Intelligent ocean front detection method based on U-Net network structure
By improving the U-Net network structure and residual learning units, and combining them with the Dice coefficient loss function, the inefficiency and adaptability problems of traditional methods are solved, achieving high-precision automatic identification of ocean fronts and adapting to complex marine environments.
Patent Information
- Application Number
- CN202511695060.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-18
- Publication Date
- 2026-02-24
AI Technical Summary
Traditional ocean front identification methods are inefficient, difficult to adapt to the needs of large-scale data processing, and exhibit poor adaptability and generalization ability in complex marine environments, which can easily lead to missed detections and misjudgments.
An improved U-Net network structure is adopted, which combines residual learning units and Dice coefficient loss function. An end-to-end ocean front detection model is constructed by preprocessing satellite remote sensing data, enhancing gradient features, and creating multi-scale front labels.
It achieves high-precision and high-efficiency automatic identification of marine fronts, improving the accuracy and robustness of detection and adapting to complex environments under different sea areas, seasons and meteorological conditions.
Smart Images
Figure CN121564564A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of marine remote sensing image processing and computer vision technology, and in particular to a method for intelligent detection of marine fronts based on a U-Net network structure. Background Technology
[0002] Oceanic fronts are narrow transitional zones formed at the interface between two bodies of water with different properties. As a typical mesoscale oceanic phenomenon, their rapid and accurate identification is of great significance for marine resource development and management, ecological environment protection, climate change research, and military navigation support.
[0003] In recent years, with the rapid development of satellite remote sensing technologies (such as infrared radiometers and microwave radiometers) and ocean numerical models, the spatiotemporal resolution and volume of globally available marine environmental data have significantly improved. This massive and continuous sea surface temperature and other remote sensing data provides a solid foundation for systematic and large-scale ocean front analysis, while also posing a serious challenge to traditional methods relying on manual interpretation and semi-automatic detection based on fixed thresholds. These traditional methods typically set gradient thresholds based on expert experience, resulting in low processing efficiency, difficulty in adapting to large-scale data processing needs, and poor adaptability and generalization ability when facing complex marine environments under different sea areas, seasons, and meteorological conditions, easily leading to missed detections and misjudgments.
[0004] With the continuous advancement of artificial intelligence technology, deep learning, especially image segmentation methods based on convolutional neural networks (CNNs), has provided a new technical path for the automatic recognition of marine veins. Among numerous deep learning architectures, the U-Net network, with its encoder-decoder structure and skip connection mechanism, can effectively fuse deep semantic features and shallow spatial information of images. It has demonstrated outstanding performance in pixel-level segmentation tasks in fields such as medical imaging, showing its potential for application in marine vein detection.
[0005] The purpose of this invention is to construct a method for detecting ocean fronts using satellite remote sensing data by specifically optimizing the U-Net network structure and its training strategy, thereby achieving high-precision and high-efficiency automatic identification of ocean fronts. Summary of the Invention
[0006] The purpose of this invention is to construct a method for detecting ocean fronts using satellite remote sensing data by specifically optimizing existing deep learning networks and their training strategies. This invention provides an intelligent ocean front detection method based on a U-Net network structure to achieve high-precision and high-efficiency intelligent identification of ocean fronts. It includes the following steps:
[0007] Step 1: Preprocess the sea surface temperature data observed by satellite remote sensing to obtain a color image of the sea surface temperature gradient;
[0008] Step 2: Based on the sea surface temperature gradient data, use an automatic front extraction algorithm that combines gradient method and edge detection method to create labels for ocean fronts and frontal zones;
[0009] Step 3: Use the sea surface temperature gradient map as the model input to construct a dataset and divide it into a training set and a test set. Input the training set into the optimized U-Net network model and train the model using ocean front labels. After training, evaluate the model using the test set.
[0010] The improved U-Net model described in this invention introduces a residual learning unit. This model uses a 3rd-order full 32-bit filter architecture and adopts a typical "encoder-decoder" network architecture. It adds the feature information of the previous residual block to the next residual unit, effectively avoiding the performance degradation and feature information loss problems caused by excessive network depth.
[0011] The encoder path has three stages, each containing a residual unit module. After residual connection, it is downsampled through a max pooling layer.
[0012] The decoder path also has three stages. In each stage, the upsampled feature map is first concatenated with the corresponding level of the encoder convolutional layer to return to the original resolution of the image, and then the corresponding feature extraction work is performed.
[0013] The final prediction network consists of a 1×1 convolutional layer followed by a Softmax activation function, which converts the feature values into pixel classification probabilities. Finally, the feature map size is restored to the original input size, and the output channel count is 2, achieving pixel-level binary classification of ocean fronts and background.
[0014] Step 4: Preprocess the sea surface temperature data of the sea area to be measured obtained from satellite remote sensing to obtain a color image of the sea surface temperature gradient, and input it into the qualified model for intelligent detection of ocean fronts and frontal zones.
[0015] In the above scheme, in step 1, the sea surface temperature data observed by satellite remote sensing is first preprocessed by median filtering to suppress salt-and-pepper noise and small-scale outliers in the data; then, the Sobel operator is used to calculate the spatial gradient value of the repaired sea surface temperature data; finally, the calculated temperature gradient amplitude data is contrast-enhanced and pseudo-color mapped to convert it into a three-channel color image.
[0016] In the above scheme, in step 1, the sea surface temperature data observed by satellite remote sensing is processed by median filtering. The window size (W, H) of the median filter is determined by the following formula:
[0017]
[0018] Where dx represents the spatial resolution of the data (unit: degrees / pixel). This represents the function for rounding up.
[0019] In the above scheme, step 1 uses Sobel calculation to calculate the spatial gradient value of the repaired sea surface temperature data, and the formula is as follows:
[0020]
[0021]
[0022] Where G represents the gradient value at the current coordinate point, G x G y These represent the gradient values of the current coordinate point in the horizontal and vertical directions, respectively. The symbol * indicates convolution operation, and I is the two-dimensional scalar field of the repaired sea surface temperature data.
[0023] In the above scheme, in step 2, based on the sea surface temperature gradient data obtained in step 1, an image segmentation method based on dual thresholds is used to identify oceanic frontal zones. Specifically, two gradient thresholds are set: a high threshold is the temperature gradient value of 90% or more pixels in the image on that day, and a low threshold is the temperature gradient value of 80% or more pixels. According to the dual thresholds, each pixel in the image is divided into two categories: areas that meet the gradient threshold conditions are identified as oceanic frontal zones, and the remaining areas are identified as background. Based on this, the pixel values of oceanic frontal zone pixels are marked as 1, and the pixel values of background pixels are marked as 0, generating a binarized oceanic frontal zone labeled image.
[0024] In the above scheme, in step 2, based on the ocean front region data obtained in step 2, the ocean front surface is extracted using a non-maximum suppression and dual-threshold edge detection algorithm. The pixels in the extracted result are divided into two categories: pixels belonging to the ocean front surface are labeled as 1, and background pixels are labeled as 0, generating a binarized ocean front surface labeled image.
[0025] In the above scheme, step 3, the residual unit module mainly includes two batch normalization processing layers, a ReLU activation function layer, a Dropout layer (Dropout=0.5) and a convolutional layer with a window size of 3×3. The above structure is repeated twice. After residual connection, downsampling is performed through a max pooling layer with a window size of 2×2.
[0026] In the above scheme, step 3, the encoder path uses a residual unit module to achieve feature extraction and dimensionality reduction, consisting of three downsampling stages. The specific implementation is as follows:
[0027] In the first stage, the input feature map of size H×W×3 first passes through a residual unit module composed of 3×3 convolutional kernels for feature extraction, and then passes through a 2×2 max pooling layer for downsampling. The output feature map size becomes H / 2×W / 2×64, which is half of the original input size.
[0028] In the second stage, the feature map output from the first stage is input into this stage; after being downsampled by a residual unit module consisting of 3×3 convolutional kernels and a 2×2 max pooling layer, the size of the output feature map becomes H / 4×W / 4×128, which is one-quarter of the original input size;
[0029] In the third stage, the feature map output from the second stage is input into this stage; after being downsampled by a residual unit module consisting of 3×3 convolutional kernels and a 2×2 max pooling layer, the size of the output feature map becomes H / 8×W / 8×256, which is one-eighth of the original input size.
[0030] Through the above three stages of processing, the original input image is gradually encoded into a high-dimensional, abstract feature representation.
[0031] In the above scheme, step 3, the decoder stage consists of three sequentially connected upsampling stages, specifically implemented as follows:
[0032] In the first stage, the input feature map, with a size of H / 8×W / 8×256, is first upsampled through a 2×2 transposed convolutional layer to double its size, and then concatenated with the feature maps of the corresponding layers in the encoder path. Subsequently, it undergoes feature fusion through a residual unit module composed of 3×3 convolutional kernels. After this stage, the output feature map size is H / 4×W / 4×128, which is one-quarter of the original input size.
[0033] In the second stage, the feature map output from the first stage is input into this stage. First, it is upsampled through a 2×2 transposed convolutional layer and then concatenated with the feature map of the corresponding layer in the encoder path. Afterwards, it undergoes feature fusion through a residual unit module composed of 3×3 convolutional kernels. After this stage, the output feature map size is H / 2×W / 2×64, which is half the original input size.
[0034] In the third stage, the feature map output from the second stage is input into this stage. First, it is upsampled through a 2×2 transposed convolutional layer and then concatenated with the feature map of the corresponding layer in the encoder path. Afterwards, it undergoes feature fusion through a residual unit module composed of 3×3 convolutional kernels. Finally, the output feature map has a size of H×W×32, which restores it to the original input size.
[0035] Through the above three stages of processing, the high-dimensional features output by the encoder are gradually upsampled and fused with the multi-scale features of the encoder, and finally restored to the original image resolution, completing the mapping and reconstruction from features to segmentation results.
[0036] In the above scheme, in step 3, during model training, the weighted Dice coefficient is used to calculate the loss function. The formula for calculating the Dice coefficient is as follows:
[0037]
[0038] Where P is the predicted region, G is the true region, and |P|+|G| represents the sum of elements in the two regions.
[0039] The formula for calculating the loss function is as follows:
[0040] D weighted =αD front +βD bg
[0041] L = 1 - D weighted
[0042] Among them, D front and D bg These are the Dice coefficients for categories belonging to the frontal region and the background region, respectively, with α and β representing their respective weights. weighted The Dice coefficient is used to weight the overall data, and L is the loss function value calculated by the model.
[0043] Beneficial Effects: Through the above technical solution, this invention provides an innovative intelligent detection method for ocean fronts based on an improved U-Net network. This method effectively solves the problems of weak generalization ability of traditional thresholding methods and the limitations of the basic U-Net model in complex feature extraction by constructing an enhanced encoder*decoder architecture that integrates residual learning. This invention realizes a complete technical process from sea surface temperature data preprocessing and gradient feature enhancement to multi-scale front label creation and end-to-end model training, significantly improving the accuracy, robustness, and automation of ocean front detection, and providing reliable technical support for marine environmental monitoring and applications. Attached Figure Description
[0044] Figure 1 Flowchart for creating the dataset according to an embodiment of the present invention
[0045] Figure 2 This is a model structure diagram of an embodiment of the present invention. Detailed Implementation
[0046] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0047] The environment used in this embodiment of the invention is as follows: the server's CPU is an Intel(R) Core(TM) i7-14700HX, the GPU is an NVIIIA GeForce RTX 4060 24G, the operating system is Ubuntu 20.04 64-bit, and the compilation environment is Python 3.8 and PyTorch 1.10.
[0048] This invention provides an intelligent detection method for ocean fronts based on a U-Net network structure, achieving high-precision and high-efficiency intelligent identification of ocean fronts. It includes the following steps:
[0049] Step 1: Preprocess the sea surface temperature data observed by satellite remote sensing to obtain a color image of the sea surface temperature gradient.
[0050] First, the sea surface temperature data observed by satellite remote sensing is preprocessed by median filtering. Then, the Sobel operator is used to calculate the spatial gradient value of the repaired sea surface temperature data. Finally, the calculated temperature gradient amplitude data is contrast-enhanced and pseudo-color mapped to convert it into a three-channel color image.
[0051] Median filtering is used to preprocess sea surface temperature data obtained from satellite remote sensing. The window size (W, H) of the median filter is determined by the following formula:
[0052]
[0053] Where dx represents the spatial resolution of the data (unit: degrees / pixel). This represents the function for rounding up.
[0054] The spatial gradient of the restored sea surface temperature data was calculated using Sobel calculation, and the formula is shown below:
[0055]
[0056] Where G represents the gradient value at the current coordinate point, G x G yThese represent the gradient values of the current coordinate point in the horizontal and vertical directions, respectively. The symbol * indicates convolution operation, and I is the 2D scalar field of the repaired sea surface temperature data.
[0057] Step 2: Based on the sea surface temperature gradient data, an automatic front extraction algorithm that combines gradient fusion and edge detection methods is used to create labels for ocean fronts and frontal zones. The algorithm flowchart for creating the sample dataset is shown below. Figure 1 As shown:
[0058] Based on the sea surface temperature gradient data obtained in step 1, a dual-threshold image segmentation method is used to identify oceanic frontal zones. Specifically, two gradient thresholds are set: a high threshold is defined as the temperature gradient value of 90% or more pixels in the image on that day, and a low threshold is defined as the temperature gradient value of 80% or more pixels. According to the dual thresholds, each pixel in the image is divided into two categories: areas that meet the gradient threshold conditions are identified as oceanic frontal zones, and the remaining areas are identified as background. Based on this, the pixel values of oceanic frontal zone pixels are labeled as 1, and the pixel values of background pixels are labeled as 0, generating a binarized oceanic frontal zone labeled image.
[0059] Next, based on the ocean front region data obtained in this step, the ocean front surface is extracted using a non-maximum suppression and dual-threshold edge detection algorithm. The extracted pixels are divided into two categories: pixels belonging to the ocean front surface are labeled as 1, and background pixels are labeled as 0, generating a binarized ocean front surface labeled image.
[0060] Step 3: Use the sea surface temperature gradient map as the model input to construct a dataset and divide it into a training set and a test set. Input the training set into the optimized U-Net network model and train the model using ocean front labels. After training, evaluate the model using the test set.
[0061] (1) Model building
[0062] The improved U-Net model described in this invention introduces residual learning units, and the model structure is as follows: Figure 2 As shown, this model uses a 3rd-order full 32-bit filter architecture and a typical encoder-decoder network architecture, which adds the feature information of the previous residual block to the next residual unit, effectively avoiding the performance degradation and feature information loss problems caused by excessive network depth.
[0063] The improved U-Net model's encoder path has three stages. Each stage's residual unit module mainly includes two batch normalization processes, a ReLU activation function layer, and a 3×3 convolutional layer. After residual concatenation, downsampling is performed through a 2×2 max-pooling layer. Dropout layers (Dropout = 0.5) are added before each max-pooling layer and before each transposed convolutional layer, as implemented below:
[0064] In the first stage, the input feature map of size H×W×3 first passes through a residual unit module composed of 3×3 convolutional kernels for feature extraction, and then passes through a 2×2 max pooling layer for downsampling. The output feature map size becomes H / 2×W / 2×64, which is half of the original input size.
[0065] In the second stage, the feature map output from the first stage is input into this stage; after being downsampled by a residual unit module consisting of 3×3 convolutional kernels and a 2×2 max pooling layer, the size of the output feature map becomes H / 4×W / 4×128, which is one-quarter of the original input size;
[0066] In the third stage, the feature map output from the second stage is input into this stage; after being downsampled by a residual unit module consisting of 3×3 convolutional kernels and a 2×2 max pooling layer, the size of the output feature map becomes H / 8×W / 8×256, which is one-eighth of the original input size.
[0067] The decoder path of the improved U-Net model also has three stages. The specific implementation is as follows:
[0068] In the first stage, the input feature map, with a size of H / 8×W / 8×256, is first upsampled through a 2×2 transposed convolutional layer to double its size, and then concatenated with the feature maps of the corresponding layers in the encoder path. Subsequently, it undergoes feature fusion through a residual unit module composed of 3×3 convolutional kernels. After this stage, the output feature map size is H / 4×W / 4×128, which is one-quarter of the original input size.
[0069] In the second stage, the feature map output from the first stage is input into this stage. First, it is upsampled through a 2×2 transposed convolutional layer and then concatenated with the feature map of the corresponding layer in the encoder path. Afterwards, it undergoes feature fusion through a residual unit module composed of 3×3 convolutional kernels. After this stage, the output feature map size is H / 2×W / 2×64, which is half the original input size.
[0070] In the third stage, the feature map output from the second stage is input into this stage. First, it is upsampled through a 2×2 transposed convolutional layer and then concatenated with the feature map of the corresponding layer in the encoder path. Afterwards, it undergoes feature fusion through a residual unit module composed of 3×3 convolutional kernels. Finally, the output feature map has a size of H×W×32, which restores it to the original input size.
[0071] The final prediction network consists of a 1×1 convolutional layer followed by a Softmax activation function, which converts the feature values into pixel classification probabilities. Finally, the feature map size is restored to the original input size, and the output channel count is 2, achieving pixel-level binary classification of ocean fronts and background.
[0072] (2) Model Training
[0073] In model training, the loss function is calculated using weighted Dice coefficients. The formula for calculating the Dice coefficients is as follows:
[0074]
[0075] Where P is the predicted region, G is the true region, and |P|+|G| represents the sum of elements in the two regions.
[0076] The formula for calculating the loss function is as follows:
[0077] D weighted =αD front +βD bg
[0078] L = 1 - D weighted
[0079] Among them, D front and D bg These are the Dice coefficients for categories belonging to the frontal region and the background region, respectively, with α and β representing their respective weights. weighted The Dice coefficient is used to weight the overall data, and L is the loss function value calculated by the model.
[0080] During training, the training and test sets were divided in a 9:1 ratio, and the epoch was set to 100. To avoid overfitting, an early stopping mechanism was implemented: the training process was automatically terminated when the loss function value on the validation set did not decrease for 10 consecutive training epochs.
[0081] Step 4: Preprocess the sea surface temperature data of the sea area to be measured obtained from satellite remote sensing to obtain a color image of the sea surface temperature gradient, and input it into the qualified model for intelligent detection of ocean fronts and frontal zones.
[0082] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for intelligent detection of ocean fronts based on a U-Net network structure, characterized in that, Includes the following steps: Step 1: Preprocess the sea surface temperature data observed by satellite remote sensing to obtain a color image of the sea surface temperature gradient; Step 2: Based on the sea surface temperature gradient data, use an automatic front extraction algorithm that combines gradient method and edge detection method to create labels for ocean fronts and frontal zones; Step 3: Use the sea surface temperature gradient map as the model input to construct a dataset and divide it into a training set and a test set. Input the training set into the optimized U-Net network model and train the model using ocean front labels. After training, evaluate the model using the test set. The improved U-Nct model described in this invention introduces a residual learning unit. This model uses a 3rd-order full 32-bit filter architecture and a typical "encoder-decoder" network architecture. It adds the feature information of the previous residual block to the next residual unit, effectively avoiding the performance degradation and feature information loss problems caused by excessive network depth. The encoder path has three stages, each containing a residual unit module. After residual connection, it is downsampled through a max pooling layer. The decoder path also has three stages. In each stage, the upsampled feature map is first concatenated with the corresponding level of the encoder convolutional layer to return to the original resolution of the image, and then the corresponding feature extraction work is performed. The final prediction network is a 1×1 convolutional layer followed by a Softmax activation function, which converts the feature values into pixel classification probabilities. Finally, the feature map size is restored to the original input size, and the number of output channels is 2, achieving pixel-level binary classification of ocean fronts and background. Step 4: Preprocess the sea surface temperature data of the sea area to be measured obtained from satellite remote sensing to obtain a color image of the sea surface temperature gradient, and input it into the qualified model for intelligent detection of ocean fronts and frontal zones.
2. The intelligent detection method for ocean fronts based on a U-Nct network structure according to claim 1, characterized in that, In step 1, the sea surface temperature data observed by satellite remote sensing is first preprocessed by median filtering to suppress salt-and-pepper noise and small-scale outliers in the data. The spatial gradient values of the repaired sea surface temperature data were then calculated using the Sobel operator. Finally, the calculated temperature gradient amplitude data were contrast-enhanced and pseudo-color mapped to convert it into a three-channel color image.
3. The intelligent detection method for ocean fronts based on a U-Net network structure according to claim 1, characterized in that, In step 1, the sea surface temperature data observed by satellite remote sensing is preprocessed using median filtering. The window size (W, H) of the median filter is determined by the following formula: Where dx represents the spatial resolution of the data (unit: degrees / pixel). This represents the function for rounding up.
4. The intelligent detection method for ocean fronts based on a U-Net network structure according to claim 1, characterized in that, In step 1, the spatial gradient value of the restored sea surface temperature data is calculated using Sobel calculation, and the formula is as follows: Where G represents the gradient value at the current coordinate point, G x G y These represent the gradient values of the current coordinate point in the horizontal and vertical directions, respectively. The symbol * indicates convolution operation, and I is the two-dimensional scalar field of the repaired sea surface temperature data.
5. The intelligent detection method for ocean fronts based on a U-Net network structure according to claim 1, characterized in that, In step 2, based on the sea surface temperature gradient data obtained in step 1, an image segmentation method based on dual thresholds is used to identify the ocean front zone. Specifically, two gradient thresholds are set: a high threshold is the temperature gradient value of 90% or more pixels in the image on that day, and a low threshold is the temperature gradient value of 80% or more pixels. According to the dual thresholds, each pixel in the image is divided into two categories: the area that meets the gradient threshold condition is determined to be the ocean front zone, and the remaining area is determined to be the background. Based on this, the pixel value of the ocean front zone pixel is marked as 1, and the pixel value of the background pixel is marked as 0, generating a binarized ocean front zone labeled image.
6. The intelligent detection method for ocean fronts based on a U-Net network structure according to claim 1, characterized in that, In step 2, based on the ocean front region data obtained in step 2, the ocean front surface is extracted using a non-maximum suppression and dual-threshold edge detection algorithm. The pixels in the extracted results are divided into two categories: pixels belonging to the ocean front surface are labeled as 1, and background pixels are labeled as 0, generating a binarized ocean front surface labeled image.
7. The intelligent detection method for ocean fronts based on a U-Nct network structure according to claim 1, characterized in that, In step 3, the residual unit module mainly includes two batch normalization processing layers, a ReLU activation function layer, a Dropout layer (Dropout=0.5) and a 3×3 convolutional layer. The above structure is repeated twice. After residual connection, downsampling is performed through a 2×2 max pooling layer.
8. The intelligent detection method for ocean fronts based on a U-Nct network structure according to claim 1, characterized in that, In step 3, the encoder path uses residual unit modules to implement feature extraction and dimensionality reduction, which consists of three downsampling stages, as detailed below: In the first stage, the input feature map of size H×W×3 first passes through a residual unit module composed of 3×3 convolutional kernels for feature extraction, and then passes through a 2×2 max pooling layer for downsampling. The output feature map size becomes H / 2×W / 2×64, which is half of the original input size. In the second stage, the feature map output from the first stage is input into this stage; after being downsampled by a residual unit module consisting of 3×3 convolutional kernels and a 2×2 max pooling layer, the size of the output feature map becomes H / 4×W / 4×128, which is one-quarter of the original input size; In the third stage, the feature map output from the second stage is input into this stage. After passing through a residual unit module consisting of 3×3 convolutional kernels and a 2×2 max pooling layer for downsampling, the size of the output feature map becomes H / 8×W / 8×256, which is one-eighth of the original input size. Through the above three stages of processing, the original input image is gradually encoded into a high-dimensional, abstract feature representation.
9. The intelligent detection method for ocean fronts based on a U-Net network structure according to claim 1, characterized in that, In step 3, the decoder stage consists of three sequentially connected upsampling stages, as specifically implemented below: In the first stage, the input feature map, which has a size of H / 8×W / 8×256, is first upsampled through a 2×2 transposed convolutional layer to double its size. Then, it is concatenated with the feature map of the corresponding layer in the encoder path. After that, it is fused through a residual unit module composed of 3×3 convolutional kernels. After this stage, the output feature map has a size of H / 4×W / 4×128, which is one-quarter of the original input size. In the second stage, the feature map output from the first stage is input into this stage. First, it is upsampled through a 2×2 transposed convolutional layer and concatenated with the feature map of the corresponding layer in the encoder path. Then, it is fused through a residual unit module composed of 3×3 convolutional kernels. After this stage, the output feature map size is H / 2×W / 2×64, which is half of the original input size. In the third stage, the feature map output from the second stage is input into this stage. First, it is upsampled through a 2×2 transposed convolutional layer and then concatenated with the feature map of the corresponding layer in the encoder path. After that, it is fused through a residual unit module composed of 3×3 convolutional kernels. Finally, the output feature map has a size of H×W×32, which restores it to the original input size. Through the above three stages of processing, the high-dimensional features output by the encoder are gradually upsampled and fused with the multi-scale features of the encoder, and finally restored to the original image resolution, completing the mapping and reconstruction from features to segmentation results.
10. The intelligent detection method for ocean fronts based on a U-Net network structure according to claim 1, characterized in that, In step 3, during model training, the loss function is calculated using weighted Dice coefficients. The formula for calculating the Dice coefficients is as follows: Where P is the predicted region, G is the true region, and |P|+|G| represents the sum of elements in the two regions; The formula for calculating the loss function is as follows: D weiqhted =αD front +βD bg L=1-D weighted Among them, D front and D bg These are the Dice coefficients for categories belonging to the frontal region and the background region, respectively, with α and β representing their respective weights. weighted The Dice coefficient is used to weight the overall data, and L is the loss function value calculated by the model.