Ocean mesoscale vortex detection method based on convolutional neural network

Abstract

The invention discloses an ocean mesoscale vortex detection method based on a convolutional neural network, and the method effectively improves the detection capability of a model for complex edge structures of vortex targets such as anti-cyclonic vortices (AE) and cyclonic vortices (CE) through multi-path convolution processing, pooling operation and multi-scale feature fusion, and improves the detection capability of the model for the complex edge structures of vortex targets such as the anti-cyclonic vortices (AE) and the cyclonic vortices (CE). And the defects of the conventional YOLOv8 convolution in the complex target detection are overcome.

Description

Technical Field

[0001] The present invention belongs to the technical field of ocean mesoscale vortex detection and relates to an ocean mesoscale vortex detection method based on a convolutional neural network. Background Art

[0002] The detection of ocean mesoscale eddies has important scientific and practical significance, involving research and applications in multiple fields. Ocean mesoscale eddies are an important part of ocean circulation, and they play a key role in the transfer of heat, salinity, and momentum. By detecting and studying ocean mesoscale eddies, scientists can better understand the circulation and dynamic processes in the ocean, thereby revealing the operating mechanism of the ocean system.

[0003] The detection methods of mesoscale eddies in the ocean are usually based on sea surface height, ocean flow field, remote sensing images, numerical simulation, machine learning, spatiotemporal filtering and other schemes. Among them, based on the convolutional neural network, the ocean surface height data or remote sensing images can be used as input, and the convolutional neural network can learn the spatial characteristics of the mesoscale eddies. During the training process, the model can extract the edges, textures and other features of the mesoscale eddies from the data by continuously adjusting the internal weights, thereby achieving automatic detection and helping to explain the formation and evolution mechanism of the eddies. By visualizing the model's attention distribution and activation maps in the data, the features that the model focuses on when identifying mesoscale eddies can be revealed, which helps to understand the eddy formation process.

[0004] At present, in the ocean eddy detection method based on convolutional neural network, the convolution module determines the feature extraction ability and overall performance of the model. Although the conventional convolution module of YOLOv8 performs well in general scenarios, when dealing with targets with complex morphology, especially vortex targets such as anticyclonic eddies (AE) and cyclonic eddies (CE), their complex edge structures put forward higher requirements on the existing convolution design and expose certain limitations. Summary of the invention

[0005] Purpose of the invention: In target detection tasks, the convolution module determines the feature extraction capability and overall performance of the model. Although the conventional convolution module of YOLOv8 performs well in general scenarios, when dealing with targets with complex morphology, especially vortex targets such as anticyclonic eddies (AE) and cyclonic eddies (CE), their complex edge structures put forward higher requirements on the existing convolution design and expose certain limitations. In order to solve this problem, through multi-path convolution processing, pooling operation and multi-scale feature fusion, the model's detection capability for the complex edge structures of AE and CE is effectively improved, and the shortcomings of conventional YOLOv8 convolution in the detection of such complex targets are improved.

[0006] Technical solution: A method for detecting mesoscale eddies in the ocean based on convolutional neural networks is proposed. This method effectively improves the model's ability to detect complex edge structures of AE and CE through multi-path convolution processing, pooling operations, and multi-scale feature fusion, and improves the shortcomings of conventional YOLOv8 convolution in detecting such complex targets. The specific construction steps are:

[0007] Step S1: Obtain multi-source data of ocean eddies, such as sea surface height, temperature, etc. Ensure that the data has high quality and high temporal and spatial resolution;

[0008] Step S2: preprocess the data. Perform missing value processing, error value correction and normalization on the vortex data;

[0009] Step S3: Divide the data into training set, validation set and test set, and perform necessary data enhancement, such as rotation, flipping, scaling, etc., to expand the data set;

[0010] Step S4: Detection model construction and comparative analysis. This method builds an experimental platform based on the YOLOv8 model. The YOLO (You Only Look Once) series of models are widely used in various target detection tasks due to their end-to-end target detection process, fast inference speed, and high detection accuracy. As the latest version of the series, YOLOv8 continues the high efficiency of the YOLO family model, and through further optimization and innovation in the architecture, it achieves better performance while maintaining real-time performance.

[0011] Step S5: Target model interpretation. A new convolution module, AConv convolution module, is added. This module effectively improves the model's detection capability for complex edge structures of AE and CE through multi-path convolution processing, pooling operation and multi-scale feature fusion, and improves the shortcomings of conventional YOLOv8 convolution in such complex target detection.

[0012] Furthermore, in order to adjust the structure of YOLOv8 to better handle the misalignment problem in target detection, BiFPN (weighted bidirectional feature pyramid network) is introduced and improved. Compared with the traditional feature fusion method, BiFPN introduces a weighting mechanism, which enables the network to assign different weights according to the importance of each feature during the fusion process, ensuring the rationality and effectiveness of feature fusion. This strategy enables the network to process features of different scales more flexibly and accurately, especially in the detection of complex structures such as AE and CE. Through weight adjustment, the key details of each input feature can be captured more accurately.

[0013] Furthermore, the LSK (Large Kernel with Stride) attention mechanism is introduced. Since the introduction of the Transformer model, the attention mechanism has gradually become an important tool for feature extraction due to its ability to focus on global features. In target detection tasks, especially small target detection and target recognition in complex scenes, the attention mechanism can effectively improve the model's ability to focus on small objects and reduce missed detections and false detections.

[0014] The present invention has the following beneficial effects:

[0015] (1) The AConv convolution module in the present invention shows significant advantages in edge feature preservation and multi-scale feature fusion. Compared with the conventional convolution module in YOLOv8, the AConv module performs better in terms of missed detection rate and false detection rate, especially in complex backgrounds and low-contrast scenes, where AConv performs more stably. Through feature partitioning, multi-path processing, pooling operations and feature fusion, AConv significantly improves the AE and CE detection accuracy, reduces the computational cost, and enhances the robustness and adaptability of the model.

[0016] (2) The structure of BiFPN in the present invention not only includes bidirectional feature fusion from top to bottom and from bottom to top, but also introduces direct jump connections. These jump connections enable the network to transfer information between feature maps at different levels, reducing the problem of information loss. In the vortex detection task, this structure improves the detection capability of cyclonic and anticyclonic vortex features by capturing a wide range of background and boundary information, reducing the risk of false detection and missed detection.

[0017] (3) The LSK attention mechanism in the present invention enhances the detection capability of the YOLOv8 model for small targets in complex backgrounds by combining large convolution kernels, multi-scale convolution operations and attention weights. In the detection of cyclonic and anticyclonic vortices, the LSK module helps the model to better extract the multi-scale features of the vortex structure and improves the recognition and classification capabilities of the vortex morphology. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of the detection method based on convolutional neural network;

[0019] Figure 2 This is the structure diagram of AConv;

[0020] Figure 3 This is the PAN-FPN structure diagram;

[0021] Figure 4 This is the working principle diagram of LSK; DETAILED DESCRIPTION

[0022] The technical solution of the present invention is described in detail below so that those skilled in the art can understand the present invention. However, the protection scope of the present invention is not limited to the scope of the embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

[0023] Embodiment 1:

[0024] like Figure 1 As shown, this embodiment provides a method for detecting ocean mesoscale vortices based on a convolutional neural network, which specifically includes the following steps:

[0025] Step S1: Obtain multi-source data of ocean eddies, such as sea surface height, temperature, etc. Ensure that the data has high quality and high temporal and spatial resolution;

[0026] Step S2: preprocess the data. Perform missing value processing, error value correction and normalization on the vortex data;

[0027] Step S3: Divide the data into training set, validation set and test set, and perform necessary data enhancement, such as rotation, flipping, scaling, etc., to expand the data set;

[0028] Step S4: Detection model construction and comparative analysis. This method builds an experimental platform based on the YOLOv8 model. The YOLO (You Only Look Once) series of models are widely used in various target detection tasks due to their end-to-end target detection process, fast inference speed, and high detection accuracy. As the latest version of the series, YOLOv8 continues the high efficiency of the YOLO family model, and through further optimization and innovation in the architecture, it achieves better performance while maintaining real-time performance.

[0029] Step S5: Target model interpretation. A new convolution module, AConv convolution module, is added. This module effectively improves the model's detection capability for complex edge structures of AE and CE through multi-path convolution processing, pooling operation and multi-scale feature fusion, and improves the shortcomings of conventional YOLOv8 convolution in such complex target detection.

[0030] First, the AConv module divides the input feature map into two parts according to the number of channels, and processes them through different convolution paths respectively, such as Figure 2 As shown. The first part of the feature map undergoes a 3×3 convolution operation, which not only performs spatial downsampling (stride 2), but also effectively extracts local spatial features. In this way, the model can obtain rich local information from the feature map and adapt to the complex and diverse morphological characteristics of AE and CE edge details. The second part of the feature map first undergoes a 1×1 convolution for channel compression, and then performs a maximum pooling operation. This design retains a lot of detail information, while capturing the maximum value of local features through pooling operations, enhancing the model's sensitivity to the edge details of vortex targets. This can more effectively capture edge features and reduce information loss when dealing with the complex forms of AE and CE.

[0031] Among them, the second path adds maximum pooling and average pooling operations, especially maximum pooling, which can effectively retain local features and is not easily disturbed by background noise. Maximum pooling improves the model's ability to recognize complex edges by extracting significant information from local features, while average pooling helps capture global information and reduce the impact of background noise on vortex edge detection. This pooling strategy enables the model to more accurately distinguish the edge and background of the vortex in the AE and CE detection tasks, especially in complex scenes with low contrast.

[0032] Furthermore, the AConv module concatenates the features processed by different convolution paths in the channel dimension through the Concat operation. This feature fusion design takes into account the local spatial features extracted by the 3*3 convolution and the local important information captured by the 1*1 convolution and maximum pooling operations, thereby improving the ability to capture the AE and CE vortex edge structures. Through this feature fusion, the AConv module reduces the amount of calculation while enhancing the model's perception of multi-scale features, especially when dealing with AE and CE targets with complex morphology and rich edge details, which can effectively improve detection accuracy.

[0033] Furthermore, in order to adjust the PAN-FPN structure of YOLOv8, such as Figure 3As shown in the figure, in order to better deal with the inaccuracy problem in target detection, BiFPN (weighted bidirectional feature pyramid network) is introduced and improved. Compared with the traditional feature fusion method, BiFPN introduces a weighting mechanism, which enables the network to assign different weights according to the importance of each feature during the fusion process, ensuring the rationality and effectiveness of feature fusion. This strategy enables the network to process features of different scales more flexibly and accurately, especially in the detection of complex structures such as AE and CE. Through weight adjustment, the key details of each input feature can be captured more accurately.

[0034] Among them, in order to further improve the detection capability of large-scale targets, a P6 detection layer is added on the basis of BiFPN. The P6 layer is mainly used to extract larger-scale features, which is crucial for detecting large-scale, complex-structured vortices (such as CE and AE). Since these vortex structures usually appear as large-scale circular or eddy morphology, the P5 layer may not provide a clear enough representation, so the added P6 layer can better capture these complex structural features and improve the accuracy of positioning and classification.

[0035] Furthermore, improved feature fusion and new detection heads are added. In addition to the BiFPN and P6 layers, the feature fusion method is further adjusted and a new detection head is added. The new detection head is particularly effective in dealing with the complex morphology of vortex structures, reducing feature misalignment problems and information loss. Compared with the original PAN-FPN structure, the feature fusion method proposed in this paper is more flexible, fully utilizes multi-level features, and improves the detection performance of AE and CE under different backgrounds.

[0036] Furthermore, the LSK (Large Kernel with Stride) attention mechanism is introduced. Since the introduction of the Transformer model, the attention mechanism has gradually become an important tool for feature extraction due to its ability to focus on global features. In target detection tasks, especially small target detection and target recognition in complex scenes, the attention mechanism can effectively improve the model's ability to focus on small objects and reduce missed detections and false detections.

[0037] Among them, the LSK attention mechanism is a spatial attention mechanism that combines large convolution kernels and multi-scale convolution operations, such as Figure 4 As shown in the figure, it aims to improve the model's ability to distinguish features of different spatial locations by expanding the receptive field and extracting multi-scale features. It mainly consists of the following parts:

[0038] Large Kernel Receptive Field Expansion: The receptive field of traditional convolution operations is limited by the size of the convolution kernel, which is usually 3*3 or 5*5. In order to improve the model's perception ability in complex backgrounds and small target detection, LSK significantly expands the receptive field by using larger convolution kernels (such as 5*5, 7*7, etc.). For example, the first layer of convolution uses a 5*5 convolution kernel to extract preliminary spatial information from the input feature map.

[0039] Multi-scale convolution operation: A single convolution kernel size cannot effectively capture features of different scales. To this end, the LSK module further expands the receptive field and extracts a wider range of contextual information by introducing multiple convolution kernels of different scales (such as 7*7 convolution kernels and convolutions with dilation rates). Dilated convolution (dilation=3) further enhances the model's ability to capture long-range features, enabling it to extract rich spatial information at multiple scales.

[0040] Feature fusion and attention calculation: After the multi-scale convolution operation, the LSK module fuses the outputs of different convolution paths. The specific operation is to weight the features in the spatial dimension by combining the average attention and the maximum attention. In this way, the model can dynamically adjust the weights of the features and pay more attention to those areas with high information content. Finally, after compression and fusion of the channel dimension, the final attention weight is generated.

[0041] In order to dynamically select the appropriate spatial kernel, the input feature map is divided into multiple sub-feature maps. Subsequently, various convolution kernels of different sizes are applied to each sub-feature map to generate multiple output feature maps:

[0042]

[0043] The concatenated feature maps undergo average pooling and maximum pooling operations along the channel dimension:

[0044]

[0045] After concatenation, they are transformed into spatial attention maps using convolutional layers, ensuring that they have the same number of depthwise convolutions:

[0046] S^A=F 2→N ([SA avg ;SA max ])

[0047] The weighted deep convolution feature map is obtained by element-wise multiplication of the weights and the corresponding deep convolution, and a convolutional layer is used to fuse these feature maps and generate the final attention feature:

[0048]

[0049]

[0050] The LSK attention mechanism enhances the YOLOv8 model's ability to detect small targets in complex backgrounds by combining large convolution kernels, multi-scale convolution operations, and attention weights. In the detection of cyclonic and anticyclonic vortices, the LSK module helps the model better extract the multi-scale features of the vortex structure and improves the ability to recognize and classify vortex morphology. In addition, improvements in the LSK module in multi-scale feature extraction, spatial feature differentiation, and detection robustness further improve the model's performance in various complex scenarios. In the future, the LSK attention mechanism is expected to continue to be optimized and expanded to target detection tasks in more complex scenarios, providing stronger performance guarantees for various target detection tasks.

[0051] The method proposed in this embodiment achieves an ideal balance between real-time and accuracy in the anticyclonic vortex (AE) and cyclonic vortex (CE) detection tasks through high precision, high recall rate, low inference time and optimized computational complexity, and becomes an excellent solution for vortex detection tasks in complex backgrounds. Compared with other mainstream models, this method not only has strong detection capabilities, but also performs well in real-time application scenarios. It is an efficient and reliable choice in AE and CE detection tasks, and is suitable for application scenarios that require both high precision and real-time response.

Claims

1. A method for detecting mesoscale eddies in the ocean based on convolutional neural networks. This method effectively improves the model's ability to detect complex edge structures of AE and CE through multi-path convolution processing, pooling operations, and multi-scale feature fusion, and improves the shortcomings of conventional YOLOv8 convolution in detecting such complex targets. The specific construction steps are: Step S1: Obtain multi-source data of ocean eddies, such as sea surface height, temperature, etc. Ensure that the data has high quality and high temporal and spatial resolution; Step S2: Preprocess the data. Perform missing value processing, error value correction and normalization on the vortex data: Step S3: Divide the data into training set, validation set and test set, and perform necessary data enhancement, such as rotation, flipping, scaling, etc., to expand the data set; Step S4: Detection model construction and comparative analysis. This method builds an experimental platform based on the YOLOv8 model. The YOLO (You Only Look Once) series of models are widely used in various target detection tasks due to their end-to-end target detection process, fast inference speed, and high detection accuracy. As the latest version of the series, YOLOv8 continues the high efficiency of the YOLO family model, and through further optimization and innovation in the architecture, it achieves better performance while maintaining real-time performance. Step S5: Target model interpretation. A new convolution module, AConv convolution module, is added. This module effectively improves the model's detection capability for complex edge structures of AE and CE through multi-path convolution processing, pooling operation and multi-scale feature fusion, and improves the shortcomings of conventional YOLOv8 convolution in such complex target detection.

2. Furthermore, in order to adjust the structure of YOLOv8 to better handle the inaccuracy problem in target detection, BiFPN (weighted bidirectional feature pyramid network) is introduced and improved. Compared with the traditional feature fusion method, BiFPN introduces a weighting mechanism, which enables the network to assign different weights according to the importance of each feature during the fusion process, ensuring the rationality and effectiveness of feature fusion. This strategy enables the network to process features of different scales more flexibly and accurately, especially in the detection of complex structures such as AE and CE. Through weight adjustment, the key details of each input feature can be captured more accurately. Furthermore, the LSK (Large Kernel with Stride) attention mechanism is introduced. Since the introduction of the Transformer model, the attention mechanism has gradually become an important tool for feature extraction due to its ability to focus on global features. In target detection tasks, especially small target detection and target recognition in complex scenes, the attention mechanism can effectively improve the model's ability to focus on small objects and reduce missed detections and false detections.

Images