Remote sensing coastline extraction method, device, equipment and medium

Abstract

The invention discloses a remote sensing coastline extraction method and device, equipment and a medium, and relates to the technical field of deep learning, and the method comprises the steps: obtaining a target remote sensing image of a target region; inputting the target remote sensing image into the trained coastline extraction model to obtain a target coastline extraction result of the target area; the coastline extraction model is an improved DeeplabV < 3 + > model; according to the improved DeeplabV3 + model, a CBAM attention mechanism module is introduced in the process of outputting the backbone network of the original DeeplabV3 + model to a decoder, the CBAM attention mechanism module is introduced before up-sampling, and the coastline extraction precision is improved.

Description

Technical Field

[0001] The present application relates to the field of deep learning technology, and in particular to a remote sensing coastline extraction method, device, equipment and medium. Background Art

[0002] Shoreline surveys and data extraction are crucial for ecological protection, coastal planning, and disaster prevention. They reveal coastal ecological conditions, identify erosion and siltation, and guide the rational planning of coastal resources. Furthermore, shoreline change data lays the foundation for geomorphological and dynamic studies, enhances our understanding of coastal evolution, and provides scientific support for responding to natural disasters and climate change.

[0003] In recent years, deep learning, as an important branch of machine learning, has achieved rapid development. In particular, convolutional neural networks (CNNs), with their excellent feature extraction capabilities, have made breakthrough progress in multiple fields such as image classification, object detection, and semantic segmentation. In the task of coastline extraction, deep learning algorithms utilize their powerful generalization and multi-dimensional information fusion capabilities to achieve high-precision coastline extraction. Their automatic threshold segmentation and automatic feature extraction capabilities provide new possibilities for automated coastline extraction. However, current deep learning network models for coastline extraction suffer from the problems of large number of parameters and high number of floating-point operations (FLOPs). This results in large computational resources, making it difficult to achieve a balance between extraction accuracy and operational efficiency, which is not conducive to real-time automated coastline management. Summary of the Invention

[0004] The purpose of this application is to provide a remote sensing coastline extraction method, device, equipment and medium, which can improve the accuracy of remote sensing coastline extraction while taking into account efficiency, and provide a lightweight solution for real-time coastline monitoring.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] In a first aspect, the present application provides a remote sensing coastline extraction method, comprising:

[0007] Acquire target remote sensing images of the target area;

[0008] The target remote sensing image is input into the trained coastline extraction model to obtain the target coastline extraction result of the target area; the coastline extraction model is an improved DeeplabV3+ model; the coastline extraction model is an improved DeeplabV3+ model; the improved DeeplabV3+ model introduces the CBAM attention mechanism module in the process of outputting the backbone network of the original DeeplabV3+ model to the decoder, and introduces the CBAM attention mechanism module before upsampling.

[0009] Optionally, before inputting the remote sensing image of the target area into the trained coastline extraction model, the remote sensing coastline extraction method further includes:

[0010] Acquire a sample set; the sample set includes a plurality of sample remote sensing images and a sample coastline extraction result corresponding to each of the sample remote sensing images;

[0011] The coastline extraction model is trained using the sample set to obtain a trained coastline extraction model.

[0012] Optionally, obtaining a sample set includes:

[0013] Obtain original remote sensing images;

[0014] Preprocessing the original remote sensing image to obtain a sample remote sensing image;

[0015] The sample remote sensing image is labeled to obtain a label corresponding to the sample remote sensing image; the label is a sample coastline extraction result corresponding to the sample remote sensing image.

[0016] Optionally, the preprocessing includes image segmentation and image enhancement processing.

[0017] Optionally, the sample set is used to train the coastline extraction model to obtain a trained coastline extraction model, specifically comprising: using a loss function and the sample set to train the coastline extraction model to obtain a trained coastline extraction model.

[0018] Optionally, the loss function is a BCEWithLogitsLoss function.

[0019] Optionally, the remote sensing coastline extraction method is applicable to the extraction of different coastline types, including bedrock coastline, sandy coastline, silty coastline and artificial coastline.

[0020] In a second aspect, the present application provides a remote sensing coastline extraction device, comprising:

[0021] Target remote sensing image acquisition module, used to: acquire target remote sensing images of the target area;

[0022] The coastline extraction module is used to: input the target remote sensing image into a trained coastline extraction model to obtain the target coastline extraction result of the target area; the coastline extraction model is an improved DeeplabV3+ model; the improved DeeplabV3+ model introduces the CBAM attention mechanism module in the process of outputting the backbone network of the original DeeplabV3+ model to the decoder, and introduces the CBAM attention mechanism module before upsampling.

[0023] In a third aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned remote sensing coastline extraction method.

[0024] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which implements the above-mentioned remote sensing coastline extraction method when executed by a processor.

[0025] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0026] The present application provides a remote sensing coastline extraction method, apparatus, equipment and medium, which introduces a CBAM attention mechanism module from the output of the backbone network of the original DeeplabV3+ model to the decoder, and introduces the CBAM attention mechanism module before upsampling, so that the coastline extraction model can adaptively focus on key areas in the image, thereby giving the coastline extraction model higher flexibility and adaptability in the feature extraction process and improving the coastline extraction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0028] Figure 1 This is an application environment diagram of a remote sensing coastline extraction method in an embodiment of the present application;

[0029] Figure 2 A schematic flow chart of a remote sensing coastline extraction method provided in one embodiment of the present application;

[0030] Figure 3 A schematic diagram of the structure of the improved DeeplabV3+ model provided in one embodiment of the present application;

[0031] Figure 4 A schematic diagram of the inverted residual structure of the MobileNetV2 network provided in one embodiment of the present application;

[0032] Figure 5 A schematic diagram of the stripe pooling layer structure provided in one embodiment of the present application;

[0033] Figure 6 A schematic diagram of the CBAM attention mechanism module structure provided in one embodiment of the present application;

[0034] Figure 7 A schematic diagram of the coastline extraction result labeling process provided in one embodiment of the present application;

[0035] Figure 8 A schematic diagram showing the comparison results of the segmentation accuracy indicators of various models provided in an embodiment of the present application;

[0036] Figure 9 A schematic diagram showing a comparison of performance indicators of various models provided in an embodiment of the present application;

[0037] Figure 10 Schematic diagram of visual comparison results of five typical shoreline extractions of various models provided in one embodiment of the present application;

[0038] Figure 11 A schematic diagram of the functional modules of a remote sensing coastline extraction device provided in one embodiment of the present application;

[0039] Figure 12 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0041] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0042] The remote sensing coastline extraction method provided in the embodiment of the present application can be applied to Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set up separately, integrated on the server 104, or placed on the cloud or other servers. The terminal 102 can send the target remote sensing image of the target area to the server 104. After the server 104 receives the target remote sensing image of the target area, the server 104 inputs the target remote sensing image into the trained coastline extraction model to obtain the target coastline extraction result of the target area. The server 104 can feed back the target coastline extraction result for the target area to the terminal 102. In addition, in some embodiments, the remote sensing coastline extraction method can also be implemented separately by the server 104 or the terminal 102. For example, the terminal 102 can directly perform coastline extraction on the target remote sensing image of the target area, or the server 104 can obtain the target remote sensing image of the target area from the data storage system and perform coastline extraction on the target remote sensing image of the target area.

[0043] The terminal 102 may be, but is not limited to, various desktop computers and laptop computers. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers, or may be a cloud server.

[0044] In an exemplary embodiment, Figure 2 As shown, a remote sensing coastline extraction method is provided. The method is executed by a computer device. Specifically, it can be executed by a computer device such as a terminal or a server alone, or by a terminal and a server together. In the embodiment of the present application, the method is applied to Figure 1 The server 104 in the example is used as an example to illustrate the process, including the following steps 201 to 202.

[0045] Step 201: Acquire a target remote sensing image of a target area.

[0046] Step 202: Input the target remote sensing image into the trained coastline extraction model to obtain the target coastline extraction result of the target area; the coastline extraction model is an improved DeeplabV3+ model; the improved DeeplabV3+ model introduces the CBAM attention mechanism module in the process of outputting the backbone network of the original DeeplabV3+ model to the decoder, and introduces the CBAM attention mechanism module before upsampling.

[0047] By implementing the above steps 201 to 202, a convolutional block attention module (CBAM) is introduced before the output of the original DeeplabV3+ model backbone network is transmitted to the decoder, and the CBAM attention mechanism module is introduced before upsampling, so that the coastline extraction model can adaptively focus on key areas in the image, thereby giving the coastline extraction model higher flexibility and adaptability in the feature extraction process. Based on the above improved DeeplabV3+ model, the coastline extraction is performed on the target remote sensing image of the target area, thereby optimizing the coastline extraction accuracy.

[0048] In another exemplary embodiment of the present application, before inputting the remote sensing image of the target area into the trained coastline extraction model, the remote sensing coastline extraction method further includes the following steps 301 and 302:

[0049] Step 301: Acquire a sample set; the sample set includes a number of sample remote sensing images and a sample coastline extraction result corresponding to each of the sample remote sensing images.

[0050] Step 302: Using the sample set to train the coastline extraction model to obtain a trained coastline extraction model.

[0051] Step 301 may specifically include the following steps 401 to 403:

[0052] Step 401: Acquire original remote sensing images;

[0053] Step 402: Preprocess the original remote sensing image to obtain a sample remote sensing image; the preprocessing includes image segmentation and image enhancement.

[0054] Step 403: labeling the sample remote sensing image to obtain a label corresponding to the sample remote sensing image; the label is a sample coastline extraction result corresponding to the sample remote sensing image.

[0055] Field surveys were conducted, image control points were deployed, flight routes were determined, and complete island coastline data was acquired. Oblique photogrammetry uses an unmanned aerial vehicle (UAV) equipped with a digital camera to capture the ground in area X from multiple angles, acquiring raw remote sensing imagery. Specifically, a CW-15 vertical take-off and landing fixed-wing UAV equipped with a CA-502R oblique-view camera was used as the primary aerial photography tool. To improve the georeferencing accuracy of the aerial imagery, 96 ground control points (GCPs) were collected during the data collection process. These control points were measured using a combination of RTK-GNSS and CORS, and used in image processing software to correct and enhance the image's georeferencing accuracy.

[0056] The remote sensing coastline extraction method provided in this application is applicable to the extraction of different coastline types, including bedrock coastline, sandy coastline, silt coastline and artificial coastline.

[0057] Coastlines are generally categorized into two main types based on their formation and characteristics: natural and artificial. Artificial coastlines in Area X primarily include artificially constructed or modified shorelines, such as docks and seawalls. Area X's natural coastlines can be further divided into three subtypes: bedrock, sandy, and muddy. Bedrock coastlines are composed of rocks such as granite, sandstone, or shale, and are characterized by high stability and resistance to wave and tidal erosion. Sandy coastlines are composed of loose particles such as sand and silt, exhibiting dynamic characteristics and susceptible to ocean dynamics and climate change. Muddy coastlines are composed of fine-grained sediments such as silt and clay, exhibiting strong sedimentation and erosion, and are often home to mangroves and coastal wetland vegetation. Given that Area X's different shoreline types vary in length and frequently alternate spatially, this application introduces a third shoreline type: composite coastline, which refers to areas where natural and artificial shorelines intersect. This classification not only enriches the diversity of coastline types, but also more comprehensively reflects the complexity of coastline types.

[0058] After data collection was completed, meticulous manual screening was performed to select 1,000 highly representative and clear photogrammetric images as the raw remote sensing image data. To ensure a balanced distribution of land and sea sample points and to avoid bias and overfitting caused by sample imbalance during model training, the raw remote sensing images were precisely cut into equal-sized image blocks of 4,000 × 4,000 pixels. These cut images provided standardized, high-quality data input for further data processing and model training.

[0059] The Labelme plug-in was used to perform detailed labeling on the sample remote sensing images. The land areas were clearly highlighted in green, while the ocean areas were distinguished in blue, ensuring the clear identification of different geographical features in the image. The coastline extraction result labels are as follows: Figure 7 The annotated raw data is stored in JSON format and converted to PNG format using the Pycharm development environment (the converted annotation file is used as the label of the sample remote sensing image), achieving seamless compatibility with subsequent analysis tools. Figure 7 (a) in the equation is the ground control point. Figure 7 (b) is the UAV route planning. Figure 7 (c) in the figure is the label making process.

[0060] Image enhancement processing: In order to further enrich the diversity of samples and improve the generalization ability of the model, this application implements a series of data enhancement technologies and comprehensively processes the sample set (the sample set includes the training set and the validation set). The specific enhancement methods implemented include random rotation of the image (90°, 180°, 270°), horizontal and vertical flipping, and the introduction of Gaussian noise. These technical means effectively expand the diversity of training samples. After image enhancement processing, a total of 7,000 images were generated, and based on a 9:1 ratio, a training set containing 6,300 sample remote sensing images and a validation set containing 700 sample remote sensing images were constructed.

[0061] The CBAM attention mechanism module structure is as follows Figure 6 As shown in (a) in the figure, it consists of a channel attention module (CAM) and a spatial attention mechanism (SAM).

[0062] The structure of the channel attention module is as follows Figure 6 As shown in (b) in the figure: The statistical information between channels is captured by global average pooling (AvgPool) and maximum pooling (MaxPool) operations, and the global average pooling features and maximum pooling features are obtained respectively. Then, the dependency between channels is learned by multi-layer perceptron (MLP), and channel weights are generated. Each channel is weighted to obtain global weighted features and maximum pooling weighted features respectively. The global weighted features and maximum pooling weighted features are added together to obtain the added features. The added features are normalized by the Sigmoid activation function to obtain the channel attention weight map. The process is shown in formula (1).

[0063] M C (F1)=σ(MLP(AvgPool(F1))+MLP(MaxPool(F1))) (1);

[0064] Where F1 is the input feature map of the channel attention module; AvgPool(F1) and MaxPool(F1) respectively represent the global information of each channel obtained by global average pooling and global maximum pooling, that is, global average pooling features and maximum pooling features respectively; MLP represents multi-layer perceptron, σ is Sigmoid activation function, M C (F1) is the channel attention weight map.

[0065] The channel attention weight map and the input of the CBAM attention mechanism module are multiplied to obtain the first multiplied feature; the spatial feature of the first multiplied feature is extracted through the spatial attention module to obtain the spatial attention weight map.

[0066] The structure of the spatial attention module is as follows Figure 6 As shown in (c), it includes the maximum pooling layer, the average pooling layer, and the 7×7 convolution layer.

[0067] The first multiplication feature extracts spatial information through the maximum pooling layer and the average pooling layer respectively to obtain the maximum pooling channel feature and the average pooling channel feature; the maximum pooling channel feature and the average pooling channel feature are spliced ​​to obtain the splicing feature; the splicing feature is processed by the 7×7 convolution layer to generate spatial weights, highlight the key areas in the image, and obtain the spatial attention weight map. The process is shown in formula (2).

[0068] M S (F2)=σ(f 7×7 ([AvgPool(F2);MaxPool(F2)])) (2);

[0069] Where F2 is the input feature map; AvgPool(F2) and MaxPool(F2) represent the spatial features captured by average pooling and maximum pooling, respectively, that is, they represent the average pooling channel features and the maximum pooling channel features respectively; f 7×7 represents 7×7 convolution, M S (F2) is the spatial attention weight map.

[0070] The first multiplied feature and the spatial attention weight map are multiplied to obtain the second multiplied feature, which is the output of the CBAM attention mechanism module.

[0071] Preferably, the improved DeeplabV3+ model also replaces the backbone network of the original DeeplabV3+ model with the MobileNetV2 network and introduces a strip pooling layer in the ASPP module. The structure of the improved DeeplabV3+ model is as follows: Figure 3 In Part A of the paper, the improved DeeplabV3+ model is introduced below:

[0072] Figure 3 Part B in the figure is the inverse residual structure of the MobileNetV2 network (linear bottleneck inverse residual structure), and its structure is as follows: Figure 4 As shown in the figure, 1×1 convolution is used to increase the dimension of the model to enhance feature representation capabilities. Depthwise Separable Convolution is then used to reduce the dimension, thereby reducing the number of parameters and computational complexity while maintaining sensitivity to key visual features. Finally, residual connections are used to fuse information with the original features. This not only significantly reduces the number of model parameters and computational complexity, but also improves the model's ability to capture key visual features.

[0073] Figure 3 The C part in the figure is the strip pooling layer, and its structure is as follows: Figure 5 As shown in Figure 1, the main process of the strip pooling layer includes: (1) Input feature map: Input feature map of size C×H×W, i.e., the output feature of the MobileNetV2 network; (2) Strip pooling: The feature map is strip pooled in the horizontal and vertical directions respectively, with the pooling kernel of 1×N or N×1, to obtain feature maps of size H×1 and 1×W. The pooling result is the average value of the pixels in the strip area; (3) Feature enhancement: The strip pooled feature map is processed by 1D convolution to restore the number of channels of the input feature map, i.e., C×H×W; (4) Feature fusion and activation: Features are fused by 1×1 convolution, and nonlinear transformation is performed using Sigmoid activation function, and finally pixel-by-pixel multiplication with the input feature map is obtained to obtain the final output. Strip pooling uses a long and narrow kernel (1×N or N×1) to expand the network's receptive field and more effectively capture long-range dependencies in the image.

[0074] Figure 3 Part D is the CBAM attention mechanism module.

[0075] Step 302 specifically includes: using a loss function and the sample set to train the coastline extraction model to obtain a trained coastline extraction model. The loss function is a BCEWithLogitsLoss function.

[0076] In order to improve numerical stability and simplify the model training process, while optimizing the performance of binary classification tasks, this application uses BCEWithLogitsLoss as the loss function. This function is a Sigmoid activation function combined with a binary cross entropy loss (BCELoss), which enables the coastline extraction model to directly learn the unnormalized predicted values ​​(logits). This method avoids the gradient vanishing or exploding problem caused by the output range of the Sigmoid function (between 0 and 1). The loss function is defined as follows:

[0077] loss=y×(-log(σ(z))+(1-y)×(-log(1-σ(z)))) (3);

[0078] Where loss represents the loss function, y is the true label, and z is the predicted output of the coastline extraction model, that is, the unnormalized predicted value (logits).

[0079] During this process, the network continuously optimizes its parameter weights through continuous learning. The difference between the generated feature maps and the original labels is quantified using a loss function. This loss value serves as a feedback signal, guiding the network to self-correct and improve performance in subsequent training iterations. Our goal is to minimize this loss so that the feature maps output by the model are as close to the true labels as possible in detail. Through this iterative optimization process, the network ultimately converges to an optimal set of parameters and weights, ensuring high performance and accuracy in the shoreline extraction task on the X-region dataset.

[0080] At the same time, this application uses four evaluation indicators to evaluate the performance of the model, namely pixel accuracy (PA), recall rate (Recall), F1 score (F1 Score) and intersection over union (IoU). The specific calculation is shown in formula (4):

[0081]

[0082] Where TP means the model predicts a positive example and it is actually a positive example, FP means the model predicts a positive example but it is actually a negative example, FN means the model predicts a negative example but it is actually a positive example, and TN means the model predicts a negative example and it is actually a negative example.

[0083] The experimental hardware environment for this application is detailed in Table 1. All experiments were conducted strictly according to the configurations listed in the table. Model training used the Adam optimization algorithm, with an initial learning rate of 0.0001, a batch size of 8, and an initial epoch of 200. To monitor the training progress and preserve the optimal model state, model parameters were saved every four epochs for subsequent evaluation and selection of the optimal model.

[0084] Table 1 Training environment hardware configuration parameters

[0085] Experimental environment project Configuration operating system Linux 5.4.0 operating system Development language 3.8 Deep Learning Framework Pytorch 1.12.0 CPU AMD EPYC 7542 GPU NVIDIA RTX3090 (24G)

[0086] To comprehensively evaluate the performance improvements of the improved DeeplabV3+ model, this application systematically compared and analyzed three advanced attention mechanisms: channel attention (CA), squeezed excitation network (SE-Net), and the CBAM attention module. These mechanisms were strategically embedded in two key stages of the model: shallow feature extraction and deep feature upsampling. The shallow layers capture the underlying visual patterns of the image, while the deep layers integrate and refine these features.

[0087] The results shown in Table 2 reveal the significant impact of integrating three different attention mechanism modules on the performance of the improved DeeplabV3+ model. The CBAM attention mechanism performs particularly well, improving the accuracy (PA) by 3.1% and the intersection over union (IoU) by 8.2% compared to the original network. Following closely behind the CBAM attention mechanism is the SE-Net attention mechanism, which improves PA and IoU by 2.9% and 7.7%, respectively. While the CA attention mechanism achieves a smaller improvement, it still achieves a 1.9% increase in PA and 6.6% in IoU.

[0088] Table 2 Attention mechanism selection results

[0089]

[0090] It should be noted that in Table 2, “√” indicates adding the corresponding module, “1” indicates adding the module after shallow feature extraction, and “2” indicates adding the module before deep feature upsampling.

[0091] In this application, the improved DeeplabV3+ model introduces an attention mechanism, especially the Convolutional Block Attention Module (CBAM), which significantly improves the model's ability to recognize coastline features, as shown in Table 2. This finding emphasizes the importance of considering both channel and spatial information in the process of capturing key image features.

[0092] Furthermore, integrating the attention mechanism before the deep feature upsampling stage can achieve even more significant performance improvements compared to integrating it only after the shallow feature extraction stage. This is because deep features contain richer semantic information, which often contains important semantic content of the image, such as the shape and texture of objects. Introducing the attention mechanism at this stage allows the model to focus more closely on key areas of the image, strengthening the representation of this semantic information. As a result, when performing feature upsampling, the model can more effectively recover high-quality, semantically rich image details.

[0093] Furthermore, the deep feature upsampling stage is a critical phase for fusing high-level semantic information with low-level pixel-level details. Applying the attention mechanism at this stage helps the model more accurately identify which low-level features are critical, thereby selectively enhancing these features during the upsampling process while suppressing noise and unimportant information.

[0094] In order to verify the effectiveness of the improved DeeplabV3+ model in the shoreline extraction task, this application conducted a detailed comparison of the improved DeeplabV3+ model with FCN-8, Segnet, U-Net and the original DeeplabV3+ model. All models were evaluated under the same training parameters and conditions to ensure the fairness of the comparison and the reliability of the results. For the convenience of representation, the original DeeplabV3+ model is denoted as DL3 and the improved DeeplabV3+ model is denoted as DL3-I. The comparison results of each model in terms of segmentation accuracy indicators are shown in the figure below. Figure 8 As shown in Table 3.

[0095] Table 3 Comparison of several different models in terms of parameter quantity, computational complexity, and other performance

[0096]

[0097]

[0098] Figure 8 Table 3 comprehensively shows the performance of each model in terms of segmentation accuracy indicators. Figure 8 Figures (a), (b), (c), and (d) show the comparison results of each model on pixel accuracy (PA), recall (Recall), F1-score (F1-Score), and intersection over union (IoU) metrics. The improved DeeplabV3+ model performs exceptionally well across all evaluation metrics, achieving 98.7% PA, 97.7% Recall, 98.0% F1-Score, and 96.2% IoU. Compared to the original DeeplabV3+ model, these metrics improved by 4.7%, 5.5%, 7%, and 12.7%, respectively. Compared with FCN-8, Segnet and U-Net, the improved DeeplabV3+ model achieves significant improvements of 4%, 10.3% and 7.1%, 6.3%, 16.2% and 10.2%, 6.0%, 15.7% and 10.7%, and 11.0%, 26.2% and 18.7% in PA, Recall, F1-Score and IoU, respectively.

[0099] In practical applications, the efficiency of the model is a crucial consideration, which directly affects the deployability and practicality of the model. Figure 9 The improved DeeplabV3+ model was compared with other existing models in terms of parameter count and floating-point computational power (GFLOPS). The improved DeeplabV3+ model, with 6.61M parameters and 6.7GFLOPS of floating-point computational power, showed the lowest resource consumption among all the models involved in the verification.

[0100] The specific comparison results are as follows Figure 9 As shown in the figure, compared to the FCN-8 model, the improved DeeplabV3+ model achieves an approximately 80% reduction in parameters and a 92% reduction in floating-point computations. Compared to the SegNet model, the reductions in parameters and floating-point computations reach 80% and approximately 86%, respectively. Compared to the U-Net model, the improved DeeplabV3+ model reduces parameters by 80% and floating-point computations by approximately 88% through its linear bottleneck inverse residual structure. Compared to the original DeepLabV3+ model, the improved DeeplabV3+ model also reduces parameters by approximately 88% and floating-point computations by approximately 67%. Figure 9 In the figure, the lower left point indicates that the model has a small number of parameters and a small amount of floating-point calculations. Tier I represents the model belongs to the first tier, Tier II represents the model belongs to the second tier, and Tier III represents the model belongs to the third tier.

[0101] The following is a comprehensive evaluation of the performance impact of the introduced CBAM attention mechanism module, MobileNetV2 network, and strip pooling layer module on the improved DeeplabV3+ model. Detailed data is shown in Table 4, where "√" indicates the addition of the corresponding module.

[0102] Table 4 Performance evaluation of the improved DeeplabV3+ model

[0103]

[0104] In terms of coastline extraction accuracy, the improved DeeplabV3+ model demonstrated excellent performance. The improved DeeplabV3+ model introduced a strip pooling layer in the ASPP module, which effectively captured long-distance dependencies through horizontal and vertical pooling operations, solving the problem of insufficient processing of detailed features by traditional pooling. Targeting the feature capture of linear features such as coastlines, the F1-Score and IoU indicators of the coastline extraction model were significantly improved by 4.4% and 7.5% respectively, greatly enhancing the accuracy of feature extraction. In addition, the CBAM attention mechanism module reduces missed detections and false detections in complex backgrounds through fine-grained channel attention and spatial attention mechanisms, giving the coastline extraction model the ability to adaptively focus on key features, significantly improving the coastline extraction model's ability to express key features, with the F1-Score and IoU indicators further improved by 2.8% and 4.8% respectively.

[0105] The improved DeeplabV3+ model also performed exceptionally well in terms of model efficiency. Using the MobileNetV2 network as its backbone, it reduced GFLOPS by 48.89 and the number of parameters by 14.19M, significantly reducing the number of parameters and floating-point computations. Compared to traditional models and other comparable models, it is more resource-efficient.

[0106] However, in this application, the improved DeeplabV3+ model shows obvious superiority over the U-Net model. This result is attributed to the higher resolution of the dataset used in this application and the richer image texture and structural levels.

[0107] By replacing the backbone network of the original DeeplabV3+ model with the MobileNetV2 network, and using 1×1 convolution for dimensionality increase and depth-wise separable convolution for dimensionality reduction, the number of parameters (reduced by 80%) and floating-point operations (reduced by 88%) are significantly reduced. At the same time, the segmentation accuracy indicators (PA, IoU, etc.) are better than the ResNet backbone network. This not only greatly reduces the number of model parameters and computational complexity, but also improves the model's ability to capture key visual features. In addition, by introducing the Strip Pooling layer in the ASPP module, the model can be trained to detect and predict key visual features. Layer), which effectively expands the network's receptive field and enables it to capture long-distance dependencies in the image. Experiments show that in complex backgrounds (such as composite coastline areas), the IoU of the improved DeeplabV3+ model is improved by 12.7%, and it can clearly distinguish the interlaced areas of natural coastlines and artificial coastlines. The CBAM attention mechanism module is introduced before the output of the backbone network to the decoder, and the CBAM attention mechanism module is introduced before upsampling, so that the coastline extraction model can adaptively focus on the key areas in the image, thereby giving the coastline extraction model higher flexibility and adaptability in the feature extraction process. Based on the above-mentioned improved DeeplabV3+ model, the target remote sensing image of the target area is extracted from the coastline, which optimizes the coastline extraction accuracy and improves the operation efficiency of the extraction process, solving the problems of low accuracy and efficiency of the existing model.

[0108] As a result, the improved DeeplabV3+ model not only achieves significant improvements in segmentation accuracy but also in model efficiency, fully demonstrating its broad application prospects and practical value in complex visual tasks such as high-precision coastline extraction. These improvements enable the coastline extraction model to not only provide more accurate coastline extraction results but also operate efficiently in resource-constrained environments, providing strong technical support for coastline monitoring and management.

[0109] To verify the performance of the improved DeeplabV3+ model in different coastline extraction scenarios, this application uses FCN-8, Segnet, U-Net and the original DeeplabV3+ model as comparison objects to perform coastline extraction on the validation set. Among them, the validation set island coastline images are carefully classified into three types: natural coastline, artificial coastline and composite coastline, and representative areas are selected from each type for in-depth analysis and comparison. The comparison results are shown in the figure below. Figure 10 shown.

[0110] Figure 10 The results show that although all five algorithms can identify the approximate location of the coastline to a certain extent, the improved DeeplabV3+ model performs well in different coastline extraction tasks, especially in terms of details and boundary accuracy. Specifically:

[0111] In bedrock-type natural shoreline extraction, the improved DeepLabV3+ model was more accurate in segmenting small pieces of bedrock, effectively reducing misclassifications and omissions, and achieving high consistency with the true labels, compared to the FCN-8, Segnet, U-Net, and original DeepLabV3+ models. In sandy shoreline extraction, the FCN-8, U-Net, and Segnet models lacked boundary information, resulting in fragmented and blurred results. The U-Net model, in particular, misidentified white caps formed by seawater impacting the beach as land, leading to significant misclassification. The improved DeepLabV3+, on the other hand, was largely consistent with the labels. In silt-type shoreline extraction, the Segnet and U-Net models experienced misdetections and omissions due to ghosting, as the spectral characteristics of silt shoals are similar to those of land. The improved model achieved the best results with the fewest misdetections and omissions.

[0112] Network extraction is more difficult for extracting artificial shorelines with complex structures. The FCN-8 model mistakenly classifies part of the seawater as land due to the loss of boundaries. In the early feature extraction stage, the Segnet and U-Net models are unable to accurately identify the detailed features of the artificial shoreline due to their small receptive fields, resulting in fragmentation and blurring. Although the DeepLabV3+ model can clearly segment the main part of the dock, it exhibits fragmentation when identifying small protruding buildings on the dock. In contrast, the improved DeeplabV3+ model performs best in extracting the main body of the dock, accurately identifying small protruding buildings on the dock without fragmentation, and with the least missed detections and false detections.

[0113] In complex coastline extraction, the FCN-8 model misidentified blue houses as seawater, resulting in poor visual quality. The U-Net model, due to its small receptive field, lost some boundary information and mistakenly classified parts of seawater as land. SegNet and the original DeepLabV3+ models exhibited blurring when processing complex coastline structures due to insufficient extraction capabilities. The improved model demonstrated the most complete segmentation results, avoiding the blurring, fragmentation, and loss of boundary information common in other models.

[0114] This application introduces the CBAM attention mechanism during the model's shallow feature extraction and deep feature upsampling stages. Compared to the SE-Net and CA attention mechanisms, the CBAM attention mechanism significantly improves shoreline extraction accuracy, with improvements in the PA metric of 0.2% and 1.2%, respectively, and in the IoU metric of 0.5% and 1.6%, respectively. Furthermore, experimental results show that integrating the attention mechanism before the deep feature upsampling stage achieves even more significant performance improvements than integrating it only after shallow feature extraction.

[0115] The improved DeeplabV3+ model demonstrates superior performance across a wide range of shoreline extraction tasks, particularly when dealing with artificial and composite shorelines with complex textures and spectral characteristics. Leveraging its advanced feature extraction capabilities, the model effectively distinguishes between shadows and water, significantly reducing missed and false detections. It avoids the blurring, fragmentation, and loss of boundary information common in other models, thereby reducing the risk of misclassification and delivering the most complete shoreline segmentation results.

[0116] The improved DeeplabV3+ model also has broad application value. It can be used to analyze disaster prevention and emergency response to coastal erosion, siltation, and storm surges, as well as for environmental monitoring and assessment of sea level rise. By providing accurate shoreline data, the model can provide fundamental data support for rational coastal planning and management, contributing to the sustainable development and utilization of coastal resources and preventing environmental damage caused by overexploitation.

[0117] After the above process, a trained remote sensing coastline extraction model can be obtained. After obtaining the target remote sensing image of the target area, the target remote sensing image is input into the trained coastline extraction model to obtain the target coastline extraction result of the target area.

[0118] The flow of image data and the interaction between its various components are as follows: The input target remote sensing image is first passed to the MobileNetV2 network for feature extraction. The MobileNetV2 network outputs shallow features and deep features through its inverted residual structure. Shallow features are mainly extracted in the first few layers of the MobileNetV2 network. They are responsible for capturing basic visual patterns in the image, such as edges and simple textures. Deep features are extracted in the later layers of the MobileNetV2 network. These features can capture more complex visual patterns and rich semantic information, providing important clues for the precise positioning of the coastline.

[0119] Next, the shallow features enter the ASPP module for further processing. The ASPP module consists of multiple submodules, including a strip pooling layer module, a 1×1 convolution module, and three 3×3 convolution modules with different dilation rates (6, 12, and 18), plus a normal pooling layer. These submodules work together to perform multi-scale processing on shallow features, thereby enhancing the coastline extraction model's ability to recognize objects of different scales. The processed shallow features are first compressed by a 1×1 convolution, then enhanced by the CBAM attention mechanism module, and finally adjusted to the image resolution through upsampling to match the size of the original input image, ensuring the clarity of the coastline details.

[0120] At the same time, deep features are first processed by the CBAM attention mechanism module to enhance the ability to capture key semantic information. They are then compressed using 1×1 convolutions to reduce the number of channels. These deep features are then concatenated with the upsampled shallow features, and the number of channels is further compressed using 3×3 convolutions. Finally, this fusion of shallow and deep features is upsampled to produce the final feature map, which is the target coastline extraction result for the target area. This provides a highly accurate segmentation map for coastline extraction in the target area.

[0121] The present application also provides an application scenario, which applies the above-mentioned remote sensing coastline extraction method. Specifically: the remote sensing coastline extraction method provided in this embodiment can be applied in the coastline extraction scenario. The coastline extraction scenario includes an image acquisition link and a coastline extraction link; the target remote sensing image of the target area enters the coastline extraction link from the image acquisition link, and the corresponding target coastline extraction result is obtained through human-computer collaboration. The remote sensing coastline extraction method provided in this embodiment belongs to the machine marking link in the coastline extraction link. Specifically, in the coastline extraction link process for the target remote sensing image, the target remote sensing image can be input into the trained coastline extraction model to obtain the target coastline extraction result of the target area.

[0122] Based on the same inventive concept, embodiments of the present application also provide a remote sensing coastline extraction device for implementing the aforementioned remote sensing coastline extraction method. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of the remote sensing coastline extraction device provided below can be found in the aforementioned limitations of the remote sensing coastline extraction method and will not be further elaborated here.

[0123] In an exemplary embodiment, Figure 11 As shown, a remote sensing coastline extraction device is provided, comprising:

[0124] The target remote sensing image acquisition module T1 is used to: acquire the target remote sensing image of the target area;

[0125] The coastline extraction module T2 is used to: input the target remote sensing image into the trained coastline extraction model to obtain the target coastline extraction result of the target area; the coastline extraction model is an improved DeeplabV3+ model; the improved DeeplabV3+ model introduces the CBAM attention mechanism module in the process of outputting the backbone network of the original DeeplabV3+ model to the decoder, and introduces the CBAM attention mechanism module before upsampling.

[0126] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 12 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store coastline extraction data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a remote sensing coastline extraction method is implemented.

[0127] Those skilled in the art will understand that Figure 12The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0128] In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0129] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0130] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0131] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0132] The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may include, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic units, data processing logic units based on quantum computing, and the like.

[0133] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0134] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A remote sensing coastline extraction method, characterized in that: The remote sensing coastline extraction method comprises: Acquire target remote sensing images of the target area; Inputting the target remote sensing image into a trained coastline extraction model to obtain a target coastline extraction result of the target area; The coastline extraction model is an improved DeeplabV3+ model; the improved DeeplabV3+ model introduces a CBAM attention mechanism module in the process of outputting the backbone network of the original DeeplabV3+ model to the decoder, and introduces the CBAM attention mechanism module before upsampling.

2. The remote sensing coastline extraction method according to claim 1, characterized in that: Before inputting the remote sensing image of the target area into the trained coastline extraction model, the remote sensing coastline extraction method further includes: Acquire a sample set; the sample set includes a plurality of sample remote sensing images and a sample coastline extraction result corresponding to each of the sample remote sensing images; The sample set is used to train the coastline extraction model to obtain a trained coastline extraction model.

3. The remote sensing coastline extraction method according to claim 2, characterized in that: Get the sample set, including: Obtain original remote sensing images; Preprocessing the original remote sensing image to obtain a sample remote sensing image; The sample remote sensing image is labeled to obtain a label corresponding to the sample remote sensing image; the label is a sample coastline extraction result corresponding to the sample remote sensing image.

4. The remote sensing coastline extraction method according to claim 3, characterized in that: The preprocessing includes image cutting and image enhancement processing.

5. The remote sensing coastline extraction method according to claim 2, characterized in that: The coastline extraction model is trained using the sample set to obtain a trained coastline extraction model, specifically including: The coastline extraction model is trained using the sample set by utilizing the loss function to obtain a trained coastline extraction model.

6. The remote sensing coastline extraction method according to claim 5, characterized in that: The loss function is the BCEWithLogitsLoss function.

7. The remote sensing coastline extraction method according to claim 1, characterized in that: The remote sensing coastline extraction method is suitable for extracting different coastline types, including bedrock coastline, sandy coastline, silty coastline and artificial coastline.

8. A remote sensing coastline extraction device, characterized in that: The remote sensing coastline extraction device comprises: The target remote sensing image acquisition module is used to: acquire the target remote sensing image of the target area; The coastline extraction module is used to: input the target remote sensing image into the trained coastline extraction model to obtain the target coastline extraction result of the target area; the coastline extraction model is an improved DeeplabV3+ model; the improved DeeplabV3+ model introduces a CBAM attention mechanism module in the process of outputting the backbone network of the original DeeplabV3+ model to the decoder, and introduces the CBAM attention mechanism module before upsampling.

9. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the remote sensing coastline extraction method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the remote sensing coastline extraction method according to any one of claims 1 to 7 is implemented.

Images