Multiband remote sensing coastline extraction method based on CCAU-Net

By introducing asymmetric convolution modules and convolution attention modules in U-Net model, the problems of leakage and misclassification of small water bodies in coastline extraction are solved, the accuracy and robustness of coastline extraction are improved, and information is solved Overload and semantic gap problems achieve more accurate sea and land boundary segmentation.

CN119048896APending Publication Date: 2024-11-29QINGDAO UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410956950.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-17
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The prior art is prone to leaking small water bodies when extracting coastlines, and mistakenly treats shadows and clouds as water bodies. It is difficult for traditional models to choose appropriate thresholds or hyperparameters to determine sea and land boundaries, and the U-Net model has problems with original information overload and semantic gaps. .

Method used

Using the CCAU-Net model, by introducing asymmetric convolution modules and convolution attention modules, replacing standard square convolutions in U-Net, the network's attention to input features is enhanced, and the fusion ability of low-level semantics and high-level features is improved. Improve the accuracy of coastline extraction in multi-band remote sensing images.

Benefits of technology

It significantly improves the accuracy and performance of coastline extraction, enhances the model's robustness to image flip, rotation and distortion, reduces the impact of background noise, improves information capture capabilities, and obtains clearer results of sea and land boundary segmentation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119048896A_ABST
    Figure CN119048896A_ABST
Patent Text Reader

Abstract

The invention discloses a CCAU-Net-based multiband remote sensing coastline extraction method, and provides an improved model and an extraction method for solving the problems of omission and misclassification of shadow targets when a small water body is extracted by an existing classical U-Net model and the problem of insufficient attention of the model to input feature key information. An asymmetric convolution module is introduced to replace standard square convolution of a U-Net model, and a convolution attention module is integrated in jump connection of the model, so that the concentration degree of the network model to an input feature key region is enhanced. According to the multiband remote sensing coastline extraction method provided on the basis of the improvement, the fusion capability of the model on low-level semantics and high-level features is improved, and the extraction precision and the overall performance of the coastline in the multiband remote sensing image are remarkably improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a multi-band remote sensing coastline extraction method based on CCAU-Net. Background Art

[0002] The coastline is the boundary between land and sea. At present, the use of remote sensing technology to extract the coastline and monitor its dynamic changes has become the mainstream technical means of studying the coastline.

[0003] In recent years, Chinese and foreign scholars have proposed a variety of methods for extracting coastlines using remote sensing images. These methods can be divided into three categories: (1) threshold-based methods; (2) image edge detection algorithms; and (3) object-oriented classification methods. With the rapid development of artificial intelligence, semantic segmentation methods based on convolutional neural networks have attracted widespread attention in the field of image processing. Compared with traditional methods, this deep learning method can better handle the complex features of images and has a strong autonomous learning ability.

[0004] Most of the above methods are aimed at natural scene images. On this basis, some scholars have proposed semantic segmentation models specifically for segmenting urban buildings, land, rivers, forests and other ground objects. Kampffmeyer et al. improved the U-Net algorithm for high-resolution remote sensing ground object information. Lin et al. designed a deep convolutional neural network for small-target remote sensing images in urban areas, realizing end-to-end image pixel-level segmentation. He et al. combined the attention mechanism with the classic U-Net network and designed a self-attention mechanism module U-Net network for extracting glacial lakes in remote sensing images. Compared with traditional models, it not only improves classification accuracy but also achieves clearer boundaries.

[0005] However, although some work has been done on the task of extracting coastlines from multispectral remote sensing images, there are still some problems, which are reflected in the following aspects: 1. The extraction effect is not good, and it is easy to miss small water bodies, and mistake shadows and clouds as water bodies; 2. It is difficult for traditional natural scene semantic segmentation models to select an optimal threshold or hyperparameter to determine the location and classification of land-sea boundaries; 3. In the U-Net model, the combination of features of different dimensions leads to the original information overload problem and the semantic gap problem. Summary of the invention

[0006] Aiming at the problems of omission of small water bodies and misclassification of shadow targets in the existing classic U-Net model, as well as the model's insufficient attention to key information of input features, the present invention proposes an improved model and extraction method.

[0007] In order to achieve the above object, the present invention provides a multi-band remote sensing coastline extraction method based on CCAU-Net, comprising the following steps: Step A: Construct the CCAU-Net (Combined CBAM and ACBlock U-Net, a U-Net that integrates the convolutional attention module and the asymmetric convolutional module) network model; Specifically, the asymmetric convolution block (ACBlock) is introduced to replace the standard square convolution of the U-Net model, and the convolutional block attention module (CBAM) is integrated in the jump connection of the model to enhance the network model's focus on the key areas of the input features, improve the model's ability to integrate low-level semantics and high-level features, and improve the accuracy and overall performance of feature extraction in multi-band remote sensing images. The convolutional attention module concatenates channel attention and spatial attention along two independent dimensions, channel and space, and uses global average pooling and maximum pooling to obtain the attention map, which is then multiplied by the input feature map for adaptive feature refinement. Step B: Obtain 5 Landsat8 multispectral remote sensing images covering the sea-land boundary area, and pre-process the coastline remote sensing images to obtain images that meet the requirements for data extraction; Step C: Create neural network training samples, construct a data set and divide it into training set, validation set and test set; Step D: Use the training set and validation set to train the CCAU-Net network and optimize the parameters; Step E: Input the test set coastline remote sensing image that needs to be extracted into the trained CCAU-Net network to obtain the binary prediction image result of land and sea segmentation; Step F: The contour in the binary prediction image result of land-sea segmentation is used as the land-sea boundary, and then the land-sea boundary binary image is vectorized to obtain the coastline extraction result as the final output result.

[0008] Further, step C specifically includes the following steps: Step C1: In the image obtained in step B, the image is irregularly cropped to retain only the land and sea parts close to the coastline; Step C2: The images obtained in step C1 are cropped into multiple images of 256×256 pixels, retaining all bands. 920 images with both ocean and land are selected, and the water bodies are marked using the MNDWI water index. Step C3: Perform data augmentation on the images obtained in step C2 to obtain a data set of 6440 images and 6440 labels. The data set is randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0009] Further, step D specifically includes the following steps: Step D1: Input the training set obtained in step C3 into the encoder composed of asymmetric convolutional layer, ReLU activation function and pooling layer in the CCAU-Net model to extract feature maps; Step D2: The resulting feature map obtained in step D1 is passed into the model decoder, and processed by its upsampling, convolution and convolution attention modules to improve the image resolution; Step D3: After a 1×1 convolution, a 2-channel land-sea segmentation feature map and a coastline distribution feature map are obtained. The model performance is evaluated through the validation set after each epoch of training. After several iterative trainings, a trained network is obtained. Step D4: Calculate the loss between the predicted result and the true label, where the loss function adopts the cross entropy loss function and the Adaptive Moment Estimation (Adam) is used as the optimizer to optimize the network parameters to achieve the purpose of optimizing the model.

[0010] Furthermore, step D1 specifically includes the following steps: Step D1-1: Input the image to the first-stage encoder, and process it through two asymmetric convolution layers consisting of three parallel layers with convolution kernels of 3×3, 1×3 and 3×1 and ReLU activation function to obtain a 16-channel feature map; Step D1-2: The 2×2 max pooling layer in the first stage extracts the intermediate features and passes them to the second stage encoder; Step D1-3: The second stage encoder performs the same convolution process twice on the incoming feature map to obtain a 32-channel feature map; Step D1-4: The 2×2 max pooling layer in the second stage extracts the intermediate features and passes them to the third stage encoder; Step D1-5: The third stage encoder performs the same convolution process on the incoming feature map twice to obtain a 64-channel feature map; Step D1-6: The 2×2 max pooling layer in the third stage extracts the intermediate features and passes them to the fourth stage encoder; Step D1-7: The fourth stage encoder performs the same convolution process on the incoming feature map twice to obtain a 128-channel feature map; Step D1-8: The 2×2 max pooling layer in the fourth stage extracts the intermediate features and passes them to the fifth stage encoder; Step D1-9: The fifth stage encoder performs two identical convolutions on the incoming feature map to obtain a 256-channel feature map.

[0011] Furthermore, step D2 specifically includes the following steps: Step D2-1: By upsampling the feature map of the fifth stage and concatenating the feature map of the fourth stage, a 256-channel feature map entering the sixth stage decoder is obtained, and a 128-channel feature map is obtained through a convolutional attention module (CBAM) and convolution operations; Step D2-2: By upsampling the feature map of the sixth stage and concatenating the feature map of the third stage, a 128-channel feature map entering the seventh stage decoder is obtained, and a 64-channel feature map is obtained through a convolutional attention module (CBAM) and convolution operations; Step D2-3: By upsampling the feature map of the seventh stage and concatenating the feature map of the second stage, a 64-channel feature map entering the eighth stage decoder is obtained, and a 32-channel feature map is obtained through the convolutional attention module (CBAM) and convolution operation; Step D2-4: By upsampling the feature map of the eighth stage and connecting the feature map of the first stage, a 32-channel feature map entering the ninth stage decoder is obtained, and a 16-channel feature map is obtained through the convolutional attention module (CBAM) and convolution operation.

[0012] The beneficial effects of this solution can be known from the description of the above solution. Compared with the prior art, it has the following beneficial effects: 1. In the U-Net network, an asymmetric convolution layer is used to replace the standard convolution layer to adjust the feature weights, so that richer semantic features can be obtained. The present invention does not use the standard square convolution in the classic U-Net network, but replaces it with an asymmetric convolution layer. In the implementation process, the fused convolution kernel parameters are used to initialize the existing network, but the network parameters use the fused convolution kernel parameters, so the amount of calculation will not be increased in the reasoning stage. And through this asymmetric convolution structure, the effect of the square convolution in the horizontal and vertical directions is enhanced, the model's robustness to image flipping, rotation and distortion is improved, and the backbone of the square convolution kernel is significantly enhanced, highlighting the importance of local features to segmentation tasks, which can effectively improve classification accuracy; 2. By introducing the convolutional attention module and adding it to the jump connection of the U-Net network model, the network pays more attention to the more noteworthy parts of the input features, thereby solving the original information overload problem and semantic gap problem caused by the combination of features of different dimensions in the jump part of the U-Net network. The present invention proposes to add the convolutional attention module to the jump connection of the U-Net network model, so that the model can capture the spatial and band information of the image, realize the successful integration of low-level semantics and high-level abstract features, eliminate background noise during feature information extraction, reduce the impact of low image contrast on the model, and improve the network's ability to capture information flow. These features improve the extraction accuracy and performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a flow chart of the present invention; Figure 2 The CCAU-Net network architecture diagram of the present invention; Figure 3(a) is a schematic diagram of the training phase of the asymmetric convolution module; Figure 3(b) is a schematic diagram of the inference phase of the asymmetric convolution module; Figure 4(a) is a schematic diagram of the overall structure of the convolutional attention module; Figure 4(b) is a schematic diagram of the channel attention module structure; Figure 4(c) is a schematic diagram of the spatial attention module structure; Figure 5 Comparison chart of coastline extraction results using four methods. DETAILED DESCRIPTION

[0014] In order to clearly illustrate the technical features of this solution, the present invention is described below through specific implementation methods.

[0015] like Figure 1 As shown in FIG. 1 , a multi-band remote sensing coastline extraction method based on CCAU-Net includes the following steps: Step A: Build the CCAU-Net network model; Specifically: by introducing an asymmetric convolution module to replace the standard square convolution of the U-Net model and integrating the convolutional attention module in the jump connection of the model, the network model's focus on the key areas of input features is enhanced, the model's ability to fuse low-level semantics and high-level features is improved, and the accuracy and overall performance of coastline extraction in multi-band remote sensing images are improved.

[0016] like Figure 2As shown, the overall structure of the CCAU-Net model proposed in the present invention continues the U-shaped symmetrical structure in the U-Net model, but since rich spatial details are important for segmentation and in order to improve the feature extraction capability of the network, this model replaces the ordinary convolution in the original U-Net with an asymmetric convolution. In addition, the classic U-Net model directly splices the shallow features in the encoder and decoder with the deep features, which easily leads to the omission of small water bodies and misclassification of shadow targets. Considering that inter-channel information and spatial position information are crucial for capturing the long-term dependence of target architecture and modeling in visual tasks, the present invention proposes to introduce a convolutional attention module into the step connection part of the U-Net network to adjust the feature weights, focus on learning water body features, enable the model to capture sensitive information such as cross-channel and position, and enhance the network's ability to extract water bodies.

[0017] The CCAU-Net model proposed in this invention not only continues the general architecture of the U-Net model, but also improves the standard square convolution of the original model. The original standard square h × h convolution kernel is replaced by three parallel h × h, h × 1, 1 × h convolution kernels. The theoretical basis behind this is that a two-dimensional convolution kernel can be equivalently converted into a series of one-dimensional convolutions. At the same time, the convolution operation satisfies the distributive law property, as shown in formula (1): (1)

[0018] Where I represents the input matrix, K(1) and K(2) represent two 2D kernels with compatible sizes, is a summation operation. The meaning of this equation is that the convolution of K(1) with K(2) and I and then adding the results is consistent with the result of first adding K(1) and K(2) point by point and then convolving with I. In other words, the additivity of two-dimensional convolution holds, even if the kernel sizes are different. This is the theoretical basis of asymmetric convolution.

[0019] Taking the most common 3×3 convolution kernel in CNN as an example, in the training phase, the asymmetric convolution is replaced by ACB consisting of three parallel layers with convolution kernels of 3×3, 1×3 and 3×1, as shown in Figure 3(a). The ACB output is the sum of the outputs of the three parallel layers. In the derivation phase, these three convolutions are fused, as shown in Figure 3(b). In the implementation process, this part uses the fused convolution kernel parameters to initialize the existing network. The network structure is exactly the same as the original network, except that the network parameters use the fused convolution kernel parameters, so the amount of calculation will not increase in the inference phase.

[0020] And through this asymmetric convolution structure, the effect of square convolution in the horizontal and vertical directions is enhanced, the robustness of the model to image flipping, rotation and distortion is improved, and the backbone part of the square convolution kernel is significantly enhanced, highlighting the importance of local features for segmentation tasks, which can effectively improve classification accuracy.

[0021] The attention mechanism is inspired by the human visual mechanism and allows the network to learn features that deserve more attention. In a neural network, the attention mechanism uses some network layers to calculate the weight values ​​corresponding to the feature map, and performs "attention mechanism" on the feature map to extract important features, filter out useless features, and improve classification accuracy. In recent years, it has been verified that introducing the attention mechanism in neural networks can improve the feature expression ability of network models.

[0022] The convolutional attention module introduced in the present invention connects channel attention and spatial attention in series along two independent dimensions, channel and space, and uses global average pooling and maximum pooling to obtain an attention map, which is then multiplied by the input feature map for adaptive feature refinement. It is a lightweight general module. Its structure is shown in Figure 4(a).

[0023] The channel attention module in the convolutional attention module is shown in Figure 4(b). First, a H×W×C feature F is input, and a spatial global average pooling and maximum pooling are performed to obtain two 1×1×C channel descriptions. Then, they are sent to a two-layer neural network, where the number of neurons in the first layer is C / r, the activation function is ReLU, and the number of neurons in the second layer is C. This two-layer neural network is shared. Then, the two features are added together and passed through a Sigmoid activation function to obtain the weight coefficient Mc. Finally, the weight coefficient is multiplied by the original feature F to obtain the scaled new feature, and finally the new feature is input into the spatial attention module. The channel attention mapping process is shown in formula (2): (2)

[0024] Among them, Mc(F) is the channel attention module, AvgPool() and MaxPool() are average pooling and maximum pooling respectively, σ is the Sigmoid activation function, and MLP is a shared network composed of multi-layer perceptrons.

[0025] The spatial attention module in the convolutional attention module is shown in Figure 4(c). First, the new feature F processed by the channel attention module is ’First, perform average pooling and maximum pooling of a channel dimension to obtain two H×W×1 channel descriptions, and then splice these two descriptions together according to the channel. Then, pass through a 7×7 convolution layer and then the activation function Sigmoid to obtain the weight coefficient Ms. Finally, multiply the weight coefficient Ms and the feature F' to obtain the scaled new feature. The spatial attention mapping process is shown in formula (3): (3)

[0026] Among them, Ms(F') is the spatial attention module, AvgPool() and MaxPool() are average pooling and maximum pooling respectively, σ is the Sigmoid activation function, and f is the convolutional layer operation.

[0027] Step B: Obtain five Landsat8 multispectral remote sensing images covering the land-sea boundary area of ​​the Yellow Sea, with a spatial resolution of 30m and 8 bands, and pre-process the coastline remote sensing images to obtain images that meet the requirements for data extraction.

[0028] Step C: Create neural network training samples, build a data set and divide it into training set, validation set and test set. The specific steps are as follows: Step C1: In the image obtained in step B, the image is irregularly cropped to retain only the land and sea parts close to the coastline; Step C2: The images obtained in step C1 are cropped into multiple images of 256×256 pixels, retaining all 8 bands. 920 images with both ocean and land are selected, and the water bodies are marked using the MNDWI water index. Step C3: Perform data augmentation on the images obtained in step C2 to obtain a data set of 6440 images and 6440 labels. The data set is randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0029] Step D: Use the training set and validation set to train the CCAU-Net network and optimize the parameters. The specific steps are as follows: Step D1: Input the training set obtained in step C3 into the encoder composed of asymmetric convolution layer, ReLU activation function and pooling layer in the CCAU-Net model to extract feature maps; Step D1 specifically includes the following steps: Step D1-1: Input the image to the first-stage encoder, and process it through two asymmetric convolution layers consisting of three parallel layers with convolution kernels of 3×3, 1×3 and 3×1 and ReLU activation function to obtain a 16-channel feature map; Step D1-2: The 2×2 max pooling layer in the first stage extracts the intermediate features and passes them to the second stage encoder; Step D1-3: The second stage encoder performs the same convolution process twice on the incoming feature map to obtain a 32-channel feature map; Step D1-4: The 2×2 max pooling layer in the second stage extracts the intermediate features and passes them to the third stage encoder; Step D1-5: The third stage encoder performs the same convolution process on the incoming feature map twice to obtain a 64-channel feature map; Step D1-6: The 2×2 max pooling layer in the third stage extracts the intermediate features and passes them to the fourth stage encoder; Step D1-7: The fourth stage encoder performs the same convolution process on the incoming feature map twice to obtain a 128-channel feature map; Step D1-8: The 2×2 max pooling layer in the fourth stage extracts the intermediate features and passes them to the fifth stage encoder; Step D1-9: The fifth stage encoder performs the same convolution process on the incoming feature map twice to obtain a 256-channel feature map; Step D2: The result feature map obtained in step D1 is passed into the model decoder, and the image resolution is improved through upsampling, convolution and convolution attention module processing; step D2 specifically includes the following steps: Step D2-1: By upsampling the feature map of the fifth stage and concatenating the feature map of the fourth stage, a 256-channel feature map entering the sixth stage decoder is obtained, and a 128-channel feature map is obtained through a convolutional attention module (CBAM) and a convolution operation; Step D2-2: By upsampling the feature map of the sixth stage and concatenating the feature map of the third stage, a 128-channel feature map entering the seventh stage decoder is obtained, and a 64-channel feature map is obtained through a convolutional attention module (CBAM) and a convolution operation; Step D2-3: By upsampling the feature map of the seventh stage and concatenating the feature map of the second stage, a 64-channel feature map entering the eighth stage decoder is obtained, and a 32-channel feature map is obtained through a convolutional attention module (CBAM) and a convolution operation; Step D2-4: By upsampling the feature map of the eighth stage and concatenating the feature map of the first stage, a 32-channel feature map entering the ninth stage decoder is obtained, and a 16-channel feature map is obtained through a convolutional attention module (CBAM) and a convolution operation; Step D3: After a 1×1 convolution, a 2-channel land-sea segmentation feature map and a coastline distribution feature map are obtained. The model performance is evaluated through the validation set after each epoch of training. After several iterative trainings, a trained network is obtained. Step D4: Calculate the loss between the predicted result and the true label, where the loss function adopts the cross entropy loss function, and uses adaptive moment estimation as the optimizer to optimize the network parameters to achieve the purpose of optimizing the model.

[0030] Step E: Input the coastline remote sensing image that needs to be extracted into the trained CCAU-Net network to obtain the binary prediction image result of land and sea segmentation.

[0031] Step F: The contour in the binary prediction result of land-sea segmentation is used as the land-sea boundary, and then the land-sea boundary binary image is vectorized to obtain the coastline extraction result as the final output result.

[0032] Table 1 Accuracy comparison of network models Model F1 / % OA / % IoU(Water) / % MIoU / % CCAU-Net 91.02 96.15 95.36 95.16 U-Net 89.78 95.22 93.43 93.52 DeepLabv3+ 88.25 94.87 93.16 92.74 SegNet 84.23 92.23 89.32 89.38 As shown in Table 1 and Figure 3 (in the label image and the extraction result image on the right side thereof, black represents land and white represents ocean), the binarization results obtained by the extraction method of the present invention are compared with the accuracy of each network model. The F1 score reaches 91.02%, the overall accuracy (ie, OA) reaches 96.15%, the category union ratio (ie, IoU) reaches 95.36%, and the average category union ratio (ie, MIoU) reaches 95.16%. Compared with the classic U-Net, DeepLabv3+ (the backbone network uses Xceotion), and SegNet, the F1 scores are increased by 1.24%, 2.77%, and 6.79%, respectively, the overall accuracy is increased by 0.93%, 1.28%, and 3.92%, respectively, the category union ratio is increased by 1.93%, 2.2%, and 6.04%, respectively, and the average category union ratio is increased by 1.64%, 2.42%, and 5.78%, respectively. The results of the present invention are obviously excellent.

[0033] Technical features not described in the present invention can be achieved through or by adopting existing technologies and will not be described in detail herein. Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by ordinary technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. A multi-band remote sensing coastline extraction method based on CCAU-Net, It is characterized in that The following steps are involved: Step A: Build the CCAU-Net network model; Specifically, an asymmetric convolution module is introduced to replace the standard square convolution of the U-Net model, and a convolutional attention module is integrated in the jump connection of the model to enhance the network model's focus on the key areas of the input features, improve the model's ability to fuse low-level semantics and high-level features, and improve the accuracy and overall performance of feature extraction in multi-band remote sensing images; the convolutional attention module concatenates channel attention and spatial attention along two independent dimensions, channel and space, and uses global average pooling and maximum pooling to obtain an attention map, which is then multiplied by the input feature map for adaptive feature refinement; Step B: Obtain 5 Landsat8 multispectral remote sensing images covering the sea-land boundary area, and pre-process the coastline remote sensing images to obtain images that meet the requirements for data extraction; Step C: Create neural network training samples, construct a data set and divide it into training set, validation set and test set; Step D: Use the training set and validation set to train the CCAU-Net network and optimize the parameters; Step E: Input the test set coastline remote sensing image that needs to be extracted into the trained CCAU-Net network to obtain the binary prediction image result of land and sea segmentation; Step F: The contour in the binary prediction image result of land-sea segmentation is used as the land-sea boundary, and then the land-sea boundary binary image is vectorized to obtain the coastline extraction result as the final output result.

2. The multi-band remote sensing coastline extraction method based on CCAU-Net according to claim 1, It is characterized in that Step C specifically includes the following steps: Step C1: In the image obtained in step B, the image is irregularly cropped to retain only the land and sea parts close to the coastline; Step C2: The images obtained in step C1 are cropped into multiple images of 256×256 pixels, retaining all bands. 920 images with both ocean and land are selected, and the water bodies are marked using the MNDWI water index. Step C3: Perform data augmentation on the images obtained in step C2 to obtain a data set of 6440 images and 6440 labels. The data set is randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:

1.

3. The multi-band remote sensing coastline extraction method based on CCAU-Net according to claim 1, It is characterized in that Step D specifically includes the following steps: Step D1: Input the training set obtained in step C3 into the encoder composed of asymmetric convolutional layer, ReLU activation function and pooling layer in the CCAU-Net model to extract feature maps; Step D2: The resulting feature map obtained in step D1 is passed into the model decoder, and processed by its upsampling, convolution and convolution attention modules to improve the image resolution; Step D3: After a 1×1 convolution, a 2-channel land-sea segmentation feature map and a coastline distribution feature map are obtained. The model performance is evaluated through the validation set after each epoch of training. After several iterative trainings, a trained network is obtained. Step D4: Calculate the loss between the predicted result and the true label, where the loss function adopts the cross entropy loss function, and uses adaptive moment estimation as the optimizer to optimize the network parameters to achieve the purpose of optimizing the model.

4. The multi-band remote sensing coastline extraction method based on CCAU-Net according to claim 3, It is characterized in that Step D1 specifically includes the following steps: Step D1-1: Input the image to the first-stage encoder, and process it through two asymmetric convolution layers consisting of three parallel layers with convolution kernels of 3×3, 1×3 and 3×1 and ReLU activation function to obtain a 16-channel feature map; Step D1-2: The 2×2 max pooling layer in the first stage extracts the intermediate features and passes them to the second stage encoder; Step D1-3: The second stage encoder performs the same convolution process twice on the incoming feature map to obtain a 32-channel feature map; Step D1-4: The 2×2 max pooling layer in the second stage extracts the intermediate features and passes them to the third stage encoder; Step D1-5: The third stage encoder performs the same convolution process on the incoming feature map twice to obtain a 64-channel feature map; Step D1-6: The 2×2 max pooling layer in the third stage extracts the intermediate features and passes them to the fourth stage encoder; Step D1-7: The fourth stage encoder performs the same convolution process on the incoming feature map twice to obtain a 128-channel feature map; Step D1-8: The 2×2 max pooling layer in the fourth stage extracts the intermediate features and passes them to the fifth stage encoder; Step D1-9: The fifth stage encoder performs two identical convolution operations on the incoming feature map to obtain a 256-channel feature map.

5. The multi-band remote sensing coastline extraction method based on CCAU-Net according to claim 3, It is characterized in that Step D2 specifically includes the following steps: Step D2-1: By upsampling the feature map of the fifth stage and connecting the feature map of the fourth stage, a 256-channel feature map entering the sixth stage decoder is obtained, and a 128-channel feature map is obtained through a convolutional attention module and a convolution operation; Step D2-2: By upsampling the feature map of the sixth stage and connecting the feature map of the third stage, a 128-channel feature map entering the seventh stage decoder is obtained, and a 64-channel feature map is obtained through a convolutional attention module and a convolution operation; Step D2-3: By upsampling the feature map of the seventh stage and connecting the feature map of the second stage, a 64-channel feature map entering the eighth stage decoder is obtained, and a 32-channel feature map is obtained through a convolutional attention module and a convolution operation; Step D2-4: By upsampling the feature map of the eighth stage and connecting the feature map of the first stage, a 32-channel feature map entering the ninth stage decoder is obtained, and a 16-channel feature map is obtained through the convolutional attention module and convolution operation.

Citation Information

Cited By

  • Remote sensing image water body and shadow classification method based on double-classification attention network

    CN121033675A