A self-adaptive edge enhancement neural network remote sensing image fish pond extraction method

Abstract

The application discloses a kind of self-adapting edge enhancement neural network remote sensing image fish pond extraction method, comprising the following steps: selecting sentinel satellite multispectral band synthetic false color image;Fish pond area annotation dataset is constructed;Establish self-adapting edge enhancement neural network architecture, the architecture fuses the edge detection module of learnable threshold and semantic feature extraction;Integrating edge perception double attention module and pyramid pooling module, enhance multiscale feature representation ability;Multi-level deep supervision and prediction fusion strategy are used to optimize training process;Finally, fish pond area is extracted by training model and generates standardized result.The application significantly improves fish pond boundary segmentation precision, enhances the adaptability to different scale fish ponds, especially improves the detection rate of small scale fish ponds, and accelerates the model convergence process through deep supervision strategy, effectively reduces the misclassification of similar water bodies such as rivers, ditches and the like, solves the technical problems of existing technology, such as edge blur, poor multiscale adaptability, complex background interference sensitivity and high small scale fish pond detection rate.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of remote sensing image processing and computer vision, and particularly relates to a remote sensing image fish pond extraction method based on an adaptive edge enhancement network. BACKGROUND

[0002] Remote sensing image interpretation is an important means of geographic information acquisition, and its core lies in the coordinated use of the spectral and spatial features of ground objects. Traditional fish pond extraction methods mainly rely on manual field investigation, which has problems of long cycle and high cost, while methods based on remote sensing technology can achieve large-scale monitoring, but are limited by the lack of feature expression and spatial modeling capability. In the prior art, although the pixel-based spectral segmentation method of Normalized Difference Water Index (NDWI) can quickly distinguish water body regions, it is difficult to capture the regular geometric features of fish ponds and is easily affected by adjacent ground objects, resulting in boundary adhesion. Although traditional machine learning models (such as random forests) can improve accuracy by manually designing morphological features, their generalization ability is limited in complex scenarios.

[0003] In recent years, deep learning methods such as full convolutional neural networks (FCN) and U-Net have made significant progress in remote sensing image segmentation, but they still have technical limitations: first, the conventional convolution operation is insufficient for modeling long-range spatial dependencies, making it difficult to effectively associate the regularity features of fish pond row distribution, resulting in edge breakage or excessive smoothing; second, the problem of insufficient multi-scale feature fusion makes the model less adaptable to mixed distribution scenarios of dense small ponds and scattered large ponds, especially in complex aquaculture areas along the Chinese coast, where the misjudgment rate of interference ground objects such as rivers and ditches is high.

[0004] To address the limitations of traditional and deep learning methods in fishpond extraction, this invention proposes a fusion architecture based on parameterizable threshold adaptive edge detection and hierarchical spatial attention. This invention achieves a technological breakthrough through a learnable edge-enhanced U-Net architecture: by introducing a Canny detection module with learnable high and low threshold parameters and edge-enhancing convolution, the edge strength of the image is adaptively optimized, and dynamic feature fusion with the semantic features of the U-Net backbone is performed through a gated fusion mechanism, significantly improving the accuracy of fishpond boundary recognition; by combining the edge-aware dual attention module (EDAM) channel and spatial dual attention mechanism with Pyramid Scene Parsing (PSP), a multi-scale attention feature field is constructed, simultaneously enhancing the recognition ability of dense small ponds and isolated large ponds, greatly improving the recall rate of small-scale fishponds (<50㎡); based on a deep supervision strategy to optimize network training, the overall accuracy on Sentinel-2 multispectral data (blue, green, red, and near-infrared bands) is significantly improved compared to U-Net, successfully solving the core technical problems of blurred fishpond boundaries, scale mixing, and complex background interference in high-resolution images. Summary of the Invention

[0005] This invention addresses the technical problems of existing remote sensing image fishpond extraction methods, such as blurred edges, poor multi-scale adaptability, sensitivity to complex background interference, and high false negative rate for small-scale fishponds. It proposes a fishpond extraction method based on an adaptive edge enhancement neural network. By integrating a learnable edge detection module with a hierarchical spatial attention mechanism, combined with multi-scale feature field construction and deep supervision strategies, the method significantly improves the accuracy of fishpond boundary segmentation and robustness in complex scenes.

[0006] 1. To achieve the above-mentioned objective of this invention, the technical solution adopted is as follows: A method for extracting fishponds from remote sensing images based on an adaptive edge enhancement neural network is constructed, specifically including the following steps: S1: Multispectral band selection and image synthesis: Selecting blue bands with a spatial resolution of 10 meters from the Sentinel-2 satellite imagery (…). , 490 nm), green ( 560 nm), red ( 665 nm), near-infrared ( Four bands (842 nm, etc.) are used to combine the green, red, and near-infrared bands to create a false-color image. (Band order:) This combination enhances the spectral differences between water bodies and vegetation, and employs histogram equalization technology. Enhance image contrast to generate a three-channel basic image of the input data. S2: Fishpond Area Labeling and Dataset Construction: On the synthesized false-color image, polygonal labels are generated for the fishpond area based on manual visual interpretation, producing binary mask labels. (1 represents the type of fishpond) 0 represents non-fishpond areas. The data for the fishpond area was cropped into 256×256 pixel sample blocks. Construct a training dataset containing typical fishpond features. D The pixel values ​​are normalized to the [0,1] range; at the same time, data augmentation methods such as random rotation and flipping are applied. To address the imbalanced sample problem and ensure the model's balanced learning of the morphological characteristics of various fishponds.

[0007] S3: Construction of an Adaptive Edge Enhancement Neural Network Architecture Model: Building an Adaptive Edge Enhancement Network Based on the U-Net Architecture An edge detection module with a learnable threshold is introduced in the encoder stage. This enables adaptive extraction of edge features from the input image; the edge detection module employs learnable parameters. and Optimize the high and low thresholds of the Canny operator and combine them with edge-enhancing convolutional blocks. Enhance the expressive power of edge features, and transform the extracted edge features Through edge feature fusion module With semantic features Multi-level fusion is performed to enhance the network's ability to perceive target boundaries.

[0008] S4: Integration of Neural Network Attention Mechanisms: Introducing an Edge-Aware Dual Attention Module in the Network Encoding and Decoding Stages This enables adaptive enhancement of key features; a pyramid pooling module is integrated into the network bottleneck layer. Through multi-scale feature extraction Enhance the model's ability to represent targets of different sizes; the edge feature fusion module employs a gating mechanism. Through learnable weights By dynamically adjusting the influence of edge information on semantic features, the recognition accuracy of water body boundaries such as fishponds and small targets can be improved.

[0009] S5: Deeply Supervised Optimization Training Strategy. The network employs a multi-level deep supervision mechanism at different stages of the upsampling decoder. Set auxiliary monitoring signal These correspond to feature representations at different scales; learnable prediction fusion weight parameters are introduced. For the main output With in-depth supervision and forecasting at all levels Adaptive fusion is performed to form a final segmentation result; a deep supervision strategy is used to relieve gradient vanishing, accelerate model convergence, enhance the network's learning ability for different scale features, and improve overall segmentation accuracy.

[0010] S6: Based on S3-S5, a model based on an adaptive edge enhancement neural network architecture is constructed.

[0011] S7: The preprocessed remote sensing image dataset is input into the model for training: the BCE-Dice joint loss function is used, wherein the main branch and the auxiliary supervision branch are combined, and wherein is the loss weight, BCE is the Dice loss weight; an optimizer is used with an initial learning rate . Adam End-to-end training is performed, and the learning rate adopts a cosine annealing strategy , wherein in each training cycle , samples with a batch size B=16 are used for iterative optimization, and after training, the verification set accuracy with the highest weight parameters is saved.

[0012] S8: Obtaining the fish pond remote sensing extraction result: the preprocessed image to be extracted is input into the trained model , and a pixel-level fish pond probability map is output, which is processed by an adaptive threshold to generate a binary mask , and the final result is a standardized raster format.

[0013] In one embodiment, S3 further includes a specific implementation of the edge detection module, which can be expressed as:

[0014]

[0015] wherein represents a Canny edge detection operation with a learnable threshold parameter, and are the low threshold and high threshold parameters, respectively, which are optimized through backpropagation; the constraint mechanism of the learnable threshold is expressed as:

[0016]

[0017]

[0018] wherein , , , ​For the threshold interval parameter, the gradient calculation of edge detection adopts the Sobel operator, and the expression is:

[0019]

[0020] wherein and are the Sobel kernels in horizontal and vertical directions respectively, denotes the convolution operation; is an edge enhancement convolution block, which contains a cascaded convolution layer, and the expression is:

[0021]

[0022] wherein denotes a residual connection; the edge feature fusion module adopts a gated attention mechanism, and the expression is:

[0023]

[0024]

[0025] wherein is an attention gating unit, is a learnable fusion weight, denotes a feature concatenation operation, denotes an element-level multiplication.

[0026] In an embodiment, the channel and spatial dual attention mechanism in step S4 The expression can be expressed as:

[0027]

[0028] The calculation expression of the channel attention module is:

[0029]

[0030]

[0031] wherein the MLP structure is , is a reduction ratio; is an activation function, denotes an element-wise multiplication in the channel dimension.

[0032] The calculation expression of the edge perception module is:

[0033]

[0034]

[0035]

[0036] wherein represents edge feature, represents feature concatenation operation, represents edge weight matrix.

[0037] The spatial attention module calculates the expression as:

[0038]

[0039]

[0040] wherein represents channel dimension concatenation, is two-dimensional convolution.

[0041] The pyramid pooling module Through multi-scale feature extraction, the expression is:

[0042]

[0043]

[0044] wherein represents different proportion of pooling operation, and the corresponding output feature size proportion is .

[0045] In an embodiment, the deep supervision mechanism expression in step S5 can be expressed as:

[0046]

[0047] wherein represents the layer decoder feature map, is activation function; different levels of features can be expressed as:

[0048]

[0049] wherein represents up-sampling operation, represents the corresponding encoder stage of the skip connection feature; the final prediction fusion adopts adaptive weight combination, and the expression is:

[0050]

[0051] wherein the weight parameter satisfies the constraint condition , which is optimized by end-to-end training; the normalized expression of the prediction fusion is:

[0052]

[0053] wherein represents the main output, represents the prediction result of the auxiliary supervision branch.

[0054] In one embodiment, the training optimization expression of step S7 can be expressed as:

[0055]

[0056] wherein the joint loss function comprises the main branch loss and the auxiliary supervision branch loss, and the weight coefficients are , and ;

[0057] BCE The loss and Dice loss are defined as:

[0058]

[0059]

[0060] wherein is the total number of pixels, is the real label, is the predicted probability, is a smoothing factor for improving numerical stability;

[0061] The auxiliary supervision branch loss function is defined as:

[0062]

[0063] wherein = =0.5 is the weight coefficient, is the prediction of the layer auxiliary branch.

[0064] The Adam optimizer is used for parameter update, wherein =0.9, , is the hyperparameter of the optimizer, and the learning rate adopts the cosine annealing strategy:

[0065]

[0066] wherein is the initial learning rate, is the minimum learning rate, is the current iteration number, is the total number of iterations.

[0067] In an embodiment, the fish pond extraction result expression of step S8 can be expressed as:

[0068]

[0069] wherein represents a model trained, is the parameter with the optimal performance on the validation set; the generation expression of the binary mask is:

[0070]

[0071] wherein is an indicator function, is an optimal threshold value; the determination expression of the adaptive threshold value is:

[0072]

[0073]

[0074] wherein represents a balance factor of precision and recall;

[0075] The beneficial effects of the present application are as follows:

[0076] 1. By introducing the edge detection module with a learnable threshold value and the edge feature fusion mechanism, the segmentation accuracy of the fish pond boundary is significantly improved, effectively solving the technical problem of blurred water body boundary in complex remote sensing scenes, making the extracted fish pond boundary more accurate and continuous.

[0077] 2. The integrated edge perception dual attention module and the pyramid pooling module construct a multi-scale feature field, which enhances the adaptability of the model to fish ponds of different sizes, especially improves the detection recall rate of small-scale fish ponds, and overcomes the limitations of traditional methods in multi-scale target recognition.

[0078] 3. The multi-level deep supervision and prediction fusion strategy is adopted to accelerate the convergence process of the model, improve the training stability, and at the same time, through adaptive fusion of prediction results of different scales, the robustness of the network to complex background environment is enhanced, effectively reducing the misclassification of similar water bodies such as rivers and ditches, and improving the overall precision of fish pond extraction. BRIEF DESCRIPTION OF DRAWINGS

[0079] Figure 1 is the system architecture structure diagram of the remote sensing image fish pond extraction method of the adaptive edge enhancement neural network.

[0080] Figure 2 This is a structural diagram of an adaptive edge detector.

[0081] Figure 3 This is a structural diagram of the edge-aware dual-attention module.

[0082] Figure 4 This is a comparison chart of experimental results from embodiments of the present invention. Detailed Implementation

[0083] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0084] Example 1

[0085] This embodiment uses an adaptive edge enhancement neural network to extract fishponds from Sentinel-2 remote sensing images, achieving high-precision fishpond boundary recognition.

[0086] This embodiment selects Sentinel-2 imagery data from Jiujiang Town, Danzao Town, and Xiqiao Town in Nanhai District, Foshan City. This region has a suitable climate for aquaculture, with an average annual temperature of approximately 22°C, abundant rainfall, and distinct seasons, all conducive to fish growth. The towns' water sources are mainly surface water and groundwater, making this area an ideal fish farming site. Considering the potential impact of spatial resolution on water analysis, four bands with a spatial resolution of 10 meters—blue, green, red, and near-red—were used to improve analysis accuracy and reduce data bias that may be caused by resampling.

[0087] The experimental design employed cross-validation, conducting three rounds of experiments. This example illustrates the use of the dataset described in Table 1.

[0088] Table 1

[0089]

[0090] S1: Multispectral Band Selection and Image Combining: In this embodiment, multispectral images of the study area in Nanhai District, Foshan City, were acquired from the Sentinel-2 satellite. A spatial resolution of 10 meters for the blue band (…) was selected. , 490 nm), green ( 560 nm), red ( 665 nm), near-infrared ( Four bands (near-infrared, 842 nm, etc.) were used in the experiment. Synthetic false color images IThis combination visually presents the water body such as fish pond as deep blue or black, and the vegetation as red, enhancing the contrast of water-land boundary. Taking the central area of Jiujiang Town as an example, the difference between the fish pond and the surrounding objects is not obvious in the original RGB true color image, while after false color synthesis, the spectral characteristics of the fish pond are more prominent. Subsequently, the embodiment applies histogram equalization processing to the false color image , further improving the image contrast, making the fish pond boundary more clear, and generating the final three-channel basic input image .

[0091] S2: Sample labeling and dataset construction: Based on the false color synthesis image, 3,856 fish ponds (area 30-5,000 square meters) were labeled in Danzao Town, Jiujiang Town and Xiqiao Town of Foshan City. The labeling results were converted into binary mask labels (fish pond area is 1, background is 0), and the global image was regularly cropped according to 256x256 pixels. After removing the boundary incomplete samples, 5,420 effective sample blocks were obtained. All samples were normalized (pixel value mapped to [0,1] interval)

[0092] S3: The embodiment constructs an adaptive edge enhancement network based on the U-Net architecture , as shown in Figure 1 . In the encoder stage, the embodiment introduces an edge detection module with a learnable threshold , which extracts edge features by applying an improved Canny operator to the input image , where and are the low and high threshold parameters learned by the model, as shown in Figure 2 . In the initial state, is set to 0.1, is set to 0.3, and during the training process, these parameters will be continuously optimized through the back propagation algorithm. For example, on the image of the dense fish pond area in Xiqiao Town, the edges extracted by the traditional fixed threshold Canny operator have a lot of breaks and noise, while the learnable threshold module of the present method optimizes the parameters to after training, significantly improving the continuity and accuracy of the edges. In addition, the present method also integrates an edge enhancement convolution block , which uses a cascade of 3x3 convolution and batch normalization layers to enhance the expression ability of edge features . Finally, through the edge feature fusion module , the edge features are dynamically fused with the semantic features of each layer of the encoder , achieving accurate enhancement of the fish pond boundary.

[0093] S4: In the encoding and decoding stages of the network, the embodiment integrates an edge-aware dual attention module , as shown inFigure 3 The module includes two sub-modules of channel attention and spatial attention, which focus on the two dimensions of "what to focus on" and "where to focus on". In addition, the edge perception mechanism enhances the model's ability to accurately delineate the contours of fish ponds through edge feature extraction and adaptive weight allocation. Taking a complex area in Jiujiang Town as an example, the channel attention mechanism adaptively highlights the feature channels that are most helpful for fish pond identification, such as edge gray gradient and water texture features; the spatial attention mechanism emphasizes the spatial distribution features of the fish pond boundaries and internal regions. The synergy of channel attention, spatial attention, and edge perception enables the model to accurately distinguish fish ponds from surrounding farmland, roads, and other features. At the network bottleneck layer, the embodiment integrates a pyramid pooling module , which obtains multi-scale context information through different scale pooling operations (1x1, 2x2, 3x3, 6x6).

[0094] S5: To improve network training efficiency and segmentation accuracy, the embodiment adopts a multi-level deep supervision mechanism, as shown in Figure 3 At different stages of the decoder l ∈{1,2,3}, the auxiliary supervision signal generates a prediction output , where represents the feature map fused by upsampling and jump connection. The embodiment introduces a learnable prediction fusion weight parameter { } , which adaptively fuses the main output with the deep supervision prediction at each level, with the expression being The weight is normalized to ensure reasonable allocation. This mechanism significantly speeds up model convergence and improves the recognition ability of small-scale fish ponds.

[0095] S6: According to the design of S3-S5, the embodiment constructs a complete adaptive edge enhancement neural network model, as shown in Figure 1 The model integrates edge detection and enhancement modules, dual attention mechanisms, pyramid pooling modules, and deep supervision mechanisms to form an end-to-end fish pond extraction framework.

[0096] S7: The embodiment uses BCE-Dice a joint loss function for model training, where BCE the loss weight , the Dice loss weight , and the auxiliary supervision branch weight . The training process uses the Adam optimizer, with an initial learning rate of and a weight decay coefficient of . The learning rate strategy adopts a cosine annealing method, gradually reducing to . The batch size B = 16 is used for each iteration, and the average Intersection over Union (IoU) of the validation set is monitored during the training process.

[0097] S8: The pre-processed image to be extracted Input the trained model , and output the pixel-level fishpond probability map . To obtain the final binary mask result, this embodiment uses an adaptive threshold method to determine the optimal segmentation threshold . Specifically, on the validation set, different threshold values (step size 0.01) are tried, and the threshold corresponding to the highest F1 score is selected as the optimal threshold . In the Jiujiang Town test area, the optimal threshold is determined to be , and the binary mask is generated . Compared with the traditional U-Net, this method has achieved significant accuracy improvement in all three test areas, especially in the accuracy of fishpond boundary and the identification of small-scale fishponds.

[0098] Experimental results:

[0099] This embodiment uses a three-round cross-validation experimental design to fully verify the adaptability and generalization ability of the method (Adaptive Edge-detection U-Net, AEU-Net) algorithm in different geographical environments. In each round of experiment, data from two regions are selected for model training, and independent testing is performed in the third region to ensure that the test data and the training data are completely separated in terms of geographical distribution.

[0100] Table 2-5 respectively shows the experimental results of three rounds of cross-validation, in which AEU-Net is superior to SegNet, U-Net, Attention U-Net, U-Net++ and U-Net3+ in each round of experiment. In the first round of experiment (Table 2), the data of Danzhao Town and Xiqie Town were used for training, and Jiujiang Town was used as the test area. The F1-score of AEU-Net reached 0.903, and the IoU reached 0.823, which was significantly higher than other comparative methods. Similarly, in the second round (Table 3) and the third round (Table 4) of experiments, AEU-Net also performed well in the test of different geographical areas, especially in the test of Xiqie Town, the F1-score reached 0.918, and the IoU reached 0.843, which proved the excellent generalization ability of the algorithm. Table 5 further summarizes the average performance of the three rounds of experiments, and AEU-Net achieves the best results in all evaluation indicators, with an average F1-score of 0.905, an average IoU of 0.825, an overall accuracy (OA) of 0.968, and an MCC of 0.885, as shown in the segmentation result comparison chart of Figure 4 The segmentation results of the comparison chart show that the proposed method has significant anti-interference ability and effectively avoids the misclassification of water interference (rivers, ditches), verifying the stability and effectiveness of the algorithm in different geographical environments.

[0101] Table 2

[0102]

[0103] Table 3

[0104]

[0105] Table 4

[0106]

[0107] Table 5

[0108]

Claims

1.A method for fish pond extraction from remote sensing images using an adaptive edge-enhanced neural network, the method comprising: The method comprises the following steps: S1: select the blue, green, red, and near-infrared four bands with a spatial resolution of 10 meters in the Sentinel-2 satellite image, combine the green, red, and near-infrared bands into a false color image, enhance the contrast of the image through histogram equalization technology, and generate a three-channel basic image of the input data; S2: based on artificial visual interpretation, a polygon is labeled on the synthesized false color image to generate a binary mask label; the fishpond region data is cropped into a sample block of 256x256 pixels to construct a training data set containing typical fishpond features, and the pixel value is normalized to the interval [0, 1]; S3: based on the U-Net architecture, an adaptive edge enhancement network is constructed, and a learnable threshold edge detection module is introduced in the encoder stage to realize adaptive extraction of edge features of the input image; The edge detection module uses learnable parameters and optimizes the high and low thresholds of the Canny operator, combines edge enhancement convolution blocks to improve the expression ability of edge features, and performs multi-level fusion of the extracted edge features and semantic features through an edge feature fusion module; S4: an edge-aware dual attention module is introduced in the network encoding and decoding stage to realize adaptive enhancement of key features; a pyramid pooling module is integrated in the network bottleneck layer to enhance the representation ability of the model for targets of different sizes through multi-scale feature extraction; S5: the network adopts a multi-level deep supervision mechanism, sets auxiliary supervision signals at different stages of the up-sampling decoder, and respectively corresponds to different scale feature representations; a learnable prediction fusion weight parameter is introduced to adaptively fuse the main output and the deep supervision prediction at each level to form the final segmentation result; S7: input the preprocessed image to be extracted into the trained model, output a pixel-level fishpond probability map, and generate a binary mask after adaptive threshold processing, and the final result is a standardized raster format. S6: input the preprocessed remote sensing image dataset into the model for training, adopt the BCE-Dice combined loss function combined with the main branch and the auxiliary supervision branch, wherein α = 0.7 is the BCE loss weight, β=0.3 is the Dice loss weight; end-to-end training is performed by using the Adam optimizer with an initial learning rate, and the learning rate adopts the cosine annealing strategy; The edge enhancement convolution block comprises cascaded convolution layers and a residual connection structure. 2.The fish pond extraction method of a remote sensing image based on a self-adaptive edge enhancement neural network according to claim 1, wherein, The edge detection module in S3 comprises a combination of a Canny edge detection operation with learnable thresholds and an edge enhancement convolutional block, where the high and low threshold parameters The edge feature fusion module adopts a gated attention mechanism to dynamically adjust the influence of edge information on semantic features through learnable weights. L and The dual attention module in S4 comprises a cascade of channel attention modules and spatial attention modules, wherein the channel attention module is based on feature aggregation of average pooling and maximum pooling, and the spatial attention module is based on feature aggregation in the channel dimension. H are optimized by backpropagation, and The pyramid pooling module extracts multi-scale features through different scale pooling operations, and the corresponding output feature size ratio is {1, 1 / 2, 1 / 4, 1 / 8}. H satisfy a constraint mechanism that the gradient computation of the edge detection employs a Sobel operator. The deep supervision mechanism in S5 adds an auxiliary prediction branch to each decoder layer to generate a supervision signal through convolution operation and activation function; L ​ 3.The fish pond extraction method of a remote sensing image based on a self-adaptive edge enhancement neural network according to claim 1, characterized in that, The feature representations at different levels comprise up-sampling operations and encoder skip connections; and the final prediction fusion adopts adaptive weight combination, and the weight parameter is optimized through end-to-end training and satisfies the normalization constraint. ​ 4.The fish pond extraction method of remote sensing image based on self-adaptive edge enhancement neural network according to claim 1, wherein, ​ 5. The method of claim 1, wherein the method further comprises: ​ 6. The method of claim 1, wherein the method further comprises: ​ ​

Images