Intelligent ocean front detection method integrating dense skip connections and seasonal enhancement

Abstract

The present invention relates to the field of ocean temperature front detection, and discloses an ocean front intelligent detection method integrating dense skip connection and seasonal enhancement, comprising the following steps: obtaining daily satellite remote sensing SST data of the research sea area, and preprocessing to obtain an SST gradient map; marking the front labels to obtain front label data; constructing a seasonal enhanced multi-scale recognition network model integrating dense skip connection; the model includes 5 encoders, 4 decoders, a head network and 6 intermediate states; dividing the SST gradient map and the front label data into a training set and a test set, and the model is trained and the accuracy is evaluated; after preprocessing the SST data of the sea area to be tested, an SST gradient map is obtained, and the SST gradient map is input into the qualified model for intelligent detection of ocean fronts. The method disclosed by the present invention can improve the overall expressiveness of the model and the accuracy of front detection.

Description

Technical Field

[0001] The present invention relates to the field of ocean temperature front detection, and in particular to an ocean front intelligent detection method integrating dense jump connection and seasonal enhancement. Background Art

[0002] In the ocean front region, the temperature, salinity, chlorophyll concentration and other physical properties of seawater change significantly, so the ocean front can be characterized by factors such as temperature, salinity, and chlorophyll concentration. As a common mesoscale phenomenon in the ocean, the ocean front is of great significance to the application fields of marine fisheries, atmospheric physical processes, and marine military. It is crucial to timely grasp the location and movement laws of the ocean front for research and practice in the above fields.

[0003] The earliest ocean front research relied on field observation methods such as station observation, ship measurement and buoy observation. However, these methods are easily restricted by factors such as weather conditions and ship navigation, and it is difficult to obtain long-term and large-scale observation data. In comparison, satellite remote sensing technology, with its advantages of wide-area coverage, all-day, all-weather, and little impact from weather changes, can obtain long-term, large-scale, and high-spatial coverage observation data, which effectively supplements the shortcomings of field observations. Therefore, more and more researchers have begun to use remote sensing data for ocean front detection. Neural networks perform well in modeling complex data and have strong spatial feature extraction capabilities. They have been applied to the automated detection of ocean fronts, but how to further improve the accuracy of front detection is still a key issue that needs to be solved in the field of front research. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides an ocean front intelligent detection method that integrates dense skip connections and seasonal enhancement, so as to improve the overall expressiveness of the model and the accuracy of front detection.

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] An intelligent ocean front detection method integrating dense skip connections and seasonal enhancement comprises the following steps:

[0007] Step 1: Obtain daily satellite remote sensing SST data of the study area and preprocess it to obtain an SST gradient map;

[0008] Step 2: Based on the SST gradient map, combined with previous ocean front research and expert experience, Labelme is used to annotate the front labels to obtain the front label data;

[0009] Step 3, construct a seasonal enhanced multi-scale recognition network model integrating dense skip connections;

[0010] The model includes 5 encoders, 4 decoders, a head network and 6 intermediate states;

[0011] The encoder is composed of two branches, the bottom branch is an improved multi-scale residual convolution block, that is, the jump connection operation is integrated into the multi-scale residual convolution block, and the ReLU nonlinear activation function is used after each layer of convolution to realize feature extraction of different scales; the upper branch of the encoder is a channel supervision unit, which is composed of 1 global average pooling layer and 2 fully connected layers. At the same time, improved seasonal coding is added inside the channel supervision unit, and the feature map obtained by the bottom branch and the upper branch is added and then passed through the sampling layer to obtain the output of the encoder;

[0012] The decoder has the same structure as the encoder, except for the last sampling layer. The last layer of the encoder is a downsampling layer, while the last layer of the decoder is an upsampling layer. The feature map is used to extract features and perform downsampling in the encoder, and the number of channels is adjusted. The adjusted feature map is passed to the decoder for feature decoding and upsampling. The feature map output by the last decoder is passed to the head network.

[0013] The head network is responsible for pixel classification and generating detection results. The head network consists of two branches. The bottom branch consists of a 1×1 convolution layer and a Softmax layer. The upper branch is a position attention module. The results of the upper branch and the bottom branch are multiplied first, and the multiplication result is added to the result of the bottom branch to obtain the final detection result.

[0014] The 6 intermediate states are 6 feature maps obtained by upsampling the outputs of the second, third, and fourth encoders, respectively, which are composed of 1 upsampling operation of the second encoder, 2 upsampling operations of the third encoder, and 3 upsampling operations of the fourth encoder; the 6 feature maps obtained are the same size as the feature maps obtained by encoders or decoders at different positions, and then the encoder part and the feature maps obtained by the 6 intermediate states are jump-connected to the decoder part, and this process realizes the dense jump connection operation;

[0015] Step 4: divide the SST gradient map and front label data into a training set and a test set, use the training set to train the model, and use the test set to evaluate the model accuracy;

[0016] Step 5: After preprocessing the SST data of the sea area to be tested, an SST gradient map is obtained, and the SST gradient map is input into the qualified model for intelligent detection of the ocean front.

[0017] In the above scheme, the multi-scale residual convolution block is composed of three 3×3 convolutional layers, and the jump connection operation fused in the improved multi-scale residual convolution block is as follows: the input data is jump-connected to the outputs of the second and third 3×3 convolutional layers respectively, and the output of the first 3×3 convolutional layer is jump-connected to the output of the third 3×3 convolutional layer; finally, the outputs of the three 3×3 convolutional layers are spliced, and finally added element-by-element with the output of the input through the 1×1 convolutional layer.

[0018] In the above scheme, in the upper branch of the encoder, improved seasonal coding is performed on each feature map, that is, the seasonal coding of the current month is assigned a value of 0.8, and the seasonal coding of the month before and after the current month are both assigned a value of 0.1.

[0019] In the above scheme, in the encoder part, the input size is The feature map of is firstly passed through the first encoder for feature extraction and dimensionality reduction, and the output size is The feature map is then passed to the second encoder for feature extraction and dimensionality reduction, and the output size is The feature map is passed through 5 encoders in sequence, and the final output size is The feature map is passed to the decoder for feature decoding and upsampling. The feature map is reversed from the feature map of the encoder. After the feature map passes through four decoders, the feature map is restored to .

[0020] In the above scheme, in the upper branch of the head network, the input size is The feature map first passes through the average pooling layer to obtain a size of The feature map is then passed to the convolution layer, and the result of the convolution is added to the feature map of the position encoding; the added feature map is passed to a series of convolution layers for feature extraction. The last convolution layer is a convolution layer with a Sigmoid activation function to improve the nonlinear expression ability of the network. After that, the feature map size is restored to the original image size through the upsampling layer.

[0021] In the above scheme, the six intermediate states are specifically as follows: adding an upsampling layer after the second encoder, and the output of the second encoder is subjected to an upsampling operation to obtain a feature map of the same size as the output of the first encoder, and the size of the feature map is also the same as the output size of the fourth decoder; adding two upsampling layers after the third encoder, and the output of the third encoder is subjected to upsampling in the first layer to obtain a feature map of the same size as the output of the second encoder, and the size of the feature map is also the same as the output size of the third decoder; after another upsampling, a feature map of the same size as the output of the first encoder is obtained, and the size of the feature map is also the same as the output size of the fourth decoder; similarly, adding three upsampling layers after the fourth encoder, and the output of the fourth encoder is subjected to upsampling in the first layer to obtain a feature map of the same size as the output of the third encoder, and the size of the feature map is also the same as the output size of the second decoder; after upsampling in the second layer, a feature map of the same size as the output of the second encoder is obtained, and the size of the feature map is also the same as the output of the third decoder; after upsampling in the third layer, a feature map of the same size as the output of the first encoder is obtained, and the size of the feature map is also the same as the output size of the fourth decoder.

[0022] In the above scheme, the cross entropy loss function is used in model training, and the calculation formula is as follows:

[0023] ;

[0024] in, The first elements, Represents model predictions Belong to The probability of the class is Class time, =1, otherwise =0, Indicates the number of sample labels.

[0025] In the above scheme, the learning rate decay strategy is adopted during model training. When the loss error does not decrease within three consecutive training rounds, the learning rate at this time is reduced to half of the original value.

[0026] Through the above technical solution, the ocean front intelligent detection method integrating dense skip connections and seasonal enhancement provided by the present invention has the following beneficial effects:

[0027] The present invention proposes an ocean front detection model that integrates multi-scale residual convolution and dense connection. The model is based on LSENet, and introduces an improved multi-scale residual convolution module while adding dense skip connection operations. In addition, the original seasonal encoding in LSENet is optimized to construct a new ocean front detection deep learning model - Seasonal Enhanced Multi-Scale Recognition Network with Dense Skip Connections (SEMNet-DC) that integrates dense skip connections. The model uses the SST gradient map generated by satellite remote sensing sea surface temperature (SST) maps to realize automatic detection of ocean fronts, further improving the accuracy of front detection.

[0028] The original LSENet increases the network depth by stacking multiple layers of convolution. The network has strong nonlinear expression ability and can learn complex features contained in the input feature map, thereby effectively improving the network's expression ability and feature extraction ability. However, as the network depth increases, the learning ability of the shallow network will decrease, which not only limits the learning efficiency of the deep network, but also causes the network to lose a large amount of small-scale information, thereby affecting the detection accuracy of the model. Shallow networks are good at extracting basic information such as edges and textures contained in the input feature map. Jump connections can directly establish connections between certain layers of the network, allowing information to be directly transmitted across different layers. In this way, the low-level features extracted by the shallow network can be directly transmitted to the deep network and integrated with the high-level features, so that the output of the deep network not only contains high-level features, but also retains the small-scale edge information in the shallow features, reducing the risk of small-scale information loss caused by increasing network depth.

[0029] In order to enable the SEMNet-DC model to fully consider the small-scale information of the edge of the ocean front, 6 intermediate states were added to the model, and these intermediate states were jump-connected to the encoder part, fusing the feature information extracted by different deep networks, further enhancing the model's detection ability in terms of the edge of the ocean front.

[0030] The model uses an improved multi-scale residual convolution block to replace the stack of original 3×3 convolution layers. The skip connection operation is also applied in the improved multi-scale residual convolution block to improve the module's ability to extract features of fronts of different scales.

[0031] In addition, the present invention further improves the seasonal coding. Specifically, when performing seasonal coding, the code of the corresponding month is assigned a value of 0.8, and the codes of the preceding and following adjacent months are respectively assigned a value of 0.1. This provides additional contextual information for the network, which helps the network to better understand, capture and utilize the continuity and transition of the seasons, thereby improving the overall expressiveness of the model and the accuracy of front detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below.

[0033] Figure 1 A schematic diagram of a flow chart of an intelligent ocean front detection method integrating dense skip connections and seasonal enhancement disclosed in an embodiment of the present invention;

[0034] Figure 2 This is the structure diagram of the SEMNet-DC model;

[0035] Figure 3 This is the encoder structure diagram in the SEMNet-DC model;

[0036] Figure 4 It is a multi-scale residual convolution block structure diagram;

[0037] Figure 5 This is the head network structure diagram;

[0038] Figure 6 A comparison chart of front detection results; (a) is a schematic diagram of front labels; (b) is the detection result of the original LSENet model; (c) is the detection result of the SEMNet-DC model. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0040] The present invention provides an intelligent ocean front detection method integrating dense skip connections and seasonal enhancement, such as Figure 1 As shown, the following steps are included:

[0041] Step 1: Obtain daily satellite remote sensing SST data of the study area and preprocess it to obtain the SST gradient map.

[0042] First, the daily SST data of the research area is obtained. The data is m×n grid data, where m represents the number of latitudinal grids and n represents the number of longitudinal grids. The gradient of the daily SST data is calculated using the Sobel operator. The calculation formula is as follows:

[0043] ;

[0044] ;

[0045] in, Respectively represent the horizontal and vertical gradient values ​​of the current coordinate point. Represents the 3×3 neighborhood matrix of the current coordinate point, represents the convolution operation, Indicates the gradient value of the current coordinate point.

[0046] After the gradient value is obtained through gradient calculation, the “jet” color mapping function in Matlab is used to draw the SST gradient map.

[0047] Step 2: Based on the SST gradient map, combined with previous ocean front research and expert experience, Labelme is used to annotate the front labels to obtain the front label data.

[0048] The labeling work divides each pixel in the image into two categories: one is the front pixel, and the other is the background pixel. If the current pixel is an ocean front pixel, it is labeled as 1; if it is a background pixel, it is labeled as 0.

[0049] Step 3: Construct a seasonal enhanced multi-scale recognition network model (SEMNet-DC) with dense skip connections.

[0050] The structure of the complete SEMNet-DC model is as follows Figure 2 As shown, it includes 5 encoders, 4 decoders, a head network and 6 intermediate states.

[0051] 1. Encoder

[0052] like Figure 3 As shown in Figure 1, the encoder consists of two branches. The bottom branch uses an improved multi-scale residual convolution block to replace the stack of two 3×3 convolution layers in the original network, in order to improve the model's ability to capture fronts of different scales. The specific structure of the improved multi-scale residual convolution block is shown in Figure 1. Figure 4 As shown in the figure. The multi-scale residual convolution block consists of three 3×3 convolutional layers, and the outputs of the second and third convolutional layers are approximately equal to the 5×5 and 7×7 convolution operations. The improved multi-scale residual convolution block incorporates the jump connection operation, which is to jump-connect the input data to the outputs of the second and third 3×3 convolutional layers, and jump-connect the output of the first 3×3 convolutional layer to the output of the third 3×3 convolutional layer. Finally, the outputs of the three 3×3 convolutional layers are concatenated and finally added element-by-element with the output of the input after the 1×1 convolutional layer. The ReLU nonlinear activation function is used after each convolution layer to realize feature extraction at different scales.

[0053] The upper branch of the encoder is a channel supervision unit, which consists of 1 global average pooling layer and 2 fully connected layers. At the same time, seasonal features are added to the channel supervision unit. The specific method is to encode the season for each feature map. The original LSENet's seasonal encoding is a simple 0-1 encoding. When learning the seasonal features of a certain month, the model only focuses on the features of that month and does not pay attention to the features of other months. However, the change of the ocean front is a process that changes slowly with the change of seasons. That is, the seasonal features of the front in the current month are not only closely related to the front features of the month, but also closely related to the front features of the adjacent months. The seasonal encoding is improved according to this feature of the front. The improved seasonal encoding assigns the seasonal code of the current month to 0.8, and the seasonal codes of the previous month and the next month of the current month are both assigned to 0.1. This approach is equivalent to using a weighted operation on the seasonal coding, which enables the model to perceive the relationship and trend between months, reduces the discontinuity of seasonal information, helps the network better capture the gradual change of seasons and the periodic characteristics of seasons, and improves the performance of the model and the accuracy of front detection. The output of the encoder is obtained by adding the feature maps obtained by the bottom branch and the upper branch and then sampling.

[0054] 2. Decoder

[0055] The decoder has the same structure as the encoder, except for the last sampling layer. The last layer of the encoder is a downsampling layer, while the last layer of the decoder is an upsampling layer. The feature map is used to extract features and downsample in the encoder, and the number of channels is adjusted. Specifically, the input image is input to the encoder with an input size of The feature map of is firstly passed through the first encoder for feature extraction and dimensionality reduction, and the output size is The feature map is then passed to the second encoder for feature extraction and dimensionality reduction, and the output size is The feature map is passed through 5 encoders in sequence, and the final output size is The feature map is passed to the decoder for feature decoding and upsampling. In contrast to the feature map transformation process of the encoder, the feature map is restored to , the feature map output by the last decoder will be passed to the head network.

[0056] 3. Head network

[0057] The head network is mainly responsible for the classification of pixels and generating detection results. The specific structure is as follows Figure 5 As shown in Figure 2, the head network consists of two branches. The bottom branch consists of a 1×1 convolutional layer and a Softmax layer. The upper branch is a position attention module with an input size of The feature map is first average pooled to obtain a size of The feature map of the image is obtained by upsampling, where r represents the downsampling coefficient. The feature map is then passed to the convolution layer, and the result of the convolution is added to the feature map of the position encoding. The added feature map is passed to a series of convolution layers for feature extraction. The last convolution layer is a convolution layer with a Sigmoid activation function to improve the nonlinear expression ability of the network. After that, upsampling is performed to restore the feature map size to the original image size. Finally, the results of the upper branch and the bottom branch are multiplied first, and the multiplication result is added to the result of the bottom branch to obtain the final detection result.

[0058] As the number of network layers increases, the encoder part gradually performs downsampling operations, and the size of the feature map gradually decreases, which will cause the loss of some small-scale information, making it impossible for the deep network to learn this information. Therefore, it is necessary to first upsample the feature map obtained by the encoder part, and then perform jump connections to complete feature fusion after obtaining the feature map of the same size as the decoder part. Jumping the feature information obtained by the shallow network to the deep network can make the deep network contain richer small-scale edge information. The jump connection of the original LSENet model jumps the output of the first encoder to the fourth decoder, the output of the second encoder to the third decoder, the output of the third encoder to the second decoder, and the output of the fourth encoder to the first decoder.

[0059] On the basis of retaining the jump connection of the original LSENet model, the improved algorithm of the present invention adds 6 intermediate states to realize dense jump connection operation. The 6 intermediate states in the improved algorithm are 6 feature maps obtained by upsampling the outputs of the second, third and fourth encoders, respectively. These 6 feature maps are respectively composed of 1 upsampling operation of the second encoder, 2 upsampling operations of the third encoder in sequence, and 3 upsampling operations of the fourth encoder in sequence. Specifically, an upsampling layer is added after the second encoder, and the output of the second encoder is upsampled once to obtain a feature map of the same size as the output of the first encoder, and the size of the feature map is also the same as the output size of the fourth decoder. Two upsampling layers are added after the third encoder, and the output of the third encoder is upsampled by the first layer to obtain a feature map of the same size as the output of the second encoder, and the size of the feature map is also the same as the output size of the third decoder; after another upsampling, a feature map of the same size as the output of the first encoder is obtained, and the size of the feature map is also the same as the output size of the fourth decoder. Similarly, three upsampling layers are added after the fourth encoder. The output of the fourth encoder is upsampled in the first layer to obtain a feature map of the same size as the output of the third encoder, which is also the same size as the output of the second decoder; the second layer is upsampled to obtain a feature map of the same size as the output of the second encoder, which is also the same size as the output of the third decoder; the third layer is upsampled to obtain a feature map of the same size as the output of the first encoder, which is also the same size as the output of the fourth decoder. The six feature maps obtained in this way are the same size as the feature maps obtained by encoders or decoders at different positions. After that, the encoder part and the feature maps obtained by the six intermediate states are jump-connected to the decoder part. This process realizes the dense jump connection operation. The sampling multiple is 2 times in the upsampling process, and the interpolation method selects the bilinear interpolation method.

[0060] Step 4: Divide the SST gradient map and front label data into a training set and a test set, use the training set to train the model, and use the test set to evaluate the model accuracy.

[0061] (1) Model training

[0062] The essence of the model's ocean front detection task is to classify pixels. Therefore, the cross entropy loss function commonly used in classification tasks is used during training. The loss function calculation formula is as follows:

[0063] ;

[0064] in, The first elements, Represents model predictions Belong to The probability of the class is Class time, =1, otherwise =0, Indicates the number of sample labels.

[0065] The parameters in the model training process are as follows: the batch size is set to 4, the training rounds are set to 80, the cross entropy loss function is used, the Adam optimizer is used, and the initial learning rate of the training is 0.001. During the training process, the learning rate decay strategy is adopted. When the loss error does not decrease within three consecutive training rounds, the learning rate at this time is reduced to half of the original.

[0066] (2) Model accuracy evaluation

[0067] The present invention aims to further improve the front detection accuracy of the model based on the LSENet model. Therefore, the front detection accuracy of SEMNet-DC is compared with that of the original LSENet model. The results are shown in Table 1. At the same time, ablation experiments are carried out to verify the effectiveness of various improvements.

[0068] Table 1 Comparison of front detection accuracy of models under different conditions

[0069]

[0070] Table 1 shows the front detection accuracy of the model under different conditions. The detection accuracy of a single category is measured by IoU (intersection over union, IoU), and the overall accuracy is reflected by the mean intersection over union (mIoU). From the data in the table, it can be seen that for the original LSENet model (serial number 1), the front IoU is 68.51% and the average IoU is 83.20%. Based on the original LSENet model, the front IoU and the average IoU are improved after replacing the two 3×3 convolutional layers in the bottom branch of the encoder with the improved multi-scale residual convolution block. Similarly, based on the original LSENet model, when only dense skip connections are performed or only the seasonal encoding is improved, the front IoU and the average IoU are also improved compared with the original LSENet model. It shows that whether replacing two 3×3 convolutional layers with improved multi-scale residual convolution blocks on the basis of the original LSENet model, adding dense skip connection operations on the basis of the original LSENet model, or improving the seasonal coding of the LSENet model, the front detection accuracy of the model can be improved. In the table, serial numbers 5 to 7 represent two improvements based on the original LSENet model. Compared with the original LSENet model (serial number 1) and an improvement based on the original LSENet model (serial numbers 2 to 4), the front detection accuracy is further improved, which further shows that the three improvements are effective and can improve the front detection accuracy of the model. In the table, serial number 8 represents the improved algorithm SEMNet-DC of the present invention. It is not difficult to see from the data in the table that the intersection-over-union ratio of the SEMNet-DC model is higher than all the above detection results, whether it is the background or the front, and the front intersection-over-union ratio is increased from 68.51% of the original LSENet model to 76.09%, and the average intersection-over-union ratio is increased from 83.20% to 87.21%. From these data, it can be seen that the improved model SEMNet-DC based on the LSENet model is effective in improving the front detection accuracy in the study area.

[0071] Step 5: After preprocessing the SST data of the sea area to be tested, an SST gradient map is obtained, and the SST gradient map is input into the qualified model for intelligent detection of the ocean front.

[0072] The embodiment of the present invention also compares the front detection result map in the study area on December 1, 2021. The front detection result on that day is as follows: Figure 6As shown in the figure. From the comparison of front detection results, it can be seen that the SEMNet-DC model has better front detection effect. The original LSENet model has missed detection of many fronts, while the SEMNet-DC model has a lower missed detection rate. In addition, the original LSENet model has many false detection problems, detecting many areas where there are no fronts as front areas, and the original LSENet model has low sensitivity to small-scale front lines and cannot effectively detect some small-scale front lines. In comparison, the performance of the SEMNet-DC model in these aspects has been improved.

[0073] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one 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 present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent ocean front detection method integrating dense skip connections and seasonal enhancement, characterized in that: The steps include: Step 1: Obtain daily satellite remote sensing SST data of the study area and preprocess it to obtain an SST gradient map; Step 2: Based on the SST gradient map, combined with previous ocean front research and expert experience, Labelme is used to annotate the front labels to obtain the front label data; Step 3, construct a seasonal enhanced multi-scale recognition network model integrating dense skip connections; The model includes 5 encoders, 4 decoders, a head network and 6 intermediate states; The encoder consists of two branches, the bottom branch is an improved multi-scale residual convolution block, that is, the jump connection operation is integrated in the multi-scale residual convolution block, and the ReLU nonlinear activation function is used after each layer of convolution to realize feature extraction of different scales; the multi-scale residual convolution block consists of three 3×3 convolution layers, and the jump connection operation integrated in the improved multi-scale residual convolution block is as follows: the input data is jump-connected to the outputs of the second and third 3×3 convolution layers respectively, and the output of the first 3×3 convolution layer is jump-connected to the output of the third 3×3 convolution layer Output; Finally, the outputs of the three 3×3 convolutional layers are concatenated and added element by element with the output of the input after the 1×1 convolutional layer; The upper branch of the encoder is a channel supervision unit, which consists of a global average pooling layer and two fully connected layers. At the same time, an improved seasonal coding is added to the channel supervision unit, that is, the seasonal coding of the current month is assigned to 0.8, and the seasonal coding of the month before and after the current month is assigned to 0.1; The feature maps obtained by the bottom branch and the upper branch are added and then passed through the sampling layer to obtain the output of the encoder; The decoder has the same structure as the encoder, except for the last sampling layer. The last layer of the encoder is a downsampling layer, while the last layer of the decoder is an upsampling layer. The feature map is used to extract features and perform downsampling in the encoder, and the number of channels is adjusted. The adjusted feature map is passed to the decoder for feature decoding and upsampling. The feature map output by the last decoder is passed to the head network. The head network is responsible for classifying pixels and generating detection results. The head network consists of two branches. The bottom branch consists of a 1×1 convolutional layer and a Softmax layer; the upper branch is the position attention module. The results of the upper branch and the bottom branch are first multiplied, and the multiplication result is added to the result of the bottom branch to obtain the final detection result; The 6 intermediate states are the 6 feature maps obtained by upsampling the outputs of the second, third, and fourth encoders, which are respectively composed of 1 upsampling operation of the second encoder, 2 upsampling operations of the third encoder, and 3 upsampling operations of the fourth encoder; The 6 feature maps obtained are of the same size as the feature maps obtained by the encoder or decoder at different positions. Then, the encoder part and the feature maps obtained by the 6 intermediate states are jump-connected to the decoder part. This process realizes the dense jump connection operation; the 6 intermediate states are specifically: adding an upsampling layer after the second encoder, and the output of the second encoder is upsampled once to obtain a feature map of the same size as the output of the first encoder, and the size of the feature map is also the same as the output size of the fourth decoder; adding two upsampling layers after the third encoder, and the output of the third encoder is upsampled by the first layer to obtain a feature map of the same size as the output of the second encoder, and the size of the feature map is also the same as the output size of the third decoder. The output size is the same; after another upsampling, a feature map of the same size as the output of the first encoder is obtained, and the size of the feature map is also the same as the output size of the fourth decoder; similarly, three upsampling layers are added after the fourth encoder, and the output of the fourth encoder is upsampled in the first layer to obtain a feature map of the same size as the output of the third encoder, and the size of the feature map is also the same as the output size of the second decoder; after upsampling in the second layer, a feature map of the same size as the output of the second encoder is obtained, and the size of the feature map is also the same as the output of the third decoder; after upsampling in the third layer, a feature map of the same size as the output of the first encoder is obtained, and the size of the feature map is also the same as the output of the fourth decoder; Step 4: divide the SST gradient map and front label data into a training set and a test set, use the training set to train the model, and use the test set to evaluate the model accuracy; Step 5: After preprocessing the SST data of the sea area to be tested, an SST gradient map is obtained, and the SST gradient map is input into the qualified model for intelligent detection of the ocean front.

2. The method for intelligent ocean front detection integrating dense skip connections and seasonal enhancement according to claim 1, characterized in that: In the encoder part, the input size is The feature map of is firstly passed through the first encoder for feature extraction and dimensionality reduction, and the output size is The feature map is then passed to the second encoder for feature extraction and dimensionality reduction, and the output size is The feature map is passed through 5 encoders in sequence, and the final output size is The feature map is passed to the decoder for feature decoding and upsampling. The feature map is reversed from the feature map of the encoder. After the feature map passes through four decoders, the feature map is restored to .

3. The method for intelligent ocean front detection integrating dense skip connections and seasonal enhancement according to claim 1, characterized in that: In the upper branch of the head network, the input size is The feature map first passes through the average pooling layer to obtain a size of The feature map is then passed to the convolution layer, and the result of the convolution is added to the feature map of the position encoding; the added feature map is passed to a series of convolution layers for feature extraction. The last convolution layer is a convolution layer with a Sigmoid activation function to improve the nonlinear expression ability of the network. After that, the feature map size is restored to the original image size through the upsampling layer.

4. The method for intelligent ocean front detection integrating dense skip connections and seasonal enhancement according to claim 1, characterized in that: In model training, the cross entropy loss function is used, and the calculation formula is as follows: ； in, The first elements, Represents model predictions Belong to The probability of the class is Class time, =1, otherwise =0, Indicates the number of sample labels.

5. The method for intelligent ocean front detection integrating dense skip connections and seasonal enhancement according to claim 1, characterized in that: During model training, a learning rate decay strategy is adopted. When the loss error does not decrease within three consecutive training rounds, the learning rate at this time is reduced to half of the original value.

Images