Segmented multi-source shoreline extraction calibration method and system based on physical prior

Abstract

The application discloses a segmented multi-source shoreline extraction calibration method and system based on physical prior, and particularly relates to the technical field of remote sensing image processing and computer vision, and integrates a consistency evaluation and reliability measurement mechanism of multi-source surface water products at a data level, and constructs a high-quality training sample set by using water body persistence characteristics; at a model level, a feature extraction architecture is developed by fusing physical prior guidance and a geometric self-adaptive operator, a multi-dimensional physical constraint and a noise tolerance loss function are constructed, and pixel-level accurate identification of a shoreline under a large-range remote sensing image and quality closed-loop calibration with a rigorous topology are realized.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and computer vision technology, specifically to a segmented multi-source shoreline extraction and calibration method and system based on physical priors. Background Technology

[0002] As a crucial transitional zone between land and water, the shoreline not only serves as a natural ecological barrier for maintaining wetland biodiversity and purifying inflows, but also as a core window for monitoring the dynamic evolution of surface water bodies. Regional shoreline studies can accurately grasp the erosion and deposition patterns of the land-water boundary, providing core support for preventing bank collapse risks, ensuring dike safety, and optimizing water conservancy project scheduling. Simultaneously, dynamic shoreline monitoring is also a vital foundation for balancing shoreline resource development and ecological protection.

[0003] Shorelines exhibit highly dynamic spatiotemporal characteristics, manifested not only in long-term landform reshaping such as port reclamation but also in the continuous influence of short-term tidal fluctuations and natural erosion, resulting in a constantly evolving and complex water-land boundary. Existing research focuses on specific local scales, limiting shoreline extraction accuracy in situations with frequent cloud cover and limited image quality. When faced with complex landforms such as natural rocky shores, mudflats, and artificial shorelines, traditional algorithms lack robustness to tidal disturbances and topographical complexity, leading to poor consistency and universality of extraction results, making it difficult to meet the high-precision monitoring requirements for shoreline changes under variable environments.

[0004] While visual interpretation of high-resolution imagery offers high accuracy, its high cost makes it difficult to scale globally. Deep learning-based semantic segmentation models excel in object extraction, but they heavily rely on large-scale, high-quality labeled samples. Limited by the significant spatiotemporal dynamics of coastlines and the complexity of terrain, and the difficulty in acquiring large-scale, high-precision samples, existing models suffer from insufficient feature mining and generalization capabilities at medium-resolution scales.

[0005] There is an urgent need to develop a systematic technical solution that integrates a consistency assessment and reliability measurement mechanism for multi-source surface water products at the data level, and utilizes the continuous characteristics of water bodies to construct a high-quality training sample set. At the model level, it is necessary to develop a feature extraction architecture that integrates physical prior guidance and geometric adaptive operators, construct multi-dimensional physical constraints and noise tolerance loss functions, and achieve pixel-level accurate identification of shorelines and topologically rigorous quality closed-loop calibration under large-scale remote sensing images. Summary of the Invention

[0006] To address these issues, the present invention provides a segmented multi-source shoreline extraction and calibration method and system based on physical priors, in order to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a segmented multi-source shoreline extraction and calibration method based on physical priors, comprising:

[0008] The original Landsat image was preprocessed to generate an annual composite base map with cloud-free cover and spectrally stable true surface reflectance.

[0009] The basic bands are extracted from the annual composite base map. Based on the selected basic bands, the NDWI index of the annual composite base map is calculated, thereby obtaining a five-channel multidimensional feature cube.

[0010] By integrating multi-source surface water product data and employing an adaptive gap-filling method and a multi-source fusion strategy, a reconstructed monthly water body dynamic dataset is obtained. Based on the monthly water body dynamic dataset, the water body frequency of each pixel at the annual scale is calculated to complete the water body type classification. The matching degree between the water body frequency of each pixel at the annual scale in the multi-source surface water product data at the same spatiotemporal location and the surface classification rules is compared, and pixel-level reliability weights are assigned to construct a multi-source consistency probability map (MCP), which serves as a confidence prior for subsequent Shore-SegFormer model training.

[0011] The generated initial shoreline binary label image is superimposed with the annual composite image to generate initial labels. After subpixel-level topological verification and geometric alignment, synchronous cropping is performed and a sample categorization screening mechanism is used to obtain a mixed training set containing features of the water-land interface.

[0012] The Shore-SegFormer model uses a hierarchical Transformer encoder as the backbone feature extraction network. The first layer of the hierarchical Transformer encoder has an overlapping patch embedding module to adapt to the input image, which includes spectral bands, normalized differential water index, and five-channel multi-dimensional feature cube. The model uses the NDWI index and boundary candidate prior feature map to pre-guide neurons to focus on the high-frequency region at the water-land interface. Through hierarchical multi-head self-attention calculation, it outputs four-level multi-scale feature maps step by step. After multi-scale feature fusion decoder, a single-channel shoreline probability map is obtained.

[0013] Specifically, after the low-level feature map, an explicit gradient-guided edge perception refinement module (EARM) is embedded, combined with the explicit guidance of the multi-source consistency probability map (MCP) to perceive feature map overlap; after the high-level feature map, a DR-LDSC module is embedded to capture the detailed curvature of local meandering shorelines and to ensure the overall logical continuity of cross-regional shorelines by utilizing long-distance receptive fields.

[0014] A reliability-weighted mechanism is introduced, and a triple-constraint composite loss function is constructed, which is combined with the consistency probability graph MCP to overcome the noise when generating the initial shoreline binary label map.

[0015] Through sliding window reasoning, AWEI nsh Exponential fusion, eight-neighbor connectivity component analysis, and vector optimization enable high-precision shoreline extraction and database construction from large-scale imagery.

[0016] The preferred annual composite base map generation process is as follows: the original Landsat image is integrated using the GEE platform, and subpixel-level alignment is achieved through geometric correction and resampling. The QA band mask is then used to remove clouds and mean aggregation is performed to generate the annual composite base map.

[0017] The preferred five-channel multidimensional feature cube construction process involves selecting the NIR, SWIR1, Red, and Green fundamental bands from the generated annual composite base map and calculating the NDWI index of the annual composite base map as follows: ;

[0018] The 3×3 Scharr operator is used to extract the first-order gradient magnitude of the NDWI exponent, capturing high-frequency abrupt change signals at the boundary to generate candidate prior feature maps of the boundary, as detailed below:

[0019] The horizontal and vertical gradient components are extracted using a horizontal convolution kernel Gx = [[-3, 0, 3], [-10, 0, 10], [-3, 0, 3]] and a vertical convolution kernel Gy = [[-3, -10, -3], [0, 0, 0], [3, 10, 3]].

[0020] Calculate gradient magnitude By capturing high-frequency abrupt change signals in the spatial distribution of NDWI, the spectral response at the water-land interface is amplified to generate a boundary candidate prior feature map;

[0021] Finally, the basic band, NDWI index, and boundary candidate prior feature maps are stitched together in multiple channels to construct a five-channel multidimensional feature cube.

[0022] Preferably, the water frequency (WIF) of each pixel on an annual scale is calculated using the following formula:

[0023] ;

[0024] Where n=12, The binary water state for month i;

[0025] The Earth's surface is divided into:

[0026] Permanent water bodies PSW, with a corresponding WIF ≥ 75%;

[0027] Intermittent water body ISW, corresponding to 5% <WIF <75%;

[0028] For land-based applications, the corresponding WIF ≤ 5%;

[0029] Simultaneously construct a multi-source consistency probability map (MCP): compare the prediction results of multi-source surface water product data at the same spatiotemporal location. If they are completely consistent, the reliability weight w = 1.0. If there are two types of conflicts, the weight w = 0.6. This is used as the confidence prior for subsequent Shore-SegFormer training.

[0030] Preferably, leveraging the distribution stability of permanent water bodies (PSWs) over long time, PSWs are defined as shoreline target classes (Label=1). Morphological gradient operators are used to extract the spatial contours of PSWs, generating initial binary shoreline labels, which are then overlaid with the annual composite base map. When dynamic geomorphological evolution causes a systematic shift between the image and the labels, the edge gradient features of the annual composite base map are first extracted. Subsequently, local normalized cross-correlation (NCC) feature matching is performed with the binary shoreline labels within a 5×5 pixel window. The peak correlation coefficient within the search window determines the sub-pixel displacement vector, achieving sub-pixel-level alignment correction between the labels and the image. The calibrated labels are uniformly mapped to a single-channel 8-bit grayscale layer. The high-precision, standard-format shoreline ground truth labels obtained through physical correction of the single-channel 8-bit grayscale layer are the fundamental basis for the Shore-SegFormer model to learn to accurately extract shorelines, ensuring computational accuracy and storage efficiency.

[0031] Preferably, the process of obtaining the training set containing features of the land-water interface is as follows: NIR, SWIR1, and Red bands, which are sensitive to water body response, are selected to construct a 3-channel feature image, and synchronously cropped to 256×256 pixels with the corresponding labels; a pixel ratio threshold T is defined. p T p This represents the proportion of pixels labeled as water bodies out of the total number of pixels in each 256×256 training set slice. Only pixels with a water body percentage less than 5% are retained. wate Slices with less than 95% coverage were used as a mixed dataset of land-water boundary features, removing redundant information from both the entire land and water areas.

[0032] The preferred Shore-SegFormer model architecture is as follows:

[0033] Backbone feature extraction network: adopts a hierarchical Transformer architecture;

[0034] Input layer: An overlapping patch embedding module is used to adapt to a five-channel multidimensional feature cube I ϵ R H×W×5 Through hierarchical multi-head self-attention computation, four-level multi-scale feature maps {C1, C2, C3, C4} are output step by step. The NDWI index and boundary candidate prior feature maps are used to pre-guide neurons to focus on the high-frequency region at the water-land interface.

[0035] The explicit gradient-guided edge-aware refinement module (EARM) is deployed after the low-level feature maps {C1, C2}, and then the generated boundary candidate prior feature map G is introduced. prior Implement explicit guidance, and will G prior As a spatial weighting operator, overlap detection is performed on the feature map, and the calculation logic is as follows: ;

[0036] Where X represents the input feature map, Y represents the output feature map, Gprior represents the boundary candidate prior feature map, Conv represents the convolution operation, and BN represents batch normalization. This indicates element-wise multiplication at the pixel level. The Sigmoid activation function is used; by directly injecting the boundary candidate prior feature map into the network feature layer, the problem of blurred edge localization caused by relying solely on spectral features under complex mudflats and cloud shadow interference is solved.

[0037] To address the geometric characteristics of coastlines in images, which are often elongated, arbitrarily oriented, and exhibit varying curvature, a DR-LDSC module is embedded after the high-level stages {C3, C4} of the Shore-SegFormer encoder. The C3 stage enhances the feature map with a received spatial resolution of 1 / 16, focusing on capturing the detailed curvature of local meandering coastlines. The C4 stage enhances the global semantic feature map with a received spatial resolution of 1 / 32, utilizing a long receptive field to ensure the overall logical continuity of the coastline across regions.

[0038] The DR-LDSC module employs a dual-path parallel structure to handle the spatial dependencies of the horizontal H-path and the vertical V-path separately: before performing strip convolution, it first passes through a lightweight convolutional layer f with 2K channels. off Dynamically learn the 2D offset of each sampling point Perform geometrically adaptive rotation and deformation; for a 1×K horizontal path, the dimension of the offset generation is R. H×W×2K ;

[0039] For asymmetric dilated convolution parameters, the kernel size K=15 is set, with a 1×15 kernel used for the horizontal path and a 15×1 kernel used for the vertical path. The dilation rate r=3 is set for layer C3 and r=5 for layer C4. The coordinates of the sampling points after offset correction are... The pixel values ​​at non-integer coordinates are obtained through bilinear interpolation, and the final output is:

[0040] ;

[0041] ;

[0042] in and These represent the output feature values ​​after convolution and aggregation via horizontal and vertical paths, respectively, and K represents the kernel size (set to 15). The convolution weights at the nth sampling point are represented by x, where x represents the pixel value on the input feature map. This represents the coordinates of the center pixel currently being calculated. This represents the original sampling offset coordinates of the standard convolution kernel on a regular grid. This represents the adaptive position offset learned by the model. This indicates the corrected coordinates of the sampling points.

[0043] Will and After pixel-level summation and fusion, a 1×1 convolutional layer restores the feature dimension to the original dimensions of each level of the encoder (d=160 or 256), and the residual structure is used to merge with the input features and output to the decoder.

[0044] The multi-scale feature fusion decoder consists of linear projection, upsampling alignment, feature concatenation, fusion convolution, and prediction layers. The MLP projection layer maps the number of channels of the four-level feature maps {C1, C2, C3, C4} to d = 256. Bilinear interpolation is used to align the resolution to 1 / 4 of the original. The aligned features are concatenated along the channel axis to form a 1024-dimensional tensor. Cross-scale feature fusion is performed by 1×1 convolution to compress it back to 256 dimensions. A single-channel shoreline probability map is output through the linear prediction layer and the Sigmoid operator.

[0045] Preferably, a triple-constraint composite loss function is constructed: ;

[0046] in, This represents the total value of the triple-constraint composite loss function. , and These represent the balance weight coefficients for different loss terms. , and These represent the noise-tolerant weighted binary cross-entropy loss, the noise-tolerant weighted Dice loss, and the physical consistency constraint loss, respectively.

[0047] Noise-tolerant weighted binary cross-entropy loss : Utilizing the generated pixel-level reliability weights Dynamically reweight the classification loss:

[0048] ;

[0049] This represents the total number of pixels in the image, where n is the pixel index. Denote the pixel-level reliability weight of the nth pixel, i.e., the pixel-level reliability weight (for dynamic reweighting), Denote the true label value of the nth pixel, Denote the probability value that the model predicts the nth pixel as the shoreline;

[0050] By reducing the gradient contribution of the multi-source conflict area (noise points), enable the model to have the ability to extract high-purity features under weakly supervised labels;

[0051] Noise-tolerant weighted Dice loss : For the sparse distribution of shoreline pixels, introduce the weighted area overlap constraint:

[0052] ;

[0053] Among them, Denote the smoothing coefficient used to prevent the denominator from being zero and stabilize the gradient backpropagation.

[0054] Physical consistency constraint loss : Introduce the geographical explicit constraint, calculate the Laplacian operator consistency error between the model prediction probability map and the physical index gradient map, and use the second-order derivative constraint to force the extraction result to conform to the physical mechanism of the surface water distribution; in this embodiment, the weight coefficients of each loss term are set to = =1.0, =0.5, and use the AdamW optimizer for iterative training until the IoU and Boundary F-score of the model on the validation set reach the steady-state threshold;

[0055] Preferably, use sliding window inference to perform pixel-level semantic prediction on the large-range images to be monitored; set the window size to 256×256 pixels, the window overlap degree to 50%, and eliminate the stitching traces at the window edges; for the overlapping areas, adopt a weighted fusion strategy based on a two-dimensional Gaussian kernel: ; among them, m is the number of windows covering this pixel, is the probability output of the kth window, W k is the Gaussian weight coefficient;

[0056] For highly interfering artificial ground objects such as aquaculture areas and salt pans, introduce the shadow-free automatic water extraction index AWEI nsh Index to implement semantic refinement; use GEE to call the green band G, near-infrared band NIR, and shortwave infrared bands SWIR1 / SWIR2 of Landsat images, and the calculation formula is: ; construct a dual-path prediction confidence fusion function to perform pixel-level correction on the model probability map: if the model prediction probability is 0.4 < y < 0.6 and If the region is a highly significant region, pixel-level recalibration is performed to force the correction of semantic ambiguity in the model at the artificial shoreline, where y represents the predicted probability value of the pixel output by the deep learning model belonging to the water body;

[0057] Spatially isolated pixel clusters are identified using eight-neighbor connected component analysis; an area threshold T is set. area = 500 pixels, automatically discarding connections with fewer than T pixels. area The patches filter out intermittent rivers, scattered pits and ponds, and discrete noise points;

[0058] The corrected binarized mask is converted into a vector format, and intersecting polygons, overhangs, and non-closed loops in the vector line segments are detected. The snapping operator is used to automatically capture and connect broken line segments within a set tolerance (0.5 pixel units). The cleaning operator is used to correct logical overlaps and non-closed structures in the vector line segments, and outputs a long-term shoreline standard database that is spatially continuous, topologically rigorous, and conforms to physical mechanisms.

[0059] The present invention also discloses a method for operating the system as described above, comprising:

[0060] The image data preprocessing and physical prior construction module is used to perform subpixel-level alignment and cloudless synthesis of the original Landsat image. It uses the Scharr operator to extract the NDWI spatial gradient magnitude and constructs a five-channel multidimensional feature cube containing spectral and physical dimensions, providing the Shore-SegFormer model with explicit edge guidance as the initial input.

[0061] The multi-source gap filling and reliability-weighted training set construction module is used to perform gap filling by integrating multi-source surface water products. It automatically extracts permanent water body edges based on water body persistence to generate labels, introduces multi-source consistency probability map (MCP) to measure pixel reliability weights, and automatically constructs a hybrid training set containing water-land boundary features after NCC sub-pixel correction.

[0062] The Shore-SegFormer training module is used to build an integrated explicit gradient-guided edge-aware refinement module EARM and a DR-LDSC module, coupled with a physical consistency and noise tolerance composite loss function, to enhance the model's robustness in capturing shoreline geometric continuity and complex semantics.

[0063] Physics-Topology Dual Calibration Extraction Module: Utilizes sliding window inference to generate probability maps, incorporating AWEI. nsh Exponential, eight-neighbor connected component analysis, and topology repair operators are used to perform geometric correction and quality closed-loop control, outputting topologically rigorous shoreline vector results.

[0064] This invention offers the following advantages: it enables automated, noise-resistant construction and sub-pixel alignment of high-purity training samples. By fusing multi-source surface water products for spatiotemporal reconstruction, it not only addresses the observational gaps in multi-cloud areas with single-source data but also introduces a consistency probability graph (MCP) and reliability weighting mechanism, achieving noise-tolerant learning in a weakly supervised environment. Furthermore, by combining the NCC normalized cross-correlation algorithm with sub-pixel alignment of labels and images, it fundamentally solves the dependence of deep learning on high-quality manual annotation, ensuring the purity and spatial accuracy of the training benchmark.

[0065] Beneficial Effects: This invention significantly enhances the accuracy of edge recognition and localization guided by physical priors, breaking through the limitations of traditional models that rely solely on pure spectral feature extraction. It constructs a five-channel multidimensional feature cube containing Scharr gradient magnitude and NDWI exponent. The Edge Aware Refinement Module (EARM) directly injects explicit physical gradient signals into the feature layer, forcing the model to generate stronger neuronal responses at physical boundaries. This effectively reduces ambiguity interference from shadows and turbid water, ensuring pixel-level accurate localization of shorelines against complex backgrounds.

[0066] This invention ensures the geometric continuity and orientation adaptability of shorelines with complex curvatures. Addressing the challenge of extracting breaks caused by the arbitrary and winding orientation of natural shorelines, this invention utilizes the deformable rotating strip convolution module DR-LDSC. By introducing learnable offsets, the convolution kernel automatically adjusts its sampling geometry according to the actual curvature of the shoreline, effectively capturing long-distance, non-axial spatial dependencies. It resolves geometric and topological breaks caused by vegetation occlusion or low contrast at the operator level, significantly improving the tracking and reconstruction capabilities of linear features.

[0067] This invention constructs a quality closed-loop system constrained by both physical mechanisms and geographic topology. This invention utilizes a physical consistency loss function. and The confidence fusion mechanism forces the prediction results to conform to the physical mechanisms of surface water distribution, accurately filtering out interference from artificial features. Combined with the snapping and cleaning topology repair operators, it systematically corrects overhangs and logical overlaps in the vector results, ensuring that the generated shoreline database possesses a rigorous topological structure and geographic logical consistency. Attached Figure Description

[0068] Figure 1 The present invention provides an overall flowchart of a multi-source gap-filling sample generation and Shore-SegFormer shoreline extraction and calibration method based on continuous water body.

[0069] Figure 2 This is a schematic diagram of the automated process for building datasets in an embodiment of the present invention.

[0070] Figure 3This is a schematic diagram of the overall architecture of the Shore-SegFormer deep learning segmentation model designed for this invention.

[0071] Figure 4 The schematic diagram shows the structural principle of the Shore-SegFormer model edge-aware refinement module (EARM) designed for this invention.

[0072] Figure 5 The schematic diagram shows the structure of the Lightweight Directed Strip Convolutional Module (LDSC) in the Shore-SegFormer model designed for this invention.

[0073] Figure 6 This is a flowchart illustrating the final implementation process of model training and extraction for this invention.

[0074] Figure 7 The structural block diagram of the multi-source remote sensing shoreline automated extraction and quality calibration system constructed for this invention is shown. Detailed Implementation

[0075] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] This invention discloses a method for generating multi-source gap-filling samples and extracting and calibrating Shoreline based on continuous water body conditions, by sequentially executing the following steps 1 to 4, as follows: Figure 1 As shown, this method enables high-precision and automated extraction of complex dynamic shorelines within the study area.

[0077] Step 1: Image data preprocessing and physical prior construction.

[0078] This study integrates Landsat surface reflectance imagery from the Google Earth Engine platform. Spatial geometric intersection calculations were performed on the study area boundaries and satellite orbits to eliminate redundant images. Pixel-level geometric correction and resampling techniques were employed to eliminate spatial displacement biases across sensor images, ensuring sub-pixel-level orientational precision alignment for long-term imagery. For all annual observation images, QA-quality band masking was used to remove cloud cover, cloud shadows, and aerosol interference pixels. A pixel-level mean aggregation strategy was used to calculate the statistical mean of valid observations for the year, generating a high-quality annual composite base map with stable spectral characteristics and no cloud cover.

[0079] Select the near-infrared (NIR), short-wave infrared (SWIR1), and red (Red) bands: Use the strong absorption and strong reflection differences of the NIR band on both sides of the water-land boundary to lock the boundary. Use the high sensitivity of the SWIR1 band to water to enhance the recognition of turbid waters and complex tidal flats. Use the Red band to refine the extraction of the contour texture of ground objects. Synchronously calculate the normalized difference water index (NDWI) of the annual composite image, and its formula is: . Use NDWI to enhance the distinction between water bodies and land backgrounds. Use the 3×3 Scharr spatial gradient operator to perform gradient calculations on the generated NDWI image: Use the horizontal convolution kernel Gx = [[-3, 0, 3], [-10, 0, 10], [-3, 0, 3]] and the vertical convolution kernel Gy = [[-3, -10, -3], [0, 0, 0], [3, 10, 3]] to extract the horizontal and vertical direction gradient components; calculate the gradient amplitude , by capturing the high-frequency mutation signals in the NDWI spatial distribution, amplifying the spectral response at the water-land junction, and generating a boundary candidate prior feature map. Finally, mosaic the above basic bands (NIR, SWIR1, Red), physical index (NDWI), and spatial gradient amplitude map in multiple channels to form a 5-channel multi-dimensional feature cube input image.

[0080] Step 2: Construction of a high-quality training set based on multi-source gap filling and reliability weighting.

[0081] As Figure 2 shown, integrate multi-source surface water products such as GSW, GLAD, and GLC_FCS30D. For the observation gaps caused by cloud cover, adopt the adaptive gap filling method and multi-source fusion strategy to realize the reconstruction of the monthly water body dynamic data set. Calculate the water body frequency (WIF) of each pixel on an annual scale, and the formula is: (where n = 12, t i is the binary water body state in the i-th month), and accurately divide the surface into permanent water bodies (PSW, WIF≥75%), intermittent water bodies (ISW, 5% <WIF <75%), and land (WIF≤5%). Synchronously construct a multi-source consistency probability map (MCP): Compare the prediction results of the three products at the same spatio-temporal position. If the three are completely consistent, the reliability weight w = 1.0. If there are conflicts between two of them, the weight w = 0.6, which is used as the confidence prior for subsequent Shore-SegFormer training.

[0082] Utilizing the long-term distribution stability of permanent water bodies (PSWs), they are defined as shoreline target classes (Label=1). Morphological gradient operators are used to extract the spatial contours of the PSWs, generating initial binary shoreline label images, which are then overlaid with the Landsat annual composite image generated in step 1. When dynamic geomorphological evolution causes systematic shifts between the image and labels, edge gradient features of the annual composite base map are first extracted. Then, local normalized cross-correlation (NCC) feature matching is performed with the binary shoreline labels within a 5×5 pixel window. The peak correlation coefficient within the search window determines the sub-pixel displacement vector, achieving sub-pixel-level alignment correction between the labels and the image. The calibrated labels are uniformly mapped to a single-channel 8-bit grayscale layer to ensure computational accuracy and storage efficiency.

[0083] Three-channel feature images were constructed using NIR, SWIR1, and Red bands, which are sensitive to water body responses. These images were then simultaneously cropped to 256×256 pixels with their corresponding labels, and a pixel ratio threshold T was defined. p Only water pixels are retained, with a percentage of less than 5%. water Slices with less than 95% coverage were used as "mixed samples" to remove redundant information from both land and water areas.

[0084] Step 3: Training the Shore-SegFormer model that integrates physical constraints and deformable operators.

[0085] like Figure 3 As shown, this embodiment constructs an enhanced semantic segmentation model, Shore-SegFormer, for optimizing complex dynamic shorelines. This model deeply adapts the input to the 5-channel feature cube generated in step 1 and introduces a reliability weighting mechanism to combat noise in the automatic annotation process of step 2. The construction process is as follows:

[0086] A hierarchical Transformer architecture is used as the backbone feature extraction network. Its first-layer Overlapping Patch Merging module is custom-designed to adapt to 5-channel input images containing spectral bands, NDWI physical indices, and Scharr gradient maps. H×W×5 Through hierarchical multi-head self-attention computation, four-level multi-scale feature maps {C1, C2, C3, C4} are output step by step. The physical index and gradient map are used to pre-guide neurons to focus on the high-frequency region at the water-land interface.

[0087] like Figure 4 As shown, the explicit gradient-guided edge-aware refinement module (EARM) is deployed on the low-level feature maps {C1, C2}, and the Scharr gradient prior map G generated in step 1 is introduced. prior Implement explicit guidance, and will G priorAs a spatial weighting operator, overlap detection is performed on the feature maps. The calculation logic is as follows: .

[0088] in This indicates element-wise multiplication at the pixel level. This is the Sigmoid activation function. By directly injecting the boundary candidate prior feature map into the network feature layer, the problem of blurred edge localization caused by relying solely on spectral features under complex mudflats and cloud shadow interference is solved.

[0089] To address the geometric characteristics of coastlines in images, which are elongated, arbitrarily oriented, and exhibit varying curvature, DR-LDSC modules are embedded after the high-level stages {C3, C4} of the Shore-SegFormer encoder. The C3 stage is enhanced by receiving feature maps with a spatial resolution of 1 / 16, focusing on capturing the detailed curvature of local meandering coastlines. The C4 stage is enhanced by receiving global semantic feature maps with a spatial resolution of 1 / 32, utilizing a long receptive field to ensure the overall logical continuity of coastlines across regions.

[0090] DR-LDSC employs a dual-path parallel structure to handle spatial dependencies in the horizontal (H-path) and vertical (V-path) directions respectively: before performing strip convolution, it first passes through a lightweight convolutional layer f with 2K channels. off Dynamically learn the 2D offset of each sampling point Geometric adaptive rotation and deformation are performed. For a 1×K horizontal path, the dimension of the offset generation is R. H×W×2K .

[0091] For asymmetric dilated convolution parameters, the kernel size is set to K=15, with a 1×15 kernel used for the horizontal path and a 15×1 kernel used for the vertical path. The dilation rate is set to r=3 in layer C3 and r=5 in layer C4. The coordinates of the sampling points after offset correction are... The pixel values ​​at non-integer coordinates are obtained through bilinear interpolation, and the final output is:

[0092] ;

[0093] ;

[0094] Will and After pixel-level summation and fusion, a 1×1 convolutional layer restores the feature dimensions to the original dimensions of each level of the encoder (d=160 or 256). The residual structure is then used to merge the feature dimensions with the input features, and the result is output to the decoder. Figure 5 As shown.

[0095] The multi-scale feature fusion decoder consists of linear projection, upsampling alignment, feature concatenation, fusion convolution, and a prediction layer. The MLP projection layer maps the four-level feature maps {C1, C2, C3, C4} to a uniform number of channels, d = 256. Bilinear interpolation is used to align the resolution to 1 / 4 of the original. The aligned features are concatenated along the channel axis into a 1024-dimensional tensor, which is then compressed back to 256 dimensions via 1×1 convolution. A single-channel shoreline probability map is output through the linear prediction layer and the Sigmoid operator.

[0096] To overcome label noise in step 2 automatic annotation and to incorporate the remote sensing physical mechanism from step 1, this invention constructs a triple-constraint composite loss function:

[0097] Noise-tolerant weighted binary cross-entropy loss : Utilize the pixel-level reliability weights generated in step 2 Dynamically reweight the classification loss:

[0098] ;

[0099] By reducing the gradient contribution of multi-source conflict regions (noise points), the model is able to extract high-purity features under weak supervision labels.

[0100] Noise tolerance weighted Dice loss To address the sparse pixel distribution along the shoreline, a weighted region overlap constraint is introduced:

[0101] ;

[0102] in The smoothing coefficient is preferably 1×10. -5 It is used for gradient backpropagation under stable extreme distributions.

[0103] Physical consistency constraint loss This method introduces explicit geographical constraints, calculates the Laplacian operator consistency error between the model's predicted probability map and the boundary candidate prior feature maps (NDWI, gradient), and uses second-order derivative constraints to force the extraction results to conform to the physical mechanism of surface water distribution. In this embodiment, the weight coefficients of each loss term are set to... = =1.0, =0.5, and the AdamW optimizer is used for iterative training until the model's IoU and Boundary F-score on the validation set reach the steady-state threshold.

[0104] Step 4: Automated shoreline extraction and calibration.

[0105] Perform pixel-level semantic prediction on a large range of images to be monitored using sliding window inference technology. Set the window size to 256×256 pixels and the window overlap to 50% to eliminate the stitching traces at the window edges. For the overlapping regions, adopt a weighted fusion strategy based on a two-dimensional Gaussian kernel: . Among them, m is the number of windows covering this pixel, and P k is the probability output of the k-th window, and W k is the Gaussian weight coefficient.

[0106] For highly interfering artificial ground objects such as aquaculture areas and salt pans, introduce the AWEI nsh index to implement semantic refinement. Use GEE to call the green band (G), near-infrared band (NIR), and short-wave infrared bands (SWIR1 / SWIR2) of Landsat images. The calculation formula is: , construct a dual-path prediction confidence fusion function to perform pixel-level correction on the model probability map: If the model prediction probability is 0.4 < y < 0.6 and AWEI nsh is a high-significance value area, then perform pixel-level recalibration to forcefully correct the semantic ambiguity of the model at the artificial shoreline.

[0107] Adopt the eight-neighborhood (8-connectivity) connected component analysis method to identify spatially isolated pixel clusters; set the area threshold T area = 500 pixels, and automatically remove patches with fewer connected pixels than T area to filter out intermittent rivers, scattered ponds, and discrete noise points.

[0108] Convert the corrected binary mask to a vector format, detect self-intersections, dangling lines, and non-closed loops in the vector line segments, and use the Snapping operator to automatically capture and connect the broken line segments within the set tolerance (0.5 pixel units); use the Cleaning operator to correct the logical overlaps and non-closed structures in the vector line segments, and output a long-term shoreline standard database with continuous space, strict topology, and conforming to physical mechanisms, as Figure 6 shown.

[0109] As Figure 7 shown, the present invention constructs a shoreline automatic extraction and quality calibration system integrating physical prior guidance and multi-source reliability measurement: including an image preprocessing module, a reliability measurement training set construction module, a Shore-SegFormer training module, and a physical-topological dual calibration extraction module. Through the collaborative coupling of the four functional modules, a full-process closed-loop processing is achieved from the input of the original multi-spectral image, automatic annotation of high-purity samples, geometric adaptive feature extraction to the output of a high-precision shoreline vector result with strict topology.

[0110] Example 1: Experimental Environment and Computational Resource Deployment. The training and inference tasks of the Shore-SegFormer model are implemented based on a general-purpose deep learning computing platform. The hardware environment is equipped with an NVIDIA processor with GPU acceleration, preferably with 12GB or more of GPU memory (to meet the computational requirements of a 5-channel multi-dimensional feature cube), and 16GB or more of RAM. The software environment uses Python 3.8, built on PyTorch 2.0, and configured with CUDA 11.7 and cuDNN libraries to achieve hardware acceleration of the DR-LDSC deformable operator and hierarchical attention mechanism.

[0111] Example 2: Model Training Parameters and Strategy Settings. The 5-channel annual cloudless synthetic feature cube (NIR, SWIR1, Red, NDWI, NDWI gradient magnitude) generated in step 1 was used as input. Image tile sizes were uniformly set to 256×256 pixels, divided in an 8:1:1 ratio. The AdamW optimizer was used during training, with an initial learning rate of 6×10⁻⁶. -5 A cosine annealing strategy is used for dynamic adjustment. The loss function is changed to a triple-constraint composite loss function, namely: the weighted binary cross-entropy guided by the weight w in step 2. ), weighted Dice Loss ( ) and physical consistency constraint loss based on physical gradient distribution ( The training process was iterated for a total of 200 epochs, and the sliding window inference mode was enabled.

[0112] Example 3: Comparison Experiment Results with Mainstream Deep Learning Models. To verify the performance advantages of the Shore-SegFormer model of this invention, it was compared with mainstream models such as SegFormer, U-Net, PSPNet, HRNetV2, and DeepLabV3+ under a unified experimental environment and dataset. The experimental accuracy evaluation index (%) is shown in the table below:

[0113] Model Name mIoU Water body F1 Land F1 Water body recall rate Land accuracy PSPNet 97.9327 97.4521 98.3514 98.2415 98.4125 HRNetV2 97.9171 97.5142 98.3125 98.412 98.5412 DeepLabV3+ 97.7387 97.3125 98.1542 97.8541 98.1254 U-Net 98.4534 98.2145 99.6124 98.5412 98.6542 SegFormer 98.5275 99.1836 99.2462 98.8826 99.6124 Shore-SegFormer 99.2146 99.5412 99.7854 99.4125 99.8541

[0114] Experimental Conclusions and Analysis: Compared to the standard SegFormer, this invention improves mIoU by 0.69 percentage points, thanks to the 5-channel input and EARM module constructed in step 1. By explicitly injecting physical gradient priors, the model achieves optimal discrimination accuracy for land-water edges in complex backgrounds, resolving ambiguity issues caused by turbid water. The significant improvement in water body recall (99.41%) verifies the effectiveness of the Deformable Rotatable Strip Convolutional Module (DR-LDSC). Compared to strip convolution with a fixed kernel size, DR-LDSC can adapt to shoreline curvature, ensuring extremely high geometric continuity in cloud shadow occlusion or extraction of long-distance, fine shoreline segments. Compared to U-Net, this invention exhibits stronger topological consistency when handling easily confused terrains such as aquaculture areas, thanks to the design of the reliability weight w in the composite loss function, enabling the model to learn high-purity features even in noisy data generated by automatic annotation, fully demonstrating the universality and robustness of this invention in practical engineering tasks.

[0115] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A segmented multi-source shoreline extraction and calibration method based on physical priors, characterized in that: include: The original Landsat image was preprocessed to generate an annual composite base map with cloud-free cover and spectrally stable true surface reflectance. The basic bands are extracted from the annual composite base map. Based on the selected basic bands, the NDWI index of the annual composite base map is calculated, thereby obtaining a five-channel multidimensional feature cube. Construct a multi-source consistency probability graph (MCP) as a confidence prior for subsequent Shore-SegFormer model training; The generated initial shoreline binary label image is superimposed with the annual composite image to generate initial labels. After subpixel-level topological verification and geometric alignment, synchronous cropping is performed and a sample categorization screening mechanism is used to obtain a mixed training set containing features of the water-land interface. The Shore-SegFormer model uses a hierarchical Transformer encoder as the backbone feature extraction network. The first layer of the hierarchical Transformer encoder has an overlapping patch embedding module to adapt to the input image, which includes spectral bands, normalized differential water index, and five-channel multi-dimensional feature cube. The model uses the NDWI index and boundary candidate prior feature map to pre-guide neurons to focus on the high-frequency region at the water-land interface. Through hierarchical multi-head self-attention calculation, it outputs four-level multi-scale feature maps step by step. After multi-scale feature fusion decoder, a single-channel shoreline probability map is obtained. Specifically, after the low-level feature map, an explicit gradient-guided edge perception refinement module (EARM) is embedded, combined with the explicit guidance of the multi-source consistency probability map (MCP) to perceive feature map overlap; after the high-level feature map, a DR-LDSC module is embedded to capture the detailed curvature of local meandering shorelines and to ensure the overall logical continuity of cross-regional shorelines by utilizing long-distance receptive fields. A reliability-weighted mechanism is introduced, and a triple-constraint composite loss function is constructed, which is combined with the consistency probability graph MCP to overcome the noise when generating the initial shoreline binary label map. Through sliding window reasoning, AWEI nsh Exponential fusion, eight-neighbor connectivity component analysis, and vector optimization enable high-precision shoreline extraction and database construction for large-scale imagery. The annual composite base map generation process involves: integrating the original Landsat images using the GEE platform, achieving sub-pixel alignment through geometric correction and resampling, calling QA band masks for cloud removal and performing mean aggregation to generate the annual composite base map. Five-channel multidimensional feature cube construction process: Select the NIR, SWIR1, Red, and Green fundamental bands from the generated annual composite base map and calculate the NDWI index of the annual composite base map as follows: ; The 3×3 Scharr operator is used to extract the first-order gradient magnitude of the NDWI exponent, capturing high-frequency abrupt change signals at the boundary to generate candidate prior feature maps of the boundary, as detailed below: The horizontal and vertical gradient components are extracted using a horizontal convolution kernel Gx = [[-3, 0, 3], [-10, 0, 10], [-3, 0, 3]] and a vertical convolution kernel Gy = [[-3, -10, -3], [0, 0, 0], [3, 10, 3]]. Calculate gradient magnitude By capturing high-frequency abrupt change signals in the spatial distribution of NDWI, the spectral response at the water-land interface is amplified to generate a boundary candidate prior feature map; Finally, the basic band, NDWI index, and boundary candidate prior feature maps are stitched together in multiple channels to construct a five-channel multidimensional feature cube.

2. The segmented multi-source shoreline extraction and calibration method based on physical priors according to claim 1, characterized in that: The water frequency (WIF) of each pixel on an annual scale is calculated using the following formula: ； Where n=12, The binary water state for month i; The Earth's surface is divided into: Permanent water bodies PSW, with a corresponding WIF ≥ 75%; Intermittent water body ISW, corresponding to 5% <WIF <75%; For land-based applications, the corresponding WIF ≤ 5%; Simultaneously construct a multi-source consistency probability map (MCP): compare the prediction results of multi-source surface water product data at the same spatiotemporal location. If they are completely consistent, the reliability weight w = 1.

0. If there are two types of conflicts, the weight w = 0.

6. This is used as the confidence prior for subsequent Shore-SegFormer training.

3. The segmented multi-source shoreline extraction and calibration method based on physical priors according to claim 2, characterized in that: Permanent water bodies (PSWs) are defined as shoreline target classes. Morphological gradient operators are used to extract the spatial contours of the PSWs, generating initial binary shoreline labels, which are then overlaid with the annual composite base map. The edge gradient features of the annual composite base map are extracted and then matched with the binary shoreline labels in a 5×5 pixel window using local normalized cross-correlation (NCC) feature matching. The peak correlation coefficient within the search window determines the sub-pixel displacement vector, achieving sub-pixel level alignment correction between the labels and the image. The calibrated labels are uniformly mapped to a single-channel 8-bit grayscale layer.

4. The segmented multi-source shoreline extraction and calibration method based on physical priors according to claim 1, characterized in that: The process of obtaining the training set containing features of the land-water interface is as follows: NIR, SWIR1, and Red bands, which are sensitive to water body response, are selected to construct a 3-channel feature image, which is then simultaneously cropped to 256×256 pixels with the corresponding labels; a pixel ratio threshold T is defined. p T p This represents the proportion of pixels labeled as water bodies out of the total number of pixels in each 256×256 training set slice. Only pixels with a water body percentage less than 5% are retained. wate Slices with less than 95% coverage were used as a mixed dataset of land-water boundary features, removing redundant information from both the entire land and water areas.

5. The segmented multi-source shoreline extraction and calibration method based on physical priors according to claim 1, characterized in that: The Shore-SegFormer model architecture is as follows: Backbone feature extraction network: adopts a hierarchical Transformer architecture; Input layer: An overlapping patch embedding module is used to adapt to a five-channel multidimensional feature cube I ϵ R H×W×5 Through hierarchical multi-head self-attention computation, four-level multi-scale feature maps {C1, C2, C3, C4} are output step by step. The NDWI index and boundary candidate prior feature maps are used to pre-guide neurons to focus on the high-frequency region at the water-land interface. The explicit gradient-guided edge-aware refinement module (EARM) is deployed after the low-level feature maps {C1, C2}, and then the generated boundary candidate prior feature map G is introduced. prior Implement explicit guidance, and will G prior As a spatial weighting operator, overlap detection is performed on the feature map, and the calculation logic is as follows: ; Where X represents the input feature map, Y represents the output feature map, and G... prior Let represent the boundary candidate prior feature map, Conv represent the convolution operation, and BN represent batch normalization. This indicates element-wise multiplication at the pixel level. The activation function is Sigmoid. To address the geometric characteristics of coastlines in images—which are elongated, arbitrarily oriented, and exhibit varying curvature—a DR-LDSC module is embedded after the high-level stages {C3, C4} of the Shore-SegFormer encoder. Stage C3 enhances the feature map with a received spatial resolution of 1 / 16, focusing on capturing the detailed curvature of locally meandering coastlines. Stage C4 enhances the global semantic feature map with a received spatial resolution of 1 / 32, utilizing a long receptive field to ensure the overall logical continuity of the coastline across regions. The DR-LDSC module employs a dual-path parallel structure to handle the spatial dependencies of the horizontal H-path and the vertical V-path separately: before performing strip convolution, it first passes through a lightweight convolutional layer f with 2K channels. off Dynamically learn the 2D offset of each sampling point Perform geometrically adaptive rotation and deformation; for a 1×K horizontal path, the dimension of the offset generation is R. H×W×2K ; For asymmetric dilated convolution parameters, the kernel size K=15 is set, with a 1×15 kernel used for the horizontal path and a 15×1 kernel used for the vertical path. The dilation rate r=3 is set for layer C3 and r=5 for layer C4. The coordinates of the sampling points after offset correction are... The pixel values ​​at non-integer coordinates are obtained through bilinear interpolation, and the final output is: ； ； in, and These represent the output feature values ​​after convolution along the horizontal and vertical paths, respectively, where K represents the kernel size. The convolution weights at the nth sampling point are represented by x, where x represents the pixel value on the input feature map. This represents the coordinates of the center pixel currently being calculated. This represents the original sampling offset coordinates of the standard convolution kernel on a regular grid. This represents the adaptive position offset learned by the model. Indicates the corrected coordinates of the sampling points; Will and After pixel-level summation and fusion, a 1×1 convolutional layer restores the feature dimensions to the original dimensions of each level of the encoder. The residual structure is then used to merge with the input features and output to the decoder. The multi-scale feature fusion decoder consists of linear projection, upsampling alignment, feature concatenation, fusion convolution, and prediction layers. The MLP projection layer maps the number of channels of the four-level feature maps {C1, C2, C3, C4} to d = 256. Bilinear interpolation is used to align the resolution to 1 / 4 of the original. The aligned features are concatenated along the channel axis to form a 1024-dimensional tensor. Cross-scale feature fusion is performed by 1×1 convolution to compress it back to 256 dimensions. A single-channel shoreline probability map is output through the linear prediction layer and the Sigmoid operator.

6. The segmented multi-source shoreline extraction and calibration method based on physical priors according to claim 1, characterized in that: Construct a triple-constraint composite loss function: ; in, This represents the total value of the triple-constraint composite loss function. , and These represent the balance weight coefficients for different loss terms. , and These represent the noise-tolerant weighted binary cross-entropy loss, the noise-tolerant weighted Dice loss, and the physical consistency constraint loss, respectively. Noise-tolerant weighted binary cross-entropy loss : Utilizing the generated pixel-level reliability weights Dynamically reweight the classification loss: ； This represents the total number of pixels in the image, where n is the pixel index. This represents the pixel-level reliability weight of the nth pixel, used for dynamic reweighting. This represents the true label value of the nth pixel. This represents the probability value that the model predicts the nth pixel to be a shoreline. Noise tolerance weighted Dice loss To address the sparse pixel distribution along the shoreline, a weighted region overlap constraint is introduced: ； in, This represents the smoothing coefficient used to prevent the denominator from being zero and to stabilize gradient backpropagation; Physical consistency constraint loss : Introduce explicit geographical constraints, calculate the consistency error of the Laplacian operator between the model's predicted probability map and the boundary candidate prior feature map, and use second derivative constraints to force the extraction results to conform to the physical mechanism of surface water distribution; use the AdamW optimizer for iterative training until the model's IoU and Boundary F-score on the validation set reach the steady-state threshold.

7. The segmented multi-source shoreline extraction and calibration method based on physical priors according to claim 1, characterized in that: A sliding window inference method is used to perform pixel-level semantic prediction on a large area of ​​the image to be monitored. The window size is set to 256×256 pixels, and the window overlap is set to 50% to eliminate stitching marks at the window edges. For the overlapping areas, a weighted fusion strategy based on a two-dimensional Gaussian kernel is adopted. Where m is the number of windows covering the pixel; W represents the probability output for the k-th window. k These are the Gaussian weighting coefficients; For highly interfering artificial features, the shadow-free Automatic Water Extraction Index (AWEI) is introduced nsh Semantic refinement is implemented for the index; by using Google Earth Engine (GEE) to call the green band (G), near-infrared band (NIR), and shortwave infrared bands (SWIR1 / SWIR2) of Landsat images, the calculation formula is: ; A dual-path prediction confidence fusion function is constructed to perform pixel-level correction on the model probability map: if the model prediction probability is 0.4 < y < 0.6 and it is a high-significance value area, pixel-level recalibration is performed to force the correction of the semantic ambiguity of the model at the artificial shoreline, where y represents the predicted probability value of the pixel output by the deep learning model belonging to water Eight-neighbor connected component analysis is used to identify spatially isolated pixel clusters; an area threshold T is set. area = 500 pixels, automatically discarding connections with fewer than T pixels. area plaques; The corrected binarized mask is converted into a vector format, and intersecting polygons, overhangs, and non-closed loops in the vector line segments are detected. The snapping operator is used to automatically capture and connect broken line segments within a set tolerance. The cleaning operator is used to correct logical overlaps and non-closed structures in the vector line segments, and outputs a long-term shoreline standard database that is spatially continuous, topologically rigorous, and conforms to physical mechanisms.

8. A segmented multi-source shoreline extraction and calibration system based on physical priors, characterized in that: The system operates the method according to any one of claims 1-7, comprising: The image data preprocessing and physical prior construction module is used to perform subpixel-level alignment and cloudless synthesis of the original Landsat image. It uses the Scharr operator to extract the NDWI spatial gradient magnitude and constructs a five-channel multidimensional feature cube containing spectral and physical dimensions, providing the Shore-SegFormer model with explicit edge guidance as the initial input. The multi-source gap filling and reliability-weighted training set construction module is used to perform gap filling by integrating multi-source surface water products. It automatically extracts permanent water body edges based on water body persistence to generate labels, introduces multi-source consistency probability map (MCP) to measure pixel reliability weights, and automatically constructs a hybrid training set containing water-land boundary features after NCC sub-pixel correction. The Shore-SegFormer training module is used to build an integrated explicit gradient-guided edge-aware refinement module EARM and a DR-LDSC module, coupled with a physical consistency and noise tolerance composite loss function, to enhance the model's robustness in capturing shoreline geometric continuity and complex semantics. Physics-Topology Dual Calibration Extraction Module: Utilizes sliding window inference to generate probability maps, incorporating AWEI. nsh Exponential, eight-neighbor connected component analysis, and topology repair operators are used to perform geometric correction and quality closed-loop control, outputting topologically rigorous shoreline vector results.

Images