Sea-land boundary extraction method based on remote sensing image

Abstract

The invention discloses a sea-land boundary extraction method based on a remote sensing image. The method comprises the following steps: step 1, obtaining a remote sensing image of a sea-land boundary to be extracted and radar data; step 2, acquiring a remote sensing image data set after data enhancement; obtaining a backscattering coefficient through the radar data; step 3, obtaining an average value of typical ground feature spectral reflectivity in the sea-land boundary to be extracted, and constructing an optimal water body index; step 4, constructing an NIWI-Net network, performing channel splicing on the optimal water body index and the output of a decoder of the U-Net network to serve as network output, and forming the NIWI-Net network; step 5, processing the real-time data through the trained NIWI-Net network; and step 6, performing convolution operation on the output of the NIWI-Net network to obtain a sea-land boundary extraction result graph after edge enhancement. According to the method, end-to-end edge enhancement is realized, the problem of inaccurate boundary region extraction is solved, and particularly in a coast region with large suspended sediment concentration difference, the definition and classification precision of the boundary are remarkably improved.

Description

Technical Field

[0001] The invention relates to the field of image processing technology, and in particular to a method for extracting land and sea boundaries based on remote sensing images. Background Art

[0002] The coastal zone, located at the junction of ocean and land, is a vital component of Earth's ecosystem and a valuable land resource, providing the foundation for marine development and economic development. A key landmark of the coastal zone is the coastline. Coastline extraction is fundamental to coastal and marine resource management, regional environmental monitoring in coastal areas, and sustainable development planning. Coastlines also provide a crucial basis for demarcating land and water resources, providing crucial information for shoreline erosion, automated navigation, and water depth monitoring. Land-sea boundary extraction is a prerequisite and foundation for research on coastline monitoring and coastline evolution.

[0003] With the continuous maturity of remote sensing technology, the extraction of land and sea boundaries from optical remote sensing imagery has flourished. Currently, extraction methods can be categorized into two main types: visual interpretation and automatic interpretation. Visual interpretation, the most basic and reliable initial method, relies on the human eye to identify the coastline in the image. While this method offers advantages in terms of simplicity, it also suffers from high workload and susceptibility to subjective factors. Automatic interpretation methods are primarily categorized as follows: threshold segmentation, edge detection, active contour models, and neural network classification. The first three methods divide the image into regions based on characteristics such as grayscale, contour, color, and texture, achieving segmentation by aligning pixels within regions with similarities and pixels between regions with differences. The other type involves land and sea segmentation methods that utilize deep learning to construct multi-level convolutional neural networks. These networks continuously learn and autonomously extract features through forward convolution, achieving land and sea segmentation. However, the complex and varied coastlines in remote sensing images, coupled with the presence of numerous elongated docks in port areas, can easily lead to inaccurate extraction of these boundary regions. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention provides a method for extracting land and sea boundaries based on remote sensing images to solve the technical problem of poor boundary area recognition effect in the existing technology.

[0005] The present invention provides a method for extracting land-sea boundaries based on remote sensing images, comprising the following steps:

[0006] Step 1: Obtain remote sensing images and radar data of the sea and land boundaries to be extracted, and preprocess the remote sensing images;

[0007] Step 2: Crop the preprocessed remote sensing image to construct a data set, perform data enhancement and standardization on the data set to obtain a data-enhanced remote sensing image data set; obtain the backscatter coefficient through radar data;

[0008] Step 3: Obtain the average spectral reflectance of typical objects within the sea-land boundary to be extracted and construct the optimal water body index;

[0009] Step 4: Construct the NIWI-Net network, specifically: construct a U-Net network, in which the multispectral image and backscatter coefficient in the remote sensing image dataset are used as multi-channel inputs; the encoder of the U-Net network is connected to the decoder of the U-Net network through a skip connection mechanism, and the optimal water index and the output of the U-Net network decoder are channel-concatenated as the network output to form the NIWI-Net network;

[0010] Step 5: Train the NIWI-Net network using the enhanced remote sensing image dataset, and process the real-time data using the trained NIWI-Net network.

[0011] Step 6: Perform a convolution operation on the output of the NIWI-Net network to obtain the land-sea boundary extraction result map after edge enhancement.

[0012] Furthermore, in step 1, Landsat 8 remote sensing images and Sentinel-1 radar data of the land and sea boundaries to be extracted are obtained, and the remote sensing images are preprocessed using the ENVI tool. The preprocessing includes at least: radiometric calibration, orthorectification, and atmospheric correction operations. The method for obtaining the backscatter coefficient is: first performing radiometric correction and terrain correction on the Sentinel-1 radar data, and then calculating the polarization of the backscatter coefficients VV and VH.

[0013] Furthermore, in the step 2, before cropping the remote sensing image, the step also includes: using the shadow water index to mark the water body and retaining all bands; cropping the remote sensing image into multiple pictures of 256×256 pixels in size.

[0014] Furthermore, in step 3, the specific formula for the optimal water index is:

[0015]

[0016] Where ρ is the reflectivity; b2, b3, b5, b6, and b7 are the OLI blue, green, near-infrared, shortwave infrared 1, and shortwave infrared 2 bands, respectively.

[0017] Furthermore, in step 4, the convolution in the encoder of the NIWI-Net network is a striped convolution.

[0018] Furthermore, in step 4, the skip connection mechanism is optimized through the channel-guided attention mechanism.

[0019] Furthermore, in step 4, the decoder of the NIWI-Net network uses a multi-scale attention feature fusion method to fuse features at different levels.

[0020] Beneficial effects of the present invention:

[0021] The model proposed in this paper is applicable to all coasts with high water content in the muddy intertidal zone and large differences in nearshore suspended sediment concentration. It also analyzes the spectral characteristics of typical coastal features and proposes a new normalized improved spectral water index, which can eliminate the misclassification caused by residual water bodies in muddy tidal flats. By enhancing the difference between suspended sediment water bodies and muddy tidal flats, the NIWI index can significantly improve the separation of muddy tidal flats and sediment-laden water bodies, thereby reducing the problem of misclassification. The present invention combines multispectral imagery and backscatter coefficients, using the UNet network as a multi-channel input to increase the fusion of multi-source data. It also uses radar data to fill in areas obscured by clouds in optical images, improving the accuracy and applicability of boundary extraction. This paper proposes channel-wise combining the optimal water index with the output of the decoder in a UNet network to construct a new NIWI-Net neural network model. This achieves multi-source feature complementarity, overcoming the shortcomings of a single data source. Furthermore, the output of the NIWI-Net neural network model is processed through a 3×3 convolution to achieve end-to-end edge enhancement. This enhanced edge-enhanced sea-land boundary extraction result map addresses the inaccurate boundary region extraction problem encountered by traditional methods, particularly in coastal areas with large variations in suspended sediment concentration. This significantly improves boundary clarity and classification accuracy. The paper optimizes the UNet network by configuring the convolution as strip convolution, capturing boundary information from four directions for each pixel. This enables targeted capture of boundary information and reduces interference from irrelevant information, thereby improving the accuracy of boundary extraction for elongated areas such as ports and docks. The CGAM module is introduced into the skip connections in the UNet network, enabling the cascaded information output by the skip connections to better reflect the characteristics of important targets and suppress other background features, thereby enhancing the overall feature extraction of the sea-land boundary. The U-Net decoder uses a multi-scale attention feature fusion method to fuse features at different levels, addressing the inconsistency of multi-scale features. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The features and advantages of the present invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the present invention in any way. In the accompanying drawings:

[0023] Figure 1 This is a flow chart of an embodiment of the present application;

[0024] Figure 2 is the average value of the spectral reflectance of typical ground objects in the embodiment of this application;

[0025] Figure 3 This is a schematic diagram of the structure of the NIWI-Net network in the embodiment of the present application;

[0026] Figure 4 Schematic diagram of the CGAM module in the NIWI-Net network according to an embodiment of the present application;

[0027] Figure 5 This is a schematic diagram of the multi-scale channel attention module in the NIWI-Net network of an embodiment of the present application;

[0028] Figure 6 Schematic diagram of strip convolution in the NIWI-Net network according to an embodiment of the present application. DETAILED DESCRIPTION

[0029] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0030] The present invention will be further described below with reference to specific examples. Those skilled in the art will appreciate that these examples are intended only to illustrate the present invention and are not intended to limit the scope of the present invention, and that modifications to various equivalent forms of the present invention fall within the scope defined by the appended claims.

[0031] like Figure 1 As shown, the present invention provides a method for extracting land and sea boundaries based on remote sensing images, comprising the following steps:

[0032] Step 1: Obtain the remote sensing image of the land and sea boundary to be extracted and preprocess the remote sensing image;

[0033] In the embodiment of the present invention, the area selected for extracting the sea-land boundary is the coastline of Jiangsu Province. Due to the complex types of the coastline of Jiangsu Province, the Lianyungang section is mainly composed of bedrock and sandy coastlines, the Yancheng section is mainly composed of silt and muddy coastlines, and the Nantong section is mainly composed of reclaimed and muddy coastlines. The muddy coast type accounts for about 93% of the total length of the province's coastline.

[0034] Obtain Landsat8 remote sensing imagery for the land and sea boundaries to be extracted and preprocess it using the ENVI tool. This preprocessing includes at least radiometric calibration, orthorectification, and atmospheric correction. Excessive cloud content in the imagery can significantly negatively impact image quality and information extraction. Radar data can be used to remove clouds from these overly cloudy images. Obtain Sentinel-1 radar data, perform radiometric and terrain correction on the radar data, and calculate the polarization-based backscatter coefficients (VV) and (VH).

[0035] Step 2: Crop the preprocessed remote sensing image, construct a data set, and perform data enhancement and standardization on the data set to obtain a data-enhanced remote sensing image data set;

[0036] Water bodies were labeled using multiple water body indices, and the shadow water index (SWI) with the best effect was selected as the annotation standard. The single-band threshold method was used to separate land and sea. Finally, the binary images after water and land separation were manually de-noised, and the boundaries were extracted to obtain the labeled images. All bands were retained, and the preprocessed images and labeled images were cropped into multiple images of 256×256 pixels to create the relevant dataset. Data enhancement was performed on the land and sea boundary dataset, including rotation at different angles and horizontal and vertical flipping. The dataset was randomly divided into training, validation, and test sets in a ratio of 7:2:1.

[0037] Step 3: Obtain the average spectral reflectance of typical objects within the sea-land boundary to be extracted and construct the optimal water body index;

[0038] Because the land-sea boundary in Jiangsu's muddy tidal flats appears and disappears intermittently, they are the most affected by tides. Residual water remains even after low tide, resulting in a large number of mixed pixels of water and tidal flats, resulting in complex spectral characteristics. The spectral reflectance values of muddy tidal flats and water are similar, reducing the applicability of traditional spectral water indexes.

[0039] The specific embodiment of the present invention optimizes the water body index. According to the ground object type classification rules, the pure pixels of the relevant ground objects in the Landsat8 image after atmospheric correction are extracted. More than 10,000 pure pixels are selected for each ground object and the average reflectivity of each band is calculated. The average reflectivity of the typical ground object spectrum is as follows: Figure 2 As shown in the figure, the blue light band is a band with relatively high reflectivity in suspended sediment water bodies, and has a certain penetrating effect on water. Based on the spectral reflectivity of typical ground objects, the blue light band and the green light band are combined into a high reflectivity band combination, and the short-wave infrared band and the near-infrared band are combined into a low reflectivity band combination. By using ratio operations and quadratic power operations, the separation of sediment-containing water bodies and muddy tidal flats is emphasized, and the water body index NIWI is constructed. The formula is:

[0040]

[0041] Where ρ is the reflectivity; b2, b3, b5, b6, and b7 are the blue light band, green light band, near infrared, shortwave infrared 1 band, and shortwave infrared 2 band, respectively.

[0042] NIWI is constructed based on the high reflectivity characteristics of suspended sediment water bodies in the blue and green bands, thereby eliminating the misclassification caused by residual water bodies in muddy tidal flats; at the same time, it increases the difference between suspended sediment water bodies and muddy tidal flats, so that the reflectivity of residual water bodies in muddy tidal flats is close to that of water bodies.

[0043] Step 4: Construct the NIWI-Net network, specifically: construct a U-Net network, in which the Landsat 8 multispectral image and Sentinel-1 backscatter coefficient (VV and VH polarization) are used as multi-channel input to make up for the deficiency of single data, thereby improving the accuracy and robustness of coastline segmentation. The encoder and decoder of the U-Net network are connected through a skip connection mechanism, and the optimal water body index is spliced with the predicted probability map output by the decoder of the U-Net network. Channel splicing is performed as the network output to form the following: Figure 3 The NIWI-Net network shown;

[0044] When extracting feature information, the encoder in the NIWI-Net network replaces the 3×3 square convolution in the encoder with strip convolution. When traversing the image, the strip convolution block captures boundary information from the four directions of each pixel and establishes boundary connection relationships. It specifically captures the neighboring pixel information associated with the central pixel to reduce the interference of irrelevant information.

[0045] The schematic diagram of strip convolution is as follows Figure 4 As shown in the figure, assuming that the input image is X, in the strip convolution block, X is sent to four parallel paths after 1×1 convolution. Each path contains strip convolutions in different directions, which can capture boundary information in a targeted manner from the horizontal, vertical and two diagonal directions, reduce the interference of irrelevant information, and help better process the edges and contours in the image; the feature maps output by these different paths are spliced, upsampled and 1×1 convolved to restore the size of the feature map to match the original input or the output size of other modules, while enabling the network to learn richer and more comprehensive feature representations.

[0046] The low-level features of the encoder are used to adjust the weights of the effective semantic information and position information of the target area in the decoder through the skip connection. The skip connection can output a more comprehensive low-level feature map with more important details.

[0047] The jump connection of the UNet network only performs a simple copy and splicing process on the feature information, and does not use any algorithm to further process the feature information. The NIWI-Net network improves the jump connection through the CGAM module, so that the cascade information output by this step can better reflect the characteristics of important targets and suppress other background feature information; at the same time, it combines texture information and abstract features to enhance the network's ability to extract the overall characteristics of land and sea boundaries.

[0048] CGAM modules such as Figure 5 As shown in the figure, it is mainly divided into two parts. The first half performs global max pooling and global average pooling on the input feature map, respectively, reducing the feature map size from C*H*W to C*1*1. The feature map is compressed based on two dimensions to obtain two feature descriptions of different dimensions. A fully connected layer compresses the number of channels to 1 / r of the original number, which is used to learn the attention weight of each channel. After the ReLU activation function, it is expanded to the original number of channels through another fully connected layer. The two output results are added element-by-element and then activated by the sigmoid function to generate channel attention weights. The normalized weights are added to the input feature map to obtain the attention-weighted channel feature map. The second half uses the same reduction ratio as the first half, compressing the number of channels to 1 / r of the original number through 7×7 convolution. Because the max pooling operation reduces information usage, spatial information fusion is directly performed through convolution to further preserve the feature map.

[0049] The low-level features and high-level features in the decoder are fused through multi-scale attention AFF features, such as Figure 6 As shown in the figure, A represents the underlying features rich in spatial information, and B represents the feature map with a large receptive field. Attention weights are extracted through two branches of different scales. One branch uses global average pooling to extract global feature attention, while the other branch directly uses point-by-point convolution to extract channel-wise attention for local features. The weights are normalized using the sigmoid activation function, and the weights of A and B are determined by taking a weighted average. AFF solves the problem of fusing features of different scales by fusing features with inconsistent semantics and scales.

[0050] Step 5: Train the NIWI-Net network using the enhanced remote sensing image dataset, and process the real-time data using the trained NIWI-Net network.

[0051] Step 6: Perform a convolution operation on the output of the NIWI-Net network to learn the two category weights and update the category probabilities to obtain the edge-enhanced land-sea boundary extraction result map.

[0052] Although the embodiments of the present invention have been described with reference to the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention. Such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A method for extracting land and sea boundaries based on remote sensing images, characterized in that: The steps include: Step 1: Obtain remote sensing images and radar data of the sea and land boundaries to be extracted, and preprocess the remote sensing images; Step 2: Crop the preprocessed remote sensing image to construct a data set, perform data enhancement and standardization on the data set to obtain a data-enhanced remote sensing image data set; obtain the backscatter coefficient through radar data; Step 3: Obtain the average spectral reflectance of typical objects within the sea-land boundary to be extracted, and construct the optimal water body index; Step 4: Construct the NIWI-Net network, specifically: construct a U-Net network, in which the multispectral image and backscatter coefficient in the remote sensing image dataset are used as multi-channel inputs; the encoder of the U-Net network is connected to the decoder of the U-Net network through a skip connection mechanism, and the optimal water index and the output of the U-Net network decoder are channel-concatenated as the network output to form the NIWI-Net network; Step 5: Train the NIWI-Net network using the enhanced remote sensing image dataset, and process the real-time data using the trained NIWI-Net network. Step 6: Perform a convolution operation on the output of the NIWI-Net network to obtain the land-sea boundary extraction result map after edge enhancement.

2. The method for extracting land and sea boundaries based on remote sensing images according to claim 1, wherein: In step 1, Landsat 8 remote sensing images and Sentinel-1 radar data of the land and sea boundaries to be extracted are obtained, and the remote sensing images are preprocessed using the ENVI tool. The preprocessing includes at least radiometric calibration, orthorectification, and atmospheric correction operations. The backscatter coefficient is obtained by first performing radiometric and terrain correction on the Sentinel-1 radar data, and then calculating the polarization of the backscatter coefficients VV and VH.

3. The method for extracting land and sea boundaries based on remote sensing images according to claim 1 or 2, wherein: In the step 2, before cropping the remote sensing image, the steps further include: marking the water body using the shadow water index and retaining all bands; and cropping the remote sensing image into multiple images of 256×256 pixels in size.

4. The method for extracting land and sea boundaries based on remote sensing images according to claim 1, wherein: In step 3, the specific formula for the optimal water index is: Where ρ is the reflectivity; b2, b3, b5, b6, and b7 are the OLI blue, green, near-infrared, shortwave infrared 1, and shortwave infrared 2 bands, respectively.

5. The method for extracting land and sea boundaries based on remote sensing images according to claim 1, wherein: In step 4, the convolution in the encoder of the NIWI-Net network is a strip convolution.

6. The method for extracting land and sea boundaries based on remote sensing images according to claim 1, wherein: In step 4, the skip connection mechanism is optimized through the channel-guided attention mechanism.

7. The method for extracting land and sea boundaries based on remote sensing images according to claim 1, wherein: In step 4, the decoder of the NIWI-Net network uses a multi-scale attention feature fusion method to fuse features at different levels.

Images