Coastal natural object remote sensing segmentation method and system based on edge information deep supervision

Abstract

This application belongs to the field of image semantic segmentation technology, specifically relating to a remote sensing segmentation method and system for coastal natural targets with edge information depth supervision. The system includes a remote sensing image preprocessing module, an image segmentation dataset generation module, a prediction module, and an output module. The core of this method is the construction of a multi-path edge supervision structure, which embeds an automatic edge supervision label generation module and designs a combination of multi-path loss functions that emphasize edge features to enhance the model's edge learning ability. This addresses problems in existing coastal target segmentation technologies, such as blurred edges, missed detection of small targets, and class imbalance, achieving high-precision segmentation of coastal natural targets and providing technical support for coastal ecological monitoring.

Description

Technical Field

[0001] This application belongs to the field of image semantic segmentation technology, specifically relating to a remote sensing segmentation method and system for coastal natural targets based on edge information depth supervision. Background Technology

[0002] Currently, remote sensing technology has become a core tool for large-scale coastal feature identification and information extraction, and deep learning-based semantic segmentation methods have been widely applied in areas such as wetland vegetation and other coastal target identification. However, existing semantic segmentation methods still have technical shortcomings for segmenting coastal salt marsh vegetation. Most existing methods rely on single-pixel classification constraints, and the loss functions they employ only focus on overall global classification accuracy, lacking specific constraints and targeted supervision mechanisms for vegetation edge features and small, fragmented patches. Traditional segmentation models generally emphasize global pixel classification performance, neglecting feature mining of small, fragmented patches and edge details in salt marsh vegetation, making it difficult to accurately capture target contours and boundary features.

[0003] Crucially, existing segmentation loss functions lack specific supervision strategies designed for the weak boundaries and fragmented distribution of natural targets in coastal zones. This makes it difficult to effectively guide the network to focus on edge details and small-scale patch features, and hinders overcoming interference from complex coastal backgrounds. Conventional networks rely solely on the backbone network to autonomously extract features, lacking targeted deep constraint mechanisms for edge information. This fails to address the root causes of vegetation boundary segmentation distortion and missed small patches, resulting in limited segmentation accuracy and stability. Consequently, they cannot meet the practical operational needs of routine coastal ecological monitoring and precise vegetation resource surveys, and also limit their application in coastal seaward segmentation scenarios involving fragmented, weakly boundary targets, such as green tides. Summary of the Invention

[0004] To address the shortcomings of existing coastal target remote sensing segmentation methods, such as the lack of specific depth supervision of natural target edge features, which leads to blurred edges and inaccurate boundary positioning of salt marsh vegetation patches, easy confusion with surrounding tidal flats and herbaceous vegetation, high false negative rates and insufficient identification accuracy for fragmented small-area salt marsh vegetation patches, this paper proposes a coastal target remote sensing segmentation method that enhances edge feature learning and improves the accuracy and integrity of salt marsh vegetation patch segmentation through depth supervision of edge information. This method aims to meet the practical technical needs of coastal wetland ecological monitoring, resource surveys, and ecological restoration assessments. The technical solution is as follows: A remote sensing segmentation method for coastal zone natural targets with edge information depth supervision includes the following steps: S1. Remote sensing image preprocessing; Coastal remote sensing image preprocessing includes normalization and image enhancement for water body mottling, weak boundaries of tidal flats, and differences in land feature scale. S2. Image segmentation dataset generation; A dedicated coastal zone segmentation dataset was constructed, and a hierarchical annotation specification was designed for elongated and blurred boundary targets such as salt marshes, tidal channels, and seaweed. Annotation verification, image tiling, and ground value storage were completed. S3. Using the U-Net network as the core feature extraction architecture, a multi-path edge supervision structure was designed and added to the internal network architecture of U-Net, including the encoder, decoder, and skip connections, to construct a complete edge information deep-supervised segmentation network model. Using U-Net as the basic feature extraction architecture, a coupled network connecting the encoder, decoder, skip connections, and multi-path edge supervision structure was built. Independent feature supervision units were configured in each layer of the decoder, and each decoding layer could calculate a layer-specific loss function based on its individual prediction results. In addition to the inverse constraint of the corresponding layer features by the independent loss of each layer during the model training phase, the prediction results of all decoding layers were also weighted and fused to calculate the global fusion loss, jointly completing the overall network optimization and constructing a remote sensing segmentation network model adapted to the weak boundaries of the coastal zone.

[0005] Preferably, in step S2, the image data is sliced ​​uniformly at a fixed size, and invalid slices with no effective targets or excessive background ratio are removed to obtain image slice data; in step S2, the preprocessed coastal remote sensing image is sliced ​​uniformly at a fixed size, and a background ratio threshold is set to filter and remove invalid slices with no land features or water background ratio exceeding the threshold, while retaining slice samples containing effective coastal land features such as salt marshes, tidal channels, and seaweed. Based on the coastal zone land cover category system, hierarchical labeling rules were developed, and target categories were labeled according to the actual land cover types and the preset rules. In a binary classification scenario, there are only two types: pixel label 0 represents the background, and pixel label 1 represents the natural target to be segmented. If there are multiple types, then: 0 represents the background. The representative category is used to distinguish various types of coastal target features. The image segmentation label map is obtained through visual interpretation and annotation. In multi-class scenarios, if there are more than two types, then: pixel label 0 is the general background, and label i takes the values ​​1, 2, 3, etc., which correspond to different types of coastal natural features. Based on visual interpretation and prior knowledge from the field, fine pixel-level annotation is completed to obtain the ground truth label. After all annotations are completed, the annotation results are verified to remove erroneous and blurry annotations, and then matched and integrated with the corresponding image slice data. Finally, a high-quality coastal zone target segmentation dataset is obtained, which is divided into training set, validation set and test set according to the proportion.

[0006] The integrated complete sample set is randomly divided into layers according to a preset ratio to generate training set, validation set and test set respectively, and a coastal target segmentation dataset adapted to the training of multi-path edge deep supervision network is constructed.

[0007] Preferably, at the encoding end: a total of 5 sets of convolutional modules are set, each layer consists of 3×3 Conv2d convolution, BN batch normalization, and ReLU activation function. The convolution stride is set to 2 to complete feature map downsampling and simultaneously extract global semantic information of the target. Multi-scale coastal land cover features from fine texture to global semantics are extracted layer by layer; the multi-scale features output by each layer will be independently linked to the corresponding decoding layer.

[0008] Preferably, the decoding end has four layers of transposed convolutional upsampling modules to gradually restore the spatial resolution of the feature map and restore the spatial size of the feature map layer by layer to restore the details of the ground features; the output features of each layer of transposed convolutional upsampling are fused with the corresponding layer features of the encoding end through skip connections to complement shallow edge information and deep semantic information, thereby alleviating the problems of coastal zone boundary fracture and blurred contours.

[0009] The preferred multi-path edge supervision structure: For the output features of each decoding layer, the segmentation result is predicted independently, and a loss function is calculated once with the segmentation label. Then, the loss functions of multiple decoding layer paths are combined as the network's loss function. Multi-loss weighted combination training is performed in each model iteration to optimize the overall network. A dedicated edge prediction unit is built for each layer of fused features at the decoding end. Each supervision unit independently predicts the coastal land cover segmentation result at the corresponding scale. The prediction result of each layer is scaled and calculated separately with the ground truth label to solve for the hierarchical independent loss function, forming a hierarchical multi-scale prediction bias. The multi-scale prediction biases of all decoding layer supervision units are aggregated and the total loss of the entire layer is calculated. During the training iteration, the independent losses of each decoding layer are weighted and combined into a global loss, forming an edge loss supervision mechanism that takes into account both hierarchical fine constraints and global contour calibration, and collaboratively completes the overall parameter optimization of the network.

[0010] The preferred optimization steps for the multi-path edge supervision structure are as follows: The first supervisory unit at the decoding end applies an activation function to the output features, and then calculates a hierarchical loss function that fuses Dice Loss and Focal Loss. ; The Dice Loss of the first layer at the decoding end. The Focal Loss is the first layer loss at the decoding end. The second-layer supervision unit at the decoding end applies an activation function to the output features and then calculates a hierarchical loss function that combines Dice Loss and Focal Loss. ; The Dice Loss is the second layer loss at the decoding end. This refers to the Focal Loss loss of the second layer at the decoding end. The third-layer supervision unit at the decoding end introduces an edge ground truth generation module based on the 3×3 Laplacian operator. This module extracts edge responses by performing Laplacian convolution on the original segmentation labels, normalizes them after taking their absolute values, sets a (0,1) threshold for binarization to generate edge supervision labels, applies an activation function to the output features, and then calculates a hierarchical loss function that fuses BCELoss and Dice Loss. ; The Dice Loss of the third layer at the decoding end; weight coefficients ; The fourth supervisory unit at the decoding end predicts temporary segmentation results based on the output features and calculates the hierarchical loss function that combines Dice Loss and Focal Loss. ; The Dice Loss is the loss of the fourth layer at the decoding end. This is the Focal Loss loss of the fourth layer at the decoding end.

[0011] Preferably, the overall network joint loss function is a weighted fusion of the layer losses corresponding to the four decoding supervision units, as shown in the following formula: ; in These are the independent weighting coefficients corresponding to the loss of the four-layer supervision unit; Used to constrain the multi-scale feature representation in the middle layer of the network, mine the shallow texture details of coastal land features, and improve the generalization ability of the model; It is used to optimize the boundary prediction effect, compensate for the defects of the edge fracture of the narrow and weak contrast of the coastal zone, and ensure the integrity and continuity of the outline. It dominates the overall land cover segmentation and alleviates the segmentation offset problem caused by the uneven distribution of coastal zone samples.

[0012] Preferably, in terms of optimizer settings, the Adam optimizer is used, with a base learning rate and weight decay coefficient set, combined with a learning rate cosine annealing scheduling strategy to achieve efficient optimization of model parameters and avoid overfitting; In terms of training iterations, it is divided into training and validation phases. After each iteration, the training loss and validation loss are calculated, model parameters are adjusted, and the optimal model is saved. During training, the training loss and validation loss are monitored in real time. If the training loss continues to decrease but the validation loss increases, it indicates that the model is overfitting, and the weight decay coefficient can be increased or data augmentation strategies can be added. If the training loss decreases slowly, the learning rate can be appropriately increased. If gradient explosion occurs, the learning rate can be decreased and gradient clipping can be increased. After each training round, the training log is saved to facilitate subsequent analysis of the model training effect and adjustment of hyperparameters. After training, the optimal model is saved for subsequent testing and practical applications.

[0013] A remote sensing segmentation system for coastal natural targets with edge information depth supervision includes a remote sensing image preprocessing module, an image segmentation dataset generation module, a prediction module, and an output module; Remote sensing image preprocessing module: Preprocesses remote sensing image data to obtain corrected images. Based on the vector boundary file of the target coastal zone, the image is cropped to remove irrelevant areas; after preprocessing, the image pixel values ​​are mapped to the [0,1] interval to obtain a normalized image. Image segmentation dataset generation module: The image data is sliced ​​uniformly at a fixed size, and invalid slices with no effective targets or excessive background are removed to obtain image slice data; the target categories are labeled according to the actual land cover types and preset rules; and the image segmentation label map is obtained through visual interpretation.

[0014] After the annotation is completed, the annotation results are verified, erroneous and blurry annotations are removed, and the data is integrated with the corresponding image tile data to form a dataset, finally obtaining a high-quality coastal zone target segmentation dataset; it is divided into training set, validation set and test set according to the proportion; Prediction Module: Based on the baseline model U-Net, an improved prediction model is constructed. A new multi-path edge supervision structure is designed on the basis of the encoder and decoder. The encoder includes 5 sets of convolutional modules and 4 pooling operations to complete feature downsampling. The decoder consists of 4 layers of transposed convolutional modules, which fuse deep and shallow features through skip connections. The classifier connected to the last layer of the decoder generates the backbone of the final segmentation result. The multi-path edge supervision structure is designed in each layer of the decoder to participate in the optimization of the model segmentation result. Output module: Outputs the prediction results.

[0015] Preferably, the model initialization uses He normal distribution to initialize the convolutional layer weights and constants to initialize the batch normalization layer parameters.

[0016] Compared with the prior art, the beneficial effects of this application are as follows: This technical solution addresses core issues in traditional segmentation methods, such as blurred edges, missed detection of small targets, and class imbalance, by designing a multi-path edge supervision structure. Compared with existing technologies, it has the following advantages: (1) In terms of edge recognition, the edge features of the target are accurately extracted by combining the edge supervision label generation module and the multi-path loss function, which solves the problem of blurred and difficult-to-define edges of natural targets such as salt marsh vegetation, and improves the integrity of edge contours; (2) In terms of small target recognition, the application of Focal Loss and weighted BCE Loss enhances the recognition of small patches and fine edges, and reduces the false detection rate of small targets; (3) In terms of stability, the effects of seawater background and lighting changes were effectively suppressed through image preprocessing and attention mechanism, and the segmentation accuracy remained stable in complex environments; (4) In terms of versatility, the scheme is not only applicable to the segmentation of multiple targets such as coastal salt marsh vegetation, tidal flats, and aquaculture areas, but also suitable for the segmentation of single-type targets such as floating green tides during nearshore outbreaks, and has a wide range of application scenarios.

[0017] The implementation of this technical solution can be completed using a Python programming environment. The operation process is standardized, which facilitates engineering implementation and promotion. It can effectively meet the practical needs of coastal ecological monitoring, invasive species control, and disaster early warning. Attached Figure Description

[0018] Figure 1 For technical flowcharts; Figure 2 To build a multi-path edge supervision structure for the model; Figure 3 Compare the segmentation results of the model; Figure 4 Comparison of the segmentation results from the model. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] The remote sensing segmentation method for coastal targets with edge information depth supervision of the present invention is divided into six stages: remote sensing image preprocessing, image segmentation dataset generation, model building and initialization, model training and parameter optimization, model testing and performance verification, and result analysis and optimization. The operation details, equipment operation, and parameter settings of each stage are described in detail below.

[0021] S1. Remote sensing image preprocessing: Download clear, cloud-free, and noise-free remote sensing imagery. Perform preprocessing steps such as radiometric calibration and atmospheric correction to obtain corrected images. Based on the vector boundary file of the target coastal zone, crop the imagery to remove irrelevant areas. The core purpose of this preprocessing is to eliminate noise interference and spectral distortion in the imagery, ensuring uniform image specifications and providing standardized image data for subsequent feature extraction and model training.

[0022] After image preprocessing, the image pixel values ​​are mapped to the [0,1] interval to obtain a normalized image. Image normalization can eliminate pixel value differences between different bands and different images, ensuring numerical stability during model training. The calculation formula is as follows: Where x is the original pixel value, , These are the minimum and maximum pixel values ​​for that band, respectively. After normalization, check the range of image pixel values ​​to ensure that all pixel values ​​are within the [0,1] interval. If there are outliers, normalization needs to be performed again.

[0023] S2. Image segmentation dataset generation: The image data is sliced ​​uniformly at a fixed size of 128*128. Invalid slices with no effective target or excessive background are removed to obtain image slice data.

[0024] Based on the actual land cover type, the target category is labeled according to the preset rules. (1) If there are only two types, then: 0 represents the background and 1 represents the target; (2) If there are multiple types, then: 0 represents the background, 1 represents category A, 2 represents category B, 3 represents category C, 4 represents category D, and so on, in order to distinguish multiple coastal target land cover types. The image segmentation label map is obtained by visual interpretation.

[0025] After annotation, the results are manually verified to remove erroneous and blurry annotations. These are then integrated with the corresponding image tile data to form a dataset, ultimately yielding a high-quality coastal zone target segmentation dataset. This dataset is then divided proportionally into training, validation, and test sets.

[0026] S3. Model Building and Initialization: 1. Model Building This model uses the U-Net network as the core feature extraction architecture. Based on the internal network architecture of U-Net, including the encoder, decoder, and skip connections, a multi-path edge supervision structure is added to construct a complete edge information deep supervision segmentation network model. The specific structure is as follows: Encoding end: A total of 5 convolutional modules are set up. Each layer consists of 3×3 Conv2d convolution, BN batch normalization, and ReLU activation function. The convolution stride is set to 2 to complete feature map downsampling and simultaneously extract global semantic information of the target.

[0027] Decoding end: A total of 4 transposed convolution upsampling modules are set up to gradually restore the spatial resolution of the feature map and restore the spatial details of the target. After each transposed convolution, the features of the corresponding layer in the encoding end are skipped and connected to fuse shallow edge features and deep semantic features to improve edge localization accuracy.

[0028] Multi-path edge supervision structure: For the output features of each decoding layer, the Sigmoid activation function is used to independently predict the segmentation result, and a loss function is calculated once with the segmentation label. Then, the loss functions of multiple decoding layers are combined as the network's loss function. Multi-loss weighted combination training is performed in each model iteration to optimize the overall network. The specific process is as follows: (1) First layer of decoding end: Apply activation function to output features, and then calculate loss function that combines Dice Loss and Focal Loss. ; The Dice Loss of the first layer at the decoding end. This is the Focal Loss loss of the first layer at the decoding end.

[0029] (2) In the second layer of the decoding end, an activation function is applied to the output features, and then the loss function that combines Dice Loss and Focal Loss is calculated. ; The Dice Loss is the second layer loss at the decoding end. This is the Focal Loss loss of the second layer at the decoding end.

[0030] (3) Third layer of the decoding end: To enhance the model's edge perception capability, an edge ground value generation module based on the 3×3 Laplacian operator is introduced. This module performs Laplacian convolution on the original segmentation labels to extract edge responses, takes the absolute value and normalizes it, sets a threshold of 0.1 for binarization to generate edge supervision labels (0 for non-edge, 1 for edge). An activation function is applied to the output features, and then the loss function that combines BCE Loss and Dice Loss is calculated. Weighting coefficient ; This is the Dice Loss of the third layer at the decoding end.

[0031] (4) Fourth layer of the decoding end: For the output features, predict the temporary segmentation result and calculate the loss function that combines Dice Loss and Focal Loss. ; The Dice Loss is the loss of the fourth layer at the decoding end. This is the Focal Loss loss of the fourth layer at the decoding end.

[0032] (5) The loss function for the joint multi-path network is calculated by weighting the loss functions of the above multi-path networks. The formula is as follows: ,in These are the weighting coefficients.

[0033] 2. Model initialization: Initialize the network parameters to ensure rapid convergence during model training and avoid gradient vanishing or exploding. Here, the weights of the convolutional and transposed convolutional layers are initialized using a He normal distribution, and the parameters of the batch normalization layer are initialized using constants. Input a set of training samples and check if the model output is normal to ensure the model is correctly built.

[0034] S4. Model Training and Parameter Optimization: 1. Model Training Through iterative optimization, the model learns the features and edge information of coastal targets, minimizing the multi-path joint loss function of the network model. Specific steps include data loading, loss function setting, optimizer setting, and training iterations. The software environment is based on Windows 10 Professional, paired with a Python 3.9.19 programming environment, using the PyTorch 2.0.1 deep learning framework (configured with CUDA 11.8 acceleration).

[0035] For data loading, Python programming was used to read remote sensing dataset images. The number of input channels was set to 4, corresponding to the four bands of the remote sensing image: blue, green, red, and near-infrared.

[0036] Regarding the loss function setting, the multi-path joint loss function designed in this invention is adopted. The loss is calculated separately from the output features of the multi-layer decoder, and the weight coefficients are fused to achieve synergistic optimization of image segmentation accuracy and edge integrity. The loss function formula is as follows: ,in, These are the loss functions for each decoding layer. Used to optimize intermediate layer features of the model and improve the model's generalization ability; Used to optimize edge segmentation accuracy and improve edge contour integrity; Used to optimize the accuracy of main body segmentation and solve the problem of class imbalance; For the weighting coefficients, experimental verification using examples shows that the optimal value is... , , , To ensure that the losses in each part are coordinated.

[0037] In terms of optimizer settings, the Adam optimizer is used, with the base learning rate set to 0.001 and the weight decay coefficient set to 0.00001. Combined with the learning rate cosine annealing scheduling strategy, efficient optimization of model parameters is achieved, avoiding overfitting.

[0038] In terms of training iterations, it is divided into training and validation phases. After each iteration, the training loss and validation loss are calculated, model parameters are adjusted, and the optimal model is saved. During training, the training loss and validation loss are monitored in real time. If the training loss continues to decrease but the validation loss increases, it indicates that the model is overfitting, and the weight decay coefficient can be increased or data augmentation strategies can be added. If the training loss decreases slowly, the learning rate can be appropriately increased. If gradient explosion occurs, the learning rate can be decreased and gradient clipping can be added. After each training round, the training log is saved to facilitate subsequent analysis of the model training effect and adjustment of hyperparameters. After training, the optimal model is saved in .pth format for subsequent testing and practical applications.

[0039] S5. Model Testing and Performance Verification: The core of model testing is to use a test dataset to objectively evaluate the model's performance metrics such as segmentation accuracy and edge integrity, and compare it with existing mainstream methods to verify the superiority of the method of this invention. The specific operation steps are as follows: (1) Test data is used as network input. The test dataset is loaded in the same way as the training and validation sets, without the need for data augmentation, thus ensuring the objectivity of the test results.

[0040] (2) The optimal model is used as the image segmentation model. The optimal model saved during training is loaded.

[0041] (3) Calculate the segmentation metrics for the test data. Image segmentation experiments were conducted using the test dataset. Segmentation performance metrics mIoU, OA, and F1-sore were calculated, and the segmentation results were saved for visualization analysis. The calculation formulas for the metrics are as follows: TP represents true positives, FP represents false positives, and FN represents false negatives.

[0042] , K This represents the total number of categories.

[0043] .

[0044] S6. Results Analysis and Optimization: By comprehensively analyzing the model test results, problems in the segmentation process are identified, and targeted optimization strategies are proposed to further improve the segmentation accuracy and edge integrity of coastal zone targets, ensuring the practicality and superiority of the technical solution of this invention. The specific operation steps are as follows: (1) Visual analysis of segmentation results: The original remote sensing image, the actual segmentation label, and the segmentation result predicted by the model are compared and visualized on the same screen to intuitively observe the segmentation effect. The analysis focuses on the integrity of the target edge, the segmentation effect of small targets, and the accuracy of category differentiation.

[0045] For typical coastal scenarios (such as the junction of salt marsh vegetation and seawater, small patch areas, and small target areas), select local magnified images, compare and analyze the differences between the predicted results and the actual labels, and identify the areas with concentrated segmentation errors, such as areas with blurred edges, areas where small targets are missed, and areas where the background is misjudged.

[0046] (2) Quantitative analysis of performance indicators: On the test set, mean Intersection over Union (mIoU), overall accuracy (OA), and F1 score were selected as quantitative evaluation indicators to comprehensively verify the semantic segmentation effect. IoU measures the spatial overlap between the predicted region and the real ground object region, OA represents the overall classification accuracy of all pixels in the image, and the F1 score comprehensively reflects the overall segmentation effect of the model.

[0047] Compared with the baseline model U-Net, using the same test dataset and evaluation metrics, the performance differences of each method are quantitatively compared to verify the superiority of the method of the present invention in coastal zone target segmentation (especially edge accuracy and small target segmentation).

[0048] (3) Results analysis and model validation: The segmentation performance metrics of different coastal targets (salt marsh vegetation, green tides) were statistically analyzed to assess the adaptability of the proposed method to different target types. For target types with poor performance, the reasons for poor performance were identified and analyzed. The model was retrained by adjusting model parameters, loss function weights, and data processing methods to obtain an optimized model. The optimized model was then validated using the same test dataset. Performance metrics and visualization results before and after optimization were compared to confirm the optimization effect, ensuring that the optimized model significantly improved segmentation accuracy and edge integrity.

[0049] Repeat the above analysis and optimization process until the model segmentation performance meets the actual application requirements of coastal target remote sensing segmentation. Finally, determine the optimal model parameters and implementation process to form a feasible and reusable coastal target remote sensing segmentation scheme.

[0050] Example 1: Target extraction from coastal salt marshes: The experimental area was selected from the coastal wetlands of Jiaozhou Bay in Qingdao, which includes salt marsh vegetation such as Spartina alterniflora, reeds, and Suaeda salsa. The background is a mixture of bare mudflats, waterlogged depressions, and silt. The vegetation is distributed in strips and patches, and the spectrum is highly confused with the silt and shallow water, which is a typical small target and complex mudflat segmentation scene.

[0051] The remote sensing image data is GF-2 PMS multispectral imagery with a spatial resolution of 3.4 meters and 4 channels (including blue, green, red, and near-infrared channels). Various land cover types were manually interpreted and labeled: 0 background, 1 aquaculture pond, 2 bare tidal flats, 3 salt marsh vegetation, 4 seawater, 5 rivers, and 6 roads. The sample dataset is divided into a training set of 1058 images, a validation set of 176 images, and a test set of 176 images. The input sample size is 128*128 pixels, with 4 input channels, and the data is normalized without random augmentation. The output categories are 7.

[0052] Through model training and hyperparameter tuning, the final configuration was as follows: Adam optimizer, learning rate 0.001, weight decay 0.00001, total training iterations 100, batch size 16, loss function Focal hyperparameters alpha=0.5, gamma=2. Taking into account quantitative metrics such as validation accuracy OA, mIoU, and F1 score, the optimal model was saved.

[0053] Taking into account quantitative indicators such as validation accuracy (OA), mIoU, and F1 score, the optimal model is saved.

[0054] After training, the optimal model in this embodiment achieved an OA of 97.12%, mIoU of 0.9291, and F1 score of 0.9629. Compared to the baseline model U-Net, this model showed improved accuracy in all aspects (see Table 1). Figure 3 The comparison of the model segmentation results shows that misclassification of aquaculture ponds, rivers, roads, etc. has been significantly reduced, the integrity of strip-shaped and fragmented salt marsh vegetation extraction has been improved, effectively alleviating the misclassification problem caused by class imbalance, and the accuracy of each category has been significantly improved.

[0055] Table 1 Comparison of Model Test Accuracy

[0056] Example 2: Target identification of green tides during nearshore outbreaks: The experimental area was selected in the Yellow Sea near Jiangsu Province, where green tides are common. The sea surface is characterized by wave flares, turbid water, floating foam, and the green tides appear in clumps and strips. There are large areas of mixed pixels, the background spectrum of the water body changes drastically, and the sea surface has strong dynamic noise.

[0057] The remote sensing image data is GF-2 PMS multispectral imagery with a spatial resolution of 3.4 meters and 4 channels (including blue, green, red, and near-infrared channels). Binary masks were manually interpreted and annotated, with pixel 0 representing pure seawater background and pixel 1 representing floating green tide. The sample dataset is divided into a training set of 1114 images, a validation set of 669 images, and a test set of 446 images. The input sample size is 128*128 pixels, with 4 input channels, and the data is normalized without random augmentation. The output class is 2.

[0058] Through model training and hyperparameter tuning, the final configuration was as follows: Adam optimizer, learning rate 0.001, weight decay 0.00001, total training iterations 100, batch size 16, loss function Focal hyperparameters alpha=0.5, gamma=2. Taking into account quantitative metrics such as validation accuracy OA, mIoU, and F1 score, the optimal model was saved.

[0059] After training, the optimal model in this example achieved an OA of 98.75%, mIoU of 0.9422, and F1 score of 0.9702. Compared to the baseline model U-Net, this model shows improvements in all accuracy metrics. During green tide outbreaks, large areas of seawater can easily be confused with the irregularly floating edges of the green tide, making accurate identification difficult. From... Figure 3 The comparison of the model segmentation results shows that this model can effectively identify large-scale contiguous, strip-shaped, and scattered small patches of green tides. Compared with the baseline model, this model enhances the discriminative power of green tide block contours through a multi-path edge supervision structure, reducing the misclassification rate of targets, indicating that the model can effectively suppress the dominance of seawater background in training. Therefore, this model provides high-precision spatial distribution data support for marine environmental early warning and green tide disaster response.

[0060] Table 2 Comparison of Model Test Accuracy

[0061] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A remote sensing segmentation method for coastal zone natural targets with edge information depth supervision, characterized in that, Includes the following steps: S1. Remote sensing image preprocessing; Coastal remote sensing image preprocessing includes normalization and image enhancement for water body mottling, weak boundaries of tidal flats, and differences in land feature scale. S2. Image segmentation dataset generation; A dedicated coastal zone segmentation dataset was constructed, and a hierarchical annotation specification was designed for elongated and blurred boundary targets such as salt marshes, tidal channels, and seaweed. Annotation verification, image tiling, and ground value storage were completed. S3. Using the U-Net network as the core feature extraction architecture, a multi-path edge supervision structure was designed and added to the internal network architecture of U-Net, including the encoder, decoder, and skip connections, to construct a complete edge information deep-supervised segmentation network model. Using U-Net as the basic feature extraction architecture, a coupled network connecting the encoder, decoder, skip connections, and multi-path edge supervision structure was built. Independent feature supervision units were configured in each layer of the decoder, and each decoding layer could calculate a layer-specific loss function based on its individual prediction results. In addition to the inverse constraint of the corresponding layer features by the independent loss of each layer during the model training phase, the prediction results of all decoding layers were also weighted and fused to calculate the global fusion loss, jointly completing the overall network optimization and constructing a remote sensing segmentation network model adapted to the weak boundaries of the coastal zone.

2. The remote sensing segmentation method for coastal zone natural targets with edge information depth supervision according to claim 1, characterized in that, Step S2: The image data is sliced ​​uniformly at a fixed size. Invalid slices with no effective targets or excessive background ratio are removed to obtain image slice data. Step S2: The preprocessed coastal remote sensing image is sliced ​​uniformly at a fixed size. A background ratio threshold is set to filter and remove invalid slices with no land features or water background ratio exceeding the threshold. Slice samples containing effective coastal land features such as salt marshes, tidal channels, and seaweed are retained. Based on the coastal zone land cover category system, hierarchical labeling rules were developed, and target categories were labeled according to the actual land cover types and the preset rules. In a binary classification scenario, there are only two types: pixel label 0 represents the background, and pixel label 1 represents the natural target to be segmented. If there are multiple types, then: 0 represents the background. The representative category is used to distinguish various types of coastal target features. The image segmentation label map is obtained through visual interpretation and annotation. In multi-class scenarios, if there are more than two types, then: pixel label 0 is the general background, and label i takes the values ​​1, 2, 3, etc., which correspond to different types of coastal natural features. Based on visual interpretation and prior knowledge from the field, fine pixel-level annotation is completed to obtain the ground truth label. After all annotations are completed, the annotation results are verified to remove erroneous and blurry annotations, and then matched and integrated with the corresponding image tile data; finally, a high-quality coastal zone target segmentation dataset is obtained, which is divided into training set, validation set and test set according to the proportion; The integrated complete sample set is randomly divided into layers according to a preset ratio to generate training set, validation set and test set respectively, and a coastal target segmentation dataset adapted to the training of multi-path edge deep supervision network is constructed.

3. The remote sensing segmentation method for coastal zone natural targets with edge information depth supervision according to claim 1, characterized in that, Encoding end: A total of 5 convolutional modules are set up. Each layer consists of 3×3 Conv2d convolution, BN batch normalization, and ReLU activation function. The convolution stride is set to 2 to complete feature map downsampling and simultaneously extract global semantic information of the target. Multi-scale coastal land cover features from fine texture to global semantics are extracted layer by layer. The multi-scale features output from each level will be independently linked to the corresponding decoding layer.

4. The remote sensing segmentation method for coastal zone natural targets with edge information depth supervision according to claim 1, characterized in that, Decoding end: A total of 4 transposed convolutional upsampling modules are set up to gradually restore the spatial resolution of the feature map and restore the spatial size of the feature map layer by layer to restore the details of the ground features; the output features of each transposed convolutional upsampling layer are fused with the corresponding layer features of the encoding end through skip connections, complementing shallow edge information and deep semantic information, and alleviating the problems of coastal zone boundary fracture and blurred outline.

5. The remote sensing segmentation method for coastal zone natural targets with edge information depth supervision according to claim 1, characterized in that, Multi-path edge supervision structure: For the output features of each decoding layer, the segmentation result is predicted independently, and a loss function is calculated once with the segmentation label. Then, the loss functions of multiple decoding layers are combined as the network's loss function. Multi-loss weighted combination training is performed in each model iteration to optimize the overall network. A dedicated edge prediction unit is built for each layer of fused features at the decoding end. Each supervision unit independently predicts the coastal land cover segmentation result at the corresponding scale. The prediction result of each layer is scaled and calculated separately with the ground truth label to solve for the hierarchical independent loss function, forming a hierarchical multi-scale prediction bias. The multi-scale prediction biases of all decoding layer supervision units are aggregated and the total loss of the entire layer is calculated. During the training iteration, the independent losses of each decoding layer are weighted and combined into a global loss, forming an edge loss supervision mechanism that takes into account both hierarchical fine constraints and global contour calibration, and collaboratively completes the overall parameter optimization of the network.

6. The remote sensing segmentation method for coastal zone natural targets with edge information depth supervision according to claim 5, characterized in that, The optimization steps for the multi-path edge supervision structure are as follows: The first supervisory unit at the decoding end applies an activation function to the output features, and then calculates a hierarchical loss function that fuses Dice Loss and Focal Loss. ; The Dice Loss of the first layer at the decoding end. The Focal Loss is the first layer loss at the decoding end. The second-layer supervision unit at the decoding end applies an activation function to the output features and then calculates a hierarchical loss function that fuses Dice Loss and Focal Loss. ; The Dice Loss is the second layer loss at the decoding end. This refers to the Focal Loss loss of the second layer at the decoding end. The third-layer supervision unit at the decoding end introduces an edge ground truth generation module based on the 3×3 Laplacian operator. This module extracts edge responses by performing Laplacian convolution on the original segmentation labels, normalizes them after taking their absolute values, sets a (0,1) threshold for binarization to generate edge supervision labels, applies an activation function to the output features, and then calculates a hierarchical loss function that fuses BCE Loss and Dice Loss. ; The Dice Loss is the third layer loss at the decoding end. Weighting coefficient ; The fourth supervisory unit at the decoding end: For the output features, predicts the temporary segmentation result and calculates the hierarchical loss function that fuses Dice Loss and Focal Loss. ; The Dice Loss is the loss of the fourth layer at the decoding end. This is the Focal Loss loss of the fourth layer at the decoding end.

7. The remote sensing segmentation method for coastal zone natural targets with edge information depth supervision according to claim 6, characterized in that, The overall joint loss function of the network is a weighted fusion of the layer losses corresponding to the four decoding supervision units, and the formula is as follows: ； in These are the independent weighting coefficients corresponding to the loss of the four-layer supervision unit; Used to constrain the multi-scale feature representation in the middle layer of the network, mine the shallow texture details of coastal land features, and improve the generalization ability of the model; It is used to optimize the boundary prediction effect, compensate for the defects of the edge fracture of the narrow and weak contrast of the coastal zone, and ensure the integrity and continuity of the outline. It dominates the overall land cover segmentation and alleviates the segmentation offset problem caused by the uneven distribution of coastal zone samples.

8. The remote sensing segmentation method for coastal zone natural targets with edge information depth supervision according to claim 1, characterized in that, In terms of optimizer settings, the Adam optimizer is adopted, and the base learning rate and weight decay coefficient are set. Combined with the learning rate cosine annealing scheduling strategy, the model parameters are efficiently optimized to avoid overfitting. In terms of training iterations, it is divided into training and validation phases. After each iteration, the training loss and validation loss are calculated, model parameters are adjusted, and the optimal model is saved. During training, the training loss and validation loss are monitored in real time. If the training loss continues to decrease but the validation loss increases, it indicates that the model is overfitting, and the weight decay coefficient can be increased or data augmentation strategies can be added. If the training loss decreases slowly, the learning rate can be appropriately increased. If gradient explosion occurs, the learning rate can be decreased and gradient clipping can be increased. After each training round, the training log is saved to facilitate subsequent analysis of the model training effect and adjustment of hyperparameters. After training is complete, the optimal model is saved for subsequent testing and practical applications.

9. A remote sensing segmentation system for coastal natural targets with edge information depth supervision, employing the method described in any one of claims 1-8, characterized in that, It includes a remote sensing image preprocessing module, an image segmentation dataset generation module, a prediction module, and an output module; Remote sensing image preprocessing module: Preprocesses remote sensing image data to obtain corrected images. Based on the vector boundary file of the target coastal zone, the image is cropped to remove irrelevant areas; after preprocessing, the image pixel values ​​are mapped to the [0,1] interval to obtain a normalized image. Image segmentation dataset generation module: The image data is sliced ​​uniformly at a fixed size, and invalid slices with no effective targets or excessive background are removed to obtain image slice data; the target categories are labeled according to the actual land cover types and preset rules. Image segmentation label maps are obtained through visual interpretation and annotation. After the annotation is completed, the annotation results are verified, erroneous and blurry annotations are removed, and the data is integrated with the corresponding image tile data to form a dataset, finally obtaining a high-quality coastal zone target segmentation dataset; it is divided into training set, validation set and test set according to the proportion; Prediction Module: Based on the baseline model U-Net, an improved prediction model is constructed. A new multi-path edge supervision structure is designed on the basis of the encoder and decoder. The encoder includes 5 sets of convolutional modules and 4 pooling operations to complete feature downsampling. The decoder consists of 4 layers of transposed convolutional modules, which fuse deep and shallow features through skip connections. The classifier connected to the last layer of the decoder generates the backbone of the final segmentation result. The multi-path edge supervision structure is designed in each layer of the decoder to participate in the optimization of the model segmentation result. Output module: Outputs the prediction results.

10. The edge information depth-supervised remote sensing segmentation system for coastal zone natural targets according to claim 9, characterized in that, The model initialization uses He normal distribution to initialize the convolutional layer weights and constants to initialize the batch normalization layer parameters.

Images