Method, system, device, medium and product for identifying marine mesoscale phenomena based on transfer learning

Abstract

The application discloses a kind of based on transfer learning's marine mesoscale phenomenon identification method, system, equipment, medium and product, it is related to marine mesoscale phenomenon identification technical field, the method includes: obtaining the satellite SAR marine image of current country;The satellite SAR marine image of current country is preprocessed, and satellite SAR marine image after processing is obtained;Satellite SAR marine image after processing is input into marine mesoscale phenomenon identification model, and identification result is obtained;Identification result includes oceanic internal wave and mesoscale eddy;Marine mesoscale phenomenon identification model is obtained using training set to YOLOv11 model migration training;Training set includes the sample satellite SAR marine image of current country and corresponding sample identification result, the accuracy of marine mesoscale phenomenon can be improved by the application.

Description

Technical Field

[0001] This application relates to the field of marine sub-mesoscale phenomenon recognition technology, and in particular to a method, system, device, medium and product for marine sub-mesoscale phenomenon recognition based on transfer learning. Background Technology

[0002] Mesoscale processes are widespread dynamic processes in the ocean's mixing layer, making significant contributions to ocean energy cascading, mass and heat transport. Oceanic mesoscale phenomena are numerous, widely distributed, highly energetic, and strongly entrained, making them ideal for studying oceanic material cycling, energy cascading, and sphere coupling. The full life-cycle tracking and observation of oceanic mesoscale phenomena has become one of the most important advances in ocean remote sensing since the 21st century.

[0003] In some cases, over the past half-century, a series of remote sensing technologies and methods for ocean sub-mesoscale phenomena have been developed using various visible light, infrared, and microwave sensors, which have greatly promoted the development of ocean science and even Earth system science. The principles include the following four categories: (1) Positive and negative anomalies in seawater temperature at the center of a vortex and significant gradient changes in seawater temperature at a front can be detected by spaceborne infrared sensors; (2) Vortexes and fronts have strong material entrainment and transport capabilities, causing physical, biological, and chemical substances such as chlorophyll, organic carbon, plankton, and even sea ice to rotate and migrate, thus being perceived by spectrometers with certain underwater penetration capabilities and synthetic aperture radar (SAR) with sea surface imaging capabilities; (3) The topology of a vortex sea surface presents a relatively stable closed structure, and the center of a warm vortex is the maximum value of sea surface height (SSH), while the center of a cold vortex is the minimum value, thus being identified and tracked by radar altimeters; (4) Seawater flows related to vortices, fronts, and internal waves can induce upwelling and downwelling flows of surface seawater divergence and convergence, changing sea surface roughness and SSH, thus being captured by sensors such as synthetic aperture radar (SAR) and radar altimeters.

[0004] However, for autonomous satellite SAR ocean images with small sample sizes, there is currently no method that can directly detect and identify sub-mesoscale phenomena in these images. Summary of the Invention

[0005] The purpose of this application is to provide a method, system, device, medium, and product for identifying marine sub-mesoscale phenomena based on transfer learning, which can improve the accuracy of identifying marine sub-mesoscale phenomena.

[0006] To achieve the above objectives, this application provides the following solution.

[0007] In a first aspect, this application provides a method for identifying sub-mesoscale ocean phenomena based on transfer learning, comprising: acquiring satellite SAR ocean images of the current country; preprocessing the satellite SAR ocean images of the current country to obtain processed satellite SAR ocean images; inputting the processed satellite SAR ocean images into a sub-mesoscale ocean phenomenon identification model to obtain identification results; the identification results include ocean internal waves and sub-mesoscale eddies; the sub-mesoscale ocean phenomenon identification model is obtained by transfer training of a YOLOv11 model using a training set; the training set includes sample satellite SAR ocean images of the current country and corresponding sample identification results.

[0008] Secondly, this application provides a marine sub-mesoscale phenomenon recognition system based on transfer learning, comprising: an acquisition module for acquiring satellite SAR marine images of the current country; a preprocessing module for preprocessing the satellite SAR marine images of the current country to obtain processed satellite SAR marine images; and a marine sub-mesoscale phenomenon recognition module for inputting the processed satellite SAR marine images into a marine sub-mesoscale phenomenon recognition model to obtain recognition results; the recognition results include internal ocean waves and sub-mesoscale eddies; the marine sub-mesoscale phenomenon recognition model is obtained by transfer training of a YOLOv11 model using a training set; the training set includes sample satellite SAR marine images of the current country and corresponding sample recognition results.

[0009] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for identifying marine sub-mesoscale phenomena based on transfer learning.

[0010] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for identifying marine sub-mesoscale phenomena based on transfer learning.

[0011] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for identifying marine sub-mesoscale phenomena based on transfer learning.

[0012] According to the specific embodiments provided in this application, this application has the following technical effects.

[0013] This application provides a method, system, device, medium, and product for identifying marine sub-mesoscale phenomena based on transfer learning. It uses the YOLOv11 model for transfer training, and can train a marine sub-mesoscale phenomenon identification model for identifying marine sub-mesoscale phenomena in satellite SAR marine images of the current country with a small sample size. Through the marine sub-mesoscale phenomenon identification model, the processed satellite SAR marine images can be accurately identified to obtain marine sub-mesoscale phenomena (i.e., internal waves and sub-mesoscale eddies), thereby improving the identification accuracy of marine sub-mesoscale phenomena. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating a method for identifying marine sub-mesoscale phenomena based on transfer learning in one embodiment of this application.

[0016] Figure 2 This is a detailed flowchart illustrating a method for identifying marine sub-mesoscale phenomena based on transfer learning, provided as an embodiment of this application.

[0017] Figure 3 This is a schematic diagram showing the conversion results of historical satellite SAR ocean images of the current country.

[0018] Figure 4 This is a schematic diagram of the standard results of historical satellite SAR ocean images of the current country.

[0019] Figure 5 This is a schematic diagram of the functional modules of a marine sub-mesoscale phenomenon recognition system based on transfer learning, provided as an embodiment of this application.

[0020] Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] The remote sensing principles for identifying and tracking oceanic mesoscale phenomena include temperature anomalies, material tracing, closed topology, and rotating flow fields. First, for eddies, warm and cold eddies exhibit positive and negative sea surface temperature anomalies centered on their cores, respectively. For fronts, there are significant gradient anomalies in sea surface temperature. Ideally, these anomalies can reach the ocean surface, forming synchronous anomalies in sea surface temperature (SST), thus enabling detection by spaceborne infrared sensors. Second, eddies and fronts possess strong material entrainment and transport capabilities, causing physical, biological, and chemical substances such as chlorophyll, organic carbon, plankton, and even sea ice to rotate and migrate, thus being detected by spectrometers with underwater penetration capabilities and synthetic aperture radar (SAR) with sea surface imaging capabilities. Third, due to the geostrophic effect of their circulation, eddies exhibit a relatively stable closed structure in their sea surface topology, with the center of warm eddies being the maximum sea surface height (SSH) and the center of cold eddies being the minimum, allowing them to be identified and tracked by radar altimeters. Fourth, for vortices, fronts, internal waves, etc., the related seawater flow can form a rotating flow field, as well as cause upwelling and downwelling flows that diverge and converge on the surface seawater, changing the sea surface roughness and SSH, which can then be captured by sensors such as SAR and radar altimeters.

[0023] Synthetic Aperture Radar (SAR) ocean images possess high spatial resolution and a wide observation range, thus providing a more comprehensive and effective source of data features for the identification and prediction of mesoscale ocean phenomena. SAR ocean data has the advantage of being unaffected by cloud and sunlight conditions, covering small-scale and mesoscale phenomena across large ocean areas. These phenomena cause convergence and divergence on the sea surface, producing larger amplitudes in convergence zones and smaller amplitudes in divergence zones. These surface wave inhomogeneities result in significant variations in sea surface roughness, manifesting as bright or dark stripes on SAR images. Therefore, based on these stripe features, SAR data can serve as a primary data source for detecting mesoscale eddies in the ocean. Algorithms related to SAR data, such as filtering, correction, interpolation, and gradient operators, have matured, providing reliable preprocessing solutions for SAR data. Furthermore, deep learning-based sample augmentation methods and recognition models have been initially implemented, further providing a theoretical foundation for SAR recognition research.

[0024] SAR is an active microwave imaging radar with wavelengths covering L, C, and X bands depending on the observation requirements. It can infer ocean hydrodynamic processes by detecting sea surface roughness and other parameters. There are few dedicated ocean-based SAR systems abroad; most systems incorporate integrated land and ocean observation needs in their design. China's first civilian marine surveillance satellite constellation, Gaofen-3, was launched in 2016. Equipped with single-frequency, multi-polarization, and multi-channel SAR, it features 12 imaging modes, large image resolution and swath width, and provides detailed and general surveys of land and sea, as well as marine environmental monitoring capabilities. Its two subsequent operational satellites, the 1-meter C-SAR satellites 01 and 02, were launched in October 2021 and April 2022, respectively, forming a three-satellite network operating high-resolution marine surveillance system.

[0025] SAR (Shockwave SAR) observes the ocean perpendicularly to its orbit, collecting signals primarily from backscattered waves from the ocean surface. When radar parameters are stable, the main factor affecting the intensity of backscattered waves is sea surface roughness: smooth sea surfaces result in weak backscattering, while rough sea surfaces result in strong backscattering. Sub-mesoscale phenomena in the ocean alter sea surface roughness by entraining tracers and influencing sea surface currents, generating unique large-area elliptical patches or stripes on SAR ocean images. This principle allows for the identification of sub-mesoscale phenomena. However, SAR detection of vortices and fronts based on the above principle is easily affected by sea surface wind fields. When wind speeds are too high or too low, sea surface roughness tends to be uniform, making vortex signals difficult for SAR to capture. In general, current methods for detecting sub-mesoscale phenomena in the ocean based on SAR ocean images can be broadly classified into three categories: (1) Tracer observation: Under certain specific conditions, vortices, fronts, etc. can be directly detected from SAR ocean images. That is, when there are natural tracers on the sea surface, such as sea ice, plankton, oil spills, etc., the damping effect of gravity waves weakens the fluctuation of the sea surface, resulting in a decrease in the backscattering intensity of SAR. If the tracer area is coupled with the sea area of ​​vortices or fronts, the tracer exhibits a specific distribution pattern under the influence of ocean sub-mesoscale phenomena and is captured by SAR. Therefore, vortices and fronts can be identified by identifying the tracers.

[0026] (2) Flow field inversion: In any sea area, the surface roughness can be inverted by SAR ocean images. It is affected by the wave-current interaction modulation caused by vortices, fronts and internal wave flow fields. This change can be used to discover and detect vortices.

[0027] (3) Intelligent Mining: In recent years, in order to further improve the efficiency and accuracy of vortex detection in SAR ocean images, new technologies such as AI and deep learning have been combined to rapidly develop vortex, front, and internal wave identification methods. Compared with traditional methods, intelligent mining extracts more information from SAR ocean image data based on human-designed features or rules. For example, methods such as deep convolutional neural networks do not require pre-definition, but instead mine robust vortex features from a large amount of SAR data and ultimately achieve rapid identification.

[0028] Unlike other sensors such as satellite altimeters, SAR offers higher resolution but a smaller coverage area, making it more targeted at detecting sub-mesoscale ocean phenomena. Compared to mesoscale eddies, SAR ocean imagery is more often used to observe sub-mesoscale (approximately 10 km) and small-mesoscale (approximately 1 km) eddies. These ocean eddies are small in scale, short in lifespan, and change more rapidly; their edges are often filamentous and irregular, and their formation causes differ from those of mesoscale eddies. In the future, we will face a series of challenges, including a surge in SAR data volume and higher requirements for sea surface wind fields in traditional eddy identification algorithms. Therefore, it will be essential to create multi-sensor fusion eddy identification algorithms and improve existing algorithms by incorporating cutting-edge technologies such as deep learning.

[0029] Mesoscale processes are widespread dynamic processes in the ocean's mixed layer, making significant contributions to ocean energy cascading, mass transport, and heat transport. These processes typically have spatiotemporal scales of 10 m–10 km and 1 hour–1 day, and typical forms include filaments, compact eddies, and fronts. As a crucial link between mesoscale and small-scale processes, mesoscale processes not only play a vital role in energy cascading but also serve as important carriers of ocean heat and mass transport, significantly impacting the marine environment.

[0030] Current research methods for studying sub-mesoscale processes in the ocean mainly include numerical simulation and satellite remote sensing observation. One challenge in using numerical simulation to study sub-mesoscale processes is achieving high resolution. Satellite remote sensing observation has emerged to address this challenge. Synthetic Aperture Radar (SAR) ocean images possess high spatial resolution and a wide observation range, thus providing a more comprehensive and effective source of data features for the identification and prediction of sub-mesoscale phenomena. SAR ocean data has the advantage of being unaffected by cloud and sunlight conditions, covering small-scale and sub-mesoscale phenomena over large ocean areas. These phenomena cause convergence and divergence on the sea surface, producing larger amplitudes in convergence zones and smaller amplitudes in divergence zones. These surface wave inhomogeneities cause significant changes in sea surface roughness, resulting in bright or dark stripe characteristics on SAR images. Therefore, based on these stripe characteristics, SAR data can serve as a primary data source for detecting sub-mesoscale processes in the ocean. With the rapid development of ocean satellite observation technology and the rapid growth of remote sensing data, deep learning methods have provided new approaches for the detection and identification of sub-mesoscale phenomena in the ocean. This application utilizes current international satellite SAR ocean images with a small sample size and employs the YOLOv11 model for transfer training to identify ocean sub-mesoscale phenomena, thereby improving the accuracy of such identification.

[0031] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, this application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] In one exemplary embodiment, such as Figure 1 and Figure 2 As shown, a method for identifying marine mesoscale phenomena based on transfer learning is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is described using a server as an example, and includes the following steps S1 to S3.

[0033] Step S1: Obtain satellite SAR ocean images of the current country.

[0034] Step S2: Preprocess the current satellite SAR ocean image of the country to obtain the processed satellite SAR ocean image.

[0035] As an optional implementation, preprocessing includes normalization.

[0036] Step S3: Input the processed satellite SAR ocean image into the ocean sub-mesoscale phenomenon recognition model to obtain the recognition results; the recognition results include ocean internal waves and sub-mesoscale eddies; the ocean sub-mesoscale phenomenon recognition model is obtained by transferring the YOLOv11 model to the training set; the training set includes sample satellite SAR ocean images of the current country and the corresponding sample recognition results.

[0037] As an optional implementation method, the training process of the marine sub-mesoscale phenomenon recognition model is as follows.

[0038] (1) The source domain model is trained using the source domain sample training set; the source domain sample training set includes: foreign satellite SAR ocean images and corresponding foreign sample detection results; the foreign satellite SAR ocean images are satellite SAR ocean images of other countries besides the current country.

[0039] (2) The network parameters of the trained source domain model are transferred to the target domain model to obtain the initial marine mesoscale phenomenon recognition model; both the source domain model and the target domain model are YOLOv11 models.

[0040] Specifically, the source domain model trained on the YOLOv11 model on the source domain sample training set (e.g., the COCO dataset, large-scale Sentinel-1 and other foreign multi-source satellite SAR marine sub-mesoscale phenomenon data) is called, and the parameters such as the number of training rounds, data augmentation parameters, optimizer, and number of training threads are fine-tuned. Then, the model is trained using the SAR marine image training set to obtain the optimal training weight file (the file saved after training the deep learning model, which contains the weights and bias parameters of each layer of the model), thereby obtaining the target domain model. Then, the weight file can be called to make predictions on the validation set.

[0041] (3) Input the sample satellite SAR ocean images into the initial ocean sub-mesoscale phenomenon recognition model using the target domain sample training set to obtain the initial prediction detection results; the target domain sample training set includes: the sample satellite SAR ocean images of the current country and the corresponding sample detection results of the current country.

[0042] As an optional implementation method, the construction of the target domain sample training set specifically includes: acquiring historical satellite SAR ocean images of the current country; using visual interpretation methods to filter out the historical satellite SAR ocean images containing internal ocean waves and sub-mesoscale eddies to obtain preliminary SAR ocean images; annotating the preliminary SAR ocean images to obtain annotated SAR ocean images; and masking the land portion in the annotated SAR ocean images to obtain sample satellite SAR ocean images and corresponding sample detection results.

[0043] Specifically, not all historical satellite SAR ocean images in the current national database are valid data. Visual interpretation is required, involving careful observation of each image to identify potential internal wave regions. These images containing internal waves are then filtered out, and the dataset is reconstructed. Furthermore, since YOLOv11 supports 3D images, single-channel grayscale images from the current national historical satellite SAR ocean images need to be converted into 3-channel images. The conversion result is as follows: Figure 3 As shown.

[0044] Specifically, during the initial screening and annotation of SAR ocean images, the roLabelImg image annotation tool is used to draw rectangular boxes and create category labels on the current country's historical satellite SAR ocean images for internal wave / mesoscale eddies. This allows for the annotation of the selected internal wave / mesoscale eddy images, determining their location and extent, and providing accurate label information for subsequent model training. The annotation results are as follows: Figure 4 As shown.

[0045] Specifically, during masking, the land portion of SAR ocean images is masked to avoid interfering with data features during subsequent model learning, thereby improving model accuracy.

[0046] The sample satellite SAR ocean images and corresponding sample detection results were divided into a target domain sample training set and a validation set in an 8:2 ratio. The target domain sample training set was used for the model training process, allowing the model to learn the features and patterns of the data; the validation set was used to evaluate the model's performance on unseen data, helping to adjust the model's hyperparameters and select the best model.

[0047] (4) The initial prediction detection results are screened using the non-maximum suppression method to obtain the prediction detection results.

[0048] (5) Based on the predicted detection results and the sample detection results corresponding to the autonomous satellite SAR ocean images, a loss function is constructed, and the parameters of the initial ocean sub-mesoscale phenomenon recognition model are iteratively optimized according to the loss function until the loss function reaches the minimum value or the iteration optimization round reaches the maximum value, and the iteration optimization is stopped to obtain the ocean sub-mesoscale phenomenon recognition model.

[0049] Specifically, taking the validation set as the model input as an example, the specific process is as follows. The marine mesoscale phenomenon recognition model receives images from the validation set. Generally, it is expected that the input images have a certain size standard, usually 640*640 pixels, in order to facilitate effective processing.

[0050] Image preprocessing: Convert the image from BGR (OpenCV's default format) to RGB format and normalize it to scale the pixel values ​​to between 0 and 1, which helps improve the model's convergence speed.

[0051] Feature Extraction: A deep convolutional neural network (CNN) is used in the marine mesoscale phenomenon recognition model to extract image features. YOLOv11 extracts key features from the image through multiple convolutional layers. These convolutional layers capture features at different levels, such as edges, shapes, and textures. Activation functions are used to increase the network's non-linearity, helping the model learn complex feature representations. Pooling layers or stride convolutions reduce the size of the feature maps while increasing their depth, extracting high-level features.

[0052] Data Prediction: After feature extraction, YOLOv11 generates predictions. The input image is divided into a 7x7 grid, with each grid cell responsible for detecting objects within its bounding area. Each grid cell predicts a fixed number of bounding boxes, each including its center coordinates (relative to the grid cell), width and height (relative to the entire image), and a confidence score (indicating whether an object exists within the box and its accuracy). Simultaneously, the grid cell predicts the probability of a class, representing the possible object class within that grid cell.

[0053] Non-Maximum Suppression (NMS) method: The NMS method retains predicted bounding boxes based on confidence scores. The overlapping parts of high-confidence boxes are compared, and if the overlap exceeds a set threshold, the overlapping boxes are removed, and only the one with the highest score is retained, ensuring that the output predicted bounding box is optimal and reducing the number of cases where multiple boxes represent the same object.

[0054] Detection and Recognition: After the marine sub-mesoscale phenomenon recognition model predicts the validation set, the recognition result is obtained. The bounding box of the recognition result is drawn on the image, and then the image is saved to a folder for subsequent analysis and viewing.

[0055] Validation: After predicting on the validation set, the marine sub-mesoscale phenomenon recognition model matches the predicted bounding boxes with the ground truth bounding boxes, generating several metrics for evaluating the classification model's performance, namely the confusion matrix. Then, using these metrics, several performance indicators for evaluating the object detection model are calculated according to formulas.

[0056] There are four metrics for evaluating the performance of object detection models in the YOLOv11 model: P_curve, R_curve, PR_curve, and F1_curve.

[0057] P_curve: The recognition rate refers to the proportion of correctly detected positive samples out of all samples predicted as positive. The formula for its calculation is as follows.

[0058] .

[0059] in, What percentage of the results predicted as "positive samples" by the model (e.g., detected targets, classified into a certain category) are actually positive samples? (True Positive) is the number of samples correctly identified as positive. (False Positive) is the number of samples that were incorrectly identified as positive (when they were actually negative samples).

[0060] R_curve: The recall curve shows how recall changes with different thresholds, and its calculation formula is as follows.

[0061] .

[0062] in, This represents the proportion of all real positive samples that were successfully identified by the model; (False Negative) represents a positive sample that was missed (it is actually a positive sample but was not identified).

[0063] PR_curve: The recognition rate-recall curve is a graph that plots the relationship between recognition rate and recall rate. Recall rate refers to the proportion of correctly detected positive samples out of all actual positive samples.

[0064] F1_curve: The F1 score is the harmonic mean of the recognition rate and recall rate, which takes into account both the accuracy and completeness of the model. Its calculation formula is as follows.

[0065] .

[0066] If precision is high but recall is low (or vice versa), F1 will still be low; therefore, F1 is a more balanced indicator, especially suitable for tasks where both false positives and false negatives need to be controlled.

[0067] The beneficial effects of the transfer learning-based marine sub-mesoscale phenomenon identification method proposed in this application are mainly manifested as follows.

[0068] 1. This application employs diverse transfer learning strategies for SAR marine phenomenon detection to address the challenge of small-sample learning. It utilizes the YOLOv11 model for transfer training, training it with foreign satellite SAR marine images and corresponding foreign sample detection results. The trained weights are then transferred, and a smaller training set (current country's sample satellite SAR marine images and corresponding sample recognition results) is used to train a marine sub-mesoscale phenomenon recognition model, thus solving the small-sample learning problem and improving the generalization ability of the recognition model.

[0069] 2. Using the frozen pre-trained shallow weights method reduces dependence on the target dataset. This method is suitable for situations where the target dataset is small and the target domain has high similarity to the source domain. In this method, the shallow feature extraction part of the pre-trained model (usually convolutional layers) is "frozen," meaning its weights remain unchanged, and only the weights near the output layer are updated. Due to the strong similarity between the target and source domains, shallow features (such as edges and textures) can be directly transferred, reducing dependence on the target dataset.

[0070] Based on the same inventive concept, this application also provides a marine sub-mesoscale phenomenon recognition system based on transfer learning. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the marine sub-mesoscale phenomenon recognition system based on transfer learning provided below can be found in the limitations of the marine sub-mesoscale phenomenon recognition method based on transfer learning described above, and will not be repeated here.

[0071] In one exemplary embodiment, such as Figure 5 As shown, a marine mesoscale phenomenon recognition system based on transfer learning is provided, comprising the following modules.

[0072] The acquisition module is used to acquire satellite SAR ocean images of the current country.

[0073] The preprocessing module is used to preprocess the current satellite SAR ocean images of the country to obtain processed satellite SAR ocean images.

[0074] As an optional implementation, the preprocessing module includes a normalization processing unit.

[0075] The ocean sub-mesoscale phenomenon recognition module is used to input the processed satellite SAR ocean image into the ocean sub-mesoscale phenomenon recognition model to obtain the recognition result; the recognition result includes ocean internal waves and sub-mesoscale eddies; the ocean sub-mesoscale phenomenon recognition model is obtained by transferring the YOLOv11 model using a training set; the training set includes sample satellite SAR ocean images of the current country and the corresponding sample recognition results.

[0076] As an optional implementation, the marine sub-mesoscale phenomenon identification module includes the following modules.

[0077] The source domain training unit is used to train the source domain model using a source domain sample training set; the source domain sample training set includes: foreign satellite SAR ocean images and corresponding foreign sample detection results; the foreign satellite SAR ocean images are satellite SAR ocean images from countries other than the current country.

[0078] The transfer unit is used to transfer the network parameters of the trained source domain model to the target domain model to obtain an initial marine mesoscale phenomenon recognition model; both the source domain model and the target domain model are YOLOv11 models.

[0079] The target domain training unit is used to input the sample satellite SAR ocean images into the initial ocean sub-mesoscale phenomenon recognition model using the target domain sample training set to obtain the initial prediction detection result; the target domain sample training set includes: sample satellite SAR ocean images of the current country and the corresponding sample detection results of the current country.

[0080] The filtering unit is used to filter the initial prediction detection results using a nonmaximum suppression method to obtain the prediction detection results.

[0081] The generation unit is used to construct a loss function based on the predicted detection results and the sample detection results corresponding to the sample autonomous satellite SAR ocean image, and to iteratively optimize the parameters of the initial ocean sub-mesoscale phenomenon recognition model according to the loss function until the loss function reaches the minimum value or the iteration optimization round reaches the maximum value, and then stop the iteration optimization to obtain the ocean sub-mesoscale phenomenon recognition module.

[0082] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 6As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores current satellite SAR ocean images of the country. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the aforementioned method for identifying sub-mesoscale ocean phenomena based on transfer learning.

[0083] Those skilled in the art will understand that Figure 6 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0084] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0085] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0086] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0087] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0088] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchain. The processors involved in the embodiments provided in this application may be, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc.

[0089] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0090] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for identifying marine mesoscale phenomena based on transfer learning, characterized in that, The method for detecting and identifying marine sub-mesoscale phenomena based on transfer learning includes: Acquire current satellite SAR ocean images of the country; Preprocessing is performed on current national satellite SAR ocean images to obtain processed satellite SAR ocean images; The processed satellite SAR ocean images are input into the ocean sub-mesoscale phenomenon recognition model to obtain recognition results; the recognition results include ocean internal waves and sub-mesoscale eddies; the ocean sub-mesoscale phenomenon recognition model is obtained by transferring the training set to the YOLOv11 model; the training set includes sample satellite SAR ocean images of the current country and the corresponding sample recognition results.

2. The method for identifying marine sub-mesoscale phenomena based on transfer learning according to claim 1, characterized in that, The preprocessing includes normalization.

3. The method for identifying marine sub-mesoscale phenomena based on transfer learning according to claim 1, characterized in that, The training process of the marine sub-mesoscale phenomenon recognition model specifically includes: The source domain model is trained using a source domain sample training set; the source domain sample training set includes: foreign satellite SAR ocean images and corresponding foreign sample detection results; the foreign satellite SAR ocean images are satellite SAR ocean images from countries other than the current country. The network parameters of the trained source domain model are transferred to the target domain model to obtain an initial marine mesoscale phenomenon recognition model; both the source domain model and the target domain model are YOLOv11 models. The sample satellite SAR ocean images are input into the initial ocean sub-mesoscale phenomenon recognition model using the target domain sample training set to obtain initial prediction and detection results; the target domain sample training set includes: sample satellite SAR ocean images of the current country and corresponding sample detection results of the current country; The initial prediction detection results are screened using the non-maximum suppression method to obtain the final prediction detection results; Based on the predicted detection results and the sample detection results corresponding to the sample autonomous satellite SAR ocean images, a loss function is constructed, and the parameters of the initial ocean sub-mesoscale phenomenon recognition model are iteratively optimized according to the loss function until the loss function reaches its minimum value or the iteration optimization round reaches its maximum value, at which point the iterative optimization stops, and the ocean sub-mesoscale phenomenon recognition model is obtained.

4. The method for identifying marine sub-mesoscale phenomena based on transfer learning according to claim 3, characterized in that, The construction of the target domain sample training set specifically includes: Acquire historical satellite SAR ocean images of the current country; The historical satellite SAR ocean images containing internal ocean waves and sub-mesoscale eddies were selected using visual interpretation methods to obtain preliminary SAR ocean images; The initial SAR ocean images are annotated to obtain an annotated SAR ocean images; The land portion of the labeled SAR ocean image is masked to obtain the sample satellite SAR ocean image and the corresponding sample detection results.

5. A marine mesoscale phenomenon recognition system based on transfer learning, characterized in that, The ocean sub-mesoscale phenomenon detection and recognition system based on transfer learning includes: The acquisition module is used to acquire satellite SAR ocean images of the current country; The preprocessing module is used to preprocess the current satellite SAR ocean images of the country to obtain processed satellite SAR ocean images; The ocean sub-mesoscale phenomenon recognition module is used to input the processed satellite SAR ocean image into the ocean sub-mesoscale phenomenon recognition model to obtain the recognition result; the recognition result includes ocean internal waves and sub-mesoscale eddies; the ocean sub-mesoscale phenomenon recognition model is obtained by transferring the YOLOv11 model using a training set; the training set includes sample satellite SAR ocean images of the current country and the corresponding sample recognition results.

6. The marine mesoscale phenomenon recognition system based on transfer learning according to claim 5, characterized in that, The preprocessing module includes a normalization processing unit.

7. The marine mesoscale phenomenon recognition system based on transfer learning according to claim 5, characterized in that, The marine sub-mesoscale phenomenon identification module includes: The source domain training unit is used to train the source domain model using a source domain sample training set; the source domain sample training set includes: foreign satellite SAR ocean images and corresponding foreign sample detection results; the foreign satellite SAR ocean images are satellite SAR ocean images from countries other than the current country. The transfer unit is used to transfer the network parameters of the trained source domain model to the target domain model to obtain an initial marine mesoscale phenomenon recognition model; both the source domain model and the target domain model are YOLOv11 models. The target domain training unit is used to input the sample satellite SAR ocean images into the initial ocean sub-mesoscale phenomenon recognition model using the target domain sample training set to obtain initial prediction and detection results; the target domain sample training set includes: sample satellite SAR ocean images of the current country and corresponding sample detection results of the current country; The filtering unit is used to filter the initial prediction detection results using the non-maximum suppression method to obtain the prediction detection results; The generation unit is used to construct a loss function based on the predicted detection results and the sample detection results corresponding to the sample autonomous satellite SAR ocean image, and to iteratively optimize the parameters of the initial ocean sub-mesoscale phenomenon recognition model according to the loss function until the loss function reaches the minimum value or the iteration optimization round reaches the maximum value, and then stop the iteration optimization to obtain the ocean sub-mesoscale phenomenon recognition module.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the marine sub-mesoscale phenomenon identification method based on transfer learning as described in any one of claims 1-4.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the marine sub-mesoscale phenomenon identification method based on transfer learning as described in any one of claims 1-4.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the marine sub-mesoscale phenomenon identification method based on transfer learning as described in any one of claims 1-4.

Images