Raft type laver culture area end-to-end extraction method based on frequency dynamic filtering

Abstract

The invention relates to the technical field of remote sensing image processing, in particular to a raft laver culture area end-to-end extraction method based on frequency dynamic filtering, and the method comprises the steps: obtaining an original remote sensing image of a raft laver culture area, carrying out the preprocessing, constructing a mask data set for high-level semantic supervision, and carrying out the extraction of the original remote sensing image; constructing a hierarchical supervision neural network model, and performing model training on the constructed hierarchical supervision neural network model, including constructing a composite loss function composed of mask loss, gravitational field loss, vertex classification loss and vertex offset loss; and performing optimization processing on a prediction result output by the hierarchical supervision neural network model, and generating a breeding area contour vector result based on the optimized binary mask and vertex coordinates. The method solves the problem that the extraction precision is suddenly reduced in a scene of relatively strong spatial spectrum characteristic heterogeneity of the breeding area in a traditional method, and improves the extraction accuracy of the breeding area. And reliable technical support is provided for large-area and high-precision monitoring of an offshore raft type laver culture area.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image processing, in particular to a raft type of seaweed cultivation area end-to-end extraction method based on frequency dynamic filtering. BACKGROUND

[0002] Raft type of seaweed has become an important characteristic aquaculture industry in coastal areas due to its short growth cycle and high economic return rate, and plays a key role in ensuring the supply of aquatic products and promoting regional economic development. However, with the uncontrolled expansion of raft type of seaweed cultivation, a large amount of waste such as discarded insertion rods, floating balls, and net curtains generated during the cultivation process is discharged into the sea, causing serious pollution to the nearshore water environment and posing a direct threat to maritime traffic safety. Therefore, achieving accurate monitoring and dynamic management of raft type of seaweed cultivation areas is not only a practical requirement for regulating the development of the aquaculture industry, but also an important measure to protect the nearshore ecological environment and ensure maritime traffic safety.

[0003] Remote sensing technology has become a core technical means for monitoring large-scale raft type of seaweed cultivation areas due to its wide coverage, short observation period, and convenient data acquisition. The accurate extraction of the cultivation area contour is a key step in remote sensing monitoring, and the extraction results can provide basic data support for subsequent applications such as cultivation area statistics, distribution pattern analysis, and illegal cultivation identification. It is of great significance to the scientific planning and ecological protection of nearshore aquaculture. Due to problems such as cultivation harvesting, sediment shading, and large differences in spectral characteristics, extracting the cultivation area contour from remote sensing images is a major challenge. Traditional cultivation area extraction methods first generate a cultivation area binary mask through a semantic segmentation network, and then use post-processing and vectorization methods to generate a vectorized cultivation area polygon. These methods often struggle with irregular contours, rounded corners, and redundant points, and require a large amount of post-processing to produce regular polygon building contours, further reducing extraction accuracy and making it difficult to meet the automation needs of large-scale and high-frequency cultivation area dynamic monitoring. Therefore, there is a need for a raft type of seaweed cultivation area end-to-end extraction method based on frequency dynamic filtering. SUMMARY

[0004] To solve the problems of low extraction accuracy and low adaptation to automation needs of traditional remote sensing extraction methods, the present application provides a raft type of seaweed cultivation area end-to-end extraction method based on frequency dynamic filtering.

[0005] The present application provides a raft type of seaweed cultivation area end-to-end extraction method based on frequency dynamic filtering, which adopts the following technical solution: A raft type of seaweed cultivation area end-to-end extraction method based on frequency dynamic filtering, comprising: S1, obtain the original remote sensing image of raft type seaweed cultivation area and perform pretreatment, construct a mask data set for high-level semantic supervision and an MS COCO format label for middle-bottom layer geometric supervision; S2, construct a hierarchical supervision neural network model, the hierarchical supervision neural network model takes HRNet48 network as a backbone network, and includes a hierarchical feature extraction branch, a frequency dynamic selection module and a multi-task prediction head; S3, model training is performed on the constructed hierarchical supervision neural network model, including constructing a composite loss function composed of mask loss, gravity field loss, vertex classification loss and vertex offset loss; S4, input the pretreated remote sensing image to be extracted of raft type seaweed cultivation area into the trained hierarchical supervision neural network model, and output mask prediction result, vertex prediction result and gravity field prediction result; S5, the prediction results output by the hierarchical supervision neural network model are optimized to obtain optimized cultivation area binary mask and sub-pixel level precision vertex coordinates; S6, based on the optimized binary mask and vertex coordinates, through connected region analysis, vertex matching and polygon simplification, the cultivation area contour vector result is generated.

[0006] Further, step S1 specifically includes: S11, obtain high-resolution multispectral original remote sensing image covering the target raft type seaweed cultivation area, and sequentially perform radiation calibration, atmospheric correction and orthorectification pretreatment on the original remote sensing image to obtain pretreated remote sensing image with unified geometric and radiation accuracy; S12, interactive labeling is performed on the pretreated remote sensing image to generate a rasterized label map, and different values are assigned to the cultivation area pixels and background pixels in the label map to realize pixel-level differentiation, forming a mask true value basis; S13, the pretreated remote sensing image and the rasterized label map are synchronously cut by a sliding window to obtain corresponding image blocks and mask blocks, and a mask data set for high-level semantic supervision is constructed; S14, obtain the cultivation area contour boundary pixel coordinate sequence of each mask block in the mask data set, organize and convert the boundary pixel coordinate sequence and the cultivation area instance information according to the MS COCO data set standard format, and construct an MS COCO format label for middle-bottom layer geometric supervision.

[0007] Furthermore, the construction of the hierarchical supervised neural network model includes performing multi-scale feature extraction on the preprocessed dataset through the backbone network HRNet48, outputting an initial feature map and a basic feature map F, constructing a hierarchical feature extraction branch, performing three 3×3 convolution operations on the initial feature map and the basic feature map F, and sequentially performing batch normalization and ReLU activation function operations after each convolution layer to obtain the corresponding gravitational field features. Vertex features and mask features The hierarchical feature extraction branch consists of a mask feature extraction branch, a vertex feature extraction branch, and a gravitational field feature extraction branch.

[0008] Furthermore, the construction of the hierarchical supervised neural network model also includes constructing a frequency dynamic selection module based on gravitational field features. Vertex features and mask features As input, and A converged guided feature map is formed by concatenating features along the channel dimension. After global adaptive average pooling to compress the spatial dimension, it is input into a multilayer perceptron. A dynamic weight vector of dimension N is generated by the Softmax activation function. The basic feature map F is then converted into frequency domain features by a two-dimensional fast Fourier transform. After element-wise complex multiplication with N learnable global filter bases, the features are weighted and fused by the dynamic weight vector. Finally, the enhanced feature map is output by a two-dimensional inverse fast Fourier transform. The expression of the enhanced feature map is as follows: , in, For two-dimensional fast Fourier transform, This is a two-dimensional inverse fast Fourier transform. For element-wise complex multiplication, For the k-th learnable global filter basis, This is the k-th element of the dynamic weight vector.

[0009] Furthermore, the training of the constructed hierarchical supervised neural network model specifically includes: S311. Construct the mask loss function: using the true values ​​of the mask in the mask dataset. As a basis for supervision, the mask prediction results output by the binary cross-entropy loss calculation model are used. With truth value The difference, the mask loss function is denoted as ; S312. Constructing the gravitational field loss function: Generating the true value of the gravitational field based on the line segment information in the MS COCO format tags. ,use Gravitational field prediction results output by the loss calculation model With truth value The deviation, the gravitational field loss function is denoted as ; S313. Input the training dataset into the hierarchical supervised neural network model, using mask loss. With gravitational field loss The weighted sum is used as the sub-objective loss, and the model parameters are updated through backpropagation.

[0010] Furthermore, the step of training the constructed hierarchical supervised neural network model further includes: S321. Constructing a vertex classification loss function: Generating vertex heatmap ground truth based on vertex position information in MS COCO format labels. The vertex heatmap prediction result H, calculated using the binary cross-entropy loss model, is compared with the true value. The classification error, the vertex classification loss function is denoted as ; S322. Construct the vertex offset loss function: based on the vertex truth value The positional deviation of the corresponding pixel generates the offset true value. ,use The offset prediction result O output by the loss calculation model and the true value The difference in the vertex region is denoted by the vertex offset loss function as follows: ; S323. Constructing a composite loss function and full model training: Incorporating vertex classification loss... Vertex offset loss With mask loss function and gravitational field loss function Weighted fusion is performed to obtain the composite loss function. ,by Given the objective function, an adaptive moment estimator optimizer is used to iteratively train the model on the training dataset until the model converges. The expression for the composite loss function is: , in, , , and The loss factor is used to predict mask loss, gravitational field loss, heatmap loss, and vertex offset loss.

[0011] Furthermore, the optimization processing of the prediction results output by the hierarchical supervised neural network model includes performing a Softmax activation operation on the masked prediction results output by the hierarchical supervised neural network model, extracting the probability feature map corresponding to the aquaculture area, and using a bilinear interpolation algorithm to adjust the resolution of the probability feature map from the resolution of the feature map output by the backbone network to the resolution of the original remote sensing image, thereby obtaining a probability map of the aquaculture area that matches the size of the original image. Set a probability threshold T and generate a probability map of the aquaculture area after resolution restoration. Perform binarization segmentation to generate an optimized binary mask for the aquaculture area: , in, For pixel coordinates, For a preset probability threshold and , This is a probability map of the aquaculture area after resolution restoration.

[0012] Further, obtaining the optimized binary mask of the aquaculture area and sub-pixel-level precision vertex coordinates includes performing a Softmax activation operation on the vertex prediction results output by the hierarchical supervised neural network model, extracting concave and convex probability maps, obtaining the vertex offset field output by the backbone network, performing a Sigmoid activation operation and subtracting 0.5 to obtain vertex offset predictions with values ​​ranging from [-0.5, 0.5), employing a local nonmaximum suppression algorithm to retain pixels with classification scores higher than a preset threshold in the concave and convex probability maps as candidate vertices, calculating a scaling factor based on the resolution ratio of the backbone network output feature map to the original image, and combining the coordinates of the candidate vertices on the feature map and the corresponding offset predictions to calculate the sub-pixel-level vertex coordinates in the original image coordinate system. The sub-pixel-level vertex coordinates satisfy the following: , in, Let be the coordinates of the candidate vertex on the feature map. , This is the resolution scaling factor. , This is a subpixel-level offset.

[0013] Furthermore, the generated aquaculture area contour vector result includes an optimized binary mask of the aquaculture area. Perform connected component analysis and use the 8-neighborhood connectivity criterion to extract all independent connected regions. Each connected region corresponds to a breeding area instance. And calculate the value of each instance. The bounded pixel set, based on the coordinates of all sub-pixel level vertices, forms the global vertex set. Set a distance threshold for each aquaculture zone instance. Given a set of boundary pixels, calculate the Euclidean distance between each boundary pixel and all vertices in the global vertex set, and filter out a subset of matching vertices based on a distance threshold. And in clockwise order of boundary pixels Sorting the vertices to form instances The initial vertex sequence.

[0014] Furthermore, the generation of the aquaculture area contour vector result also includes processing data based on each aquaculture area instance. The initial vertex sequence is calculated by finding the index difference between adjacent vertices. An index difference threshold is set to optimize the initial vertex sequence. Based on the optimized initial vertex sequence, the vector angle between adjacent vertices is calculated for every three consecutive vertices. An angle threshold is set to filter the vertices, resulting in a final simplified vertex sequence. This simplified sequence is then applied to each aquaculture area instance. The final simplified vertex sequence is connected in clockwise order to form a closed aquaculture area outline polygon. The vertex coordinates, instance numbers and corresponding confidence scores of the polygon are stored in the geographic information vector data format to generate the aquaculture area outline vector result.

[0015] In summary, the present invention has the following beneficial technical effects: 1. This invention constructs a two-layer training data system consisting of a high-level semantically supervised mask dataset and a mid-to-low-level geometrically supervised MS COCO format label. Combined with the multi-task collaborative learning mechanism of a hierarchical supervised neural network model, it changes the traditional method's single mode of relying solely on semantic segmentation to generate masks. This enables the generated masks to possess both complete semantic region segmentation capabilities and precise geometric boundary details, effectively solving the problem of topological errors in vector reconstruction caused by smooth corners and irregular contours in traditional masks. It achieves accurate and reversible conversion from masks to regular vector contours.

[0016] 2. This invention designs a composite loss function that includes mask loss, gravitational field loss, vertex classification loss, and vertex offset loss. By balancing the contribution weights of the supervision signals of each task through preset loss factors, compared with the single loss design of traditional models that only focus on the semantics of the region, this invention significantly improves the model's geometric perception ability of the aquaculture area grid structure, avoids the extraction bias caused by insufficient geometric feature learning in traditional models, and significantly improves the accuracy of boundary and vertex capture for regular grids in aquaculture areas.

[0017] 3. In the prediction result optimization stage, this invention achieves sub-pixel level accuracy improvement by performing probability filtering and offset correction on vertex prediction results. This process breaks through the accuracy limitations of traditional pixel-level vertex positioning, provides a high-precision geometric benchmark for subsequent polygon contour reconstruction, and further ensures the regularity and accuracy of the aquaculture area contour.

[0018] 4. This invention constructs a complete end-to-end processing architecture from input of raw remote sensing images to output of aquaculture area contour vector results. It does not rely on additional manual post-processing operations. In the inference stage, it automatically segments independent aquaculture area instances and extracts boundary pixels through connected component analysis, which greatly reduces the manual intervention steps. This not only improves the automation efficiency of aquaculture area extraction, but also ensures that the output vector results retain the inherent regularity of the aquaculture area grid structure while reducing the number of vertices through polygon simplification rules, thus meeting the needs of nearshore aquaculture monitoring for high-precision and high-efficiency data output.

[0019] 5. This invention uses HRNet48 as the backbone network. Its parallel multi-resolution branch and cross-branch feature fusion structure can effectively preserve the detailed information in high-resolution remote sensing images and avoid the detail loss problem in the downsampling process of traditional backbone networks. At the same time, the frequency dynamic selection module can dynamically adjust the enhancement method of geometric features to semantic features according to the actual distribution characteristics of aquaculture areas in the input image through an adaptive weight generation mechanism. This enables the model to adapt to the aquaculture area extraction needs under different complex scenarios, significantly improves the robustness to complex environments, and solves the problem of sharp drop in extraction accuracy of traditional methods in scenarios with strong heterogeneity of spatial spectral features in aquaculture areas. This provides reliable technical support for large-area, high-precision monitoring of nearshore raft laver aquaculture areas. Attached Figure Description

[0020] Figure 1 This is a flowchart of the end-to-end extraction method for raft-type laver aquaculture areas that integrates hierarchical supervision and frequency dynamic filtering according to the present invention. Figure 2 This is a schematic diagram of the training phase of the multilayer supervised network model designed in this invention; Figure 3 This is a schematic diagram of the inference stage of the multi-layer supervised network model designed in this invention; Figure 4 A schematic diagram of the composition structure of the frequency dynamic selection module designed for this invention; Figure 5 This is a training set image of a raft-type laver aquaculture area after cropping from remote sensing images in an embodiment of the present invention; Figure 6 This is the training set ground truth image after cropping the remote sensing image of the raft-type laver aquaculture area in this embodiment of the invention; Figure 7 This is a visualization of the gravitational field of a raft-type laver aquaculture area from a remote sensing image, as described in this embodiment of the invention. Figure 8 This is a visualization of the MS COCO format labels in the raft-type laver cultivation area according to an embodiment of the present invention. Figure 9 This is a remote sensing image prediction result of a raft-type laver aquaculture area in an embodiment of the present invention; Figure 10This is a remote sensing image prediction result of a raft-type laver aquaculture area in an embodiment of the present invention.

[0021] Figure 11 The figures show the experimental results of each model in Embodiment 2 of the present invention. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the accompanying drawings.

[0023] Example 1 Reference Figure 1 This embodiment of an end-to-end extraction method for raft-type laver aquaculture areas based on frequency dynamic filtering includes: S1. Obtain the original remote sensing images of the raft-type laver aquaculture area and preprocess them to construct a masked dataset for high-level semantic supervision and MS COCO format labels for mid-to-low-level geometric supervision. S2. Construct a hierarchical supervised neural network model. The hierarchical supervised neural network model uses the HRNet48 network as the backbone network and includes a hierarchical feature extraction branch, a frequency dynamic selection module, and a multi-task prediction head. S3. Train the completed hierarchical supervised neural network model, including constructing a composite loss function consisting of mask loss, gravitational field loss, vertex classification loss and vertex offset loss; S4. Input the preprocessed remote sensing image of the raft-type laver aquaculture area to be extracted into the trained hierarchical supervised neural network model, and output the mask prediction result, vertex prediction result and gravitational field prediction result. S5. Optimize the prediction results output by the hierarchical supervised neural network model to obtain the optimized binary mask of the breeding area and the vertex coordinates with sub-pixel precision. S6. Based on the optimized binary mask and vertex coordinates, the contour vector result of the aquaculture area is generated through connected component analysis, vertex matching and polygon simplification.

[0024] Specifically, an end-to-end extraction method for raft-type laver aquaculture areas based on frequency dynamic filtering includes the following steps: S1. Obtain the original remote sensing images of the raft-type laver aquaculture area and preprocess them to construct a masked dataset for high-level semantic supervision and MS COCO format labels for mid-to-low-level geometric supervision. like Figure 1As shown, firstly, high-resolution multispectral raw remote sensing images covering the target raft-type laver aquaculture area are acquired. In this embodiment, the remote sensing images used are from the PMS sensor carried by the GF1 satellite. The images were taken in March 2024 (a critical period for laver aquaculture growth, with high distinction between the aquaculture area and the background). The geographical coverage range is 34°50′-35°10′ north latitude and 119°15′-119°40′ east longitude. The original spatial resolution of the images is 2 meters, and they include four spectral bands: blue, green, red, and near-infrared. The original image size is 10240×10240 pixels.

[0025] The following preprocessing operations were performed sequentially on the above raw remote sensing images: Radiometric calibration: The digital quantization value (DN value) of the image is converted into the top-of-atmosphere reflectance (TOA reflectance) using the radiometric calibration coefficients built into the sensor, eliminating the radiometric error caused by differences in sensor response. Specifically, this is achieved through the radiometric calibration formula provided by the satellite, ensuring that the radiometric values ​​of different pixels are physically comparable. Atmospheric correction: The dark target method combined with the 6S radiative transfer model is used to remove the influence of atmospheric scattering and absorption on the image, convert TOA reflectance to surface reflectance, and focus on correcting the brightness deviation in the near-infrared band caused by atmospheric molecular scattering, so that the spectral characteristics of the aquaculture area net curtain and seawater background in the image are closer to the real surface conditions. Orthorectification: Using the 1:50,000 digital elevation model (DEM) of the region as the elevation benchmark, the image is orthorectified using the rational function model (RFM) to eliminate geometric distortions caused by terrain undulations, satellite attitude deviations, and the curvature of the earth, so that each pixel of the corrected image corresponds to the real geographic coordinates (WGS84 coordinate system), ensuring that the geometric dimensions of aquaculture areas in different regions are consistent with reality.

[0026] Through the above preprocessing, a preprocessed remote sensing image with unified geometric and radiometric accuracy is obtained. Then, professional remote sensing image annotation software, such as eCognitionDeveloper, is used to interactively annotate the preprocessed remote sensing image. Labeling scope determination: Based on the distribution range of the target aquaculture area obtained from the field survey (including the ownership boundary of the aquaculture area and the actual aquaculture grid layout), the area to be labeled is defined on the image to ensure that the labeling scope completely covers all raft laver aquaculture areas and does not include near-shore facilities (such as fishing ports and ships) in non-aquaculture areas. Pixel-level differentiation and labeling: The aquaculture area and background in the image are labeled at the pixel level by combining manual outlining with automatic edge capture. The pixels where the aquaculture area is located are assigned a value of 1, and the background pixels (including seawater, exposed mudflats, non-aquaculture facilities, etc.) are assigned a value of 0. A rasterized label map is generated. During the labeling process, for the blurred areas at the edge of the aquaculture area caused by mixed pixels, the boundary direction of the adjacent clear aquaculture area is referenced for smoothing, so as to ensure that the outline of the aquaculture area in the label map is consistent with the geometry of the actual aquaculture grid. Labeling quality verification: The generated rasterized label images are sampled and verified. 10% of the labeled areas are randomly selected and the label images are superimposed and compared with the original preprocessed images to check the accuracy of the division between the breeding area and the background, ensuring that the labeling error rate (such as missing labels or mislabeling) is ≤2%, thus forming a reliable mask truth base.

[0027] To accommodate the input size requirements of hierarchical supervised neural network models and avoid insufficient training memory due to excessively large image sizes, a simultaneous sliding window cropping process is performed on the preprocessed remote sensing images and rasterized label images. The specific steps are as follows: Cropping parameter settings: The cropping window size is set to 512×512 pixels, which can balance model training efficiency and feature detail preservation. The sliding step size is 384 pixels, that is, the overlap ratio of adjacent windows is 25%. The setting of the overlapping area can avoid the loss of the boundary information of the breeding area during the cropping process, and ensure that each breeding area instance is completely contained in at least one cropping window. Synchronous cropping execution: Using the Python programming language combined with the OpenCV library, an automated cropping script is written to perform synchronous sliding cropping on the preprocessed remote sensing image and the rasterized label image. With the upper left corner of the image as the origin, the script slides row by row and column by column according to the set window size and step size. Each slide generates a set of image block-mask block. The image block is a local area of ​​the preprocessed image, and the mask block is the rasterized label of the corresponding area. Dataset selection and organization: The cropped image patches and mask patches are selected, and invalid samples with a pixel ratio of ≤5% in the aquaculture area are removed. The valid samples are retained and divided into training set, validation set and test set in a ratio of 7:2:1. The divided samples are indexed according to the correspondence between the image patch storage path and the mask patch storage path to form a mask dataset for high-level semantic supervision of the model. This dataset can be directly input into the semantic segmentation branch of the model to learn the regional segmentation features between the aquaculture area and the background.

[0028] To provide accurate geometric information such as line segments and vertices for low-level geometric supervision in the model, geometric features need to be extracted from the mask dataset and converted into the MS COCO standard format: First, the boundary pixel coordinate sequence of the aquaculture area is extracted. The Canny edge detection algorithm is used to extract the edge pixels of the aquaculture area in the mask block. Then, the boundary pixel coordinate sequence of the aquaculture area contour is obtained by the contour tracking algorithm (such as the Suzuki algorithm). The coordinates are sorted in a clockwise direction with the upper left corner of the mask block as the origin. Each aquaculture area instance corresponds to one boundary pixel coordinate sequence, which contains the (x,y) coordinates of all boundary pixels of the aquaculture area contour. Next, geometric information was organized and formatted. Referring to the standard format of the MS COCO dataset (which includes three core fields: images, annotations, and categories), the extracted geometric information was organized and converted. The images field records the basic information of each cropped image patch, including image patch ID, file name, width (512 pixels), height (512 pixels), and geographic coordinate range (inherited from the preprocessed image, corresponding WGS84 coordinates of the region). The “annotations” field records the geometric annotation information for each aquaculture area instance, including annotation ID, corresponding image patch ID, category ID (set the category ID of the raft-type laver aquaculture area to 1), bounding box (the smallest bounding rectangle calculated from the boundary pixel coordinate sequence, in the format [x,y,width,height], where x and y are the coordinates of the top left corner of the bounding box, and width and height are the width and height of the bounding box), and segmentation information, converting the boundary pixel coordinate sequence into a one-dimensional array according to the format required by MS COCO; The categories field defines category information, including category ID, category name "raft-type laver farming area", and parent category ID set to 0, indicating no parent category. Format validation and dataset generation: The organized geometric information is converted into a JSON format file. The JSON file is validated using the format validation tool provided by MS COCO to ensure the integrity of the fields and the correctness of the format. After the validation is passed, the JSON file is combined with the corresponding cropped image patches to form MS COCO format labels for low-level geometric supervision in the model. These labels can provide accurate geometric supervision signals to the line segment detection branch and vertex detection branch of the model, which are used to learn the line segment connection relationship and vertex spatial distribution characteristics of the breeding area.

[0029] S2. Construct a hierarchical supervised neural network model. The hierarchical supervised neural network model uses the HRNet48 network as the backbone network and includes a hierarchical feature extraction branch, a frequency dynamic selection module, and a multi-task prediction head. By constructing a hierarchical supervised neural network model with HRNet48 as the backbone network and integrating hierarchical feature extraction branches, frequency dynamic selection modules, and multi-task prediction heads, the model achieves accurate learning and multi-dimensional information prediction of multi-level features of raft-type laver aquaculture areas in remote sensing images. The training dataset (masked dataset and image patches corresponding to MS COCO format labels) constructed in step S1 is input into HRNet48. After the image patches are standardized by the network input layer, they flow through the four feature extraction stages of HRNet48 in sequence. Each stage consists of multiple residual blocks, and parallel branches with different resolutions are set in each stage (such as the first stage outputting a single-resolution feature map, and subsequent stages gradually adding low-resolution branches). The branches achieve feature fusion through cross-resolution connections, ensuring that the high-resolution branch can always obtain the semantic information of the low-resolution branch, and the low-resolution branch can also supplement the detailed information of the high-resolution branch. Ultimately, HRNet48 outputs two types of features: one is the initial feature map, including the initial feature map corresponding to the gravitational field task and the initial feature map corresponding to the vertex task, both of which are used for subsequent geometric feature extraction; the other is the basic feature map F, used for subsequent semantic feature extraction and frequency domain enhancement processing. The number of channels in both feature maps is set to 256, and the spatial resolution is 1 / 4 of the input image patch resolution. That is, if the input image patch size is H×W, then the feature map size is... ,in, , These represent the height and width of the input image block, respectively.

[0030] Based on the feature map output by HRNet48, a hierarchical feature extraction branch is constructed, consisting of a mask feature extraction branch, a vertex feature extraction branch, and a gravitational field feature extraction branch. This branch takes the basic feature map F output by HRNet48 as input and contains three 3×3 convolution operations. The number of kernels in each convolution layer is set to 256. After each convolution layer, batch normalization and ReLU activation function operations are performed sequentially: Batch normalization normalizes the mean and scales the variance of each channel of the convolution output feature map to avoid gradient vanishing or gradient exploding problems during model training; the ReLU activation function sets the pixel values ​​less than 0 in the feature map to 0, retaining positive pixel values ​​and enhancing the model's ability to learn non-linear features. After processing by this branch, the output dimension remains 256 channels. Resolution mask features This feature focuses on enhancing the semantic distinction between the breeding area and the background, providing support for subsequent mask prediction.

[0031] Next, a vertex feature extraction branch is constructed, using the initial feature map output by HRNet48. The input is the same as the mask feature extraction branch, containing a combination of three 3×3 convolutional layers, batch normalization, and ReLU activation functions. The number of convolutional kernels and the feature map dimensions are also the same as in the mask feature extraction branch. This branch enhances the geometric details related to the vertices of the aquaculture area in the initial feature map through targeted convolution and activation operations, ultimately outputting 256 channels. Vertex features of resolution This is used for subsequent vertex prediction tasks.

[0032] Finally, a gravitational field feature extraction branch is constructed, using the initial feature map output by HRNet48. As input, the same three-layer 3×3 convolution, batch normalization, and ReLU activation function structure as the two branches mentioned above is used. Convolution operations capture geometric structure information related to the aquaculture area line segments in the initial feature map. After activation and normalization processing, a 256-channel output is generated. Gravitational field characteristics at high resolution Furthermore, in the construction of the gravitational field feature extraction branch, it is necessary to introduce an alpha gravitational field (AFM) to achieve pixel-level region encoding of the aquaculture area line segments. This encoding mechanism is a key geometric representation method connecting the bottom-level vertices and the high-level masks. The specific encoding process is as follows: Gravitational field (AFM) is a region-partition-based geometric representation method. Its core principle is to encode the geometric information of line segments into pixel-level 2D projection vector fields, ensuring that the geometric structure of the line segments can be accurately recovered and can interact with vertex and mask features. Specifically, it first defines the set of line segments for polygon instances in the aquaculture area: let all line segments of a polygon instance constitute the set... , where each line segment Defined by the coordinates of two endpoints, i.e. ,in, Let the coordinates be the starting point of the line segment. Let be the coordinates of the endpoint of the line segment, and m be the total number of line segments. Then, based on this set of line segments, the image lattice (i.e., the feature map space of the gravitational field feature extraction branch, denoted as ) is processed. Perform non-overlapping region division and uniquely assign each pixel to the nearest line segment: for image lattice For any pixel x in the array, calculate its distance to all line segments. The Euclidean distance, the distance calculation formula is: , Where t is a parameter on the line segment, t=0 corresponds to the starting point, and t=1 corresponds to the ending point. Essentially, this formula calculates the distance from pixel x to the line segment. The square of the Euclidean distance from the nearest point; if pixel x to line segment The distance satisfies Then pixel x will be assigned to the line segment. Corresponding area All regions constitute a partition set. Furthermore, each region does not overlap and completely covers the image lattice. .

[0033] Based on the above region division, a pixel-level gravitational field vector field A is constructed for the region. For any pixel x within the range, calculate the position of that pixel on the corresponding line segment. Projection point on (i.e., the line segment closest to pixel x) For a point on the graph whose coordinates are determined by the optimal solution of t, the gravitational field vector of that pixel is defined as follows: ,in, For pixel x on the corresponding line segment The projection points on the vector represent the projection deviations in the x and y directions, respectively. This encoding method ensures that the gravitational field vector of each pixel points to the projection point of its corresponding line segment, thus fully encoding the geometric information of the line segment into the vector field. Simultaneously, when the gravitational field map is divided into R-coded line segments based on regions, the regional information of the higher-level mask is implicitly encoded (i.e., the mask's contour range can be inferred from the region boundaries), laying the foundation for the interaction between geometric and semantic features in the subsequent frequency dynamic selection module.

[0034] The gravitational field feature extraction branch extracts gravitational field features that match the gravitational field vector field A from the initial feature map through three layers of 3×3 convolution, batch normalization, and ReLU activation. This feature is essentially a characteristic representation of the geometric information of line segments encoded by AFM. Subsequent input into the gravitational field prediction head can output the gravitational field prediction result. This enables the prediction and supervision of line segment geometry, while providing mid-level geometric guidance signals for the frequency dynamic selection module, ensuring that the model can accurately capture the line segment connection relationships in the aquaculture area, thus solving the problem of insufficient attention to geometry in traditional methods.

[0035] After the hierarchical feature extraction branch is completed, the frequency dynamic selection module is configured. This module uses gravitational field features. Vertex features As a geometric guidance signal, the fundamental feature map F is dynamically modulated and enhanced in the frequency domain. First, the gravitational field features are... With vertex features The features are concatenated along the channel dimension to form an aggregated guided feature map. Its dimensions are The mathematical expression is: , in, This is a channel-level splicing operation, followed by... Perform a global adaptive average pooling operation, reducing its spatial dimension from Compress to This yields a 512-dimensional feature vector containing global geometric context information. This operation extracts global geometric feature statistics by averaging all spatial pixels in each channel, thus avoiding interference from local geometric noise.

[0036] Will The input is a multilayer perceptron consisting of two fully connected layers. Let the weight matrix of the first fully connected layer be... Paranoia is Then the first layer output is ,in, The activation function is used to achieve feature dimensionality reduction and nonlinear transformation from 512 dimensions to 128 dimensions; let the weight matrix of the second fully connected layer be... Paranoia is Then the output of the second layer is Finally, for Execute the Softmax activation function to generate a dynamic weight vector of dimension N. The calculation formula is: , The sum of the elements in this vector is 1. The importance weight represents the k-th learnable global filter. The weight is adaptively adjusted by the geometric features (line segment distribution, vertex density) of the input image. If a certain type of geometric feature (such as sharp line segments) has a high proportion in the input image, the corresponding weight is increased to ensure that the filtering process focuses more on the enhancement guided by that type of geometric feature.

[0037] like Figure 3 , Figure 4 As shown, frequency domain filtering and feature modulation are then set up. First, a two-dimensional Fast Fourier Transform (FFT) is performed on the basic feature map F output by HRNet48 to transform it from the spatial domain to the frequency domain, obtaining the complex form of the frequency domain feature FFT(F). This transformation can decompose the feature information in the spatial domain into components of different frequencies, which facilitates customized filtering for different frequency features. At the same time, N learnable global filter bases are preset. The dimension of each filter basis is consistent with the frequency domain feature FFT(F), and it is continuously optimized through the model training process to adapt to the frequency domain distribution pattern of the aquaculture area features; the frequency domain feature FFT(F) is then integrated with each filter basis. Element-wise complex multiplication is performed on each feature map, which is mathematically equivalent to convolving the base feature map F with the filter basis in the spatial domain, resulting in N frequency domain feature maps filtered at different frequencies. Then, the generated dynamic weight vector is used... The N filtered frequency domain feature maps are weighted and summed (i.e., each filtered feature map is multiplied by its corresponding weight and then summed) to obtain a combined frequency domain feature map that incorporates geometric guidance information. Finally, a two-dimensional inverse fast Fourier transform (IFFT) is performed on the combined frequency domain feature map to transform it back from the frequency domain to the spatial domain, resulting in the enhanced feature map. This feature map retains the semantic information of the basic feature map while incorporating boundary detail enhancement information guided by geometric features. It is directly used as input to the mask feature extraction branch to optimize the quality of the mask features. The enhanced feature map expression is as follows: , in, For two-dimensional fast Fourier transform, This is a two-dimensional inverse fast Fourier transform. For element-wise complex multiplication, For the k-th learnable global filter basis, This is the k-th element of the dynamic weight vector.

[0038] Finally, a multi-task prediction head is constructed, using the mask features extracted from the mask feature extraction branch. As input, the prediction head consists of two convolutional layers: the first layer is a 3×3 convolution, compressing the input 256-channel feature map to 64 channels, and then performing a ReLU activation function to enhance the non-linear expression; the second layer is a 1×1 convolution, further compressing the 64-channel feature map to 2 channels, corresponding to the probability outputs of the aquaculture area and the background, respectively. The final output size is [size missing]. Mask prediction results , is used to characterize the probability that each pixel belongs to the breeding area or the background.

[0039] Vertex features extracted from the branch output using vertex features For input, the same two-layer convolutional structure as the mask prediction head is used: the first 3×3 convolution reduces the 256-channel feature map to 64 channels and activates it with ReLU; the second 1×1 convolution reduces the 64-channel feature map to 3 channels, corresponding to the probability outputs of background, concave points, and convex points, respectively. The final output size is [size missing]. Vertex prediction results This is used to characterize the probability that each pixel belongs to the background, concave point, or convex point.

[0040] Gravitational field features extracted from the branch output using gravitational field features For input, the structure is consistent with the mask prediction head: the first 3×3 convolution reduces the 256-channel feature map to 64 channels and activates it with ReLU; the second 1×1 convolution reduces the 64-channel feature map to 2 channels, corresponding to the two dimensions of the gravitational field (vector components in the x and y directions). The final output size is... Gravitational field prediction results The geometric information of the aquaculture area line segments is used to encode the geometric information. Through the above construction, the hierarchical supervised neural network model can realize end-to-end processing from remote sensing image input to parallel prediction of three types of information: aquaculture area mask, vertices, and gravitational field.

[0041] S3. Train the completed hierarchical supervised neural network model, including constructing a composite loss function consisting of mask loss, gravitational field loss, vertex classification loss and vertex offset loss; like Figure 2 As shown, the mask ground truth values ​​in the mask dataset constructed in step S1 For the purpose of supervision, For rasterized labels, the dimensions and the mask prediction results of the model output are... Consistent, among which, To calculate the probability distribution of the model's mask prediction head output, the binary cross-entropy (BCE) loss is used to calculate the difference between the two. This loss effectively measures the deviation between the predicted probability and the ground truth label in a binary classification task. The specific calculation process is as follows: First, for... Performing softmax activation yields a probability map of the aquaculture area and the background. ,in, c=0,1 represents the channel index, and i,j represents the spatial pixel coordinates. Then, the binary cross-entropy loss is calculated, expressed as follows: , in, This is the function for calculating the binary cross-entropy loss. The loss is calculated by summing the classification probability of each pixel with the logarithmic error of the ground truth label and averaging the results. This forces the model to learn the semantic distinguishing features between the breeding area and the background, ensuring the accuracy of the region segmentation predicted by the mask.

[0042] Gravitational field loss function Construction: First, the true value of the gravitational field is generated based on the line segment information in the MS COCO format tags constructed in step S1. , The generation process for the pixel-level 2D projected vector field is as follows: Based on the boundary pixel coordinate sequence of the aquaculture area outline in the MS COCO tag, each line segment instance is determined. Then, each pixel is assigned to the nearest line segment, and the vector rule pointing to the projection point of the line segment is calculated to generate the gravitational field vector for each pixel, which serves as the basis for the generation of the gravitational field vector. The model outputs the gravitational field prediction results. This represents the vector prediction value output by the gravitational field prediction head. (Using...) The difference between the two loss calculations Loss is less sensitive to outliers than Loss, which provides a more robust measure of vector prediction error, is expressed as follows: , in, Let gravitational field loss function be used. for The loss calculation function, expanded as follows: N is the total number of elements. The gravitational field prediction result output by the model's gravitational field prediction head is a pixel-level 2D projected vector field. The loss is the true vector field of the gravitational field. It is obtained by summing and averaging the absolute errors between the predicted vector value and the true value of each pixel, forcing the model to learn the geometric structure information of the line segments in the breeding area, and ensuring that the direction and magnitude of the predicted vector of the gravitational field conform to the actual distribution of line segments.

[0043] Vertex classification loss function Construction: Generate vertex heatmap truth values ​​based on vertex position information in the MS COCO format tags constructed in step S1. , For dimension The raster image is generated according to the following rules: at the vertex pixel position, the corresponding category channel is assigned a value of 1, and the other channels are assigned a value of 0; all channels of non-vertex pixels are assigned a value of 0; the vertex heatmap prediction result output by the model. To calculate the classification error based on the class probability distribution output by the vertex prediction head, binary cross-entropy (BCE) loss is used. Specifically, softmax activation is applied to H to obtain the probability map. ,in, Then calculate the loss: , in, The vertex heatmap prediction results output by the model's vertex prediction head. The loss function provides ground truth labels for vertex heatmaps. This loss focuses on optimizing the classification accuracy of vertex categories, ensuring that the model can accurately distinguish between background, concave, and convex pixels.

[0044] Vertex offset loss function Construction: Based on vertex truth values The positional deviation of the corresponding pixel generates the offset true value. , For dimension The vector field is generated according to the following rules: for For each pixel marked as a vertex, calculate the deviation between its coordinates on the feature map and the sub-pixel level true vertex coordinates, and use this deviation as... Vector values ​​at corresponding positions; non-vertex pixels. The vector value is set to 0, and the model outputs the offset prediction result. The offset vector output by the auxiliary branch of the vertex prediction head is used. The loss is calculated by the difference between the two in the vertex region, and by... The expression for masking non-vertex regions (calculating only the loss of vertex pixels) is as follows: , in, The vertex offset prediction result is output by the auxiliary branch of the model vertex prediction head. This loss can optimize the sub-pixel level localization accuracy of vertices and compensate for the vertex position deviation caused by feature map downsampling.

[0045] The training dataset divided in step S1 is input into the hierarchical supervised neural network model, using mask loss. With gravitational field loss The weighted sum is used as the sub-target loss. ,Right now ,in, , The loss factor is used, and the gradient of the sub-objective loss with respect to the model parameters is calculated using the backpropagation algorithm. The parameters are then updated using a stochastic gradient descent (SGD) optimizer, with the initial learning rate of the optimizer set to [value missing]. Momentum is set to 0.9, and weight decay is set to... To prevent the model from overfitting, this stage of training iterates for 10 epochs, with the learning rate decreasing to 0.5 of the original rate every 2 epochs. The goal is to allow the model to first master the basic features of the regional semantics and line segment geometry of the aquaculture area.

[0046] Finally, a composite loss function is constructed and full training is performed, including vertex classification loss. Vertex offset loss With mask loss function and gravitational field loss function Weighted fusion is performed to obtain the composite loss function. The calculation formula is: , in, , , and To predict the loss factors for mask loss, gravitational field loss, heatmap loss, and vertex offset loss, the values ​​of all loss factors are dynamically adjusted based on the convergence speed of each loss term during training. If a certain type of loss decreases slowly, the corresponding loss factor is appropriately increased to ensure balanced convergence of losses across all tasks. During the full training phase, the training dataset is still used as the input to the model. For the objective function, the optimizer is changed to the Adaptive Moment Estimator (Adam) optimizer, and the initial learning rate of the Adam optimizer is set to... Weight decay is set to During training, an early stopping strategy is adopted, using the composite loss of the validation set as the monitoring metric. If the validation set loss does not decrease for 5 consecutive epochs, training is stopped to avoid model overfitting. At the same time, the model parameters are saved every 3 epochs, and the checkpoint with the smallest validation set loss is finally selected as the completed model. This ensures that the model has good generalization ability and can accurately learn the semantic, line segment, and vertex multi-dimensional features of the breeding area.

[0047] S4. Input the preprocessed remote sensing image of the raft-type laver aquaculture area to be extracted into the trained hierarchical supervised neural network model, and output the mask prediction result, vertex prediction result and gravitational field prediction result. The extracted remote sensing images of the raft-type laver aquaculture area undergo the same preprocessing operation as in step S1 to ensure that the geometric and radiometric accuracy of the images meet the model input requirements, resulting in preprocessed remote sensing images. If the size of the preprocessed remote sensing image exceeds the input size during model training, a sliding window cropping strategy is used to crop the preprocessed remote sensing image into several image blocks. The size of each image block is consistent with the model training input size. At the same time, the spatial position index of each image block in the original preprocessed remote sensing image is recorded for subsequent stitching and reconstruction of the prediction results. If the size of the preprocessed remote sensing image does not exceed the model input size, it is directly used as a single block input without cropping.

[0048] Subsequently, the cropped image patches are standardized according to the input format used during model training. Using the same mean and standard deviation as in the training phase, each spectral band of the image patch is normalized to eliminate the influence of differences in radiance values ​​between different bands on model inference, resulting in standardized model input data. This input data is then input into the hierarchical supervised neural network model trained in step S3. The model performs inference according to the following process: The input data first enters the HRNet48 backbone network. Through the network's four feature extraction stages and multi-resolution branch fusion mechanism, multi-scale features are extracted, outputting initial feature maps and basic feature maps F with the same dimensions as in the training phase. Then, the initial feature maps are input into the corresponding gravitational field feature extraction branch and vertex feature extraction branch, respectively. After three layers of 3×3 convolution, batch normalization, and ReLU activation operations, the gravitational field features are obtained. Vertex features The basic feature map F is then input to the frequency dynamic selection module, in which it is compared with the gravitational field features. Vertex features Interactive fusion is performed, and enhanced feature maps are output through dynamic weight generation and frequency domain filtering modulation. and will The input mask feature extraction branch undergoes the same convolution, normalization, and activation processes to obtain the mask features. Finally, mask features The input mask prediction header is processed through two convolutional layers, and the output is a mask prediction result containing the probability distributions of the aquaculture area and the background. Vertex features The input vertex prediction head, after passing through the same convolutional structure, outputs vertex prediction results containing three probability distributions: background, concave points, and convex points. Gravitational field characteristics The input gravitational field prediction head is processed through the same convolutional structure, and the output is a gravitational field prediction result containing the x and y direction vector components. .

[0049] If the preprocessed remote sensing image to be extracted is cropped into multiple image patches, the above reasoning process is performed on each image patch to obtain the mask prediction result, vertex prediction result, and gravitational field prediction result corresponding to each image patch. Then, according to the spatial location index recorded during cropping, the prediction results of each image patch are stitched together. For the mask prediction result and the gravitational field prediction result, the strategy of averaging the overlapping areas is adopted to stitch the prediction results of all image patches into a complete prediction result with the same size as the original preprocessed remote sensing image. For the vertex prediction result, the vertex prediction probability distribution in each image patch is stitched into a complete probability map according to the spatial location index. At the same time, the preliminary position of each candidate vertex in the original image coordinate system is recorded to ensure that the stitched prediction result can completely cover the range of the raft-type laver aquaculture area to be extracted.

[0050] Ultimately, the model outputs three complete prediction results: First, the mask prediction result, with the same dimension as the original preprocessed remote sensing image, includes the probability distribution of the aquaculture area and the background, which is used to generate a binary mask for the aquaculture area. Second, the vertex prediction result, also with the same dimension as the original preprocessed remote sensing image, includes the probability distribution of the background, concave points, and convex points, which is used to extract vertex coordinates with sub-pixel accuracy. Third, the gravitational field prediction result, with the same dimension as the original preprocessed remote sensing image, includes vector components in the x and y directions, which is used to help verify the accuracy of the geometric structure of the aquaculture area's line segments and provide geometric reference for subsequent optimization processing.

[0051] S5. Optimize the prediction results output by the hierarchical supervised neural network model to obtain the optimized binary mask of the breeding area and the vertex coordinates with sub-pixel precision. First, analyze the mask prediction result output in step S4. To process, The optimization process for the feature map output by the model mask prediction head is as follows: First, perform the Softmax activation operation, and... Converting to a probability distribution within the 0-1 interval, the calculation formula is: , in, Corresponding background channel, Corresponding passageway to the breeding area Using the pixel coordinates of the feature map, this operation extracts the probabilistic feature map corresponding to the aquaculture area: The second step is to use bilinear interpolation to restore the resolution, because... The resolution is the feature map resolution output by the backbone network, which needs to be adjusted to the resolution of the original preprocessed remote sensing image. Bilinear interpolation improves the resolution by calculating the weighted average of the four neighboring pixels around the pixel to be interpolated. The weights are determined by the distance between pixels, ensuring the spatial continuity of the interpolated probability map. After interpolation, a probability map of the aquaculture area matching the size of the original image is obtained. .

[0052] The third step is to set the probability threshold T. This embodiment sets Perform binarization segmentation according to the formula. ,in, For pixel coordinates, For a preset probability threshold and , Generate an optimized binary mask for the aquaculture area based on the resolution-restored probability map of the aquaculture area. ,in, These are the pixel coordinates of the original image. This indicates that the pixel is a breeding area. This indicates that the pixel is the background. Threshold filtering can effectively remove pixels with low probability of being suspected breeding areas, ensuring the accuracy of the region division of the binary mask.

[0053] Regarding the vertex prediction results output in step S4 and the vertex offset field output by the backbone network The first step is to Perform the Softmax activation operation to obtain the class probability distribution. , As background, It is a concave point. For convex points, extract the probability map of concave points. With convex point probability map The second step is to offset the vertex field. To perform correction, first execute the Sigmoid activation operation to map the offset values ​​to the [0,1] interval, then subtract 0.5 to obtain the vertex offset prediction with a value range of [-0.5, 0.5). The calculation formula is: , This correction ensures the offset remains within a single pixel, providing a basis for compensation for sub-pixel-level positioning; the third step uses a local nonmaximum suppression algorithm to filter candidate vertices, in order to Pixels are a sliding window, for and Process them separately, retaining only those with the highest probability value within the window and a classification score higher than a preset threshold. The pixels are selected as candidate vertices, and their coordinates on the feature map are recorded. and the corresponding offset value .

[0054] The fourth step is to calculate the sub-pixel level vertex coordinates in the original image coordinate system. First, the scaling factor is determined based on the resolution ratio between the backbone network output feature map and the original image. According to the formula , in, Let be the coordinates of the candidate vertex on the feature map. , This is the resolution scaling factor. , As the offset is at the sub-pixel level, through the above optimization process, two types of core data are finally obtained: one is a binary mask of the aquaculture area with clear boundaries and accurate regions. First, it can be directly used for instance segmentation of aquaculture areas; second, it is a set of sub-pixel level vertex coordinates with high positioning accuracy, which lays a key foundation for generating regular aquaculture area contour vector results through connected component analysis and vertex matching.

[0055] S6. Based on the optimized binary mask and vertex coordinates, the contour vector result of the aquaculture area is generated through connected component analysis, vertex matching and polygon simplification.

[0056] like Figures 5-10 As shown, the optimized binary mask for the aquaculture area in step S5. The analysis of connected components is performed using the 8-neighborhood connectivity criterion (i.e., if a pixel has 8 neighboring pixels (value 1) in its eight directions (up, down, left, right, and diagonal), it is considered to be in the same connected region). The analysis is then performed using either the scan-line algorithm or the seed-fill algorithm. For all pixels, extract all independent connected regions. Each connected region corresponds to a complete instance of a raft-type laver aquaculture area, denoted as . , Where n is the total number of instances in the aquaculture zone, after extracting the connected components, the calculation is further performed for each instance in the aquaculture zone. boundary pixel set This embodiment employs edge detection algorithms, specifically Canny edge detection or a boundary extraction algorithm based on neighborhood difference, to identify edge detection. For all pixels that satisfy the condition of being a breeding area pixel and having at least one adjacent pixel that is a background pixel (with a value of 0), sort the coordinates of these pixels in a clockwise direction to form a boundary pixel sequence. ,in, For example The number of boundary pixels, The coordinates of the boundary pixels in the original remote sensing image are used to provide a boundary location reference for subsequent vertex matching.

[0057] Summarize all sub-pixel level vertex coordinates obtained in step S5, remove duplicate coordinates, and then set a coordinate error threshold. pixels, if the Euclidean distance between the coordinates of two vertices is less than 1 pixel. If a vertex is found to be a duplicate, the vertex with the higher probability score is retained to form the global vertex set. ,in, The total number of vertices globally. These are sub-pixel level vertex coordinates. A distance threshold is then set. , Based on the preset grid density of the aquaculture area, this embodiment typically uses 2-5 original image pixels to ensure that only vertices near the boundary are matched for each aquaculture area instance. boundary pixel set Calculate each boundary pixel With all vertices in the global vertex set V The Euclidean distance, the distance calculation formula is: , If there exists a vertex v such that Then the vertex and the boundary pixel Association, and thus each instance Filter out the matching subset of vertices That is, all of them The distance of any boundary pixel in the middle is ≤ (vertices). To ensure the orderliness of subsequent polygon construction, it is done by boundary pixel set. The clockwise sorting order for the vertex subset Reorder the vertices in the array and traverse them. The boundary pixel sequence is used to record the vertices associated with each boundary pixel in turn, and duplicate vertices are removed to form an instance. initial vertex sequence , where t is the length of the initial vertex sequence, which ensures that the vertices are distributed in a clockwise direction according to the outline of the breeding area, providing an ordered vertex basis for the construction of the closed polygon.

[0058] Polygon simplification and vector result generation first involves the initial vertex sequence. Perform redundant vertex removal: Calculate the index difference between adjacent vertices in the sequence. , Set an index difference threshold for the index of the p-th vertex in the sequence. ,like If adjacent vertices are too densely packed, remove the middle vertices and keep only the first and last endpoints to obtain a preliminary simplified vertex sequence. This step can remove local redundant vertices caused by excessively high vertex detection accuracy, reducing the amount of subsequent computation.

[0059] Next, perform angle constraint simplification: for Take three consecutive vertices in sequence. Calculate vector with vector The included angle The formula for calculating the included angle is: ,in, The dot product of two vectors. Set an angle threshold for the magnitude of the vector. This embodiment is (Radius, adapted to the angular characteristics of the regular grid in raft-type laver farming areas) If a vertex is found to be redundant and its adjacent edges are nearly collinear, removing it will not affect the contour shape, so it is deleted from the sequence; this operation is repeated iteratively until... The included angle of all three consecutive vertices in the equation satisfies This yields the final simplified vertex sequence. Finally, the outline vector result of the aquaculture area is generated: each aquaculture area instance is... The final simplified vertex sequence Connecting the first and last vertices in a clockwise order forms a closed polygonal outline of the aquaculture area (ensuring the coordinates of the first and last vertices of the sequence are consistent to achieve polygon closure). Simultaneously, the confidence score for each aquaculture area instance is calculated (taking the average of the mask prediction probabilities corresponding to that instance, or the minimum vertex prediction probability, used to characterize the reliability of the extraction results), and associated with the instance number i. Following standard geographic information vector data formats (such as SHP or GeoJSON), the vertex coordinates (including the geographic coordinates corresponding to the original remote sensing image, obtained through geolocation information conversion during preprocessing in step S1), instance number, confidence score, and other information of each closed polygon are organized and stored, ultimately generating a complete vector outline result for the raft-type laver aquaculture area. This vector result can be directly imported into a Geographic Information System (GIS) for subsequent applications such as aquaculture area statistics, distribution analysis, and dynamic monitoring.

[0060] Example 2 The difference between this embodiment and Embodiment 1 is that this embodiment provides a specific experimental method: Table 1 Quantitative Comparison Results of Haizhou Bay Dataset

[0061] In quantitative experiments on spectral blurring in semi-submerged scenes, the proposed method achieved the best performance (Table 1). Its recall, F1 score, and intersection-over-union (IoU) ratio all outperformed all comparison models. Compared to strong baseline models such as UANet, the F1 score and IoU ratio were improved by 2.39% and 2.94%, respectively, fully demonstrating the effectiveness of the model's hierarchical supervision and frequency domain modulation mechanism in restoring degraded boundaries and ensuring the integrity of the regional structure. The slight difference in accuracy reflects its innovation in directly constructing vector boundaries based on point, line, and surface extraction tasks: this design abandons the traditional method's reliance on irregular pixel boundaries, instead generating regular structures based on the geometric features of the aquaculture area. This difference makes our results more valuable in downstream applications.

[0062] Qualitative results show that, for example Figure 10 , Figure 11 As shown, traditional methods such as U-Net generate a large number of broken and stuck error masks in spectrally blurred regions, making it difficult to form effective vector contours; while the method in this paper can still infer geometrically regular and topologically complete polygons even when the target is severely obscured by water, demonstrating a strong geometric structure reasoning ability.

[0063] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An end-to-end extraction method for raft-type laver aquaculture areas based on frequency dynamic filtering, characterized in that, include: S1. Obtain the original remote sensing images of the raft-type laver aquaculture area and preprocess them to construct a masked dataset for high-level semantic supervision and MS COCO format labels for mid-to-low-level geometric supervision. S2. Construct a hierarchical supervised neural network model. The hierarchical supervised neural network model uses the HRNet48 network as the backbone network and includes a hierarchical feature extraction branch, a frequency dynamic selection module, and a multi-task prediction head. S3. Train the completed hierarchical supervised neural network model, including constructing a composite loss function consisting of mask loss, gravitational field loss, vertex classification loss and vertex offset loss; S4. Input the preprocessed remote sensing image of the raft-type laver aquaculture area to be extracted into the trained hierarchical supervised neural network model, and output the mask prediction result, vertex prediction result and gravitational field prediction result. S5. Optimize the prediction results output by the hierarchical supervised neural network model to obtain the optimized binary mask of the aquaculture area and the vertex coordinates with sub-pixel precision. S6. Based on the optimized binary mask and vertex coordinates, the contour vector result of the aquaculture area is generated through connected component analysis, vertex matching and polygon simplification.

2. The end-to-end extraction method for raft-type laver aquaculture areas based on frequency dynamic filtering according to claim 1, characterized in that, Step S1 specifically includes: S11. Acquire high-resolution multispectral raw remote sensing images covering the target raft laver aquaculture area, and sequentially perform radiometric calibration, atmospheric correction and orthorectification preprocessing on the raw remote sensing images to obtain preprocessed remote sensing images with unified geometric and radiometric accuracy. S12. Interactively annotate the preprocessed remote sensing image to generate a rasterized label map, and assign different values ​​to the pixels of the aquaculture area and the background pixels in the label map to achieve pixel-level differentiation, forming the basis of the mask true value. S13. Perform synchronous sliding window cropping on the preprocessed remote sensing image and the rasterized label image to obtain the corresponding image blocks and mask blocks, and construct a mask dataset for high-level semantic supervision. S14. Obtain the boundary pixel coordinate sequence of the aquaculture area outline of each mask block in the mask dataset, organize and convert the boundary pixel coordinate sequence and aquaculture area instance information according to the MS COCO dataset standard format, and construct MS COCO format labels for mid-to-low-level geometric supervision.

3. The end-to-end extraction method for raft-type laver aquaculture areas based on frequency dynamic filtering according to claim 1, characterized in that, The construction of the hierarchical supervised neural network model includes performing multi-scale feature extraction on the preprocessed dataset using the backbone network HRNet48, outputting initial feature maps and basic feature maps F, constructing a hierarchical feature extraction branch, performing three 3×3 convolution operations on the initial feature maps and basic feature maps F, and sequentially performing batch normalization and ReLU activation function operations after each convolution layer to obtain the corresponding gravitational field features. Vertex features and mask features The hierarchical feature extraction branch consists of a mask feature extraction branch, a vertex feature extraction branch, and a gravitational field feature extraction branch.

4. The end-to-end extraction method for raft-type laver aquaculture areas based on frequency dynamic filtering according to claim 1, characterized in that, The construction of the hierarchical supervised neural network model also includes constructing a frequency dynamic selection module based on gravitational field features. Vertex features and mask features As input, and A converged guided feature map is formed by concatenating features along the channel dimension. After global adaptive average pooling to compress the spatial dimension, it is input into a multilayer perceptron. A dynamic weight vector of dimension N is generated by the Softmax activation function. The basic feature map F is then converted into frequency domain features by a two-dimensional fast Fourier transform. After element-wise complex multiplication with N learnable global filter bases, the features are weighted and fused by the dynamic weight vector. Finally, the enhanced feature map is output by a two-dimensional inverse fast Fourier transform. The expression of the enhanced feature map is as follows: ， in, For two-dimensional fast Fourier transform, This is a two-dimensional inverse fast Fourier transform. For element-wise complex multiplication, For the k-th learnable global filter basis, This is the k-th element of the dynamic weight vector.

5. The end-to-end extraction method for raft-type laver aquaculture areas based on frequency dynamic filtering according to claim 1, characterized in that, The training of the constructed hierarchical supervised neural network model specifically includes: S311. Construct the mask loss function: using the true mask values ​​in the mask dataset. As a basis for supervision, the mask prediction results output by the binary cross-entropy loss calculation model are used. with truth value The difference, the mask loss function is denoted as ; S312. Constructing the gravitational field loss function: Generating the true value of the gravitational field based on the line segment information in the MS COCO format tags. ,use Gravitational field prediction results output by the loss calculation model with truth value The deviation, the gravitational field loss function is denoted as ; S313. Input the training dataset into the hierarchical supervised neural network model, using mask loss. With gravitational field loss The weighted sum is used as the sub-objective loss, and the model parameters are updated through backpropagation.

6. The end-to-end extraction method for raft-type laver aquaculture areas based on frequency dynamic filtering according to claim 1, characterized in that, The step of training the constructed hierarchical supervised neural network model further includes: S321. Constructing a vertex classification loss function: Generating vertex heatmap ground truth based on vertex position information in MS COCO format labels. The vertex heatmap prediction result H, calculated using the binary cross-entropy loss model, is compared with the true value. The classification error, the vertex classification loss function is denoted as ; S322. Construct the vertex offset loss function: based on the vertex truth value The positional deviation of the corresponding pixel generates the offset true value. ,use The offset prediction result O output by the loss calculation model and the true value The difference in the vertex region is denoted by the vertex offset loss function as follows: ; S323. Constructing a composite loss function and full model training: Incorporating vertex classification loss... Vertex offset loss With mask loss function and gravitational field loss function Weighted fusion is performed to obtain the composite loss function. ,by Given the objective function, an adaptive moment estimator optimizer is used to iteratively train the model on the training dataset until the model converges. The expression for the composite loss function is: ， in, , , and The loss factor is used to predict mask loss, gravitational field loss, heatmap loss, and vertex offset loss.

7. The end-to-end extraction method for raft-type laver aquaculture areas based on frequency dynamic filtering according to claim 1, characterized in that, The optimization process for the prediction results output by the hierarchical supervised neural network model includes performing a Softmax activation operation on the masked prediction results output by the hierarchical supervised neural network model, extracting the probability feature map corresponding to the aquaculture area, and using a bilinear interpolation algorithm to adjust the resolution of the probability feature map from the resolution of the feature map output by the backbone network to the resolution of the original remote sensing image, thereby obtaining a probability map of the aquaculture area that matches the size of the original image. Set a probability threshold T and generate a probability map of the aquaculture area after resolution restoration. Perform binarization segmentation to generate an optimized binary mask for the aquaculture area: ， in, For pixel coordinates, For a preset probability threshold and , This is a probability map of the aquaculture area after resolution restoration.

8. The end-to-end extraction method for raft-type laver aquaculture areas based on frequency dynamic filtering according to claim 1, characterized in that, The optimized binary mask of the aquaculture area and sub-pixel-level precision vertex coordinates are obtained by performing a Softmax activation operation on the vertex prediction results output by the hierarchical supervised neural network model, extracting concave and convex probability maps, obtaining the vertex offset field output by the backbone network, performing a Sigmoid activation operation and subtracting 0.5 to obtain vertex offset predictions with values ​​in the range [-0.5, 0.5). A local nonmaximum suppression algorithm is used to retain pixels with classification scores higher than a preset threshold in the concave and convex probability maps as candidate vertices. A scaling factor is calculated based on the resolution ratio of the backbone network output feature map to the original image. Combining the coordinates of the candidate vertices on the feature map and the corresponding offset predictions, sub-pixel-level vertex coordinates in the original image coordinate system are calculated. The sub-pixel-level vertex coordinates satisfy the following: , in, Let be the coordinates of the candidate vertex on the feature map. , This is the resolution scaling factor. , This is a subpixel-level offset.

9. The end-to-end extraction method for raft-type laver aquaculture areas based on frequency dynamic filtering according to claim 1, characterized in that, The generated aquaculture area contour vector result includes the optimized aquaculture area binary mask. Perform connected component analysis and use the 8-neighborhood connectivity criterion to extract all independent connected regions. Each connected region corresponds to a breeding area instance. And calculate the value of each instance. The bounded pixel set, based on the coordinates of all sub-pixel level vertices, forms the global vertex set. Set a distance threshold for each aquaculture zone instance. Given a set of boundary pixels, calculate the Euclidean distance between each boundary pixel and all vertices in the global vertex set, and filter out a subset of matching vertices based on a distance threshold. And in clockwise order of boundary pixels Sorting the vertices to form instances The initial vertex sequence.

10. The end-to-end extraction method for raft-type laver aquaculture areas based on frequency dynamic filtering according to claim 1, characterized in that, The generated aquaculture area contour vector result also includes data generated based on each aquaculture area instance. The initial vertex sequence is calculated by finding the index difference between adjacent vertices. An index difference threshold is set to optimize the initial vertex sequence. Based on the optimized initial vertex sequence, the vector angle between adjacent vertices is calculated for every three consecutive vertices. An angle threshold is set to filter the vertices, resulting in a final simplified vertex sequence. This simplified sequence is then applied to each aquaculture area instance. The final simplified vertex sequence is connected in clockwise order to form a closed aquaculture area outline polygon. The vertex coordinates, instance numbers and corresponding confidence scores of the polygon are stored in the geographic information vector data format to generate the aquaculture area outline vector result.

Images