Cross-domain vortex detection method suitable for high-frequency ground wave radar
By constructing the CDEDNet neural network and combining domain adaptation and parameter transfer learning, the problem of insufficient labeled data in high-frequency ground wave radar ocean vortex detection is solved, achieving efficient and accurate cross-domain vortex detection and improving the model's adaptability and detection accuracy.
Patent Information
- Application Number
- CN202510803460.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-17
- Publication Date
- 2025-11-21
AI Technical Summary
Existing high-frequency ground wave radar methods for detecting ocean eddies suffer from problems such as insufficient labeled data, low detection efficiency, and poor model versatility and adaptability, especially with poor detection performance under different data sources and scenarios.
A cross-domain vortex detection method is adopted. By constructing a CDEDNet neural network and combining domain adaptation, parameter-based transfer learning and instance cross-domain adaptation techniques, the generalization ability of the model is improved. The domain adaptation method of samples is used to map data to the same feature space, and parameter-based transfer learning inherits detection experience and expands the training sample set to achieve cross-domain adaptation.
It improves the accuracy and efficiency of vortex detection in the field of high-frequency ground wave radar ocean detection, enhances the adaptability and versatility of the model in different scenarios, reduces the false detection rate, and provides more reliable support for ocean research and navigation safety.
Smart Images
Figure CN120995249A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application provides a cross-domain vortex detection method suitable for high-frequency ground wave radar, and belongs to the technical field of ocean exploration. BACKGROUND
[0002] In high-frequency ground wave radar ocean exploration, vortex detection is of great significance to ocean dynamics research, ocean disaster warning, and maritime navigation safety, etc. At present, the closed contour method is generally used for small and medium scale vortex detection. Although this method can achieve vortex detection to a certain extent and has high accuracy, it has the problem that the false detection rate is difficult to control. With the wide application of deep learning technology in image recognition and detection fields, it has also been gradually introduced into ocean vortex detection.
[0003] Deep learning can be divided into supervised learning, semi-supervised learning and unsupervised learning according to the completeness of training data labels. Among them, the supervised learning algorithm relies on a large number of labeled data to train the network, but in the actual application of ocean vortex detection, there are many unlabeled vortex data, which leads to a serious lack of accurately labeled data, which brings great difficulty to the training of vortex detection model based on deep learning.
[0004] In order to solve the data labeling problem, the automatic detection, positioning and pattern detection of small and medium scale vortexes are realized, and the related sea area rules are studied. Although the model has achieved certain detection effect and small error, the execution time is longer than the existing method, which is difficult to meet the demand of detection efficiency in actual application. In addition, the research on ocean vortex detection based on various remote sensing data is widely carried out, but due to the huge differences in the characteristics of ocean vortex information obtained by different sensors, the universality and adaptability of the detection model in different data sources and scenes are poor.
[0005] Therefore, in view of the above problems, the existing structure is studied and improved, and a cross-domain vortex detection method suitable for high-frequency ground wave radar is proposed to solve the above problems. SUMMARY
[0006] The present application aims to overcome the shortcomings of the prior art and provide a cross-domain vortex detection method suitable for high-frequency ground wave radar. Through the technical means of sample domain adaptation, parameter-based transfer learning and instance-based cross-domain adaptation, the generalization ability of the vortex detection model is improved, the problems of insufficient labeled data and detection experience inheritance are solved, and efficient and accurate detection of vortexes in the field of high-frequency ground wave radar ocean exploration is realized.
[0007] The application is proposed to solve the above problems, and the technical scheme is: a cross-domain vortex detection method suitable for high-frequency ground wave radar, comprising constructing a CDEDNet neural network following an end-to-end framework, improving the model generalization ability through domain adaptation of samples and parameter-based transfer learning, and being used for vortex detection in the field of high-frequency ground wave radar ocean detection, an instance-based cross-domain adaptation method is used to increase the training sample set and solve the problem of insufficient labeled data, and the parameter-based transfer learning is used for vortex detection by sharing a general feature layer in multiple regions.
[0008] Further, the CDEDNet neural network comprises a CDEDNet-0 architecture, which is a neural network based on FCN and is composed of 8 convolutional layers and an upsampling operator; the CDEDNet-0 architecture is used for vortex detection of radar data of observation points in Nansan and Xuwen regions, the input flow field map is extracted through convolutional layers, the feature map size is changed through pooling operation, and finally the background, cyclone and anticyclone are represented at the pixel level through upsampling.
[0009] Further, the domain adaptation method of the sample specifically maps the source domain and target domain data to the same feature space, when the edge distribution difference and the conditional distribution difference tend to 0, the trained prediction classifier completes the function of predicting the label of the target domain data to improve the model performance.
[0010] Further, the CDEDNet neural network further comprises a CDEDNet-2 architecture, which directly uses the first seven convolutional layers of FCN-1 and randomly changes the last layer hyperparameters; through parameter-based transfer learning, the general vortex detection ability of FCN-1 is inherited to CDEDNet-2, and the last convolutional layer of CDEDNet-2 is trained using the image dataset to adapt to the vortex feature information of the dataset, so as to realize the recognition of the remaining flow field map vortex in the dataset.
[0011] Further, the instance-based cross-domain adaptation method maps the source domain data and the target domain data to a high-dimensional space, so that the edge distribution difference and the conditional distribution difference of the two tend to 0, and a classifier is trained and predicted in this space to expand the label quantity of the target domain data.
[0012] Further, the multi-scene vortex detection network based on parameter transfer learning solves the problem of inheritance of previous vortex detection experience.
[0013] Further, the high-frequency ground wave radar data establishes a vortex detection network based on the full convolutional neural network skeleton in the high-frequency ground wave radar ocean exploration field, the input flow field map size is determined by the radar monitoring range, and linear interpolation is adopted to generate; the full convolutional neural network comprises a plurality of convolutional layers and an upsampling operator, flow field map features are extracted through convolution and pooling operations, and a feature map with the same size as the original flow field map is obtained through upsampling for vortex detection.
[0014] Further, the high-frequency ground wave radar data establishes a vortex detection network based on the full convolutional neural network skeleton, the input flow field map size is determined by the radar monitoring range, and linear interpolation is adopted to generate; the full convolutional neural network comprises a plurality of convolutional layers and an upsampling operator, flow field map features are extracted through convolution and pooling operations, and a feature map with the same size as the original flow field map is obtained through upsampling for vortex detection.
[0015] Further, the CDEDNet neural network is a cross-domain vortex detection neural network suitable for high-frequency ground wave radars, which can effectively utilize different regional data for vortex detection, and improve the adaptability and universality of the detection network in different scenarios.
[0016] Due to the adoption of the above technical solutions, the present application has the following beneficial effects:
[0017] 1. The present application can adapt to data of different sources and characteristics through field adaptation and parameter-based transfer learning of samples, effectively improving the generalization ability of the model in different scenarios in the high-frequency ground wave radar ocean exploration field, and enhancing the universality and adaptability of the model.
[0018] 2. The instance-based cross-domain adaptation method of the present application effectively increases the training sample set, solves the problem of insufficient labeled data, provides more sufficient data for model training, and thus improves the training effect and detection accuracy of the model, and reduces detection errors and false detection situations caused by data scarcity.
[0019] 3. The multi-scenario vortex detection network design based on parameter transfer learning of the present application successfully inherits the experience of previous vortex detection, so that the model can fully utilize the detection experience accumulated in different regions and different periods, quickly and accurately perform vortex detection in new detection scenarios, and improve the detection performance and reliability of the model in multiple scenarios.
[0020] 4. The CDEDNet neural network constructed by the present application combines a variety of innovative technical means, realizes efficient and accurate detection of ocean eddies, reduces the false detection rate while improving the detection efficiency compared with the prior art, and can more accurately identify the type, position and range of eddies, providing more reliable technical support for ocean research, disaster warning and maritime navigation safety, etc.BRIEF DESCRIPTION OF DRAWINGS BRIEF DESCRIPTION OF DRAWINGS
[0021] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and are meant to explain the application without limiting the application to the embodiments shown. In the drawings:
[0022] Figure 1 It is a schematic diagram of the vortex detection network technology in the cross-domain vortex detection method suitable for high-frequency ground wave radar of the application.
[0023] Figure 2 It is a schematic diagram of the vortex detection network technology based on sample domain adaptation in the cross-domain vortex detection method suitable for high-frequency ground wave radar of the application.
[0024] Figure 3 It is a schematic diagram of the vortex detection network technology based on parameter transfer learning architecture in the cross-domain vortex detection method suitable for high-frequency ground wave radar of the application.
[0025] Figure 4 It is a diagram of the detection effect comparison of three vortex detection network technologies in the cross-domain vortex detection method suitable for high-frequency ground wave radar of the application. DETAILED DESCRIPTION
[0026] The technical solutions in the embodiments of the application will be described below in a clear and complete manner. Obviously, the described embodiments are only some of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.
[0027] Embodiment one: marine vortex detection in the South Three Area
[0028] When marine vortex detection is performed in the South Three Area, first, the flow field map data of high-frequency ground wave radar in the area within one month is collected, and 2000 original flow field maps are obtained. According to the radar monitoring range, linear interpolation is used to generate all flow field maps into unified specification image data with a size of 512x512 pixels. 1500 flow field maps are randomly selected for manual labeling, and background, cyclone and anticyclone information are marked to construct an initial training sample set, and the remaining 500 flow field maps are used as a test sample set.
[0029] The training sample set is input into the CDEDNet-0 architecture, and 8 convolutional layers are used to extract features from the input flow field map in turn. During the convolution process, convolution kernels of different sizes and parameters capture vortex features of different scales in the flow field map. After convolution, a pooling operation is performed to reduce the data dimension. Finally, the background, cyclone and anticyclone are represented at the pixel level by upsampling operators. During the training process, the learning rate is set to 0.001, and the stochastic gradient descent algorithm is used. After 50 training cycles, the accuracy of the model on the training set reaches 92%.
[0030] The test sample set is input into the trained CDEDNet-0 architecture, and 468 flow field maps are successfully detected, with a detection accuracy of 93.6%. The vortex situation in the ocean flow field of the South Three Area is effectively identified.
[0031] Example Two: Joint Detection of Xuwen Area and South Three Area Data
[0032] 1200 flow field map data of high-frequency ground wave radar in the Xuwen area for half a month are collected and processed to 512x512 pixel specifications. The Xuwen area data is used as the target domain data, and the South Three Area labeled data is used as the source domain data. The sample domain adaptation method is used to map the source and target domain data to the same feature space. Through iterative calculation, the mapping parameters are adjusted. When the edge distribution difference and conditional distribution difference meet the set threshold, a prediction classifier is trained in the feature space to generate predicted labels for the Xuwen area's partially unlabeled data.
[0033] The CDEDNet-2 architecture is constructed, and the first seven convolutional layers of FCN-1 are used, with the last layer's hyperparameters randomly changed. Through parameter-based transfer learning, the general vortex detection ability of FCN-1 trained in the South Three Area is inherited to CDEDNet-2. Then, the dataset containing the Xuwen area labeled data and predicted label data is used to train the last convolutional layer of CDEDNet-2, with a training period of 30 times and a learning rate of 0.0005.
[0034] The remaining flow field maps in the Xuwen area that did not participate in the training are input into the trained CDEDNet-2 architecture, and 890 flow field maps are detected, with a detection accuracy of 91.7%. Efficient vortex detection under the joint of different area data is achieved.
[0035] Example Three: Comprehensive Detection of Multi-Region Data
[0036] The high-frequency ground wave radar flow field map data of Nansan, Xuwen and two other coastal areas are collected, a total of 5000, and the image specifications are uniformly processed. The data of the four areas are mapped to a high-dimensional space by using an instance-based cross-domain adaptive method, the edge distribution difference and the conditional distribution difference of the data of different areas are made to tend to 0 by an optimization algorithm, a classifier is trained in the high-dimensional space, and labels are generated for a large amount of unlabeled data, so that the training sample set is expanded to 4000.
[0037] Based on the expanded training sample set, the CDEDNet neural network is comprehensively trained, network parameters are adjusted by combining the advantages of the CDEDNet-0 architecture and the CDEDNet-2 architecture, the training period is 80 times, and the learning rate is dynamically adjusted in the training process.
[0038] The remaining 1000 flow field maps are used as test data, input into the trained CDEDNet neural network, and 945 flow field maps are accurately detected, with a detection accuracy of up to 94.5%, which fully embodies the strong ability and good adaptability of the method in the accurate detection of ocean eddies in the multi-region data comprehensive detection scene.
[0039] The above describes the present application and its embodiments, which are not limited, and in general, if a person skilled in the art is inspired by it, without departing from the purpose of the present application, similar structural methods and embodiments can be designed without creative design, which should belong to the protection scope of the present application.
Claims
1. A method for detecting transdomain vortices suitable for high-frequency ground wave radar, characterized in that, This includes constructing a CDEDNet neural network following an end-to-end framework, improving the model's generalization ability through sample domain adaptation and parameter-based transfer learning, and applying it to vortex detection in the field of high-frequency ground wave radar ocean detection. An instance-based cross-domain adaptation method is used to increase the training sample set to address the problem of insufficient labeled data. Parameter-based transfer learning is applied to vortex detection by sharing a common feature layer across multiple regions.
2. The method for cross-domain vortex detection suitable for high-frequency ground wave radar according to claim 1, characterized in that: The CDEDNet neural network includes the CDEDNet-0 architecture, which is an FCN-based neural network consisting of 8 convolutional layers and an upsampling operator. The CDEDNet-0 architecture is used to perform vortex detection on radar data from observation points in the Nansan and Xuwen areas. The input flow field map is processed by convolutional layers to extract features, pooling operations to change the size of the feature map, and finally upsampling to represent the background, cyclone, and anticyclone at the pixel level.
3. The method for cross-domain vortex detection suitable for high-frequency ground wave radar according to claim 1, characterized in that: The domain adaptation method for the sample specifically maps the source domain and target domain data to the same feature space. When the difference between the edge distribution and the difference between the conditional distribution tend to be 0, the predictive classifier is trained to complete the function of predicting the label of the target domain data, so as to improve the model performance.
4. The method for cross-domain vortex detection suitable for high-frequency ground wave radar according to claim 1, characterized in that: The CDEDNet neural network also includes the CDEDNet-2 architecture, which directly adopts the first seven convolutional layers of FCN-1 and randomly changes the hyperparameters of the last layer. Through parameter-based transfer learning, the general vortex detection capability of FCN-1 is inherited to CDEDNet-2, and the last convolutional layer of CDEDNet-2 is trained using dataset images to adapt to the vortex feature information of the dataset, thereby realizing the recognition of vortices in the remaining flow field map in the dataset.
5. The method for cross-domain vortex detection suitable for high-frequency ground wave radar according to claim 1, characterized in that: The instance-based cross-domain adaptive method maps source domain data and target domain data to a high-dimensional space, making the difference in their marginal distribution and conditional distribution approach zero. A classifier is trained and predictions are made in this space to increase the number of labels in the target domain data.
6. The method for cross-domain vortex detection suitable for high-frequency ground wave radar according to claim 1, characterized in that: The multi-scenario vortex detection network based on parameter transfer learning is designed to address the issue of inheriting previous vortex detection experience.
7. The method for cross-domain vortex detection suitable for high-frequency ground wave radar according to claim 1, characterized in that: The vortex detection network based on a fully convolutional neural network skeleton is established using high-frequency ground wave radar data in the field of high-frequency ground wave radar marine detection. The size of the input flow field map is determined by the radar monitoring range and is generated by linear interpolation. The fully convolutional neural network contains multiple convolutional layers and upsampling operators. It extracts flow field map features through convolution and pooling operations, and upsamples them to obtain a feature map of the same size as the original flow field map for vortex detection.
8. The method for cross-domain vortex detection suitable for high-frequency ground wave radar according to claim 7, characterized in that: The vortex detection network based on a fully convolutional neural network skeleton is established using high-frequency ground wave radar data. The size of the input flow field map is determined by the radar monitoring range and is generated by linear interpolation. The fully convolutional neural network contains multiple convolutional layers and upsampling operators. It extracts flow field map features through convolution and pooling operations, and upsamples the feature map to obtain a feature map of the same size as the original flow field map for vortex detection.
9. A transdomain vortex detection method suitable for high-frequency ground wave radar according to claim 1, characterized in that: The CDEDNet neural network is a cross-domain vortex detection neural network suitable for high-frequency ground wave radar. It can effectively utilize data from different regions for vortex detection, improving the adaptability and versatility of the detection network in different scenarios.