Coastal aquaculture pond extraction method based on DBCE-Net and Sentinel-2 images
Through the dual-branch structure and dynamic feature fusion module of the DBCE-Net framework, the problems of complex threshold setting, strong data dependence and insufficient generalization capabilities in coastal aquaculture pond extraction are solved, and efficient and accurate aquaculture pond extraction is achieved, improving the improvement of the aquaculture pond extraction. Convenient application and stability.
Patent Information
- Application Number
- CN202411159468.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-22
- Publication Date
- 2025-01-10
AI Technical Summary
The existing coastal aquaculture pond extraction methods have problems such as complex threshold setting, strong data dependence, difficulty in distinguishing similar water bodies, insufficient generalization ability and limited context information acquisition, resulting in low classification errors and application efficiency.
Using a two-branch structure deep learning framework based on DBCE-Net, combining the delicate features of high-resolution small images and the global features of downsampled large images, the dynamic feature fusion module reduces context information loss, and realizes accurate extraction of coastal aquaculture ponds. There is no need to set multiple parameters for different regions.
It significantly improves the accuracy and generalization ability of coastal aquaculture pond extraction, reduces classification errors, realizes efficient and automated aquaculture pond extraction, and reduces manpower and time costs.
Smart Images
Figure CN119295794A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of remote sensing image processing, and in particular to a Sentinel-2 coastal aquaculture pond extraction method based on an improved double-branch network structure. Background Art
[0002] As a key component of the global food supply, aquaculture plays an irreplaceable role in maintaining biodiversity, ecological balance and food safety. With the growth of the global population and the gradual depletion of marine resources, aquaculture plays an increasingly important role. According to the Food and Agriculture Organization of the United Nations (FAO), global fishery and aquaculture production is at an all-time high, and the industry will play an increasingly important role in providing food and nutrition in the future. China is already the world's largest aquaculture country, supplying more than 60% of aquaculture products worldwide. However, the rapid expansion of coastal pond aquaculture has also caused a series of environmental problems, such as disrupting the balance of ecosystems, damaging important ecosystems such as coastal wetlands, coral reefs and seagrass beds, and adversely affecting local biodiversity and habitats. The interaction between aquaculture ponds and the surrounding marine ecosystems, such as the discharge of aquaculture wastewater, may threaten the survival and reproduction of marine organisms, leading to problems such as marine pollution. ). With the continuous expansion of aquaculture scale, traditional manual on-site monitoring and statistical survey methods can no longer meet the needs of efficient and precise management, and there is an urgent need to develop efficient and reliable aquaculture monitoring methods.
[0003] In recent years, more and more researchers have used automatic extraction methods such as machine learning and deep learning to extract and map aquaculture ponds. In the application of methods for mapping coastal aquaculture ponds, random forest algorithms and decision tree models based on multiple features such as spectral indicators (such as normalized difference water index NDWI and modified normalized difference water index MNDWI), morphology and texture are widely used. For example, studies have used random forest algorithms to conduct extensive application research in China and other large-scale coastal areas using Sentinel-2 satellite images. However, this method requires artificially setting multiple different thresholds and parameters for different regions, multiple parameters need to be set across the country, and the effectiveness of this method depends on the support of other data sets. Secondly, the method combining edge detection and morphological processing can identify and analyze the edge and morphological features of the pond to extract the information of the aquaculture pond, but this method has certain limitations in distinguishing coastal aquaculture ponds from similar surface water bodies (such as salt pans). Because these water bodies have high similarities in size, appearance and spectrum, they may cause classification errors. Deep learning methods have performed well in the accurate identification and classification of aquaculture ponds due to their powerful feature extraction capabilities. However, existing deep learning methods still face the shortcomings of insufficient generalization ability and limited ability to obtain contextual information when applied on a large scale across the country. In view of the above problems, it is urgent to propose a method with strong generalization ability and high degree of automation to achieve accurate extraction of large-scale coastal aquaculture ponds.
[0004] Existing coastal aquaculture pond extraction methods, such as random forest algorithm combined with object-oriented method, edge detection and morphological processing, and U2-Net algorithm, have the following problems:
[0005] (1) Complex threshold setting: Multiple thresholds and parameters need to be set for different regions, which makes it difficult to apply nationwide.
[0006] (2) Strong data dependence: The effectiveness of the method depends on other data sets, and its accuracy decreases when relevant data is lacking.
[0007] (3) Difficulty in distinguishing similar water bodies: It is difficult to distinguish between aquaculture ponds and similar surface water bodies (such as salt pans), which leads to classification errors.
[0008] (4) Insufficient generalization ability: Deep learning methods perform very differently in different regions, making it difficult to ensure consistent high accuracy.
[0009] (5) Limited context information acquisition: Context information is easily lost in large-scale image processing, which affects the extraction of boundaries and details. Summary of the invention
[0010] The purpose of the present invention is to provide a coastal aquaculture pond extraction method based on DBCE-Net and Sentinel-2 images in view of the deficiencies in the prior art. The present invention proposes a DBCE-Net deep learning framework based on an improved dual-branch structure. On the one hand, considering the problem of missing context and boundary information of traditional single-branch structure networks in deep learning, DBCE-Net (Dual-Branch Context Enhanced Network) uses a dual-branch structure, combining the detailed features of high-resolution small images and the global features of downsampled large images to reduce the classification error of remote sensing extraction of coastal aquaculture ponds caused by the loss of context information in large-scale image processing; on the other hand, the method based on deep learning has strong generalization ability and high degree of automation compared with the traditional method combined with machine learning, and there is no need to set multiple parameters across the country, which is more convenient and efficient to implement, and the convergence effect of the model is also better.
[0011] To achieve the above object, the present invention provides the following technical solutions:
[0012] The extraction method of coastal aquaculture ponds based on DBCE-Net and Sentinel-2 images includes the following steps:
[0013] Step 1: Download Sentinel-2 images on the GEE platform and label them;
[0014] Step 2: Crop the images and labels obtained in step 1 to make images of size 512x512, organize the cropped image blocks and label blocks into a training data set, and divide them into a training set and a validation set according to the set ratio for subsequent model training and evaluation.
[0015] Step 3: Build the DBCE-Net network structure:
[0016] The DBCE-Net uses a dual-branch structure, including a main branch and a context edge supplement branch (CESB). The feature maps extracted by the two branches are adaptively fused through a dynamic feature fusion module. Finally, the spatial resolution of the feature maps is gradually restored through a decoding module to achieve mapping from high-level features to original resolution images.
[0017] Step 4: Train the DBCE-Net network according to the set training strategy.
[0018] Step 5: Establish evaluation indicators including total loss and F1 score, and finally obtain the optimal model through evaluation indicators.
[0019] Step 6: Input the Sentinel-2 image of the required extraction range into the optimal model to extract coastal aquaculture ponds.
[0020] Furthermore, the specific method of step 1 is as follows:
[0021] 1-1 Select the time, study area and desired band.
[0022] The study area should preferably be selected from March to December each year, during which time fishermen will not clean or dry the ponds. The study area should preferably be selected from coastal areas with typical characteristics and dense aquaculture ponds to ensure the representativeness of the data and the wide applicability of the research results.
[0023] The following bands are preferred for remote sensing data collection:
[0024] ●b2 (blue light band): used for water body detection and vegetation analysis.
[0025] ●b3 (green light band): helps monitor vegetation health and suspended matter in water.
[0026] ●b4 (red light band): often used to distinguish vegetation and soil.
[0027] b8 (near infrared band): has a significant effect on distinguishing vegetation and water bodies.
[0028] ●b11 (short-wave infrared band): used for soil and vegetation moisture monitoring.
[0029] 1-2 Use the QA60 mask layer to remove the cloud from the Sentinel-2 image.
[0030] 1-3 Select the median image with a resolution of 10 meters to download.
[0031] 1-4 Label the downloaded images and use the Google Earth Engine (GEE) platform to divide the images into two categories: background and aquaculture ponds.
[0032] Furthermore, the specific method of step 3 is as follows:
[0033] 3-1 Select Resnest-200e as the backbone network or main branch of the DBCE-Net network to extract the global features of the image.
[0034] 3-2 selects Resnest-50 as the context edge supplement branch (CESB) to extract contextual features and global features of the image.
[0035] 3-3 Design dynamic feature fusion module (DFF):
[0036] First, three dilated convolution operations are performed on the context boundary features output by the context edge supplementary branch, using dilation rates of 2, 3, and 4 respectively. Subsequently, these features are integrated together through channel stacking and continue to be channel stacked with the feature map output by the main branch to finally generate a fused feature map. Next, the fused feature map is further processed by two branches. The first branch first performs a 1x1 convolution operation on the fused feature map, and then applies the Sigmoid activation function. After that, the result is element-wise multiplied with the feature map before the input branch. In the second branch, the fused feature map is first pooled, and then two 1x1 convolutions and Sigmoid activation functions are performed in sequence, and finally element-wise multiplication is performed with the feature map before the input branch. The pooling operation aggregates feature information by reducing the spatial dimension, allowing the network to better capture the global structure of the feature. The next two 1x1 convolutions and Sigmoid activation functions further refine and weight the information in the feature map, ensuring that the network can extract features more accurately. Finally, the feature maps generated by the two branches are combined through element-wise addition to obtain the output feature map of the dynamic feature fusion module DFF module.
[0037] 3-4 Design a decoding module to gradually restore the spatial resolution of the feature map. The feature map output by the dynamic feature fusion module is processed by four upsampling modules and finally generates a 2-channel feature map through a convolution layer.
[0038] Specifically, the input of the upsampling module is the feature map output by the dynamic feature fusion module, or the output of the previous upsampling module. Then, the input feature map is concatenated with the corresponding main branch layer feature map using skip connections.
[0039] On the concatenated feature map, the upsampling module first performs two 3x3 convolution operations, each followed by a ReLU activation function to introduce nonlinearity. Subsequently, the convolution-processed feature map is upsampled by bilinear interpolation.
[0040] Furthermore, the specific method of step 4 is as follows:
[0041] 4-1. When building the data loader, set random parameters. Before training each batch, rotate, shear, and scale the samples with set probabilities to generate enhanced samples in real time.
[0042] 4-2. Training parameter settings: 200 training cycles, batch size of 24, Adam optimizer, initial learning rate of 1e-4, and ReLU activation function.
[0043] 4-3. Establish DiceLoss and FocalLoss loss functions, optimize the network by minimizing the loss function, and optimize the network parameters.
[0044]
[0045] Focal Loss = -α(1-p t ) γ log(p t ),
[0046] Total Loss=Dice Loss+Focal Loss;
[0047] Among them, p in Dice Loss i represents the predicted probability of the model at the i-th pixel, ranging from 0 to 1; g i Indicates the actual category of the i-th pixel, which is a binary value (0 or 1); i p i Indicates the summation operation of all pixels. Dice Loss is used to measure the overlap between the predicted result and the true result. The smaller the value, the better the model performance.
[0048] For the focal loss function, α is a balancing factor used to balance the weights of positive and negative samples, ranging from 0 to 1, to control the contribution of positive and negative samples to the loss; γ is a modulation factor used to adjust the weights of difficult and easy samples, with a value of 2 to enhance the focus on difficult-to-classify samples; p t is the model's predicted probability for the correct category. For positive samples, p t =p, where p represents the probability of the model predicting a positive class. For negative samples, p t =1-p, where 1-p represents the probability of the model predicting a negative class; log(p t ) is the logarithmic loss, which is used to calculate the difference between the predicted probability and the true label. Focal Loss reduces the impact of easy-to-classify samples on the loss and focuses on difficult-to-classify samples, thereby improving the performance of the model on unbalanced datasets.
[0049] The total loss function is the sum of Dice Loss and Focal Loss. By minimizing this loss function, the network parameters are optimized, so that the network gradually converges. 4-4. Train the network until the number of iterations reaches the preset total number.
[0050] Furthermore, the specific method of step 6 is as follows:
[0051] 6-1. Obtain Sentinel-2 images within the area to be extracted on the GEE platform.
[0052] 6-2. The image is automatically cropped into 512x512 image blocks and input into the optimal model for aquaculture pond extraction, then spliced according to the geographic location and finally converted into vector format.
[0053] 6-3. If there is no other water body within 60 meters around a water body, it will be regarded as an independent reservoir and excluded from the aquaculture pond vector.
[0054] The beneficial effects of the present invention are:
[0055] 1) The present invention proposes a DBCE-Net framework with a dual-branch structure, which combines the detailed features of high-resolution small images and the global features of downsampled large images, significantly reducing the classification error caused by the loss of contextual information, thereby improving the accuracy of coastal aquaculture pond extraction;
[0056] 2) The method based on deep learning proposed in this invention has stronger generalization ability than the traditional method combined with machine learning. There is no need to set multiple thresholds and parameters for different regions, making the application of this method more convenient and efficient across the country;
[0057] 3) Through the dynamic feature fusion module, the features extracted by the main branch and the auxiliary branch are effectively fused, which improves the robustness and stability of the model and ensures that better extraction effects can be achieved in different regions and under different imaging conditions.
[0058] 4) The method of the present invention can achieve efficient and accurate extraction from coastal aquaculture ponds without human intervention, has a high degree of automation, and greatly reduces manpower and time costs. BRIEF DESCRIPTION OF THE DRAWINGS
[0059] Figure 1 A flowchart of the basic steps of the method according to the embodiment of the present invention;
[0060] Figure 2 A research area for selecting training samples in an embodiment of the present invention;
[0061] Figure 3 This is a schematic diagram of the DBCE-NET network structure according to an embodiment of the present invention;
[0062] Figure 4 This is a schematic diagram of a dynamic feature fusion module according to an embodiment of the present invention;
[0063] Figure 5 is a loss curve during the network training process of an embodiment of the present invention;
[0064] Figure 6Comparison of the aquaculture ponds extracted for the present invention with the manually drawn sample polygons based on Google Earth;
[0065] Figure 7 Comparison of the results extracted for the present invention with other 10-m resolution datasets for coastal aquaculture ponds. DETAILED DESCRIPTION
[0066] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention is described in detail below in conjunction with specific embodiments. Specific embodiments are described below to simplify the present invention. However, it should be recognized that the present invention is not limited to the illustrated embodiments, and various modifications of the present invention are possible without departing from the basic principles, and these equivalent forms also fall within the scope defined by the claims attached to the present application.
[0067] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0068] The present invention proposes a coastal aquaculture pond extraction method based on DBCE-Net and Sentinel-2 images, and proposes an improved dual-branch DBCE-Net deep learning framework, which combines the detailed features of high-resolution small images and the global features of downsampled large images. Through this framework, a coastal aquaculture pond extraction network consisting of a main branch network, a context edge supplementary branch network and a dynamic feature fusion module is constructed, and the network loss function is designed by comprehensively considering classification error, context information, boundary information, etc. The designed aquaculture pond extraction algorithm can accurately extract aquaculture ponds while considering the loss of context information.
[0069] like Figure 1 As shown in FIG. 1 , a basic flow chart of the invented coastal aquaculture pond extraction method based on DBCE-Net and Sentinel-2 images is shown. Figure 2 The figure shows the schematic diagram of the network device structure of the invented DBCE-Net.
[0070] Step 1: Download the 10m resolution image of Sentinel-2 on the GEE platform. Select Bohai Bay, Hangzhou Bay and Sanmen Bay in Zhejiang as the geographical range, select the time from March 1 to November 30, 2021, use QA60 to remove the cloud, and finally select the median image of the entire time series for download. Select the following bands for remote sensing data collection:
[0071] ●b2 (blue light band): used for water body detection and vegetation analysis.
[0072] ●b3 (green light band): helps monitor vegetation health and suspended matter in water.
[0073] b4 (red light band): often used to distinguish vegetation and soil.
[0074] ●b8 (near infrared band): has a significant effect on distinguishing between vegetation and water bodies.
[0075] ●b11 (short-wave infrared band): used for soil and vegetation moisture monitoring.
[0076] Step 2: Create training data and validation sets, and study the selected areas such as Figure 2 Shown are Bohai Bay and Zhejiang Province and Hangzhou Bay and Sanmen Bay.
[0077] (1) The median image and the corresponding label image downloaded from the GEE platform were cropped into image blocks of size 512x512.
[0078] (2) The cropped images and labeled images are divided into training set and validation set in a ratio of 8:2.
[0079] Step 3: Build the DBCE-Net network structure.
[0080] like Figure 3 As shown in the design of DBCE-Net (Dual-Branch Context Enhanced Network), the network structure mainly includes the main branch, the context edge supplement branch (CESB), the dynamic feature fusion module and the decoding module. Through this dual-branch design, DBCE-Net can reduce the loss of context information and edge information caused by cropping of remote sensing images, and can also improve the multi-scale capability of the network, thereby realizing efficient and accurate extraction of coastal aquaculture ponds.
[0081] (1) In the feature extraction stage, the main branch of DBCE-Net uses ResNest200e as an encoder to extract high-resolution detail information from the original 10m resolution Sentinel-2 image. The main branch is mainly responsible for capturing fine features and detail information in the image, ensuring that the final output segmentation map has high-resolution accuracy.
[0082] (2) The image input by the context edge supplement branch is aligned with the center point of the image input by the main branch, and its size is twice that of the latter. After downsampling, an image with a resolution of 30m is finally generated. The main purpose is to supplement the missing context and boundary information caused by the cropped image of the main branch. The traditional single-branch structure is prone to lose information when processing the edge part of the image, resulting in incomplete extraction of aquaculture ponds or misidentification of multi-branch rivers as aquaculture ponds. To avoid these problems, the context edge supplement branch adopts a lightweight encoder ResNest50. By capturing macro features and relationships covering a large spatial range, CESB helps DBCE-Net to more accurately recognize and process complex scenes in images.
[0083] (3) In order to more fully integrate the trunk features extracted by the main branch of the dual-branch network and the context boundary features extracted by the context edge supplementary branch, we designed a dynamic feature fusion module (DFF) to learn the complementarity between the high-resolution detail information in the trunk feature map and the context boundary features in the context edge supplementary branch, so as to obtain an enhanced feature representation that contains both local details and global semantics. Figure 4 As shown in Figure 1, in the design of the DFF module, we first perform three dilated convolution operations on the context boundary features output by the context edge supplementary branch, using dilation rates of 2, 3, and 4, respectively. Subsequently, these features are integrated together through channel stacking and channel stacked with the feature map output by the main branch to finally generate the fused feature map. Next, we further process the fused feature map through two branches. The first branch first performs a 1x1 convolution operation on the fused feature map, and then applies the Sigmoid activation function. After that, the result is element-wise multiplied with the feature map before the input branch. This process enhances the key features using the output of the Sigmoid function while suppressing the less important parts, thereby achieving effective feature selection and weight assignment. In the second branch, we first perform a pooling operation on the fused feature map, then perform two 1x1 convolutions and Sigmoid activation functions in sequence, and finally perform element-wise multiplication with the feature map before the input branch. The pooling operation aggregates feature information by reducing the spatial dimension, allowing the network to better capture the global structure of the feature. The next two 1x1 convolutions and Sigmoid activation functions further refine and weight the information in the feature map, ensuring that the network can extract features more accurately. Finally, the feature maps generated by the two branches are combined through element-wise addition to obtain the output feature map of the DFF module.
[0084] (4) In the decoding module, DBCE-Net gradually restores the feature map output by the dynamic feature fusion module to its original size. The detailed steps are as follows:
[0085] The feature map output by the dynamic feature fusion module is processed by four upsampling modules and finally generates a 2-channel feature map through a convolution layer. This step converts the high-resolution feature map into a target segmentation map, ensuring that the output resolution is consistent with the original input image.
[0086] Specifically, the input of the upsampling module is the feature map output by the dynamic feature fusion module or the output of the previous upsampling, and is concatenated with the feature map of the corresponding main branch layer. The jump connection is designed to solve the gradient vanishing problem in deep neural networks and accelerate training. By directly passing the input information to the subsequent layers, the gradient can be back-propagated more smoothly, thereby maintaining gradient stability and making network training more efficient.
[0087] Two 3x3 convolution operations are performed on the concatenated feature maps to further extract and refine the features. These convolution operations help integrate features of different scales and generate more refined segmentation results. Each convolution operation is followed by a ReLU activation function to introduce nonlinearity.
[0088] Subsequently, the convolutional feature map is upsampled by bilinear interpolation. The purpose of the upsampling operation is to gradually increase the spatial resolution of the feature map to restore it to the size of the original image. Bilinear interpolation is a non-learning parameter upsampling method that calculates new pixel values by weighted averaging the neighboring pixels of the input feature map, thereby smoothly increasing the resolution of the feature map.
[0089] Step 4: Train the DBCE-Net network according to the set training strategy;
[0090] 4-1. Python and Pytorch were selected as the development language and main framework. CUDA was used for GPU acceleration in the experiment, and the GDAL library was used to process geographic data in TIF format to ensure the reading and processing of geographic information data.
[0091] 4-2. When building Dataloader, we set random parameters and rotate, shear, and scale samples with a set probability of 30% before training each batch to generate enhanced samples in real time.
[0092] 4-3. Training parameters The training cycle is 200 times, the batch size is 24, the Adam optimizer is used, the initial learning rate is set to 1e-4, and the ReLU activation function is used.
[0093] 4-4. Establish DiceLoss and FocalLoss loss functions, and optimize network parameters by minimizing the loss functions.
[0094]
[0095] Focal Loss = -α(1-(p t ) γ log(p t ),
[0096] Total Loss=Dice Loss+Focal Loss;
[0097] Among them, p in Dice Loss i represents the predicted probability of the model at the i-th pixel, ranging from 0 to 1; g i Indicates the actual category of the i-th pixel, which is a binary value (0 or 1); i p i Indicates the summation operation of all pixels. Dice Loss is used to measure the overlap between the predicted result and the true result. The smaller the value, the better the model performance.
[0098] For the focal loss function, α is a balancing factor used to balance the weights of positive and negative samples, ranging from 0 to 1, to control the contribution of positive and negative samples to the loss; γ is a modulation factor used to adjust the weights of difficult and easy samples, with a value of 2 to enhance the focus on difficult-to-classify samples; p t is the model's predicted probability for the correct category. For positive samples, p t =p, where p represents the probability of the model predicting a positive class. For negative samples, p t =1-p, where 1-p represents the probability of the model predicting a negative class; log(p t ) is the logarithmic loss, which is used to calculate the difference between the predicted probability and the true label. Focal Loss reduces the impact of easy-to-classify samples on the loss and focuses on difficult-to-classify samples, thereby improving the performance of the model on unbalanced datasets.
[0099] The total loss function is the sum of Dice Loss and Focal Loss. By minimizing this loss function, the network parameters are optimized, so that the network gradually converges.
[0100] 4-5: Train the network until the number of iterations reaches the preset total number. The training loss curve is as follows: Figure 5 shown.
[0101] Step 5: In order to obtain the optimal model, we established evaluation indicators including F1 score and total loss. The criterion for model selection is to give priority to the model with the smallest total loss. If the total losses are the same, their F1 scores are further compared and the model with the highest F1 score is selected. Through this systematic evaluation and selection process, we ensure that the model we finally obtain is the best in both performance and accuracy.
[0102] Step 6: Input the Sentinel-2 imagery within the area where coastal aquaculture ponds need to be extracted into the optimal model:
[0103] 6-1 Obtain Sentinel-2 images within the vector range of coastal counties on the GEE platform;
[0104] 6-2 The image is automatically cropped into image blocks of size 512x512 and input into the trained optimal model for aquaculture pond extraction. After model prediction, the predicted aquaculture pond data is spliced according to the geographical location to ensure complete coverage of the entire study area. Finally, the spliced aquaculture pond data is converted into vector format for further analysis and application.
[0105] 6-3 It is necessary to distinguish aquaculture ponds from independent reservoirs of regular shape, which are similar to aquaculture ponds in shape and size. These independent reservoirs cannot be distinguished from aquaculture ponds using only spectral and geometric features. Therefore, considering the spatial characteristics of the clustered distribution of aquaculture ponds, the distance between the nearest neighbors of each potential aquaculture pond object is calculated. Then a threshold is set to 60m to eliminate independently distributed water bodies. If there is no other water body within 60m near the water body, it is considered to be an independent reservoir.
[0106] In order to verify the accuracy of the contours and shapes of the aquaculture ponds predicted by this method, this paper compared the aquaculture ponds generated by this method with the polygons manually drawn based on high-resolution images of Google Earth. Figure 6 The extracted contours and shapes are highly consistent with the manually drawn high-precision results, as shown in Figure 1. This comparison verifies the reliability of the method in this study and demonstrates its effectiveness in extracting information about aquaculture ponds.
[0107] We observed that the aquaculture ponds extracted using the method in this study were highly consistent with the visual interpretation results in most cases, showing good overlap and extraction effects. However, we also noticed that there were certain errors in the extraction process of the pond edge, which resulted in the aquaculture pond area extracted based on this study being slightly smaller than the actual area overall. This phenomenon is mainly due to the limitation of spatial resolution, which affects the accurate extraction of the aquaculture pond boundary.
[0108] In order to verify the effect of the present invention, the results of this study were compared with other 10-meter resolution land cover datasets, including the FROM-GLC10 dataset, the CWaC dataset, and the ESA land cover dataset.
[0109] Although these datasets are not specifically classified for aquaculture ponds, aquaculture ponds are usually classified as a type of water body in the land cover classification system. Therefore, by comparing the extraction results of water body categories in these datasets, we can indirectly evaluate their performance in identifying aquaculture ponds.
[0110] pass Figure 7 The comparison shows that they can all roughly extract the location of aquaculture ponds, but the accuracy is relatively low, and there are generally a lot of misclassification and omission problems. Among them, the boundaries of aquaculture ponds in the ESA dataset are fuzzy, the ponds are closely connected, it is difficult to distinguish the shape and boundaries of a single pond, and the aquaculture ponds and nearby land cannot be effectively distinguished, which is inconsistent with the expected performance of 10m high-resolution data. The FROM-GLC10 dataset performs well in extracting aquaculture ponds and can effectively distinguish aquaculture ponds from land, but fails to distinguish dams from aquaculture ponds. The CWaC dataset is relatively superior in pond boundary recognition, but the extraction accuracy of small aquaculture ponds is low. In addition, the above datasets cannot effectively distinguish between salt pans and aquaculture ponds. Overall, compared with existing land cover datasets, the proposed method is more accurate in identifying the boundaries of aquaculture ponds, can more completely extract small aquaculture ponds, and successfully distinguishes the difference between salt pans and aquaculture ponds.
[0111] In order to further verify the effect of the present invention, we choose to use the single-branch structure UNet and UNet commonly used in existing research. 2 Net model and trained it on the same training set. We then used the trained model to extract coastal aquaculture ponds in China's coastal counties in 2023. After completing the extraction, we evaluated the accuracy of the extraction results through the generated validation sample points.
[0112] To ensure the accuracy of the sample points, we first randomly generated various water body sample points based on the 2023 ESA dataset, and then combined these sample points with Google Earth high-definition images to remove misclassified sample points, and finally generated accurate sample points for verification.
[0113] Table 1 Comparison of the accuracy of different models tested for coastal aquaculture ponds
[0114]
[0115] According to the comparison of the accuracy of different models for coastal aquaculture ponds, the coastal aquaculture pond extraction model (DBCE-Net) proposed in this paper is superior to the existing UNet and U 2 Net model can achieve better technical results.
[0116] The present invention describes specific embodiments to simplify the present invention. However, it should be recognized that the present invention is not limited to the described embodiments, and various modifications of the present invention are possible without departing from the basic principles, and these equivalent forms also fall within the scope defined by the appended claims of the present application.
Claims
1. A coastal aquaculture pond extraction method based on DBCE-Net and Sentinel-2 images, characterized in that: The steps include: Step 1: Download Sentinel-2 images on the GEE platform and label them; Step 2: Crop the images and labels obtained in step 1 to make images of size 512x512, organize the cropped image blocks and label blocks into training data sets, and divide them into training sets and validation sets according to the set ratio for subsequent model training and evaluation; Step 3: Build the DBCE-Net network structure: The DBCE-Net uses a dual-branch structure, including a main branch and a context edge supplementary branch. The feature maps extracted by the two branches are adaptively fused through a dynamic feature fusion module. Finally, the spatial resolution of the feature map is gradually restored through a decoding module to achieve mapping from high-level features to original resolution images. Step 4: Train the DBCE-Net network according to the set training strategy; Step 5: Establish evaluation indicators including total loss and F1 score, and finally obtain the optimal model through evaluation indicators; Step 6: Input the Sentinel-2 image of the required extraction range into the optimal model to extract coastal aquaculture ponds.
2. The coastal aquaculture pond extraction method based on DBCE-Net and Sentinel-2 images according to claim 1 is characterized in that: Step 1: 1-1 Select the time, study area and required band; The study area should preferably be selected from March to December each year; the study area should preferably be selected from coastal areas with typical characteristics and dense aquaculture ponds to ensure the representativeness of the data and the wide applicability of the research results; The following bands are preferred for remote sensing data collection: b2 blue light band: used for water detection and vegetation analysis; b3 green band: helps monitor vegetation health and suspended matter in water; b4 red light band: often used to distinguish vegetation and soil; b8 near-infrared band: has a significant effect on distinguishing vegetation and water bodies; b11 short-wave infrared band: used for soil and vegetation moisture monitoring; 1-2 Use the QA60 mask layer to remove the cloud from the Sentinel-2 image; 1-3 Select the median image with a resolution of 10 meters for download; 1-4 Label the downloaded images and use the GEE platform to divide the images into two categories: background and aquaculture ponds.
3. The coastal aquaculture pond extraction method based on DBCE-Net and Sentinel-2 images according to claim 1 is characterized in that: Step 3: 3-1 Select Resnest-200e as the backbone network of DBCE-Net network, i.e. the main branch, to extract the global features of the image; 3-2 Select Resnest-50 as the context edge supplement branch CESB to extract contextual features and global features of the image; 3-3 Design dynamic feature fusion module DFF: First, three dilated convolution operations are performed on the context boundary features output by the context edge supplementary branch, with dilation rates of 2, 3, and 4 respectively; then, these features are integrated together through channel stacking, and continue to be channel stacked with the feature map output by the main branch to finally generate a fused feature map; Next, the fused feature map is further processed through two branches; the first branch first performs a 1x1 convolution operation on the fused feature map, and then applies the Sigmoid activation function; after that, the result is element-wise multiplied with the feature map before the input branch; in the second branch, the fused feature map is first pooled, and then two 1x1 convolutions and Sigmoid activation functions are performed in sequence, and finally element-wise multiplication is performed with the feature map before the input branch; the pooling operation aggregates feature information by reducing the spatial dimension, so that the network can better capture the global structure of the feature; the next two 1x1 convolutions and Sigmoid activation functions further refine and weight the information in the feature map to ensure that the network can extract features more accurately; finally, the feature maps generated by each of the two branches are combined through point-by-point addition to obtain the output feature map of the dynamic feature fusion module DFF module; 3-4 Design a decoding module to gradually restore the spatial resolution of the feature map. After the feature map output by the dynamic feature fusion module is processed by four upsampling modules, it finally generates a 2-channel feature map through a convolution layer; Specifically, the input of the upsampling module is the feature map output by the dynamic feature fusion module, or the output of the previous upsampling module; then, the input feature map is concatenated with the corresponding main branch layer feature map using skip connections; On the concatenated feature map, the upsampling module first performs two 3x3 convolution operations, each followed by a ReLU activation function to introduce nonlinearity; then, the convolution-processed feature map is upsampled by bilinear interpolation.
4. The coastal aquaculture pond extraction method based on DBCE-Net and Sentinel-2 images according to claim 3 is characterized in that: Step 4: 4-1. When building the data loader, set random parameters, rotate, shear, and scale the samples with set probabilities before training each batch, and generate enhanced samples in real time; 4-2. Training parameter settings: 200 training cycles, batch size of 24, Adam optimizer, initial learning rate of 1e-4, and ReLU activation function; 4-3. Establish DiceLoss and FocalLoss loss functions, optimize the network by minimizing the loss function, and optimize the network parameters; Focal Loss=-α(1-p t ) γ log(p t ), Total Loss=Dice Loss+Focal Loss; Among them, p in Dice Loss i represents the predicted probability of the model at the i-th pixel, ranging from 0 to 1; g i Indicates the actual category of the i-th pixel, which is a binary value; ∑ i p i Indicates the summation operation of all pixels; Dice Loss is used to measure the overlap between the predicted result and the true result. The smaller the value, the better the model performance. For the focal loss function, α is a balancing factor used to balance the weights of positive and negative samples, ranging from 0 to 1, to control the contribution of positive and negative samples to the loss; γ is a modulation factor used to adjust the weights of difficult and easy samples, with a value of 2 to enhance the focus on difficult-to-classify samples; p t is the model's predicted probability for the correct category. For positive samples, p t =p, where p represents the probability of the model predicting a positive class. For negative samples, p t =1-p, where 1-p represents the probability of the model predicting a negative class; log(p t ) is the logarithmic loss, which is used to calculate the difference between the predicted probability and the true label; Focal Loss reduces the impact of easy-to-classify samples on the loss and focuses on difficult-to-classify samples, thereby improving the performance of the model on unbalanced datasets; The total loss function is the sum of Dice Loss and Focal Loss. By minimizing this loss function, the network parameters are optimized, so that the network gradually converges. 4-4. Train the network until the number of iterations reaches the preset total number.
5. The coastal aquaculture pond extraction method based on DBCE-Net and Sentinel-2 images according to claim 1 or 4, characterized in that: Step 6: 6-1. Obtain Sentinel-2 images within the area to be extracted on the GEE platform; 6-2. Automatically crop the image into 512x512 image blocks and input the optimal model to extract the aquaculture ponds, then splice them according to the geographical location and finally convert them into vector format; 6-3. If there is no other water body within 60 meters around a water body, it will be regarded as an independent reservoir and excluded from the aquaculture pond vector.