A coastline segmentation and extraction method and system based on remote sensing data

Abstract

The present invention relates to the technical field of coastline segmentation and extraction, and discloses a coastline segmentation and extraction method and system based on remote sensing data. Multi-source remote sensing image data of the target coastal area is acquired; a dual-path U-Net model is constructed, and the dual-path U-Net model includes a natural coastline coding path and an artificial coastline coding path; the output features of the natural coastline coding path and the artificial coastline coding path are fused through a dynamic weight fusion module to generate a fusion feature map; based on the fusion feature map, land and sea segmentation and morphological optimization are performed to segment and extract the coastline. By constructing a dual-path U-Net model to process the differentiated features of the natural coastline and the artificial coastline respectively, the dynamic weight fusion module is combined to realize the adaptive feature fusion of multi-source remote sensing data. Through an end-to-end land and sea segmentation and morphological optimization process, automated and efficient extraction from multi-source remote sensing data to coastline vectorization results is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of coastline segmentation and extraction, and in particular to a coastline segmentation and extraction method and system based on remote sensing data. Background Art

[0002] The coastal zone is an area where land and ocean interact. It is both a "golden belt" for economic development and a "fragile zone" for the ecological environment. The coastline is the core division basis and an important component of the coastal zone. Its complexity, activity and marginality make it a key element in marine surveys.

[0003] Based on morphological characteristics, coastlines can be divided into two main types: natural and artificial. Natural coastlines are typically composed of sandy beaches, mudflats, reefs, or biomes (such as mangroves), and their boundaries exhibit gradients and fuzzy features. For example, the intertidal zone alternates between exposed and submerged states during high and low tides, resulting in natural coastlines appearing as weak edge features with gentle grayscale transitions and poor texture continuity in remote sensing imagery. Artificial coastlines, on the other hand, include hardened structures formed by human activities, such as breakwaters, docks, and land reclamation projects. These coastlines often exhibit strong edge features with high contrast and regular geometric shapes in imagery. These two types of coastlines differ significantly in spatial distribution, boundary clarity, and spectral response. Deep learning-based semantic segmentation models (such as U-Net) have achieved some improvement in the automation of coastline extraction through multi-level feature learning. However, they fail to adaptively adjust to the morphological differences between coastline types. Optimizing for one type of coastline often results in a significant decrease in the extraction accuracy of another type, failing to meet the requirements for both weak (natural) and strong (artificial) boundary extraction.

[0004] Furthermore, the complex lighting conditions and terrain coverage in coastal areas further complicate segmentation. Shadows in intertidal zones (such as those obstructed by cliffs and cloud cover) have similar spectral characteristics to water in the visible light band. Traditional methods can easily misinterpret shadows as water, causing the shoreline to shift landward. While SAR imagery offers the advantage of all-weather observation, its inherent speckle noise and geometric deformation make segmentation results from a single data source less robust.

[0005] In summary, existing coastline segmentation methods still face significant bottlenecks in addressing the morphological differences between natural and artificial coastlines and integrating multi-source information. Building a segmentation model that can adapt to coastline types, preserve context and detail, and effectively integrate multi-source remote sensing data to enhance the ability to distinguish between shadows and water bodies has become a pressing challenge in coastal segmentation and extraction. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides a coastline segmentation and extraction method based on remote sensing data, comprising the following steps:

[0007] Step S1: Acquire multi-source remote sensing image data of the target coastal area;

[0008] Step S2: constructing a dual-path U-Net model, wherein the dual-path U-Net model includes a natural shoreline coding path and an artificial shoreline coding path;

[0009] Step S3: The natural shoreline coding path and the artificial shoreline coding path are used to obtain a natural shoreline feature map and an artificial shoreline feature map according to the multi-source remote sensing image data;

[0010] Step S4: fusing the natural shoreline feature map and the artificial shoreline feature map through a dynamic weight fusion module to generate a fused feature map;

[0011] Step S4: performing land and sea segmentation and morphological optimization based on the fused feature map, and segmenting and extracting the coastline.

[0012] Furthermore, the dual-path U-Net model in step S2 includes:

[0013] The natural shoreline encoding path uses a convolution kernel to extract edge features of the natural shoreline to obtain a natural shoreline feature map;

[0014] The artificial shoreline encoding path uses depthwise separable convolution to extract edge features of the artificial shoreline to obtain an artificial shoreline feature map;

[0015] The natural shoreline feature map and the artificial shoreline feature map are transmitted to the decoder through a jump connection.

[0016] Furthermore, in step S4, the natural shoreline feature map and the artificial shoreline feature map are fused by a dynamic weight fusion module to generate a fused feature map, including:

[0017] splicing the natural shoreline feature map output by the natural shoreline coding path and the artificial shoreline feature map output by the artificial shoreline coding path along the channel dimension;

[0018] Perform spatial attention calculation and channel attention calculation on the spliced ​​feature maps respectively to obtain spatial attention weight and channel attention weight;

[0019] Generate a dynamic weight α based on the spatial attention weight and the channel attention weight;

[0020] The natural shoreline feature map and the artificial shoreline feature map are weightedly fused according to the dynamic weight α to generate the fused feature map.

[0021] Furthermore, the spatial attention calculation uses dilated convolution to extract context information to obtain spatial attention weights;

[0022] The channel attention calculation generates channel attention weights through global average pooling and maximum pooling.

[0023] Furthermore, the calculation formula of the dynamic weight α is:

[0024] ;

[0025] Where, W s (i,j) is the spatial position in the concatenated feature map (i,j) The spatial attention weight at , W c (i,j) is the spatial position in the concatenated feature map (i,j) The channel attention weights at , H and W are the height and width of the fused feature map respectively.

[0026] Furthermore, in step S5, performing land and sea segmentation based on the fused feature map includes:

[0027] Step S5A1: performing multimodal feature splicing on the fusion feature map and the multi-source remote sensing image data to obtain multimodal input data;

[0028] Step S5A2: inputting the multimodal input data into a cascade ensemble learning module to obtain natural shoreline classification results and artificial shoreline classification results;

[0029] Step S5A3: spatially stitch the natural coastline classification results and the artificial coastline classification results to generate an initial coastline.

[0030] Furthermore, in step S5A1, multimodal feature splicing is performed on the fused feature map and the multi-source remote sensing image data to obtain multimodal input data, including:

[0031] Step S5A11: obtaining a normalized water index based on the green band and the near-infrared band in the multi-source remote sensing image data;

[0032] Step S5A12: extracting VV polarization data of synthetic aperture radar data from the multi-source remote sensing image data;

[0033] Step S5A13: splicing the normalized water index, VV polarization data and the fusion feature map along the channel dimension to obtain the multimodal input data.

[0034] Furthermore, in step S5A2, when the multimodal input data is input into the cascade ensemble learning module to obtain the natural shoreline classification results and the artificial shoreline classification results, the multimodal input data is input into the cascade ensemble learning module to perform regional classification, specifically including:

[0035] For the natural coastline area, the random forest classifier in the cascade ensemble learning module is used to perform threshold segmentation on the normalized water index and the near-infrared band to obtain the natural coastline classification result;

[0036] For the artificial shoreline area, the support vector machine classifier in the cascade ensemble learning module is used to perform edge enhancement on the VV polarization data and the red light band to obtain the artificial shoreline classification result.

[0037] Furthermore, the morphological optimization includes:

[0038] Step S5B1: De-noising the initial coastline by opening operation;

[0039] Step S5B2: performing a closing operation on the denoised initial coastline to obtain an optimized coastline;

[0040] Step S5B3: performing boundary tracking processing based on the optimized coastline to obtain the final coastline.

[0041] Another aspect of the present invention provides a coastline segmentation and extraction system based on remote sensing data, which executes any of the above-mentioned coastline segmentation and extraction methods based on remote sensing data, and the system includes: a data acquisition module, a model construction module, a fusion feature map generation module and a result output module;

[0042] Data acquisition module, used to obtain multi-source remote sensing image data of the target coastal area;

[0043] A model construction module is used to construct a dual-path U-Net model, wherein the dual-path U-Net model includes a natural shoreline coding path and an artificial shoreline coding path;

[0044] a fusion feature map generation module, configured to fuse the output features of the natural shoreline coding path and the artificial shoreline coding path through a dynamic weight fusion module to generate a fusion feature map;

[0045] The result output module is used to perform land and sea segmentation and morphological optimization based on the fused feature map, and to segment and extract the coastline.

[0046] The embodiments of the present invention have the following technical effects:

[0047] The present invention provides a coastline segmentation and extraction method based on remote sensing data. By constructing a dual-path U-Net model to process the differentiated features of natural coastlines and artificial coastlines respectively, and combining it with a dynamic weight fusion module to realize adaptive feature fusion of multi-source remote sensing data, it finally effectively solves the technical problem of the difficulty in balancing the weak boundary fuzziness of natural coastlines and the strong boundary features of artificial coastlines in coastline segmentation. Through independently designed natural coastline coding paths and artificial coastline coding paths, it is possible to perform differentiated feature extraction for the gradual fuzzy characteristics of natural coastlines and the regular hardened structure of artificial coastlines, overcoming the defect of poor adaptability of traditional single models to different types of coastline features; the dynamic weight fusion module is used to adaptively weighted fuse the dual-path features, which not only retains the contextual continuity characteristics of natural coastlines, but also enhances the edge sharpness characteristics of artificial coastlines, significantly improving the overall segmentation accuracy of mixed coastlines in complex coastal environments; through end-to-end land and sea segmentation and morphological optimization processes, it effectively suppresses intertidal zone shadow interference and multi-source data noise, and realizes automated and efficient extraction from multi-source remote sensing data to coastline vectorization results, providing high-precision technical support for coastal monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0049] Figure 1 This is a flowchart of the steps of a coastline segmentation and extraction method based on remote sensing data provided by an embodiment of the present invention;

[0050] Figure 2 This is a module diagram of a coastline segmentation and extraction system based on remote sensing data provided by an embodiment of the present invention;

[0051] Figure 3 This is a comparison chart of the coastline segmentation results of the single-path U-Net and dual-path U-Net provided by an embodiment of the present invention for coastal areas. DETAILED DESCRIPTION

[0052] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are also within the scope of protection of the present invention.

[0053] Based on the problems of existing coastline segmentation methods in dealing with the morphological differences between natural coastlines and artificial coastlines and multi-source information fusion, the present invention provides a coastline segmentation and extraction method based on remote sensing data, such as Figure 1 As shown, the following steps are included:

[0054] Step S1: Acquire multi-source remote sensing image data of the target coastal area;

[0055] Step S2: constructing a dual-path U-Net model, which includes a natural shoreline coding path and an artificial shoreline coding path;

[0056] In some implementations, the dual-path U-Net model in step S2 includes:

[0057] Step S21: The natural shoreline encoding path uses a convolution kernel to extract edge features of the natural shoreline to obtain a natural shoreline feature map.

[0058] For example, a standard convolution kernel is used to construct a natural shoreline encoding path. Its larger receptive field is conducive to capturing the context-related features of natural landforms such as beaches and mudflats. When the three-dimensional convolution kernel slides in the spatial dimension, it gradually abstracts the intertidal zone gradient features through multi-level downsampling, and uses the semantic information contained in the deep features to identify low-contrast boundaries.

[0059] Step S22: The artificial shoreline encoding path uses depthwise separable convolution to extract edge features of the artificial shoreline to obtain an artificial shoreline feature map;

[0060] For example, the artificial shoreline encoding path uses depthwise separable convolution, which decomposes the standard convolution into two steps: depthwise convolution and pointwise convolution. This significantly reduces the number of parameters while maintaining feature extraction capabilities. This lightweight design is particularly suitable for extracting regular edge features of artificial structures such as breakwaters and docks. By increasing the frequency of local convolution operations, it can capture detailed information about high-contrast edges in shallow networks.

[0061] Step S23: The natural shoreline feature map and the artificial shoreline feature map are transmitted to the decoder via a jump connection.

[0062] For example, the feature maps output by the two encoding paths are directly passed to the decoder via skip connections, preserving the spatial correspondence of the original features before fusion. While the decoder gradually restores spatial resolution through upsampling, it simultaneously receives skip-connected features from both paths, achieving complementary enhancement of multi-scale features. This dual-stream architecture enables the model to process significantly different shoreline types in parallel, avoiding the feature compromises that occur with single-path models when processing mixed shorelines.

[0063] There are significant differences between natural and artificial coastlines in morphological, spectral and textural characteristics:

[0064] Natural shorelines (such as beaches, mudflats, and mangroves) typically exhibit gradual, fuzzy, and weak edge features, significantly affected by tides, lighting, and shadows. This requires a model with stronger contextual awareness to capture continuous but low-contrast boundaries. Artificial shorelines (such as docks, breakwaters, and reclamation projects) exhibit strong, high-contrast, geometrically regular edge features, requiring the model to accurately extract sharp contours and avoid the edge blurring caused by traditional convolution operations. Using a single-path U-Net with a shared encoder forces the model to learn a compromise between the two types of shorelines. This results in weak natural shoreline boundaries losing detail due to oversmoothing, while strong artificial shoreline edges become blurred due to feature aliasing. To address this issue, the present invention constructs a dual-path U-Net model with two independent encoding paths: a natural shoreline encoding path employs standard convolution kernels to enhance contextual information extraction and adapt to fuzzy boundaries; and a artificial shoreline encoding path employs depthwise separable convolutions to reduce parameter redundancy and focus on local edge enhancement. Finally, dynamic weight fusion adaptively adjusts the contributions of the two feature types, ensuring that the segmentation results preserve both the continuity of the natural shoreline and the clarity of the artificial shoreline.

[0065] As an example, during implementation of the present invention, the results of coastline segmentation for a single-path U-Net and a dual-path U-Net were compared. The single-path U-Net was configured with a shared encoder and unified convolution kernels, while the dual-path U-Net was configured with independent encoding paths and dynamic weight fusion. Intersection over Union (IoU) and F1-score were used as evaluation metrics: IoU measures the degree of overlap across the entire segmented area, while F1-score specifically assesses the accuracy of shoreline boundary pixels (natural shorelines rely more on continuity, while artificial shorelines rely more on sharpness).

[0066] Depend on Figure 3 Clearly, the dual-path U-Net achieved an 18% improvement in IoU and a 21% improvement in F1-score for natural shoreline segmentation. This dual-path model, through context-aware convolution, better captures gradient features, and dynamic weight fusion effectively enhances the continuity of weak edges and reduces discontinuities. The dual-path U-Net achieved a 14% improvement in IoU and a 15% improvement in F1-score for artificial shoreline segmentation, demonstrating that depthwise separable convolutions accurately extract regularized structures and significantly improve edge sharpness, avoiding the blurring effect of traditional convolution. Therefore, the unique dual-path design outperforms the single-path design in all overall metrics and demonstrates superior generalization.

[0067] Step S3: The natural shoreline coding path and the artificial shoreline coding path are used to obtain a natural shoreline feature map and an artificial shoreline feature map according to the multi-source remote sensing image data;

[0068] Through independently designed natural shoreline coding paths and artificial shoreline coding paths, differentiated feature extraction can be performed based on the gradual blurring characteristics of natural shorelines and the regular hardened structure of artificial shorelines, overcoming the defect of poor adaptability of traditional single models to different types of shoreline characteristics.

[0069] Step S4: fusing the natural shoreline feature map and the artificial shoreline feature map through a dynamic weight fusion module to generate a fused feature map; including:

[0070] Step S41: splicing the natural shoreline feature map output by the natural shoreline coding path and the artificial shoreline feature map output by the artificial shoreline coding path along the channel dimension;

[0071] Step S42: performing spatial attention calculation and channel attention calculation on the spliced ​​feature map to obtain spatial attention weight and channel attention weight;

[0072] In some implementations, spatial attention calculation uses dilated convolution to extract context information to obtain spatial attention weights;

[0073] Exemplarily, the spatial attention weight calculation uses a dilated convolution operation, setting convolution kernels with different dilation rates to process input features in parallel, capturing nearby details and long-distance context information respectively. After adding the output features of each branch element by element, a spatial weight map is generated through the Sigmoid activation function. This design enables the model to adaptively focus on key areas of different scales. Preferably, setting a dilation rate greater than 1 can expand the receptive field of the convolution kernel and capture a wider range of context information without increasing the amount of computation. This design is particularly suitable for processing the spatial continuity characteristics of natural shoreline areas, and enhances the recognizability of weak edges by perceiving the correlation of surrounding pixels.

[0074] Channel attention calculation generates channel attention weights through global average pooling and maximum pooling.

[0075] Exemplarily, the channel attention calculation simultaneously implements global average pooling and maximum pooling operations. Global average pooling captures the overall distribution characteristics of the feature channel, and maximum pooling extracts the significant response area of ​​the feature map. After the two pooling results are spliced ​​in the channel dimension, the nonlinear relationship between channels is learned through a two-layer fully connected network, and the channel attention weight is generated using a compression-excitation network structure.

[0076] This dual-path attention mechanism effectively overcomes the limitations of a single attention model. Spatial attention focuses on the spatial distribution characteristics of features, while channel attention optimizes the information utilization of feature channels. The two work together to improve the discriminative ability of feature fusion and effectively identify feature channels that are important to the current classification task, such as enhancing feature channels with strong edge responses in artificial shoreline feature maps.

[0077] Step S43: Generate a dynamic weight α based on the spatial attention weight and the channel attention weight;

[0078] In some implementations, the dynamic weight α is calculated as follows:

[0079] ;

[0080] Where, W s (i,j) is the spatial position in the concatenated feature map (i,j) The spatial attention weight at , W c (i,j) is the spatial position in the concatenated feature map (i,j) The channel attention weights at , H and W are the height and width of the fused feature map respectively.

[0081] Spatial attention weights reflect the importance of each pixel location to the task at hand. For example, intertidal zones require a higher weight due to the spectral similarity between shadows and water. Channel attention weights indicate the information contribution of different feature channels, such as the differences in channel weights between near-infrared bands for water recognition and radar data for artificial structure detection. The two are coupled through point-by-point multiplication to achieve attention coupling in the spatial and channel dimensions, and then globally averaged to obtain the overall fusion coefficient, ensuring that the α value incorporates both local feature response intensity information and global statistical properties. When processing areas dominated by natural shorelines, spatial attention weights exhibit a continuous distribution at gradient boundaries, while channel attention weights enhance feature channels that reflect texture continuity, automatically biasing α values ​​toward natural shoreline characteristics. When processing artificial shorelines, spatial attention weights are concentrated at sharp edges, increasing the weights of relevant channels and encouraging the fused features to enhance the characteristics of artificial structures. When the feature map contains gradient boundaries such as mangroves and beaches, spatial attention captures a wide range of context, channel attention strengthens low-contrast features, α approaches 1, and the weight of natural shoreline paths is higher; when the feature map contains sharp edges such as ports and dams, channel attention suppresses noise bands (such as shadows), spatial attention focuses on local details, α approaches 0, and the weight of artificial shoreline paths is higher.

[0082] Step S44: performing weighted fusion on the natural shoreline feature map and the artificial shoreline feature map according to the dynamic weight α to generate a fused feature map.

[0083] Based on the adaptive weight distribution of input features, the segmentation accuracy of artificial and natural coastlines is significantly improved.

[0084] Step S5: Perform land and sea segmentation and morphological optimization based on the fused feature map, and segment and extract the coastline.

[0085] Based on the fusion feature map, sea and land segmentation is performed, artificial coastlines and natural coastlines are accurately classified and identified, and the morphological integrity of the coastline is achieved through morphological optimization, thereby achieving accurate segmentation and extraction of the coastline.

[0086] In some implementations, in step S5, performing land and sea segmentation based on the fused feature map includes:

[0087] Step S5A1: performing multimodal feature splicing on the fused feature map and multi-source remote sensing image data to obtain multimodal input data;

[0088] In some implementations, in step S5A1, performing multimodal feature splicing on the fused feature map and multi-source remote sensing image data to obtain multimodal input data includes:

[0089] Step S5A11: obtaining a normalized water index based on the green band and the near-infrared band in the multi-source remote sensing image data;

[0090] ;

[0091] Among them, NDWI stands for Normalized Difference Water Index, Green stands for green light band, and NIR stands for near infrared band.

[0092] Step S5A12: extracting VV polarization data of synthetic aperture radar data from the multi-source remote sensing image data;

[0093] Step S5A13: Concatenate the normalized water index, VV polarization data, and fusion feature map along the channel dimension to obtain multimodal input data.

[0094] For example, multi-source remote sensing image data includes optical imagery, SAR imagery, and vegetation indices. First, the multi-source remote sensing image data is preprocessed, such as performing radiometric and geometric correction on optical imagery, performing speckle noise reduction and terrain correction on SAR imagery, and then registering the multi-source data to the same coordinate system and resolution.

[0095] Optical imagery can include green, red, blue, and near-infrared bands. Synthetic aperture radar (SAR) imagery includes both vertical-wavelength (VV) and vertical-wavelength (VH) polarization data. VV polarization data represents the backscattered signal from ground objects collected by the radar system using both vertical transmit (V) and vertical receive (VH) polarizations. It is used to enhance the edges of artificial shorelines. VV polarization data is highly sensitive to artificial structures. For example, the geometric structure (right angles and metal surfaces) of artificial shorelines (such as piers and breakwaters) produces strong specular reflections. The VV backscatter value of such targets is significantly higher than that of natural shorelines. Typically, the VV backscatter value of a concrete dam is 10–15 dB higher than that of the adjacent water. It is also sensitive to water. Calm water surfaces reflect most radar waves, resulting in extremely weak VV signals (appearing as dark areas). Rough water (such as waves) enhances scattering, making it useful for contrast with natural shorelines. The VV echo intensity of mangroves or beaches is between that of water and artificial structures. The support vector machine classifier in the subsequent cascade integrated learning module uses VV polarization data and the red light band for edge enhancement to obtain accurate artificial shoreline classification results.

[0096] VH polarization data, where the radar transmits vertically and receives horizontally, is used to help distinguish areas of vegetation cover. Vegetation index data is used to help distinguish between vegetation cover and bare ground along natural coastlines.

[0097] The vegetation index is normalized to obtain the normalized vegetation index NDVI, which can be calculated using the following formula:

[0098] ;

[0099] RED represents the red light band. The Normalized Difference Vegetation Index (NDVI) can be used to distinguish vegetation-covered areas (such as mangroves and salt marshes) from bare surfaces, avoiding misidentification of vegetation along natural coastlines.

[0100] For example, the process of concatenating multimodal input data along the channel dimension is as follows:

[0101] (1) Input alignment

[0102] Resample the fused feature map and multi-source data to a uniform resolution (e.g. 10m);

[0103] (2) Channel-level splicing

[0104] ;

[0105] Among them, F multi Represents the spliced ​​multimodal input data, if F fusedIf the number of channels is C, the number of channels after splicing is C+3 (NDWI, VV, and NDVI each occupy 1 channel); Concat represents the splicing operation along the channel dimension (Channel Axis), stacking the feature maps of different modalities on the channel dimension to retain the consistency of spatial resolution (height H and width W remain unchanged); F fused It represents the fused feature map obtained by the dual-path U-Net, which is obtained by the weighted fusion of the features of the natural shoreline path and the artificial shoreline path. It contains spatial context and edge detail information, and provides high-level semantic features (such as shoreline morphology and texture continuity). NDWI represents the normalized water index, VV represents VV polarization data, and NDVI represents the normalized vegetation index, which are used to provide physical features (such as the spectral-scattering characteristics of water bodies, edges, and vegetation), respectively. NDWI suppresses the misjudgment of shadows in optical images, VV compensates for the loss of optical data in cloudy and rainy weather, and NDVI avoids confusion between vegetation coverage areas and shorelines.

[0106] Step S5A2: inputting the multimodal input data into the cascade ensemble learning module to obtain natural shoreline classification results and artificial shoreline classification results;

[0107] In some embodiments, in step S5A2, when the multimodal input data is input into the cascade ensemble learning module to obtain the natural shoreline classification results and the artificial shoreline classification results, the multimodal input data is input into the cascade ensemble learning module to perform regional classification, specifically including:

[0108] For natural coastline areas, the random forest classifier in the cascade ensemble learning module is used to perform threshold segmentation on the normalized water index and near-infrared band to obtain the natural coastline classification results.

[0109] For artificial shoreline areas, the support vector machine classifier in the cascade ensemble learning module is used to perform edge enhancement using VV polarization data and the red light band to obtain the artificial shoreline classification results.

[0110] Step S5A3: spatially stitch the natural coastline classification results and the artificial coastline classification results to generate an initial coastline.

[0111] For example, the classification result in step S5A2 is a binary mask, and the initial coastline is generated through pixel-by-pixel logical operations during spatial stitching. If a pixel is classified as both a natural coastline and an artificial coastline, it can be marked according to a preset priority. The preset priority can be set based on actual conditions, such as giving artificial coastlines a higher priority than natural coastlines to prevent artificial structures from being covered by natural coastlines.

[0112] Through priority overlay and logical operations, the two types of coastline classification results are merged into a conflict-free initial coastline, providing input for subsequent morphological optimization (denoising and closure), ensuring the integrity of artificial structures and the continuity of natural coastlines.

[0113] The cascaded ensemble learning module achieves precise regional classification through a combination of classifiers. For natural coastline areas, the random forest classifier utilizes an ensemble voting mechanism of multiple decision trees to construct multiple threshold segmentation planes to address the nonlinear relationship between the normalized water index and the near-infrared band. This algorithm has good adaptability to irregular decision boundaries in feature space and can effectively handle the spectral gradient characteristics of natural landforms such as beaches and mudflats. For artificial coastline areas, a support vector machine classifier is used. A kernel function is used to map the VV polarization data and the red light band into a high-dimensional space to construct an optimal separating hyperplane. Edge enhancement uses a gradient operator to enhance the contour features of artificial structures, enabling the support vector machine to accurately identify the geometric features of linear features such as breakwaters and docks. This regional classification strategy leverages the strengths of different algorithms: the random forest classifier addresses the fuzzy classification of natural coastlines, while the support vector machine is suitable for the clear boundary identification of artificial coastlines. The two are spatially spliced ​​to form a complete coastline.

[0114] In some embodiments, morphological optimization comprises:

[0115] Step S5B1: The initial coastline is subjected to denoising processing by an opening operation, which is used to eliminate small noise and burrs in the initial coastline.

[0116] For example, the initial coastline is a binary image, in which the land pixel value is recorded as 0 and the water pixel value is recorded as 1. The denoising includes: first performing an erosion operation to eliminate isolated land pixels whose area is smaller than a first threshold; then performing a dilation operation to restore the original shape of the retained land area;

[0117] Step S5B2: Perform a closing operation on the denoised initial coastline to obtain an optimized coastline, which is used to fill small holes and connect adjacent areas.

[0118] Exemplarily, a closing operation is used to process the binary image after noise removal, wherein: a dilation operation is first performed to fill water holes whose area is smaller than a second threshold; and then an erosion operation is performed to restore the original boundary of the water area.

[0119] The first threshold and the second threshold are dynamically adjusted according to the regional characteristics of the classification results of each coastline in the initial coastline:

[0120] For natural coastline areas, a larger first threshold is set to retain small features;

[0121] For artificial shoreline areas, a smaller second threshold is set to ensure edge continuity.

[0122] Step S5B3: Based on the optimized coastline, boundary tracking is performed, and the closed polygons obtained by tracking are converted into coastline vector data to obtain the final coastline.

[0123] During coastline segmentation, the initial coastline may have inaccuracies due to interference from remote sensing imagery, such as waves, clouds, and sensor noise. For example, wave foam in the intertidal zone may form isolated white pixels in the image, which are considered noise. Artificial coastline structures may also have breaks or holes due to reflections or occlusion.

[0124] The purpose of denoising through the opening operation is to remove these small noise points, such as isolated pixels or small protrusions, making the coastline smoother and reducing incorrect segmentation areas. For example, a beach area may have sporadic non-coastline points due to reflections. The opening operation can eliminate these points and prevent them from being mistakenly identified as coastline in subsequent processing.

[0125] The purpose of the closing operation is to fill small holes in the coastline caused by discontinuous segmentation or to connect sections that are broken due to occlusion. For example, cloud cover may cause a section of the coastline to be interrupted during segmentation. The closing operation can fill these interruptions by expanding the coastline area through dilation and then eroding to restore the approximate shape, thus making the coastline continuous.

[0126] The serialized execution of opening and closing operations provides a dual guarantee of noise suppression and structural enhancement:

[0127] Noise sensitivity difference: The opening operation is sensitive to positive noise (falsely increased pixels), while the closing operation focuses on negative noise (falsely lost pixels). The two form a complementary filtering system.

[0128] Edge conformal constraint: Compared with single erosion or dilation, the combination of opening and closing operations can maximize the preservation of shoreline curvature characteristics during the denoising and filling process, avoiding distortion such as straightening artificial shorelines or over-smoothing natural shorelines.

[0129] The present invention provides a coastline segmentation and extraction method based on remote sensing data. By constructing a dual-path U-Net model to process the differentiated features of natural coastlines and artificial coastlines respectively, and combining it with a dynamic weight fusion module to realize adaptive feature fusion of multi-source remote sensing data, it finally effectively solves the technical problem of the difficulty in balancing the weak boundary fuzziness of natural coastlines and the strong boundary features of artificial coastlines in coastline segmentation. Through independently designed natural coastline coding paths and artificial coastline coding paths, it is possible to perform differentiated feature extraction for the gradual fuzzy characteristics of natural coastlines and the regular hardened structure of artificial coastlines, overcoming the defect of poor adaptability of traditional single models to different types of coastline features; the dynamic weight fusion module is used to adaptively weighted fuse the dual-path features, which not only retains the contextual continuity characteristics of natural coastlines, but also enhances the edge sharpness characteristics of artificial coastlines, significantly improving the overall segmentation accuracy of mixed coastlines in complex coastal environments; through end-to-end land and sea segmentation and morphological optimization processes, it effectively suppresses intertidal zone shadow interference and multi-source data noise, and realizes automated and efficient extraction from multi-source remote sensing data to coastline vectorization results, providing high-precision technical support for coastal monitoring.

[0130] Another aspect of the present invention provides a coastline segmentation and extraction system based on remote sensing data, which executes any of the above coastline segmentation and extraction methods based on remote sensing data, such as Figure 2 As shown, the system includes: a data acquisition module, a model building module, a fusion feature map generation module and a result output module;

[0131] Data acquisition module, used to obtain multi-source remote sensing image data of the target coastal area;

[0132] A model building module is used to build a dual-path U-Net model, which includes a natural shoreline coding path and an artificial shoreline coding path;

[0133] A fusion feature map generation module is used to fuse the output features of the natural shoreline coding path and the artificial shoreline coding path through a dynamic weight fusion module to generate a fusion feature map;

[0134] The result output module is used to perform land and sea segmentation and morphological optimization based on the fusion feature map, and to segment and extract the coastline.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. A coastline segmentation and extraction method based on remote sensing data, characterized in that: The steps include: Step S1: Acquire multi-source remote sensing image data of the target coastal area; Step S2: constructing a dual-path U-Net model, wherein the dual-path U-Net model includes a natural shoreline coding path and an artificial shoreline coding path; Step S3: The natural shoreline coding path and the artificial shoreline coding path are used to obtain a natural shoreline feature map and an artificial shoreline feature map according to the multi-source remote sensing image data; Step S4: fusing the natural shoreline feature map and the artificial shoreline feature map through a dynamic weight fusion module to generate a fused feature map; Step S5: performing land-sea segmentation and morphological optimization based on the fused feature map, and segmenting and extracting the coastline; The dual-path U-Net model in step S2 includes: The natural shoreline encoding path uses a convolution kernel to extract edge features of the natural shoreline to obtain a natural shoreline feature map; The artificial shoreline encoding path uses depthwise separable convolution to extract edge features of the artificial shoreline to obtain an artificial shoreline feature map; The natural shoreline feature map and the artificial shoreline feature map are transmitted to a decoder via a skip connection; In step S4, the natural shoreline feature map and the artificial shoreline feature map are fused by a dynamic weight fusion module to generate a fused feature map, including: splicing the natural shoreline feature map output by the natural shoreline coding path and the artificial shoreline feature map output by the artificial shoreline coding path along the channel dimension; Perform spatial attention calculation and channel attention calculation on the spliced ​​feature maps respectively to obtain spatial attention weight and channel attention weight; Generate a dynamic weight α based on the spatial attention weight and the channel attention weight; The natural shoreline feature map and the artificial shoreline feature map are weightedly fused according to the dynamic weight α to generate the fused feature map.

2. The coastline segmentation and extraction method based on remote sensing data according to claim 1, characterized in that: The spatial attention calculation uses dilated convolution to extract context information and obtain spatial attention weights; The channel attention calculation generates channel attention weights through global average pooling and maximum pooling.

3. The coastline segmentation and extraction method based on remote sensing data according to claim 1, characterized in that: The calculation formula of the dynamic weight α is: ； Where, Ws(i,j) is the spatial position in the concatenated feature map (i,j) The spatial attention weight at , Wc(i,j) is the spatial position in the concatenated feature map (i,j) The channel attention weights at , H and W are the height and width of the fused feature map respectively.

4. The coastline segmentation and extraction method based on remote sensing data according to claim 1, characterized in that: In step S5, performing land and sea segmentation based on the fused feature map includes: Step S5A1: performing multimodal feature splicing on the fusion feature map and the multi-source remote sensing image data to obtain multimodal input data; Step S5A2: inputting the multimodal input data into a cascade ensemble learning module to obtain natural shoreline classification results and artificial shoreline classification results; Step S5A3: spatially stitch the natural coastline classification results and the artificial coastline classification results to generate an initial coastline.

5. The coastline segmentation and extraction method based on remote sensing data according to claim 4, characterized in that: In step S5A1, multimodal feature splicing is performed on the fused feature map and the multi-source remote sensing image data to obtain multimodal input data, including: Step S5A11: obtaining a normalized water index based on the green band and the near-infrared band in the multi-source remote sensing image data; Step S5A12: extracting VV polarization data of synthetic aperture radar data from the multi-source remote sensing image data; Step S5A13: splicing the normalized water index, VV polarization data and the fusion feature map along the channel dimension to obtain the multimodal input data.

6. The coastline segmentation and extraction method based on remote sensing data according to claim 4, characterized in that: In step S5A2, the multimodal input data is input into a cascade ensemble learning module to obtain natural shoreline classification results and artificial shoreline classification results. In order to input the multimodal input data into the cascade ensemble learning module, regional classification is performed, specifically including: For natural coastline areas, the random forest classifier in the cascade ensemble learning module is used to perform threshold segmentation on the normalized water index and near-infrared band to obtain the natural coastline classification results. For artificial shoreline areas, the support vector machine classifier in the cascade ensemble learning module is used to perform edge enhancement on the VV polarization data and the red light band to obtain the artificial shoreline classification results.

7. The coastline segmentation and extraction method based on remote sensing data according to claim 1, characterized in that: The morphological optimization includes: Step S5B1: De-noising the initial coastline by opening operation; Step S5B2: performing a closing operation on the denoised initial coastline to obtain an optimized coastline; Step S5B3: performing boundary tracking processing based on the optimized coastline to obtain the final coastline.

8. A coastline segmentation and extraction system based on remote sensing data, which executes the coastline segmentation and extraction method based on remote sensing data according to any one of claims 1 to 7, characterized in that: The system includes: a data acquisition module, a model building module, a fusion feature map generation module and a result output module; Data acquisition module, used to obtain multi-source remote sensing image data of the target coastal area; A model construction module is used to construct a dual-path U-Net model, wherein the dual-path U-Net model includes a natural shoreline coding path and an artificial shoreline coding path; a fusion feature map generation module, configured to fuse the output features of the natural shoreline coding path and the artificial shoreline coding path through a dynamic weight fusion module to generate a fusion feature map; The result output module is used to perform land and sea segmentation and morphological optimization based on the fused feature map, and to segment and extract the coastline.

Images