retained, thus obtaining the linear floating-point remote sensing data after land masking, completing the preliminary filtering of land source noise.

[0023] Understandably, by generating a land-sea mask using a high-precision surface water dataset and extending the nearshore buffer zone, and by combining morphological processing to ensure the continuous and smooth boundary of the mask, the interference of strong nearshore scattering clutter and mixed land-sea pixels can be effectively eliminated, and the influence of land-source signals on the extraction of internal solitary wave features can be significantly reduced.

[0024] Specifically, see Figure 3 In this embodiment, the radiation normalization processing is performed pixel-by-pixel based on the CMOD5N geophysical model function. The aim is to eliminate the deviation in sea surface backscattering intensity caused by differences in observed incident angles and to unify the radiation benchmark across the entire map. First, the local observed incident angle parameters corresponding to each effective sea area pixel in the linear floating-point remote sensing data after land masking are extracted pixel-by-pixel to obtain the actual observed angle information for each pixel. This incident angle parameter is the geometric metadata layer inherent in the Sentinel-1 SAR data and can be directly matched and called pixel-by-pixel. Second, the CMOD5N geophysical model function is used, with a fixed standard sea surface wind speed parameter as the input condition. In this embodiment, the reference wind speed is set to 10 m / s. A preset reference incident angle is first input, using the median incident angle of 37.5° from the Sentinel-1 data. The model calculation obtains the benchmark simulated value of the sea surface backscattering coefficient under the reference incident angle condition. Third, the actual observed incident angle of the corresponding pixel is input pixel-by-pixel to obtain the measured simulated value of the sea surface backscattering coefficient under each pixel's observation angle. The formula for calculating the equivalent sea surface roughness is as follows: , in, For equivalent sea surface roughness, The backscattering intensity at the linear scale of the original sea surface in SAR. The simulated backscattering intensity from the sea surface. For reference wind speed, For reference incident angle, To observe the incident angle, the ratio of the baseline simulated value to the measured simulated value of each pixel is obtained, and this ratio is used as the radiometric correction coefficient for the corresponding pixel. The original backscattering intensity of the corresponding pixel is corrected pixel-by-pixel using this radiometric correction coefficient, unifying the backscattering data from different observation angles to the same radiometric reference, and generating an equivalent intermediate image of sea surface roughness with radiometric consistency. Furthermore, this embodiment incorporates an adaptive quality control step during batch processing: for data with a low proportion of effective ocean pixels or a percentile difference close to 0 in a single image, the image grayscale range is conservatively scaled to 0-0.025 using empirical values ​​to reduce differences caused by variations in the quality of different images themselves, ensuring stable contrast of bright and dark fringes of internal isolated waves in different images.

[0025] Understandably, pixel-by-pixel radiative normalization correction based on the CMOD5N model can effectively eliminate the backscattering intensity deviation caused by differences in the observed incident angle, unify the radiative benchmark of images from different orbits and perspectives, significantly improve the consistency of the inner isolated fringe features, and provide a stable and reliable image input for subsequent edge detection and density determination.

[0026] Specifically, see Figure 3 Resampling is performed after radiometric normalization to unify the spatial resolution of the entire imagery while suppressing speckle noise. First, the equivalent sea surface roughness intermediate image generated by radiometric normalization is projected onto a unified geographic coordinate system to match the spatial reference benchmark of the target detection area, eliminating differences in projection coordinate systems between different images and ensuring spatial consistency in subsequent global-scale processing. Second, center point resampling is performed at 150m intervals using bilinear interpolation. The values ​​of each pixel after resampling are calculated using a weighted average interpolation method based on neighboring pixels. After resampling, the raster center point is strictly aligned with the geographic coordinates, physically reducing the interference of speckle noise on weak fringe features and avoiding jagged edge distortion caused by nearest-neighbor resampling. During resampling, the continuous morphological features of isolated inner fringes are preserved. The smoothing properties of bilinear interpolation maintain a natural transition of fringe boundaries, avoiding fringe breaks and loss of detail. Finally, a standardized equivalent sea surface roughness image with consistent spatial resolution is obtained, serving as a unified input benchmark for subsequent fringe feature extraction.

[0027] Understandably, bilinear resampling with uniform resolution not only achieves spatial scale standardization of the entire image, but also effectively suppresses SAR speckle noise, while preserving the continuous morphology of the inner isolated fringes, providing a standardized input with uniform scale and controllable noise for subsequent cloud edge detection and block determination.

[0028] Specifically, see Figure 3 This embodiment employs the Canny edge detection operator for stripe feature extraction, aiming to extract the stripe boundaries of internal isolated waves from a standardized image and generate binarized edge features. First, Gaussian smoothing filtering is applied to the standardized equivalent sea surface roughness image. In this embodiment, the Gaussian smoothing parameter is set to 1.2 pixels. The weighted smoothing effect of the Gaussian kernel suppresses random noise and speckle residue in the image, balancing edge detail preservation with noise suppression, and avoiding the loss of weak stripe features due to over-smoothing, resulting in a smoothed sea surface roughness image. Second, gray-level gradient information is extracted from the smoothed image along both the horizontal and vertical directions. The gray-level change rate of each pixel in both directions is calculated using a first-order differential operator, obtaining the gray-level gradient magnitude and gradient direction information of the corresponding pixel, accurately representing the intensity and direction of the stripe edges. The Canny calculation formula is as follows: , in, coordinates gradient magnitude at that point This represents the rate of change of grayscale in the image along the x-direction. This represents the rate of grayscale change in the y-direction of the image. Based on the obtained grayscale gradient magnitude and gradient direction information, non-maximum suppression and double threshold segmentation are performed sequentially: non-maximum suppression refines the edge width, retaining the pixels corresponding to the local gradient maximum, thus refining the edges to a single pixel width; high and low double thresholds are used to retain continuous strong edge pixels and remove isolated weak noise edges, ultimately selecting the continuous stripe boundaries corresponding to isolated inner waves, generating a structured binary image of edge features.

[0029] Understandably, the optimized Canny edge detection can accurately extract the fringe boundaries of internal solitary waves, balancing noise suppression and detail preservation. The generated binary edge images have clear structures and continuous boundaries, providing a reliable feature basis for subsequent block edge density determination.

[0030] Specifically, see Figure 4This embodiment uses grid segmentation and edge density thresholding to perform initial cloud screening, aiming to quickly remove background areas without targets and reduce data volume. First, the binary image of the edge features is divided into multiple non-overlapping grid sub-blocks with a fixed spatial size. In this embodiment, the spatial scale of a single grid sub-block is set to 10.5 km, corresponding to 70×70 pixels in a 150m resolution Sentinel-1 image. This scale matches the typical spatial distribution characteristics of isolated waves within an image and can completely cover the range of a single isolated wave train. Sub-blocks with insufficient size at the image edges are filled with zero values ​​to unify the spatial size of all sub-blocks and ensure standardized density calculations. Second, the proportion of edge pixels within each grid sub-block is statistically analyzed. The edge density value of the corresponding sub-block is obtained by calculating the ratio of the number of edge pixels to the total number of pixels. The formula for calculating the block edge density is as follows: , in, G represents the grid edge density, and G represents the target detection grid. coordinates Binary markers at the edges, This represents the total number of pixels within the grid. In a binary image of edge features, background pixels have a value of 0, while edge pixels have a value that is non-zero. Therefore, this density value reflects both the proportion of edge pixels within a sub-block and the edge intensity. In this embodiment, the preset edge density judgment threshold is 0.014. When the edge density of a sub-block is higher than the preset edge density judgment threshold, the sub-block is judged as a candidate image block containing internal solitary waves and included in the subsequent processing queue. When the edge density of a sub-block is lower than or equal to the edge density judgment threshold, the sub-block is judged as a background block without a target and is removed. This completes the lightweight target initial screening in the cloud, and the selected candidate image blocks are downloaded to the local terminal.

[0031] Understandably, the standardized grid segmentation combined with the density determination method based on a fixed threshold has low computational complexity and strong adaptability to cloud deployment. It can quickly complete the initial screening and reduction of large-scale images, significantly reduce the amount of data transmitted to areas without targets, and effectively reduce the data processing pressure and transmission bandwidth consumption on the local end while ensuring the target recall rate.

[0032] Specifically, see Figure 5 The EVA02-XGBoost hybrid model executes a three-stage process: feature extraction, feature preprocessing and fusion, and classification. Its purpose is to refine the authenticity of candidate regions initially screened in the cloud, eliminating artifacts and false positives. First, the EVA02 visual encoding module extracts multi-scale features from the candidate image blocks. The network then sequentially acquires shallow texture features, mid-level structural features, and high-level semantic features through layer-by-layer convolution operations, generating a multi-channel depth feature image to fully mine the multi-dimensional feature information of the candidate image.

[0033] Secondly, the depth feature image is statistically reduced using a feature preprocessing fusion module. The reduced depth features and the pre-extracted texture statistical features (i.e., shallow artificial visual features) for the candidate image patches are then numerically normalized and scaled to ensure uniformity in feature scale. The first step uses Principal Component Analysis (PCA) for feature dimensionality reduction, retaining high-contribution feature dimensions to reduce subsequent computation. The calculation formula is as follows: , , In the formula, The global image features output by EVA02 The global mean of the global image features. These are the dimensionality-reduced deep semantic features; The covariance matrix of visual features. To preserve the projection matrix of principal components with high contribution rates, the covariance matrix is... The unit eigenvectors corresponding to the first k largest eigenvalues, where k is the feature dimension after dimensionality reduction of the target.

[0034] The second step uses the Min-Max linear normalization method to perform global numerical scaling on the deep semantic features and texture statistical features, mapping the two types of features to a unified numerical range of [0,1], thereby achieving uniformity in the scale of all input features. The calculation formula is as follows: , in, These are the normalized standard eigenvalues. For any original feature value, The maximum value of the feature across the entire sample. The minimum value of the feature across the entire sample.

[0035] The third step uses a channel-dimensional concatenation method to achieve bidirectional complementary fusion of deep semantic features and texture statistical features. The calculation formula is as follows: , in, The feature set after splicing and fusion. (·) represents the channel dimension splicing operation. For deep semantic features, These are texture statistical features.

[0036] Finally, the fused feature set is input into the XGBoost decision module of the EVA02-XGBoost hybrid model for true / false detection. Through feature splitting and weight learning of the decision tree, it distinguishes internal solitary wave targets from background interference such as sea surface clutter and ocean current textures, removes artifacts and false detection samples, and outputs refined internal solitary wave image data. The effective region of the internal solitary wave after true / false detection and artifact removal is shown below. Figure 6 As shown.

[0037] Understandably, the fusion of deep features and manually statistical features can achieve complementary advantages between the two types of features. Combined with the strong nonlinear fitting capability of the XGBoost decision module, it can effectively eliminate various artifact interferences, significantly improve the accuracy of internal solitary wave detection, and maintain a fast inference speed, thus meeting the application requirements of local batch fine detection.

[0038] To verify the detection performance of the EVA02-XGBoost hybrid model in this embodiment, a five-fold cross-validation experiment was conducted beforehand based on the labeled internal solitary wave sample dataset. The sample set was randomly divided into five groups proportionally, with four groups selected as the training set and one group as the validation set each time. This process was repeated five times, and the average performance index was taken to ensure the generalization and reliability of the validation results. During the training phase, to address the issue of limited sample quantity, data augmentation methods such as random horizontal flipping and multi-scale random pruning were used to expand the training samples and improve the robustness of the model. To address the issues of low positive sample ratio and class imbalance in internal solitary waves, a weighted loss function was used to allocate sample weights, weakening the interference of uneven sample distribution on model convergence and ensuring the model's ability to identify rare targets.

[0039] The quantitative evaluation results of this experiment are shown in Table 1.

[0040] Table 1. Quantitative evaluation results based on the EVA02-XGBoost model Experimental results show that the model can achieve an inference capacity of 32 images per batch on a local workstation with conventional computing power, and the average inference time for a single candidate image block is about 15ms, which can meet the needs of rapid processing of batch data. The average detection accuracy, precision, recall and F1 score of the five-fold cross-validation are all excellent, with a significant improvement in precision, which can effectively reduce the false detection rate of artifacts. It ensures the accuracy of target recognition while taking into account detection efficiency, and is suitable for rapid and precise detection scenarios with massive candidate data.

[0041] Specifically, see Figure 7The improved UTNet model in this embodiment consists of an input layer, an initial convolutional module, an encoder, a bottleneck layer, a semantic graph fusion module, a decoder, and an output layer. First, the input layer receives the refined internal solitary wave image data, performs dimension alignment and size normalization of the input tensor, and outputs image data matching the model's input specifications. The encoder then includes a neighborhood convolutional construction module and a bidirectional attention module, executing a two-level feature aggregation process to enhance the local neighborhood association and long-range spatial continuity of the stripe features. First, the initial convolutional module performs preliminary feature extraction on the refined internal solitary wave image data, capturing the edge, texture, and stripe features of the internal solitary waves without reducing the feature map resolution, while suppressing image noise and obtaining a fine-grained basic feature map. Secondly, the neighborhood convolution construction module, namely the 3×3 semantic map module in this embodiment, performs pixel-by-pixel neighborhood feature aggregation on the basic feature map using 3×3 convolution kernels. It clusters the texture, edge, and grayscale correlation information within a 3×3 area surrounding each pixel, capturing the neighborhood correlation information of local stripes, constructing an initial global semantic feature map, and widening the receptive field of the shallow network. The calculation formula for semantic map convolution is as follows: , In the formula, For the initial global semantic graph, (·) represents a 3×3 convolution operation. The base feature map output by the initial convolutional module. For semantic graph convolution kernel weights, This is the bias term. Furthermore, the bidirectional attention module uses the initial global semantic feature map as the key and value, and calculates the attention weights through a scaled dot product, achieving bidirectional information interaction and complementarity between the basic feature map and the initial global semantic feature map. This transmits global contextual information to each local pixel, strengthens the long-range spatial continuity of the stripes, and outputs a multi-scale semantic feature map. The calculation formula for two-dimensional bidirectional multi-head attention is as follows: , The single-head fusion feature is calculated as follows: .

[0042] The final bidirectional fusion enhancement feature output by the module is as follows: .

[0043] in, Guide semantic graph attention for the i-th feature map. To guide feature map attention for the i-th semantic map, Softmax is the normalized activation function. The query vector for the i-th feature map. For semantic graph query vectors, For semantic graph key vectors, For feature map key vectors, Let i be the value vector of the i-th feature map. For semantic graph value vectors, For transpose operation, This is the scaling factor; It is a single-head fusion feature. This represents the i-th independent attention head. For the number of attention heads, For feature splicing; The final output of the module is a multi-scale semantic feature map. This is a layer normalization operation.

[0044] Understandably, the initial semantic map constructed by neighborhood convolution, combined with the bidirectional attention mechanism, can simultaneously strengthen the association of local neighborhood features and the modeling of long-range spatial dependencies, effectively improving the continuity and integrity of isolated wavy stripe features and solving the problem of insufficient modeling of long stripe features by the original model.

[0045] Specifically, the decoding and generation process of the improved UTNet model is executed in three stages: bottleneck layer global long-range modeling, semantic graph fusion, and decoder-level reconstruction. The aim is to generate pixel-level peak line masks and output accurate peak line position information. First, the multi-scale semantic feature map output by the encoder is input to the bottleneck layer of the improved UTNet model. Through global long-range dependency modeling, higher-order global semantic information is further refined to accurately distinguish between target, background, and clutter regions, generating a global semantic feature map. Second, through the semantic graph fusion module of the improved UTNet model, the multi-scale semantic features of the encoder and the global semantic features of the bottleneck layer are dynamically fused based on learnable weights. The weight ratio of local detail information and global context information is adaptively balanced to generate a fused semantic map that takes into account both local details and global context. The calculation formula for semantic graph fusion is as follows: , in, The final output of the semantic graph fusion module is the fused semantic graph. For layer normalization operation, These are learnable weight coefficients. This is the global semantic feature map output by the bottleneck layer. The encoder outputs a multi-scale semantic feature map. Finally, the decoder of the improved UTNet model combines the fused semantic map with the multi-level skip connection features retained by the encoder, gradually restoring the image resolution through progressive upsampling. During decoding, it fuses the shallow detail features retained by the encoder to compensate for the loss of detail information caused by downsampling. Then, it performs channel mapping and activation processing at the final output layer to achieve high-precision generation of the inner isolated wave crest image. Finally, it generates a pixel-level wave crest mask image spatially registered with the original image, simultaneously outputting the wave crest position information. The final pixel-level wave crest mask detection result is as follows: Figure 8 As shown.

[0046] Understandably, the semantic graph fusion combined with the decoding structure of multi-level skip connections can simultaneously preserve the local edge details and global spatial orientation of the stripes. The resulting pixel-level mask has continuous edges and high spatial registration accuracy, which can directly support subsequent marine dynamic analysis and operational visualization, effectively improving the overall accuracy of wave crest line extraction.

[0047] To verify the peak line segmentation accuracy of the improved UTNet model in this embodiment, three mainstream segmentation models in this field—U-Net, SegFormer, and DeepLabV3—were selected as baseline models. A comparative experiment was conducted using the same test dataset and evaluation criteria. The core evaluation metrics were uniformly set as F1 score, precision, recall, and mean intersection-over-union ratio (mIoU). mIoU measures the spatial overlap between the segmented region and the labeled ground truth, while the F1 score comprehensively reflects the completeness and accuracy of the segmentation.

[0048] The evaluation index results for each model are shown in Table 2.

[0049] Table 2 Evaluation metrics for each model The original U-Net employs a typical encoder-decoder symmetric structure, which has the advantage of preserving shallow detail features through skip connections. Experimental results show that the model achieves a recall rate of 91.76%, which is relatively high among the four types of models. It has good detection capability for isolated wavy stripe targets and can capture the target region relatively completely. However, U-Net mainly relies on convolutional operations and does not introduce a global context modeling mechanism. In complex ocean backgrounds, it is difficult to effectively distinguish between background clutter and real stripes, resulting in an accuracy of only 69.59% and a relatively significant false detection phenomenon.

[0050] The improved UTNet model in this embodiment achieves optimal results across all evaluation metrics, with an F1 score of 84.89%, precision of 82.00%, recall of 88.00%, and mean intersection-over-union (MIU) of 73.74%. These data demonstrate that this model strengthens local feature aggregation through neighborhood convolutional construction modules and achieves bidirectional information interaction between global semantics and local features through a bidirectional attention mechanism. It specifically addresses the issues of insufficient long-range continuity of long stripes and ambiguous edge localization, significantly reducing the false detection rate while maintaining target detection capability, thus verifying the effectiveness and superiority of this technical solution.

[0051] In a specific embodiment of this application, the above steps are implemented in the following ways: The scenario involved routine remote sensing observations of isolated waves in marginal sea areas. Researchers used the COPERNICUS / S1_GRD_FLOAT dataset on the GEE cloud platform to select interferometric wide-swath mode VV-polarized linear floating-point SAR data covering the target sea area within the target observation period, using this data as the input data source for the entire processing workflow. Subsequently, standardized preprocessing was performed in the cloud: a land-sea boundary mask was generated using the Hansen Global Surface Water dataset; a 1km buffer was extended along the coastline towards the sea and morphologically dilated to remove interference from mixed pixels between land and nearshore areas; the CMOD5N geophysical model function was used to calculate the radiometric correction coefficient pixel-by-pixel with a reference incident angle of 37.5° and a reference wind speed of 10m / s, and the sea surface backscattering intensity was normalized to generate an equivalent intermediate sea surface roughness image; the corrected image was then reprojected to a unified geographic coordinate system, and bilinear resampling was performed at a resolution of 150m to suppress speckle noise and unify the data spatial scale, resulting in a standardized equivalent sea surface roughness image.

[0052] After preprocessing, a lightweight initial screening is performed in the cloud: the Canny operator with a Gaussian smoothing parameter of 1.2 pixels is used to extract stripe edge features and generate a binary image of edge features; the image is divided into non-overlapping grids at a spatial scale of 10.5km, and zero-value padding is performed on sub-blocks with insufficient edge size. The edge density is calculated grid by grid and a threshold of 0.014 is used to screen out candidate image blocks containing internal solitary wave features. Only the reduced candidate data is downloaded to the local processing terminal.

[0053] The local terminal first performs a true / false detection using the EVA02-XGBoost hybrid model, extracts multi-scale depth features and normalizes and stitches them with texture statistical features, and then removes artifact interference such as wind stripes and ship wakes. Then, it inputs the improved UTNet model to complete pixel-level segmentation, generates a wave crest mask image that is spatially registered with the original image, and outputs the corresponding geographic coordinates simultaneously, finally forming the observation results of the inner solitary wave in the target sea area.

[0054] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0055] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0056] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0057] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0058] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A cloud-based and local collaborative method for detecting isolated waves in the global ocean, characterized in that, Includes the following steps: Obtain the COPERNICUS / S1_GRD_FLOAT dataset from the GEE cloud platform, and filter to obtain linear floating-point remote sensing data of the VV polarization channel in Sentinel-1 SAR interferometric wide swath mode; The linear floating-point remote sensing data is preprocessed using land masking, radiometric normalization, and resampling on the GEE cloud platform to generate a standardized equivalent sea surface roughness image. The edge detection operator is used on the GEE cloud platform to extract stripe features from the standardized equivalent sea surface roughness image and generate a binary image of edge features. Based on grid segmentation and edge density threshold determination, candidate image blocks containing internal solitary waves are selected and downloaded to the local terminal; The candidate image blocks are input into the pre-trained EVA02-XGBoost hybrid model. After multi-scale feature extraction and classification decision, artifacts and false detection samples are removed to obtain the refined internal solitary wave image data. The refined internal solitary wave image data is input into a pre-trained improved UTNet model to complete the internal solitary wave crest feature detection, and simultaneously generate a pixel-level internal solitary wave crest mask image and output the crest position information.

2. The cloud-based and local collaborative global ocean solitary wave detection method according to claim 1, characterized in that, The specific execution process of the screening is as follows: Sentinel-1 SAR image data with interferometric wide-swath imaging mode were obtained by filtering from the COPERNICUS / S1_GRD_FLOAT dataset; The backscattering coefficient data of the VV polarization channel is extracted from the Sentinel-1 SAR image data of the interferometric wide swath mode; The original linear floating-point numerical format and geographic coordinate information of the backscattering coefficient data are retained to obtain linear floating-point remote sensing data for subsequent preprocessing.

3. The cloud-based and local collaborative method for detecting isolated waves in the global ocean as described in claim 1, characterized in that, The specific steps for implementing the land mask are as follows: In the GEE cloud platform, the Hansen Global Surface Water dataset is called to extract the binary distribution data of water bodies in the target detection area, and the initial land-sea boundary line is generated based on the binary distribution data of water bodies. A buffer zone of a predetermined width is extended along the initial land-sea boundary line toward the sea area to cover the interference range of nearshore land-sea mixed pixels; Morphological dilation is performed on the boundary data after the extended buffer area to fill the scattered holes in the boundary and smooth the boundary contour, resulting in a continuous and closed effective sea area binary mask. The effective sea area binary mask is multiplied with the linear floating-point remote sensing data at the pixel level to remove interfering pixels in the land area, retaining only the effective remote sensing data within the sea area, thus obtaining the linear floating-point remote sensing data after land masking, completing the initial filtering of land-source noise.

4. The cloud-based and local collaborative method for detecting isolated waves in the global ocean as described in claim 1, characterized in that, The specific steps for implementing radiation normalization are as follows: The local observation incident angle parameters corresponding to each effective sea area pixel in the linear floating-point remote sensing data after land masking are extracted pixel by pixel to obtain the actual observation angle information of each pixel. Using the CMOD5N geophysical model function, with fixed standard sea surface wind speed parameters as input conditions, first input the preset reference incident angle, and obtain the reference simulated value of the sea surface backscattering coefficient under the reference incident angle condition; Then, input the corresponding actual observed incident angle pixel by pixel to obtain the measured simulated value of the sea surface backscattering coefficient at each pixel observation angle; Obtain the ratio of the baseline simulated value to the measured simulated value of each pixel, and use the ratio as the radiometric correction coefficient for the corresponding pixel; The original backscattering intensity of the corresponding pixel is corrected pixel by pixel by using the radiation correction coefficient, and the backscattering data under different observation angles are unified to the same radiation reference to generate an equivalent intermediate image of sea surface roughness with radiation consistency.

5. The cloud-based and local collaborative method for detecting isolated waves in the global ocean as described in claim 1, characterized in that, The resampling is performed after the radiation normalization process is completed. The specific steps are as follows: The equivalent sea surface roughness intermediate image is projected onto a unified geographic coordinate system to match the spatial reference benchmark of the target detection area. Bilinear resampling is performed on the projected image at a fixed spatial resolution, and speckle noise is suppressed by weighted averaging of neighboring pixels. By preserving the continuous morphological features of isolated inner wavy fringes, a standardized equivalent sea surface roughness image with consistent spatial resolution is obtained, which serves as a unified input benchmark for subsequent wavy feature extraction.

6. The cloud-based and local collaborative method for detecting isolated waves in the global ocean as described in claim 1, characterized in that, The edge detection operator is the Canny edge detection operator, and the specific steps for its stripe feature extraction are as follows: Gaussian smoothing filtering is applied to the standardized equivalent sea surface roughness image to suppress random noise in the image, resulting in a smoothed sea surface roughness image. The gray-level gradient information is extracted along the horizontal and vertical directions of the smoothed image to obtain the gray-level gradient magnitude and gradient direction information of the corresponding pixels. Based on the gray-level gradient magnitude and gradient direction information, the continuous stripe boundaries corresponding to the inner solitary wave are screened out through non-maximum suppression and double threshold segmentation, generating a structured binary image of edge features.

7. The cloud-based and local collaborative method for detecting isolated waves in the global ocean as described in claim 1, characterized in that, The specific steps for determining the grid segmentation and edge density threshold are as follows: The binary image of the edge features is divided into multiple non-overlapping grid sub-blocks according to a fixed spatial size. Sub-blocks with insufficient size at the image edge are filled with zero values ​​to unify the spatial size of all sub-blocks. Calculate the percentage of edge pixels within each grid sub-block and obtain the edge density value of the corresponding sub-block; When the edge density of a sub-block is higher than the preset edge density judgment threshold, the sub-block is judged as a candidate image block containing internal solitary waves and included in the subsequent processing queue; when the edge density of a sub-block is lower than or equal to the edge density judgment threshold, the sub-block is judged as a background block without a target and is removed, thus completing the cloud-based lightweight target screening.

8. The cloud-based and local collaborative method for detecting isolated waves in the global ocean as described in claim 1, characterized in that, The EVA02-XGBoost hybrid model includes a feature preprocessing fusion module and an XGBoost decision module. The specific processing flow of this hybrid model is as follows: The candidate image block is subjected to multi-scale feature extraction by the EVA02 visual encoding module, and shallow texture features, mid-level structural features and high-level semantic features are obtained in sequence to generate a multi-channel depth feature image. The feature preprocessing fusion module performs statistical dimensionality reduction on the depth feature image, and performs numerical normalization and scaling on the dimensionality-reduced depth features and the texture statistical features pre-extracted for the candidate image blocks to unify the feature scale of the two; the normalized depth features and texture statistical features are concatenated by channels to obtain a multi-source complementary fusion feature set. The fused feature set is input into the XGBoost decision module in the EVA02-XGBoost hybrid model for true / false detection, artifacts and false detection samples are removed, and the refined internal solitary wave image data is output.

9. The cloud-based and local collaborative method for detecting isolated waves in the global ocean as described in claim 1, characterized in that, The improved UTNet model is configured with an input layer, an initial convolutional module, an encoder, a bottleneck layer, a semantic graph fusion module, a decoder, and an output layer. The initial convolutional module is located at the input front of the encoder, which includes a neighborhood convolutional construction module and a bidirectional attention module. The specific processing procedure is as follows: The input layer receives the finely examined internal solitary wave image data, completes the dimension alignment and size standardization of the input data, and outputs it to the initial convolution module. The initial convolutional module performs preliminary feature extraction on the image data output from the input layer to obtain a basic feature map. The neighborhood convolution construction module performs pixel-by-pixel neighborhood feature aggregation on the basic feature map using a 3×3 convolution kernel, captures the neighborhood association information of local stripes, and generates an initial global semantic feature map. The bidirectional attention module uses the initial global semantic feature map as the key and value, and calculates attention weights to achieve bidirectional information interaction and complementarity between the basic feature map and the initial global semantic feature map, thereby strengthening the long-range spatial continuity of the stripes and outputting a multi-scale semantic feature map.

10. The cloud-based and local collaborative method for detecting isolated waves in the global ocean as described in claim 9, characterized in that, The multi-scale semantic feature map output by the encoder is input into the bottleneck layer, and high-order global semantic information is extracted through global long-range dependency modeling to generate a global semantic feature map. The semantic graph fusion module of the improved UTNet model fuses the multi-scale semantic features of the encoder with the global semantic features of the bottleneck layer based on learnable weights to generate a fused semantic graph that takes into account both local details and global context. The decoder combines the fused semantic graph with the skip connection features retained by the encoder to upsample and restore the resolution step by step. After being output by the output layer, it generates a pixel-level peak line mask image that is spatially registered with the original image and outputs the peak line position information simultaneously.

Images