Seawater buoyant raft cultivation information extraction method based on interpretable polarization deep learning network

Abstract

The invention provides a seawater buoyant raft cultivation information extraction method based on an interpretable polarization deep learning network, and belongs to the technical field of ocean remote sensing and artificial intelligence crossing. Wherein the interpretable polarization deep learning network comprises three key components: a polarization scattering type activation mapping module, an interpretable segmentation consistency learning module and a refined type activation and edge guiding network. The polarization scattering class activation mapping module generates a high-quality class activation mapping graph by combining polarization scattering characteristics and gradient weights of the fully-polarized SAR image; the interpretable segmentation consistency learning module uses the generated class activation mapping graph to guide the model to pay attention to the characteristics of the breeding area in the training process; the refined class activation and edge guidance network adopts a double-branch architecture, semantic segmentation and edge supervision are performed respectively, and an accurate segmentation result is generated. According to the method, background misjudgment can be remarkably reduced, and a clear and detailed segmentation result is ensured; and the method has relatively high applicability.

Description

Technical Field

[0001] The present invention belongs to the cross-technical field of marine remote sensing and artificial intelligence, and relates to a method for extracting seawater floating raft aquaculture information by interpreting a polarization deep learning network. Background Art

[0002] Marine aquaculture is of great significance to global seafood demand, coastal economic support, and the relief of pressure on wild fish stocks. However, disorderly expansion may lead to environmental problems such as harmful algal blooms, eutrophication, and biodiversity degradation. Traditional field survey methods are time-consuming and inefficient, while remote sensing technology can provide large-scale observations and timely data. Polarimetric Synthetic Aperture Radar (PolSAR) is a powerful tool with all-weather operation capabilities, high spatial resolution, and strong penetration characteristics, which can effectively carry out marine target monitoring. In addition, PolSAR can accurately distinguish water bodies from aquaculture structures such as floating rafts and cages due to its inherent multi-polarization channels. The polarization scattering characteristics of these aquaculture structures provide valuable prior information for subsequent classification tasks. Therefore, it is crucial to develop effective PolSAR image processing techniques to achieve accurate extraction of marine aquaculture.

[0003] In the marine aquaculture monitoring task, due to the complex and changeable environment of the aquaculture area, including the influence of multiple factors such as waves, wind and waves, and ocean currents, as well as the interaction between aquaculture facilities and the surrounding environment, the scattering characteristics of radar signals are complex and diverse, which poses a challenge to the extraction performance and reliability of the model. The Chinese invention patent (publication number CN110910494B) proposes a method for generating ISAR images by simulating the distribution of the sea surface under different wind speeds, combining Rhino software for three-dimensional modeling, and using electromagnetic simulation software FEKO to solve the backscattering field. This method is based on the Freeman and Yamaguchi decomposition theory, analyzes the polarization scattering characteristics of the target, distinguishes different scattering mechanisms of the sea surface and underwater parts, such as surface scattering, dihedral scattering, etc., and evaluates the results through equivalent visual numbers and Fresnel reflection coefficients, thereby effectively analyzing the backscattering characteristics of seawater raft aquaculture under random sea conditions. However, this method only analyzes the backscattering characteristics of seawater raft aquaculture, and does not actually apply the backscattering characteristics of seawater raft aquaculture to the deep learning model to constrain the behavior of the network model. The data-driven deep learning model can mine the characteristics of marine aquaculture targets from the original SAR data. The Chinese invention patent (publication number CN119398540A) is based on the original UNet neural network model, modifies its structure to increase complexity, forms a nested UNet neural network, and adds a channel attention mechanism (SE module) to it, constructs the U2-Net_Se2 model, and finally achieves excellent raft segmentation results. However, due to the "black box" characteristics of the model, it cannot feedback important decision areas and features to people. Summary of the invention

[0004] In view of the above problems existing in the prior art, the present invention proposes a method for extracting seawater raft aquaculture information with an interpretable polarimetric deep learning network. The present invention can solve the problems of poor interpretability and reliability of the intelligent interpretation method of seawater aquaculture, improve the interpretability of the extraction process while ensuring the extraction accuracy of raft aquaculture, and clearly point out which features play a key role in the final segmentation result. The core innovation of the present invention lies in: proposing an interpretable polarimetric deep learning network for semantic segmentation of marine aquaculture areas. The interpretable polarimetric deep learning network can achieve high-precision segmentation and interpretability analysis of aquaculture areas by integrating polarimetric scattering prior knowledge and deep learning technology. Specifically, the interpretable polarimetric deep learning network includes three key components: a polarimetric scattering class activation mapping module, an interpretable segmentation consistency learning module, and a refined class activation and edge guidance network. The polarimetric scattering class activation mapping module generates a high-quality class activation map by combining the polarimetric scattering features and gradient weights of the fully polarimetric SAR image, which can effectively highlight the features related to the aquaculture area while suppressing the noise in the complex background. The interpretable segmentation consistency learning module uses the generated class activation map to guide the model to pay more attention to the characteristics of the aquaculture area during the training process, improve the interpretability of the model, and enhance the model's attention to the aquaculture area. The refined class activation and edge guidance network adopts a dual-branch architecture to perform semantic segmentation and edge supervision respectively, which can simultaneously focus on the semantic features and edge details of the aquaculture area and generate accurate segmentation results. The synergistic effect of these three enables the interpretable polarization deep learning network to achieve high-precision segmentation of aquaculture areas in complex marine environments, while significantly improving the interpretability of the model, providing an efficient, transparent and reliable solution for marine aquaculture monitoring tasks.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A method for extracting seawater raft aquaculture information that can interpret polarization deep learning networks, the method comprising the following steps:

[0007] The first step is to collect PolSAR image data from remote sensing satellites, preprocess the PolSAR image data, and prepare a data set, as follows:

[0008] Step 1.1, first collect PolSAR image data of remote sensing satellites through satellite resource websites. These image data usually contain echo signals in different polarization states (HH, HV, VH, VV polarization). Then preprocess the PolSAR image data.

[0009] Furthermore, the preprocessing refers to using PolSARpro V.6.0.3 software to perform radiation correction, geometric correction and denoising preprocessing on the image data to ensure the quality and accuracy of the image data.

[0010] Step 1.2, using the covariance matrix of the PolSAR image data obtained in step 1.1 to represent the polarization information, polarization decomposition is performed on the polarization information, and different types of scattering components are extracted from the PolSAR image data, including dihedral scattering, volume scattering, and surface scattering.

[0011] Furthermore, the polarization decomposition method includes the Yamaguchi four-decomposition method or the Freeman-Durden decomposition method.

[0012] Step 1.3, the dihedral scattering, volume scattering and surface scattering obtained in step 1.2 are combined to obtain a PolSAR pseudo-color image, and a marine aquaculture dataset and labels are obtained based on the PolSAR pseudo-color image.

[0013] Furthermore, labelme software was used to manually annotate the PolSAR pseudo-color images to obtain the marine aquaculture dataset and labels.

[0014] The second step is to design a refined class activation and edge guidance network to solve the problem that the scattering characteristics of the sea surface affect the identification of aquaculture rafts, and when multiple rafts are close to each other or move due to wave action, the raft targets are adhered and the jagged boundaries in the image are caused. The refined class activation and edge guidance network adopts an encoder-dual decoder architecture (see step 2.1), including a convolutional backbone network, a semantic decoder, and an edge decoder; the semantic decoder uses a refined class activation attention module to enhance the semantic features of the raft area in the PolSAR pseudo-color image (see step 2.2); the edge decoder uses an edge-driven feature enhancement module to focus on accurately capturing the boundary information of the raft area in the PolSAR pseudo-color image, thereby further optimizing the boundary consistency of the segmentation result (see step 2.3). These two parts work together to generate the final segmentation output through feature fusion and convolutional segmentation head. The details are as follows:

[0015] Step 2.1, based on the encoder-dual decoder architecture, a refined class activation and edge guidance network is designed, including a convolutional backbone network, a semantic decoder, and an edge decoder, to achieve accurate semantic segmentation through feature fusion. The convolutional backbone network uses ResNet50 to extract feature maps of PolSAR image data and gradually capture multi-level spatial information. The refined class activation and edge guidance network uses a learnable parameter To weight the feature maps of the convolutional backbone network, semantic decoder, and edge decoder, the formula is expressed as:

[0016] (1);

[0017] Among them, EF represents the features extracted by the convolutional backbone network, DF represents the features obtained by the semantic decoder or edge decoder step by step decoding, and F represents the features after weighted fusion. Represents an upsampling operation.

[0018] After the PolSAR pseudo-color image is input into the refined class activation and edge guidance network, the convolutional backbone network ResNet50 first extracts its features, and then sends the extracted features to the semantic decoder and edge decoder respectively. After the features pass through the two decoders, semantic features and edge features will be generated respectively, and the two features will be fused by pixel-level summation. The fused features are further processed in the final convolution segmentation head to output the final segmentation result. In order to promote the collaborative optimization of the network, the refined class activation and edge guidance network adopts a multi-task loss function that includes semantic segmentation loss, edge loss and fusion loss.

[0019] Step 2.2, design a refined class activation attention module, the goal is to dynamically adjust the importance of each channel in the feature map by combining local and global context information, thereby improving the ability of the refined class activation and edge guidance network to focus on the raft area and suppress background interference. The core function of the refined class activation attention module is to enhance the feature expression of the target category through the attention mechanism and recalibrate the channel using global context information. Specifically:

[0020] Step 2.2.1, given feature map , where C, H, and W represent the total number of feature channels, height, and width, respectively. First, global context information Z is extracted through global average pooling. Then, channel weights are dynamically generated through a two-layer fully connected network and an activation function, as shown in formula (2):

[0021] (2);

[0022] in, , are the weights of the fully connected layer; is the ReLU activation function; is the Sigmoid activation function; For Channel The weight is used to adjust the importance of the channel; is the calibrated feature map; is the feature map before calibration. This process can dynamically adjust the importance of each channel in the feature map, strengthen category-related features, and suppress irrelevant information.

[0023] Step 2.2.2: In order to improve the category activation ability of the feature map in the spatial dimension, the feature map is divided into non-overlapping patches, represented by , the local category feature in each patch is calculated by formula (3):

[0024] (3);

[0025] in, Represents patch Local category features of represents the category weight; Represents spatial softmax, which is used to highlight the salient areas within the patch; Represents patch Local feature aggregation can enhance feature expression ability in a small range, which is particularly suitable for semantic segmentation in complex backgrounds.

[0026] In step 2.2.3, the local features of all patches are integrated to generate global category features, as shown in formula (4):

[0027] (4);

[0028] in, Represents global category features; For patch The weights are used to dynamically adjust the contribution of different patches to the global features.

[0029] In step 2.2.4, a set of attention mechanisms are used to optimize the global features obtained in step 2.2.3 to make them more suitable for category distinction in segmentation tasks. The final output of the refined class activation attention module is obtained by calculating the similarity of each pixel to the category, as shown in formula (5):

[0030] (5);

[0031] And combined with the global features to generate the final enhanced feature map, as shown in formula (6):

[0032] (6);

[0033] in, Represents a pixel block similarity of categories; is the linear transformation matrix used for feature alignment, A learnable linear transformation matrix representing the input features, A learnable linear transformation matrix representing local category features, A learnable linear transformation matrix representing global category features; Represents category softmax; Represents a patch of input features; is a patch in the enhanced input feature.

[0034] Step 2.3, design an edge-driven feature enhancement module to deal with the segmentation problem of complex or weak boundaries, especially when processing densely distributed marine aquaculture rafts in PolSAR pseudo-color images. Since edge information plays a key role in semantic segmentation, this edge-driven feature enhancement module effectively captures and strengthens boundary features by combining channel attention, spatial attention and Sobel edge detection to generate more accurate and clear segmentation results, as follows:

[0035] Step 2.3.1: The channel attention mechanism dynamically adjusts the importance of each channel in the input feature map. Specifically, the input feature map is defined as , generate channel descriptors through global average pooling , and then generate channel weights through two layers of full connection operations , and the weights are used for element-wise weighted channel enhancement. The formula is:

[0036] (7);

[0037] in, and are all convolution weights; and They are ReLU and Sigmoid activation functions respectively; Indicates the feature map Sub-feature graph; Indicates The weight of each sub-feature map; Indicates The result after weighting of sub-feature maps; In this way, the channel attention mechanism can highlight the feature expression of significant channels while suppressing the information of irrelevant channels, and finally combine all weighted sub-feature maps to obtain channel attention enhanced features. .

[0038] Step 2.3.2, the spatial attention mechanism enhances the information of key spatial regions in the feature map by generating a spatial weight map. Specifically, the input feature map first generates two spatial descriptors through the maximum pooling and average pooling operations on the channel dimension. and The two spatial descriptors are then concatenated in the channel dimension and a spatial weight map is generated through a convolution operation. Finally, the spatial weight map Multiply the input feature map element by element to generate a spatial enhanced feature map .

[0039] Step 2.3.3, in order to further extract explicit boundary information, the edge-driven feature enhancement module combines the classic Sobel edge detection method by calculating the input feature map The horizontal and vertical gradients of are used to generate preliminary boundary features. First, the Sobel convolution kernels in the horizontal and vertical directions are used. and Calculate the gradient as shown in formula (8):

[0040] (8);

[0041] in, represents the horizontal gradient, Represents the vertical gradient;

[0042] Then, calculate the gradient magnitude , as shown in formula (9):

[0043] (9) ;

[0044] In order to further refine the edge features, we first Normalize and activate with ReLU function. Then, the feature is further processed by depthwise separable convolution to obtain enhanced edge features, which are recorded as Sobel boundary enhancement features. .

[0045] Step 2.3.4: The edge-driven feature enhancement module enhances the channel attention feature , spatial attention enhancement feature 0 and Sobel edge enhancement features The final output feature map is generated by element-by-element addition, as shown in formula (10):

[0046] (10);

[0047] The third step is to design a polarization scattering class activation mapping module. The design of the polarization scattering class activation mapping module aims to improve the interpretability and target area positioning ability of the neural network in the PolSAR pseudo-color image segmentation task. The polarization scattering class activation mapping module combines the polarization scattering prior knowledge with the gradient weight information of deep learning to not only generate a class activation map with clear physical meaning. It also directly participates in the training process, guiding the convolution backbone network to extract more representative features and significantly improve the segmentation interpretability. Aquaculture areas in PolSAR pseudo-color images usually exhibit significant dihedral scattering characteristics, which are characteristic signals formed by multiple reflections of radar waves between rafts or buoys and the water surface. The polarization scattering class activation mapping module uses this physical property to optimize the generation of activation maps through dihedral scattering prior information, focus on the target area, and suppress background interference in complex marine environments. The details are as follows:

[0048] Step 3.1, the polarization scattering class activation mapping module first generates an initial segmentation mask for the input PolSAR pseudo-color image through the refined class activation and edge-guided network of step 2.1 , mark potential target areas (such as raft aquaculture areas). At the same time, use the dihedral scattering components extracted from Step 1.2 after Freeman decomposition , which is combined with the segmentation mask to generate the modified correction mask , as shown in formula (11):

[0049] (11);

[0050] in, Represents an element-by-element multiplication operation. The correction mask The dihedral scattering characteristics in the region of interest are highlighted to strengthen the focus on the target area. Resolution and dihedral scattering component different, so it is necessary to Perform downsampling to obtain a mask consistent with the feature map resolution , as shown in formula (12):

[0051] (12);

[0052] in, The function adjusts the mask to the spatial size of the feature map .

[0053] Step 3.2, the polarization scattering class activation mapping module calculates the feature map With correction mask The cosine similarity of is used to measure the relevance of the feature map to the target area, as shown in formula (13):

[0054] (13);

[0055] in, and The feature maps are and correction mask The vector representation of is the Euclidean norm. Cosine similarity It reflects the degree of alignment between the feature map and the target area.

[0056] Step 3.3: To further emphasize the importance of the target area, the polarization scatter class activation map module calculates the category by global average pooling. The prediction score For feature maps The gradient weight of is shown in formula (14):

[0057] (14);

[0058] Through the gradient weight It captures the importance of the feature map to the target category. In order to balance the contribution of similarity and gradient weight, the two are normalized separately, as shown in formula (15):

[0059] (15);

[0060] in: Represents a vector of correlation score weights for all channels, Represents a vector of gradient weights for all channels, It means finding the minimum term in the weight vector. Represents the maximum term in the weight vector. The normalized correlation score weight and gradient weight are combined to generate the final channel weight, as shown in formula (16):

[0061] (16);

[0062] in, represents the channel weight; It is a hyperparameter used to adjust the overall influence of the weight. Channel weight It is used to weight the feature map to generate the final class activation map, as shown in formula (17):

[0063] (17);

[0064] Among them, the ReLU activation function ensures that only features that contribute positively to the target area are retained.

[0065] The class activation maps generated by the polarimetric scattering class activation mapping module have significant physical significance and interpretability, and can clearly highlight the floating raft structures in the aquaculture area while effectively suppressing background interference.

[0066] The fourth step is to design an interpretable segmentation consistency learning module. In order to improve the interpretability of the refined class activation and edge-guided network and enhance the consistency of segmentation feature extraction, the present invention designs an interpretable segmentation consistency learning module. The design goal of this interpretable segmentation consistency learning module is to use the high-quality class activation map generated by the polarization scattering class activation mapping module to guide model learning, so that it can focus more on the target area (such as the aquaculture raft area) and effectively suppress the interference of the background area, thereby improving the interpretability of the refined class activation and edge-guided network itself. The details are as follows:

[0067] Step 4.1, generate a class activation map CAM through the polarized scattering class activation mapping module to highlight the focus area of ​​the convolutional backbone network of the refined class activation and edge-guided network on the target category. The generation formula of the class activation map is as follows:

[0068] (18);

[0069] in, Indicates category In spatial position The class activation value of It is feature maps of channels; For Category The corresponding The weight of each channel.

[0070] In step 4.2, the feature maps extracted from the convolutional backbone network of the refined class activation and edge-guided network are summed across the channel dimension to generate a comprehensive activation map SAM to comprehensively reflect the feature strength of the target area, as shown in formula (19):

[0071] (19);

[0072] in, Indicates that all channel feature maps are in spatial position In order to ensure that the spatial distribution of the combined activation map (SAM) and the class activation map (CAM) is consistent, the explainable segmentation consistency learning module constrains the two through a consistency loss function. The consistency loss is calculated based on the mean squared error (MSE), as shown in formula (20):

[0073] (20);

[0074] in, is the normalized comprehensive activation map, is the normalized class activation map, and Represent the height and width of the feature map respectively.

[0075] The introduction of the interpretable segmentation consistency learning module effectively improves the interpretability of the refined class activation and edge-guided networks, making the feature extraction of the target area more accurate while suppressing the interference of the background area. In addition, in complex aquaculture scenarios (such as dense raft areas or complex water surface reflections), the interpretable segmentation consistency learning module can significantly enhance the model's ability to focus on the target area, improve segmentation accuracy and boundary consistency.

[0076] Step 5: Design multi-optimization objective loss function The entire network is optimized to simultaneously improve segmentation accuracy, edge detail depiction capabilities, and model decision interpretability. With fusion loss Optimize pixel-level classification and use the cross entropy function to accurately distinguish between targets and backgrounds. Edge detection loss combining weighted binary cross entropy and intersection-over-union loss , strengthen the spatial consistency modeling of complex boundaries. At the same time, introduce an interpretable consistency loss , constraining the alignment of the convolution backbone network features with the class activation map, promoting the features to focus on the target subject and suppressing noise interference, as shown in formula (21):

[0077] (twenty one);

[0078] in, The weight used to control the loss of interpretability. Through end-to-end optimization, the accuracy of floating raft aquaculture extraction is guaranteed while improving the interpretability of the extraction process.

[0079] The beneficial effects of the present invention are:

[0080] (1) Existing PolSAR image segmentation methods have problems such as obvious background interference, difficult boundary identification, and insufficient interpretability when dealing with marine aquaculture monitoring tasks. These methods are difficult to effectively utilize the physical characteristics of polarization scattering, resulting in inaccurate positioning of the target area, blurred boundaries, and low reliability of the segmentation results, especially in complex environments with densely distributed aquaculture rafts. In response to the above problems, the present invention proposes an innovative interpretable polarization deep learning network, which significantly improves the segmentation accuracy and interpretability of the segmentation task process. First, through the design of the polarization scattering class activation mapping module, the polarization scattering prior knowledge is integrated into the segmentation network, so that the features extracted by the convolutional backbone network can accurately focus on the aquaculture raft area, while effectively suppressing background interference, fundamentally solving the problem of inaccurate positioning of the target area. Secondly, the interpretable segmentation consistency learning module guides network learning through high-quality class activation maps, which is more in line with the characteristics of the target area and significantly reduces the occurrence of background misjudgment. In addition, in order to solve the problem of difficult boundary identification, a refined class activation and edge guidance network is proposed to accurately extract target boundary information and ensure clear and detailed segmentation results.

[0081] (2) This invention not only effectively solves the limitations of existing technologies, but also achieves comprehensive improvements in segmentation accuracy, boundary refinement capability, and model interpretability in complex marine environments. Its technical framework has strong applicability and not only provides an efficient and reliable solution for precise monitoring of aquaculture areas, but can also be widely used in target recognition and segmentation tasks in other remote sensing scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 This is an overall schematic diagram of the polarization-explainable segmentation network. Figure 1 In the figure, the first part is the structure diagram of the polarization scattering class activation mapping module, the second part is the structure diagram of the interpretable segmentation consistency learning module, and the third part is the structure diagram of the refined class activation and edge guidance network;

[0083] Figure 2 Schematic diagram of the refined category activation and edge guidance network;

[0084] Figure 3 Schematic diagram of activating attention modules for refined classes;

[0085] Figure 4 It is a schematic diagram of the edge driven feature enhancement module;

[0086] Figure 5PolSAR marine aquaculture dataset, where (a) is a pseudo-color image of large-scale PolSAR data after a series of preprocessing processes and Freeman decomposition, (b) is a manually drawn true value image, (c) is a cropped small block image, (d) is the dihedral scattering component of the area, (e) is the volume scattering component of the area, and (f) is the surface scattering component of the area;

[0087] Figure 6 Schematic diagram of the segmentation results of PolSAR image data according to the present invention, wherein (a) is a PolSAR pseudo-color image, (b) is a true value image corresponding to the PolSAR pseudo-color image, (c) is a floating raft aquaculture segmentation result image in the PolSAR pseudo-color image, and (d) is a floating raft aquaculture edge extraction result image in the PolSAR pseudo-color image;

[0088] Figure 7 Visual results of interpretable class activation maps generated by the polarimetric scattering class activation mapping module, where (a) is a PolSAR pseudo-color image and (b) is a class activation map of raft aquaculture in the PolSAR pseudo-color image;

[0089] Figure 8 Comparison of the visual results of the interpretable class activation maps produced by the polarization scattering class activation mapping module before and after model training guided by the interpretable segmentation consistency learning module, where (a) is the PolSAR pseudo-color image, (b) is the interpretable class activation map without using this module, and (c) is the interpretable class activation map with using this module. DETAILED DESCRIPTION

[0090] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0091] like Figure 1 As shown, the present invention provides a method for extracting seawater raft aquaculture information using an interpretable polarization deep learning network. This embodiment is compiled in python3.8.10, pytorch1.11.1 and cuda11.3 under the windows11 system and runs using the GPU of RTX2080ti. PolSARproV.6.0.3 software is used for preprocessing PolSAR images. The following steps are included:

[0092] The first step is to collect PolSAR image data from remote sensing satellites, preprocess the PolSAR image data, and prepare a data set, as follows:

[0093] Step 1.1, radiometric calibration: convert digital image data into actual physical quantities such as reflectivity to eliminate the radiometric bias of PolSAR image data.

[0094] Step 1.2, image enhancement: Lee filtering with a window size of 7×7 is used to alleviate the influence of coherent speckle noise in PolSAR image data.

[0095] Step 1.3, Polarization decomposition: Calculate the coherence matrix of the PolSAR image data to capture the correlation between polarization states and channels. Then perform Freeman decomposition on the coherence matrix to separate the radar backscatter into three main scattering mechanisms: dihedral scattering, surface scattering, and volume scattering. These three scattering components are combined to generate a PolSAR pseudo-color image, such as Figure 5 As shown in (a) in .

[0096] Step 1.4, make a data set: After the above preprocessing, the PolSAR pseudo-color images were cropped into small blocks of 512×512 pixels, and a total of 190 small blocks were obtained. These small blocks of PolSAR pseudo-color images were manually annotated using labelme software. Then these image blocks and their true value images were pre-enhanced by vertical and horizontal flipping, so that the total number of image blocks reached 570. All image blocks were divided into training set, validation set and test set, with a ratio of 6:1:3, respectively. The data samples are as follows: Figure 5 As shown in (c) in .

[0097] The second step is to design a refined class activation and edge guidance network to solve the problem that the scattering characteristics of the sea surface affect the identification of aquaculture rafts, and when multiple rafts are close to each other or move due to wave action, the raft targets are adhered and the jagged boundaries in the image are caused. The refined class activation and edge guidance network adopts an encoder-dual decoder architecture (see step 2.1), including a convolutional backbone network, a semantic decoder, and an edge decoder; the semantic decoder uses a refined class activation attention module to enhance the semantic features of the raft area in the PolSAR pseudo-color image (see step 2.2); the edge decoder uses an edge-driven feature enhancement module to focus on accurately capturing the boundary information of the raft area in the PolSAR pseudo-color image, thereby further optimizing the boundary consistency of the segmentation result (see step 2.3). These two parts work together to generate the final segmentation output through feature fusion and convolutional segmentation head. The details are as follows:

[0098] Step 2.1, based on the encoder-dual decoder architecture, a refined class activation and edge guidance network is designed, including a convolutional backbone network, a semantic decoder, and an edge decoder, to achieve accurate semantic segmentation through feature fusion. The convolutional backbone network uses ResNet50 to extract feature maps of PolSAR image data and gradually capture multi-level spatial information. The refined class activation and edge guidance network uses a learnable parameter To weight the feature maps of the convolutional backbone network, semantic decoder, and edge decoder, the formula is expressed as:

[0099] (1);

[0100] Among them, EF represents the features extracted by the convolutional backbone network, DF represents the features obtained by the semantic decoder or edge decoder step by step decoding, represents the weighted fusion feature, Represents an upsampling operation.

[0101] After the PolSAR pseudo-color image is input into the refined class activation and edge guidance network, the convolutional backbone network ResNet50 first extracts its features, and then sends the extracted features to the semantic decoder and edge decoder respectively. After the features pass through the two decoders, semantic features and edge features will be generated respectively, and the two features will be fused by element-wise summation. The fused features are further processed in the final convolution segmentation head to output the final segmentation result. In order to promote the collaborative optimization of the network, the refined class activation and edge guidance network adopts a multi-task loss function that includes semantic segmentation loss, edge loss and fusion loss. Figure 2 The figure shows the structure diagram of the refined class activation and edge-guided network. Figure 6 The figure shows the segmentation results and edge extraction results of the network, where (a) is a PolSAR pseudo-color image, (b) is the true value image corresponding to the PolSAR pseudo-color image, (c) is the floating raft aquaculture segmentation result image in the PolSAR pseudo-color image, and (d) is the floating raft aquaculture edge extraction result image in the PolSAR pseudo-color image;

[0102] Step 2.2, design a refined class activation attention module, the goal is to dynamically adjust the importance of each channel in the feature map by combining local and global context information, thereby improving the ability of the refined class activation and edge guidance network to focus on the raft area and suppress background interference. The core function of the refined class activation attention module is to enhance the feature expression of the target category through the attention mechanism and recalibrate the channel using global context information. Specifically:

[0103] Step 2.2.1, given feature map ,in , , Represent the total number of feature channels, height and width respectively. First, global context information is extracted through global average pooling. Next, the channel weights are dynamically generated through two layers of fully connected networks and activation functions, as shown in formula (2):

[0104] (2);

[0105] in, , are the weights of the fully connected layer; is the ReLU activation function; is the Sigmoid activation function; For Channel The weight is used to adjust the importance of the channel; is the calibrated feature map; is the feature map before calibration. This process can dynamically adjust the importance of each channel in the feature map, strengthen category-related features, and suppress irrelevant information.

[0106] Step 2.2.2: In order to improve the category activation ability of the feature map in the spatial dimension, the feature map is divided into non-overlapping patches, represented by , the local category feature in each patch is calculated by formula (3):

[0107] (3);

[0108] in, Represents patch Local category features of represents the category weight; Represents spatial softmax, which is used to highlight the salient areas within the patch; Represents patch Local feature aggregation can enhance feature expression ability in a small range, which is particularly suitable for semantic segmentation in complex backgrounds.

[0109] In step 2.2.3, the local features of all patches are integrated to generate global category features, as shown in formula (4):

[0110] (4);

[0111] in, is the global category feature; For patch The weights are used to dynamically adjust the contribution of different patches to the global features.

[0112] In step 2.2.4, a set of attention mechanisms are used to optimize the global features obtained in step 2.2.3 to make them more suitable for category distinction in segmentation tasks. The final output of the refined class activation attention module is obtained by calculating the similarity of each pixel to the category, as shown in formula (5):

[0113] (5);

[0114] And combined with the global features to generate the final enhanced feature map, as shown in formula (6):

[0115] (6);

[0116] in, Represents a pixel block similarity of categories; is the linear transformation matrix used for feature alignment, A learnable linear transformation matrix representing the input features, A learnable linear transformation matrix representing local category features, A learnable linear transformation matrix representing global category features; Represents category softmax; Represents a patch of input features; is a patch in the enhanced input feature. Figure 3 Shown is a schematic diagram of the structure of the refined class activation attention module.

[0117] Step 2.3, design an edge-driven feature enhancement module to deal with the segmentation problem of complex or weak boundaries, especially when processing densely distributed marine aquaculture rafts in PolSAR pseudo-color images. Since edge information plays a key role in semantic segmentation, this edge-driven feature enhancement module effectively captures and strengthens boundary features by combining channel attention, spatial attention and Sobel edge detection to generate more accurate and clear segmentation results, as follows:

[0118] Step 2.3.1: The channel attention mechanism dynamically adjusts the importance of each channel in the input feature map. Specifically, the input feature map is defined as , generate channel descriptors through global average pooling , and then generate channel weights through two layers of full connection operations , and the weights are used for element-wise weighted channel enhancement. The formula is:

[0119] (7);

[0120] in, and are all convolution weights; and They are ReLU and Sigmoid activation functions respectively; Indicates the feature map Sub-feature graph; Indicates The weight of each sub-feature map; Indicates The result after weighting of sub-feature maps; In this way, the channel attention mechanism can highlight the feature expression of significant channels while suppressing the information of irrelevant channels, and finally combine all weighted sub-feature maps to obtain channel attention enhanced features. .

[0121] Step 2.3.2, the spatial attention mechanism enhances the information of key spatial regions in the feature map by generating a spatial weight map. Specifically, the input feature map first generates two spatial descriptors through the maximum pooling and average pooling operations on the channel dimension. and The two spatial descriptors are then concatenated in the channel dimension and a spatial weight map is generated through a convolution operation. Finally, the spatial weight map Multiply the input feature map element by element to generate a spatial enhanced feature map .

[0122] Step 2.3.3, in order to further extract explicit boundary information, the edge-driven feature enhancement module combines the classic Sobel edge detection method by calculating the input feature map The horizontal and vertical gradients of are used to generate preliminary boundary features. First, the Sobel convolution kernels in the horizontal and vertical directions are used. and Calculate the gradient as shown in formula (8):

[0123] (8);

[0124] in, represents the horizontal gradient, Represents the vertical gradient;

[0125] Then, calculate the gradient magnitude , as shown in formula (9):

[0126] (9) ;

[0127] To further refine the edge features, Normalize and activate with ReLU function. Then, the feature is further processed by depthwise separable convolution to obtain enhanced edge features, which are recorded as Sobel boundary enhancement features. .

[0128] Step 2.3.4: The edge-driven feature enhancement module enhances the channel attention feature , spatial attention enhancement feature 0 and Sobel edge enhancement features The final output feature map is generated by element-by-element addition, as shown in formula (10):

[0129] (10);

[0130] like Figure 4 Shown is a schematic diagram of the edge-driven feature enhancement module structure.

[0131] The third step is to design a polarization scattering class activation mapping module. The design of the polarization scattering class activation mapping module aims to improve the interpretability and target area positioning ability of the neural network in the PolSAR pseudo-color image segmentation task. The polarization scattering class activation mapping module combines the polarization scattering prior knowledge with the gradient weight information of deep learning to not only generate a class activation map with clear physical meaning. It also directly participates in the training process, guiding the convolution backbone network to extract more representative features and significantly improve the segmentation interpretability. Marine aquaculture areas in PolSAR pseudo-color images usually exhibit significant dihedral scattering characteristics, which are characteristic signals formed by multiple reflections of radar waves between rafts or buoys and the water surface. The polarization scattering class activation mapping module uses this physical property to optimize the generation of activation maps through dihedral scattering prior information, focus on the target area, and suppress background interference in complex marine environments. The details are as follows:

[0132] Step 3.1, the polarization scattering class activation mapping module first generates an initial segmentation mask for the input PolSAR pseudo-color image through the refined class activation and edge-guided network of step 2.1 , mark potential target areas (such as raft aquaculture areas). At the same time, use the dihedral scattering components extracted from Step 1.2 after Freeman decomposition , which is combined with the segmentation mask to generate the modified correction mask , as shown in formula (11):

[0133] (11);

[0134] in, Represents an element-by-element multiplication operation. The correction mask The dihedral scattering characteristics in the region of interest are highlighted to strengthen the focus on the target area. Resolution and dihedral scattering component different, so it is necessary to Perform downsampling to obtain a mask consistent with the feature map resolution , as shown in formula (12):

[0135] (12);

[0136] in, The function adjusts the mask to the spatial size of the feature map .

[0137] Step 3.2, the polarization scattering class activation mapping module calculates the feature map With correction mask The cosine similarity of is used to measure the relevance of the feature map to the target area, as shown in formula (13):

[0138] (13);

[0139] in, and The feature maps are and correction mask The vector representation of is the Euclidean norm. Cosine similarity It reflects the degree of alignment between the feature map and the target area.

[0140] Step 3.3: To further emphasize the importance of the target area, the polarization scatter class activation map module calculates the category by global average pooling. The prediction score For feature maps The gradient weight of is shown in formula (14):

[0141] (14);

[0142] Through the gradient weight It captures the importance of the feature map to the target category. In order to balance the contribution of similarity and gradient weight, the two are normalized separately, as shown in formula (15):

[0143] (15);

[0144] in: Represents a vector of correlation score weights for all channels, Represents a vector of gradient weights for all channels, It means finding the minimum term in the weight vector. Represents the maximum term in the weight vector. The normalized correlation score weight and gradient weight are combined to generate the final channel weight, as shown in formula (16):

[0145] (16);

[0146] in, represents the channel weight; It is a hyperparameter used to adjust the overall influence of the weight. Channel weight It is used to weight the feature map to generate the final class activation map, as shown in formula (17):

[0147] (17);

[0148] Among them, the ReLU activation function ensures that only features that contribute positively to the target area are retained.

[0149] The class activation maps generated by the polarimetric scattering class activation mapping module have significant physical significance and interpretability, and can clearly highlight the floating raft structures in the aquaculture area while effectively suppressing background interference. Figure 1 The first part of the diagram shows the design of the polarization scattering class activation mapping module. Figure 7 The effective activation of this class activation method is demonstrated, which can focus on the raft area well, where (a) is a PolSAR pseudo-color image, and (b) is a class activation map of raft aquaculture in the PolSAR pseudo-color image.

[0150] The fourth step is to design an interpretable segmentation consistency learning module. In order to improve the interpretability of the refined class activation and edge-guided network and enhance the consistency of segmentation feature extraction, the present invention designs an interpretable segmentation consistency learning module. The design goal of this interpretable segmentation consistency learning module is to use the high-quality class activation map generated by the polarization scattering class activation mapping module to guide model learning, so that it can focus more on the target area (such as the aquaculture raft area) and effectively suppress the interference of the background area, thereby improving the interpretability of the refined class activation and edge-guided network itself. The details are as follows:

[0151] Step 4.1, generate a class activation map CAM through the polarized scattering class activation mapping module to highlight the focus area of ​​the convolutional backbone network of the refined class activation and edge-guided network on the target category. The generation formula of the class activation map is as follows:

[0152] (18);

[0153] in, Indicates category In spatial position The class activation value of It is feature maps of channels; For Category The corresponding The weight of each channel.

[0154] In step 4.2, the feature maps extracted from the convolutional backbone network of the refined class activation and edge-guided network are summed across the channel dimension to generate a comprehensive activation map SAM to comprehensively reflect the feature strength of the target area, as shown in formula (19):

[0155] (19);

[0156] in, Indicates that all channel feature maps are in spatial position In order to ensure that the spatial distribution of the combined activation map (SAM) and the class activation map (CAM) is consistent, the explainable segmentation consistency learning module constrains the two through a consistency loss function. The consistency loss is calculated based on the mean squared error (MSE), as shown in formula (20):

[0157] (20);

[0158] in, is the normalized comprehensive activation map, is the normalized class activation map, and Represent the height and width of the feature map respectively.

[0159] The introduction of the interpretable segmentation consistency learning module effectively improves the interpretability of the refined class activation and edge-guided networks, making the feature extraction of the target area more accurate while suppressing the interference of the background area. In addition, in complex aquaculture scenarios (such as dense raft areas or complex water surface reflections), the interpretable segmentation consistency learning module can significantly enhance the model's ability to focus on the target area and improve segmentation accuracy and boundary consistency. Figure 1 The third part of the paper shows the design of the explainable segmentation consistency learning module, as shown in Figure 8 The figure shows a visual comparison of the interpretable class activation maps produced by the polarization scattering class activation mapping module before and after the model training is guided by the interpretable segmentation consistency learning module, where (a) is a PolSAR pseudo-color image, (b) is an interpretable class activation map without using this module, and (c) is an interpretable class activation map with using this module.

[0160] Step 5: Design multi-optimization objective loss function The entire network is optimized to simultaneously improve segmentation accuracy, edge detail depiction capabilities, and model decision interpretability. With fusion loss Optimize pixel-level classification and use the cross entropy function to accurately distinguish between targets and backgrounds. Edge detection loss combining weighted binary cross entropy and intersection-over-union loss , strengthen the spatial consistency modeling of complex boundaries. At the same time, introduce an interpretable consistency loss , constraining the alignment of the convolution backbone network features with the class activation map, promoting the features to focus on the target subject and suppressing noise interference, as shown in formula (21):

[0161] (twenty one);

[0162] in, The weight used to control the loss of interpretability. Through end-to-end optimization, the accuracy of floating raft aquaculture extraction is guaranteed while improving the interpretability of the extraction process.

[0163] The above-described embodiments merely express the implementation methods of the present invention, but they should not be understood as limiting the scope of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A method for extracting seawater raft aquaculture information using an interpretable polarization deep learning network, characterized in that: The interpretable polarimetric deep learning network in the method for extracting information from seawater floating raft aquaculture includes three key components: a polarimetric scattering class activation mapping module, an interpretable segmentation consistency learning module, and a refined class activation and edge guidance network; the polarimetric scattering class activation mapping module generates a class activation map by combining the polarimetric scattering features and gradient weights of the fully polarized SAR image; the interpretable segmentation consistency learning module uses the generated class activation map to guide the model to focus on the features of the aquaculture area during training; the refined class activation and edge guidance network adopts a dual-branch architecture to perform semantic segmentation and edge supervision respectively, and can focus on the semantic features and edge details of the aquaculture area at the same time to generate accurate segmentation results; the method for extracting information from seawater floating raft aquaculture includes the following steps: The first step is to collect PolSAR image data from remote sensing satellites and preprocess them; The second step is to design a refined class activation and edge guidance network; the refined class activation and edge guidance network adopts an encoder-dual decoder architecture, including a convolutional backbone network, a semantic decoder and an edge decoder; the semantic decoder uses a refined class activation attention module to enhance the semantic features of the raft area in the PolSAR pseudo-color image; the edge decoder uses an edge-driven feature enhancement module to accurately capture the boundary information of the raft area in the PolSAR pseudo-color image; they work together to generate the final segmentation output through feature fusion and convolutional segmentation head; The third step is to design a polarization scattering class activation mapping module. The polarization scattering class activation mapping module combines the polarization scattering prior knowledge with the gradient weight information of deep learning to generate a class activation map with clear physical meaning. At the same time, it directly participates in the training process and guides the convolution backbone network to extract representative features. The polarization scattering class activation mapping module optimizes the generation of activation maps through dihedral angle scattering prior information. The fourth step is to design an interpretable segmentation consistency learning module. The design goal of the interpretable segmentation consistency learning module is to use the class activation map generated by the polarization scattering class activation mapping module to guide model learning. Step 5: Design multi-optimization objective loss function Optimize the entire network.

2. The method for extracting seawater raft aquaculture information using an interpretable polarization deep learning network according to claim 1, characterized in that: The first step is as follows: Step 1.1, collect PolSAR image data of remote sensing satellites through satellite resource websites, and preprocess the PolSAR image data including radiation correction, geometric correction, and denoising preprocessing; Step 1.2, using the covariance matrix of the PolSAR image data obtained in step 1.1 to represent the polarization information; performing polarization decomposition on the polarization information, and extracting different types of scattering components from the PolSAR image data, including dihedral scattering, volume scattering, and surface scattering; Step 1.3, the dihedral scattering, volume scattering and surface scattering obtained in step 1.2 are combined to obtain a PolSAR pseudo-color image, and a marine aquaculture dataset and labels are obtained based on the PolSAR pseudo-color image.

3. The method for extracting seawater raft aquaculture information using an interpretable polarization deep learning network according to claim 2, characterized in that: In the step 1.2, the polarization decomposition method includes the Yamaguchi four-decomposition method or the Freeman-Durden decomposition method.

4. The method for extracting seawater raft aquaculture information using an interpretable polarization deep learning network according to claim 1, characterized in that: The second step is as follows: Step 2.1, based on the encoder-dual decoder architecture, a refined class activation and edge guidance network is designed, including a convolutional backbone network, a semantic decoder, and an edge decoder, to achieve accurate semantic segmentation through feature fusion; the convolutional backbone network uses ResNet50 to extract feature maps of PolSAR image data and gradually capture multi-level spatial information; the refined class activation and edge guidance network uses a learnable parameter To weight the feature maps of the convolution backbone network, semantic decoder, and edge decoder, the formula is expressed as: (1)； Among them, EF represents the features extracted by the convolutional backbone network, DF represents the features obtained by the semantic decoder or edge decoder step by step decoding, represents the weighted fusion feature, represents an upsampling operation; After the PolSAR pseudo-color image is input into the refined class activation and edge-guided network, the convolutional backbone network ResNet50 first extracts its features, and then sends the extracted features to the semantic decoder and edge decoder respectively. After the features pass through the two decoders, semantic features and edge features are generated respectively, and the two features are fused by element-level summation; the fused features are processed in the final convolution segmentation head to output the final segmentation result; Step 2.2, design a refined class activation attention module, enhance the feature expression of the target category through the attention mechanism, and recalibrate the channel using global context information; specifically: Step 2.2.1, given feature map ,in , , Represent the total number of feature channels, height and width respectively. First, global context information is extracted through global average pooling. ; Then, the channel weights are dynamically generated through two layers of fully connected networks and activation functions, as shown in formula (2): (2)； in, , are the weights of the fully connected layer; is the ReLU activation function; is the Sigmoid activation function; For Channel The weight is used to adjust the importance of the channel; is the calibrated feature map; is the feature map before calibration; Step 2.2.2, split the feature map into non-overlapping patches, represented by , the local category feature in each patch is calculated by formula (3): (3)； in, Represents patch Local category features of represents the category weight; Represents spatial softmax, which is used to highlight the salient areas within the patch; Represents patch Category activation maps within ; In step 2.2.3, the local features of all patches are integrated to generate global category features, as shown in formula (4): (4)； in, is the global category feature; For patch The weight of In step 2.2.4, a set of attention mechanisms are used to optimize the global features obtained in step 2.2.3; the final output of the refined class activation attention module is obtained by calculating the similarity of each pixel to the category, as shown in formula (5): (5)； And combined with the global features to generate the final enhanced feature map, as shown in formula (6): (6)； in, Represents a pixel block similarity of categories; is the linear transformation matrix used for feature alignment, A learnable linear transformation matrix representing the input features, A learnable linear transformation matrix representing local category features, A learnable linear transformation matrix representing global category features; Represents category softmax; Represents a patch of input features; is a patch in the enhanced input feature; Step 2.3, design the edge-driven feature enhancement module to generate more accurate and clear segmentation results, as follows: Step 2.3.1: The channel attention mechanism dynamically adjusts the importance of each channel in the input feature map; specifically, the input feature map is defined as , generate channel descriptors through global average pooling , and then generate channel weights through two layers of full connection operations , and the weights are used for element-wise weighted channel enhancement; the formula is: (7)； in, and are all convolution weights; and They are ReLU and Sigmoid activation functions respectively; Indicates the feature map Sub-feature graph; Indicates The weight of each sub-feature map; Indicates The result after weighting of sub-feature maps; Represents the total number of feature channels; all weighted sub-feature maps are combined to obtain channel attention enhanced features ; Step 2.3.2, the spatial attention mechanism enhances the information of key spatial regions in the feature map by generating a spatial weight map; specifically, the input feature map first generates two spatial descriptors through the maximum pooling and average pooling operations on the channel dimension and ; These two spatial descriptors are then concatenated in the channel dimension and a spatial weight map is generated through a convolution operation ; Finally, the spatial weight map Multiply the input feature map element by element to generate a spatial enhanced feature map ; Step 2.3.3: The edge-driven feature enhancement module combines the Sobel edge detection method to calculate the input feature map The horizontal and vertical gradients of are used to generate preliminary boundary features; first, Sobel convolution kernels are used in the horizontal and vertical directions. and Calculate the gradient as shown in formula (8): (8)； in, represents the horizontal gradient, Represents the vertical gradient; Then, calculate the gradient magnitude , as shown in formula (9): (9) ； To further refine the edge features, we first Normalize and activate with ReLU function; then, use depth separable convolution to process the features to obtain enhanced edge features, which are recorded as Sobel boundary enhancement features. ; Step 2.3.4: The edge-driven feature enhancement module enhances the channel attention feature , spatial attention enhancement feature 0 and Sobel edge enhancement features The final output feature map is generated by element-by-element addition, as shown in formula (10): (10)。 5. The method for extracting seawater raft aquaculture information using an interpretable polarization deep learning network according to claim 4, characterized in that: The third step is as follows: Step 3.1, the polarization scattering class activation mapping module first generates an initial segmentation mask for the input PolSAR pseudo-color image through the refined class activation and edge-guided network of step 2.1 , mark the potential target area; at the same time, use the dihedral scattering component extracted by Freeman decomposition in step 1.2 , which is combined with the segmentation mask to generate the modified correction mask , as shown in formula (11): (11)； in, Represents an element-by-element multiplication operation; the correction mask Focus on the dihedral scattering characteristics in the region of interest to strengthen the focus on the target area; due to the feature map extracted by the convolutional backbone network in step 2.1 Resolution and dihedral scattering component different, so it is necessary to Perform downsampling to obtain a mask consistent with the feature map resolution , as shown in formula (12): (12)； in, The function adjusts the mask to the spatial size of the feature map ; Step 3.2, the polarization scattering class activation mapping module calculates the feature map With correction mask The cosine similarity of is used to measure the relevance of the feature map to the target area, as shown in formula (13): (13)； in, and The feature maps are and correction mask The vector representation of is the Euclidean norm; cosine similarity It reflects the degree of alignment between the feature map and the target area; Step 3.3: To further emphasize the importance of the target area, the polarization scatter class activation map module calculates the category by global average pooling. The prediction score For feature maps The gradient weight of is shown in formula (14): (14)； Through the gradient weight It captures the importance of the feature map to the target category; in order to balance the contribution of similarity and gradient weight, the two are normalized separately, as shown in formula (15): (15)； in: Represents a vector of correlation score weights for all channels, Represents a vector of gradient weights for all channels, It means finding the minimum term in the weight vector. Represents the maximum term in the weight vector; the normalized correlation score weight and gradient weight are combined to generate the final channel weight, as shown in formula (16): (16)； in, represents the channel weight; It is a hyperparameter used to adjust the overall influence of the weight; channel weight It is used to weight the feature map and generate the final class activation map, as shown in formula (17): (17)； Among them, the ReLU activation function ensures that only features that contribute positively to the target area are retained.

6. The method for extracting seawater raft aquaculture information using an interpretable polarization deep learning network according to claim 5, characterized in that: The fourth step is as follows: Step 4.1, generate a class activation map CAM through the polarization scattering class activation mapping module, which is used to highlight the focus area of ​​the convolutional backbone network of the refined class activation and edge guidance network on the target category; the generation formula of the class activation map is as follows: (18)； in, Indicates category In spatial position The class activation value of It is feature maps of channels; For Category The corresponding The weight of each channel; In step 4.2, the feature maps extracted from the convolutional backbone network of the refined class activation and edge-guided network are summed across the channel dimension to generate a comprehensive activation map SAM to comprehensively reflect the feature strength of the target area, as shown in formula (19): (19)； in, Indicates that all channel feature maps are in spatial position The sum of In order to ensure that the comprehensive activation map and the class activation map are consistent in spatial distribution, the interpretable segmentation consistency learning module constrains the two through a consistency loss function; the consistency loss is calculated based on the mean square error MSE, as shown in formula (20): (20)； in, is the normalized comprehensive activation map, is the normalized class activation map, and Represent the height and width of the feature map respectively.

7. The method for extracting seawater raft aquaculture information using an interpretable polarization deep learning network according to claim 6, characterized in that: The fifth step is as follows: Through semantic segmentation loss With fusion loss Optimize pixel-level classification and use the cross entropy function to accurately distinguish between targets and backgrounds; combine the weighted binary cross entropy and the intersection-over-union loss to detect edge loss , strengthen the spatial consistency modeling of complex boundaries; at the same time introduce an explainable consistency loss , constraining the alignment of the convolution backbone network features with the class activation map, promoting the features to focus on the target subject and suppressing noise interference, as shown in formula (21): (21)； in, Weight used to control the interpretability loss.

Images