Optical image salt pan extraction method based on optimal transport random walk

Abstract

The present application relates to the technical field of image processing, and is an optical image coastal salt pan extraction method based on optimal transmission random walk, comprising: generating a coastal water mask with complete boundary information; selecting a typical coastal land cover type sample, performing spectral curve analysis on different land cover type samples, and constructing an improved correction salt pan index; based on the principle of spatial proximity, establishing an undirected graph model of the coastal water, and converting the spatial adjacency relationship between water segments into topological connections between graph nodes; constructing a fusion form weight between nodes; using a marked seed point to represent a uniform discrete probability measure formed by all coastal water nodes; obtaining an optimal transmission coupling matrix by solving the feasible solution set of the coupling measure; constructing a coupling form weight matrix of the random walk model, and introducing an absorption intensity to construct an absorbing Markov random walk model, thereby realizing extraction of the coastal salt pan in the optical image.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for extracting coastal salt fields from optical images based on optimal transport random walk. Background Technology

[0002] Coastal salt pans are a special type of artificial wetland. Typically composed of evaporation ponds and crystallization ponds, they produce salt through seawater evaporation, becoming an important source of edible salt for humans and playing an irreplaceable role in social production and daily life. In addition, coastal salt pans possess extremely high ecological value for the Earth's ecosystem and the survival of life, specifically in three aspects: First, the salt pan waters and tidal flats nurture extremely rich food resources, and the relatively stable habitat structure makes them an important alternative habitat for waterbirds in the context of natural wetland degradation, providing crucial resting, feeding, and recuperation sites for various migratory waterbird species during their migrations; second, the salt pan waters not only contain a rich variety and abundant reserves of natural food, but also have a stable and controllable high-salinity water environment, creating an ideal aquaculture environment for shrimp and clam polyculture with suitable water quality, abundant food, and a stable habitat; third, the salt produced by the evaporation and crystallization of seawater in salt pans is an indispensable basic substance in the human diet, maintaining the stability of various physiological functions and ensuring the orderly conduct of human health and life activities.

[0003] However, in recent years, driven by both increasingly intense human activities and climate change as a global environmental challenge, the spatial distribution and dynamic evolution of coastal salt pans have been directly or indirectly affected. Furthermore, the disorderly development and spatial expansion of salt pans have had a series of negative impacts on coastal ecosystems, primarily manifested in the rapid shrinkage of natural wetland areas, increased soil salinization, and a significant decline in regional species diversity. The reduction in wetland area not only weakens the carbon storage function of the coastal zone but also exacerbates biodiversity loss. Simultaneously, soil salinization inhibits soil microbial biomass and enzyme activity, reduces the rate of soil organic carbon sequestration, and hinders normal vegetation growth and development. These various impacts are often interconnected and mutually reinforcing, forming a complex feedback loop, thus highlighting the urgent need for efficient and timely adaptive management of climate change and human activities.

[0004] To address the shortcomings of existing salt field extraction methods, such as insufficient ability to distinguish easily confused features and poor coordination between global and local features, a method for outlining salt field edges using optical images is needed. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for extracting coastal salt pans from optical images based on optimal transport random walk. This invention integrates an improved Modified Normalized Difference Water Index (MNDWI) water body detection method with a dam enhancement technique based on the Frangi filter to generate a coastal water mask with complete boundary information. By constructing an improved Amendatory Saltpan Index (IASI) and combining it with the temporal spectral characteristics of salt pans, accurate identification of crystallization ponds is achieved. An undirected graph model of the coastal water body is established based on the spatial proximity principle, transforming the spatial adjacency relationships between water body segments into topological connections between graph nodes. An optimal transport-based random walk (OT-RW) model is proposed, which optimizes the local dependency limitations of traditional random walks through a globally optimal coupling strategy, thereby achieving accurate delineation of salt pans.

[0006] The technical means employed in this invention are as follows:

[0007] A method for extracting coastal salt pans from optical images based on optimal transport random walk includes: generating a coastal water mask with complete boundary information using an improved normalized differential water index detection method and a dam enhancement technique based on a Frangi filter; the Frangi filter is used to enhance the linear features of the dam, highlighting the continuous linear structure and suppressing speckle-like noise; selecting typical coastal land cover type samples, performing spectral curve analysis on samples of different land cover types, and constructing an improved modified salt pan index, which, combined with the temporal spectral characteristics of the salt pans, enables the identification of crystallization ponds; and establishing an undirected graph model of the coastal water body based on the spatial proximity principle to connect water body segments. The spatial adjacency relationship is transformed into topological connections between graph nodes; the morphological differences and category associations between water bodies are described by the local structural features, overall morphological features, and geometric regularity features of the node's own contour, respectively, and the fusion morphological weights between nodes are constructed; the seed points are used to represent the uniform discrete probability measure composed of all coastal water body nodes, and are used as the target measure; the coupling measure is used to describe the transmission process between the source measure and the target measure, and the optimal transmission coupling matrix is ​​obtained by solving the feasible solution set of the coupling measure; the coupling morphological weight matrix of the random walk model is constructed, and the absorption intensity is introduced to construct the absorption Markov random walk model to realize the extraction of coastal salt fields in optical images.

[0008] Furthermore, the typical coastal land cover type samples include crystallization ponds, aquaculture ponds, tidal flats, buildings, and sediment; sentinel-2 multispectral images are obtained based on the land cover type samples, and... The reflectance of a certain wavelength band is used as a spectral indicator to distinguish between crystallization ponds and aquaculture ponds. The reflectance of a certain band serves as a spectral indicator for distinguishing crystallization pools, sediment, and tidal flats. The reflectance of a wavelength band serves as a spectral indicator for identifying buildings; The improved modified salt pan index (IASI) is used to maximize the difference between crystallization ponds and other similar water bodies and disturbing features. The improved modified salt pan index is defined as follows:

[0009] in, These represent the reflectance of bands 3, 4, 8, and 12 in Sentinel-2, respectively. Indicates the whole scene image The arithmetic mean of the band reflectivity.

[0010] Furthermore, the construction of the undirected graph specifically includes: Assuming there exists inland regions of the coastal zone An independent water body unit ,in For the first Each water body unit will be divided into several water body units. Treat it as a node This will affect the coastal waters. Using undirected graphs This indicates that the set of nodes is... ; An undirected graph The edge set, where each edge connects two adjacent nodes; the edge set and the weight of each edge are described from the perspective of the neighboring nodes; For any water body node , with geometric center Construct a circular neighborhood using the reference point. The radius is defined as:

[0011] in, for The length of the longest side of the smallest bounding rectangle. The scale parameter and This is used to control the extent of the circular neighborhood, including the nodes. Neighborhood The collection of water bodies is represented as:

[0012] in, Represents any and The Euclidean distance between the centroids of the two water bodies; Set constraints to ensure that the neighborhood contains a sufficient number of other nodes: Neighborhood Internal The number of other water body nodes besides itself should not be less than the preset value. That is, satisfying Then the node and The set of edges for other nodes is represented as follows:

[0013] Undirected graph The edge set is the union of all edges, represented as: .

[0014] Furthermore, the local structural features are constructed based on the shape context contour structure similarity of the nodes. The shape context describes the local structural features of the contour by constructing a spatial distribution histogram of contour points in logarithmic polar coordinates; for any two water body nodes... and Extract its contour sampling point set, denoted as and ,in and The number of sampling points for the outlines of the two water bodies; for Establish a logarithmic-polar coordinate system with any point in the graph as the origin, and divide the surrounding area into... Radial intervals and A total of angle intervals, Each statistical unit, through statistics The rest The number of points falling into each unit is normalized to construct a shape context descriptor; the shape context descriptor is essentially a histogram of the local spatial distribution of contour points, reflecting the local structural features of the contour centered on that point; for any two points , The shape context descriptor differences are measured using chi-square distance:

[0015] in, For histogram interval index, express and Chi-square distance between them express Statistical units A histogram representing the distribution of relative positions; .

[0016] Construct a set of contour points based on the chi-square distance. and Cost matrix between , Point With point Differences in shape context between them:

[0017] in, express and Chi-square distance operator for shape context differences between point sets. and The matching is formalized as finding a from and injective mapping The mapping is Each point in The unique matching point assigned in the middle, the optimal matching is represented as:

[0018] in, Indicates all from arrive A single mapping set, Point Mapped to Point index in; Represents a point set The Middle Points Through mapping In point set Matched points The shape context difference value between the two pairs, i.e., the cost of a single match for a matching pair. Minimum cost. This reflects the degree of difference between the two contours in their overall structure; Based on the overall matching cost, the similarity metric between two nodes in terms of shape context features is defined as follows:

[0019] in, Represents a point set and Similarity measure based on shape context features.

[0020] Furthermore, the overall morphological features introduce Hu invariant moments. By calculating the second and third central moments of the image region, seven moment invariants that are invariant to translation, rotation, and scaling are constructed. Mahalanobis distance is used to eliminate the correlation between feature components through the covariance matrix, and components with different dimensions are normalized to reflect the similarity between feature vectors. For any two water body nodes... and The 7-dimensional eigenvectors of the Hu moments are and The difference is measured as:

[0021] in, The covariance matrix of the Hu moment feature matrices for all water samples is used to characterize the covariance relationship between the feature components of the Hu moment. Water body node and Mahalanobis distance, a feature of the Hu moment. The similarity metric for Hu moments obtained from the Mahalanobis distance is as follows:

[0022] in, Representing water body nodes and Similarity measure based on Hu moment features.

[0023] Furthermore, the geometric regularity features introduce a length-width similarity index (LSI). For a body of water, the LSI value is:

[0024] in, The circumference of the water body The area of ​​the water body, coefficient For the empirical scaling factor, Two water body nodes and Similarity in terms of regularity can be measured by their LSI values. and Perform measurement:

[0025] in, This indicates the absolute difference in LSI values ​​between the two water bodies. Represented as water body nodes and Similarity measure based on geometric regularity features (LSI value).

[0026] node and The fusion form weight between Weighted fusion of the similarity of each component feature:

[0027] in, Represents a node and In the The degree of difference in morphological features; These are the corresponding weighting coefficients, used to balance the contributions of different features in the similarity measurement; These represent three types of morphological features, corresponding to shape context, Hu moments, and length-width similarity index, respectively.

[0028] Furthermore, the source distribution corresponds to the labeled seed nodes, while the target distribution is all coastal water nodes; assuming the labeled seed points are... ,in, This represents the water body node in the crystallization pool obtained from the salt field index extraction. This indicates that the water body nodes are obtained after certain screening of non-salt paddy water bodies; the remaining unmarked water body nodes are denoted as... And satisfy ,in This represents all water features on the land side within a 15-kilometer radius of the coastline. In the graph model built based on water body nodes, the labeled seed points are regarded as the source distribution, and the probability measure is... Defined as feature space The discrete measure on the vector evenly distributes the probability mass across the feature vector corresponding to each labeled node:

[0029] in, Indicates the number of marked water body nodes. Let be the corresponding uniform probability mass, and satisfy . ; node eigenvectors, and ; It is located at the feature point Dirac measure on; similarly, Treating all coastal water nodes as the target distribution, probability measure Defined as a uniform discrete probability measure consisting of all coastal water nodes:

[0030] in, This indicates the total number of nodes in all water bodies. For nodes Uniform probability mass on, and satisfying ; For located at feature points Dirac measure on.

[0031] Furthermore, the description of the transmission process from the source measure to the target measure using coupled measures specifically includes: Source measurement With target measure The support sets are the labeled seed node sets. and the set of all water body nodes ,in , To mark the number of nodes and all water body nodes; measured from the source. to target measurement The transmission process is governed by coupling measures that satisfy mass conservation. The coupling measure is described as satisfying the edge distribution condition. and ,in, and Coupling measures exist and Edge measure; the edge distribution condition guarantees that the total mass flowing out from the source node is equal to the edge measure; The total mass flowing into the target node is equal to All coupling measures that satisfy the aforementioned edge distribution conditions Construct a feasible solution set :

[0032] in, Indicates all and Given the set of joint probability distributions of marginal distributions, the goal of optimal transmission is to find a coupling measure. This minimizes the total transmission cost, described as in Linear programming problem on:

[0033] in, For the source node To the target node The transmission cost between them is defined as a weighted combination of the distances to multiple morphological features:

[0034] in, The total number of feature types, These are the weight coefficients of each feature, satisfying... , For nodes and In the A measure of difference on each feature; Using the squared Euclidean distance as a feature difference measure, the squared Euclidean distance is expressed as:

[0035] in, and Representing nodes respectively and In the Vector representation on class features, when hour, This represents the shape context feature vector corresponding to the node; hour, This represents the Hu moment eigenvector corresponding to the node; hour, Represents the LSI feature corresponding to the node; symbol The Euclidean norm of the vector is denoted by the transmission cost, which makes the optimal transmission problem equivalent to minimizing the source measure in the sense of the weighted feature space. and target measure Between Wasserstein distance, will the above Wasserstein distance is expressed as:

[0036] in, It is the first Weight coefficients of class features Indicates in middle The lower bound is obtained; the optimal transmission coupling matrix is ​​obtained by solving.

[0037] Furthermore, the coupled morphological weight matrix will be a fused morphological weight matrix based on local features. With the global correlation matrix extracted from optimal transmission The coupling morphological weight matrix is ​​combined. Represented as:

[0038] in, It is a fusion morphology weight matrix; The global incidence matrix is ​​obtained by optimal transmission coupling calculation, and its elements are... Reflects the nodes and In the global transport plan, the strength of the association between each node and the marked seed node; Using the optimal transmission coupling matrix Define nodes Transmission confidence To mark the sum of coupling quality transmitted from the water body node to this node:

[0039] in, Indicates the first 1 marker seed point; Introducing absorption strength The absorption intensity is defined as a nonlinear mapping of transmission confidence:

[0040] Based on absorption strength and fusion weight Construct a Markov random walk model with node set Divided into transient sets and absorption state set ,Right now Among them, the salt field absorption end The corresponding marked salt field seed node, i.e. Non-salt field absorption end For the corresponding marked non-saline water body seed node, i.e. Under the absorbed Markov random walk model, the node arrive State transition probability Represented as:

[0041] in, It is a node and The boundary weight between; It is a node The degree; It is a node Normalized transport confidence for salt field water body categories, and and They are nodes Transmission confidence levels for salt pan and non-salt pan water body categories. It is a very small constant used to ensure that the denominator is not zero; correspondingly, For nodes Normalized transport confidence for non-saline water body categories; make Represents a node Absorbed end The probability of absorption, based on the principle of total probability in Markov chains, is determined by the nodes. Absorbed end The probability of absorption is determined by both the probability of transfer to other transient nodes and the probability of direct transfer to the foreground absorption end, that is:

[0042] If the absorption probability of a node is written in vector form Then the vector expression of equation (35) is:

[0043] in, Represent a The identity matrix; It is a diagonal matrix, where the diagonal elements are the absorption intensities of each node. Transition matrix Adjacency matrix The row-normalized matrix whose elements ; Given the normalized confidence vector of a node, the absorption probability of a node has a unique solution:

[0044] According to the maximum absorption probability criterion, for any transient node Its categories are:

[0045] in, Represents transient nodes The final classification result under the maximum absorption probability criterion is a vector of absorption probabilities. In the middle, node The category index corresponding to the maximum absorption probability component is the category to which the node belongs, as determined by the random walk model.

[0046] Compared with the prior art, the present invention has the following advantages: This invention constructs an OT-RW fusion mathematical model, aiming to provide a more robust algorithmic framework for salt field extraction by combining the global distribution modeling capability of optimal transmission with the local structure capture advantage of random walk.

[0047] This invention completes the theoretical derivation of the OT-RW method in the salt field extraction scenario, clarifies the mathematical principles and applicable boundaries of the model in distinguishing between salt fields and easily confused land features, and provides solid theoretical support for the practical application of the model.

[0048] This invention establishes a global-local collaborative walk mechanism between OT and RW, which not only solves the problem of traditional random walk relying too much on local information and being susceptible to noise interference, but also makes up for the shortcomings of optimal transmission in capturing local details, thus achieving a balance between global consistency and local accuracy in the salt field extraction process. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a framework diagram of the optical image coastal salt field extraction method based on optimal transmission random walk in this invention.

[0051] Figure 2 This is a spectral comparison of different land cover types in this invention.

[0052] Figure 3 The results are the Sentinel-2 data experiments in the embodiments of the present invention. Detailed Implementation

[0053] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. 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.

[0055] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0056] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0057] like Figure 1 As shown, this invention provides a method for extracting coastal salt fields from optical images based on optimal transport random walk, comprising: generating a coastal water mask with complete boundary information by using an improved normalized differential water index water body detection method and a dam enhancement technique based on a Frangi filter; the Frangi filter is used to enhance the linear features of the dam, highlight the continuous linear structure, and suppress speckle-like noise.

[0058] Typical coastal land cover types were selected as samples, and spectral curve analysis was performed on samples of different land cover types. An improved modified salt pan index was constructed, and crystallization ponds were identified by combining the temporal spectral characteristics of salt pans. To intuitively analyze the spectral characteristics of the selected typical land cover types, spectral response curves were plotted based on surface reflectance data of five typical coastal land cover samples, such as... Figure 2 As shown in the figure, each point represents the average reflectance predicted by support vector machine (SVM) regression of all samples within a specific spectral band. The curves reflect the spectral variation trends of various features in the visible to near-infrared range. In a specific implementation, as a preferred embodiment of the present invention, typical coastal land cover type samples include crystallization ponds, aquaculture ponds, tidal flats, buildings, and sediment; sentinel-2 multispectral images are obtained based on the land cover type samples, and... The reflectance of a certain wavelength band is used as a spectral indicator to distinguish between crystallization ponds and aquaculture ponds. The reflectance of a certain band serves as a spectral indicator for distinguishing crystallization pools, sediment, and tidal flats. The reflectance of a wavelength band serves as a spectral indicator for identifying buildings; The Improved Modified Salt Field Index (IASI) is used to maximize the difference between crystallization ponds and other similar water bodies and disturbing features. The Improved Modified Salt Field Index is defined as follows:

[0059] in, These represent the reflectance of bands 3, 4, 8, and 12 in Sentinel-2, respectively. Indicates the whole scene image The arithmetic mean of the band reflectance. This index achieves efficient extraction in the crystallization pool through a two-stage design: the first utilizes... Scale up the crystallization pool The positive slope characteristics of the band, and by introducing The first term suppresses interference from aquaculture ponds and some buildings; the second term serves as a spatial adaptive correction term, avoiding invalid responses by taking a non-negative value, and utilizing... The difference between the band and the global mean further suppresses the high reflectance interference from buildings in the shortwave infrared, while maintaining the high reflectance characteristics of the crystallization pool due to salt crystal enrichment. This index makes the crystallization pool area exhibit a high response value, while the response of interfering features such as aquaculture ponds, tidal flats, silt, and buildings is significantly reduced, thereby achieving high-precision automated extraction from the crystallization pool.

[0060] For the crystallization pool, its reflectance value is from band to The band gradually increases, and in The band reached its peak and then began to decline. However, it's worth noting that... Within the specified spectral range, the spectral curve of the crystallization pond exhibits a distinct positive spectral slope, while the reflectance of the aquaculture pond shows a different slope. band to There was a slight increase between the bands, but... It then turned into a downward trend, and in The band exhibits a negative spectral slope. Therefore, it can be used... As a key spectral indicator for distinguishing crystallization ponds from aquaculture ponds. However, because the spectral reflectance of buildings in specific wavelength bands is much greater than that of other ground features, and in It shows an upward trend within the band range, therefore, when passing through During the process of suppressing aquaculture ponds, interference from buildings may be introduced.

[0061] Secondly, for tidal flats, their reflectivity is... band to The overall trend is a slow decline across different bands, but... band to The decrease in spectral reflectance between bands is much smaller than that in the crystallization pool. Furthermore, the spectral reflectance of sediment in this range shows a slow upward trend, a stark contrast to the rapid decrease in the crystallization pool. For buildings, the overall trend within this band is also slow. Therefore, through… It can be used to distinguish crystallization pools from silt and tidal flats, and can also suppress the interference of buildings to a certain extent.

[0062] Finally, based on the spectral curves of typical ground features obtained, the study found that buildings exhibit high reflectance across the entire visible to near-infrared range, especially in the shortwave infrared region. In the spectral band, its reflectance is significantly higher than that of other ground features, exhibiting a unique spectral response. Therefore, The reflectivity of a band can serve as an important basis for identifying buildings and distinguishing them from other wetland or water features.

[0063] An undirected graph model of the coastal water body is established based on the principle of spatial proximity, transforming the spatial adjacency relationships between water body segments into topological connections between graph nodes. Specifically, as a preferred embodiment of this invention, the construction of the undirected graph includes: Assuming there exists inland areas of the coastal zone An independent water body unit ,in For the first Each water body unit will be divided into several water body units. Treat it as a node This will affect the coastal waters. Using undirected graphs This indicates that the set of nodes is... ; An undirected graph The edge set, where each edge connects two adjacent nodes; the edge set and the weight of each edge are described from the perspective of the neighboring nodes; For any water body node , with geometric center Construct a circular neighborhood using the reference point. The radius is defined as:

[0064] in, for The length of the longest side of the smallest bounding rectangle. The scale parameter and This is used to control the extent of the circular neighborhood, including the nodes. Neighborhood The collection of water bodies is represented as:

[0065] in, Represents any and The Euclidean distance between the centroids of the two water bodies; Set constraints to ensure that the neighborhood contains a sufficient number of other nodes: Neighborhood Internal The number of other water body nodes besides itself should not be less than the preset value. That is, satisfying Then the node and The set of edges for other nodes is represented as follows:

[0066] Undirected graph The edge set is the union of all edges, represented as: .

[0067] This invention describes the morphological differences and category associations between water bodies by using the local structural features, overall morphological features, and geometric regularity features of the node's own contour, and constructs a fusion morphological weight between nodes. In a preferred embodiment, the local structural features are constructed based on the shape context contour structural similarity of the nodes. The shape context describes the local structural features of the contour by constructing a spatial distribution histogram of contour points in logarithmic polar coordinates. This method exhibits good robustness to non-rigid deformation. For any two water body nodes... and Extract its contour sampling point set, denoted as and ,in and The number of sampling points for the outlines of the two water bodies; for Establish a logarithmic-polar coordinate system with any point in the graph as the origin, and divide the surrounding area into... Radial intervals and A total of angle intervals Each statistical unit, through statistics The rest The number of points falling into each unit is normalized to construct a shape context descriptor; the shape context descriptor is essentially a histogram of the local spatial distribution of contour points, reflecting the local structural features of the contour centered on that point; for any two points , The shape context descriptor differences are measured using the chi-square distance metric:

[0068] in, For histogram interval indexing, express and The chi-square distance between them express Statistical units Histogram representing the distribution of relative positions The magnitude of this metric reflects the similarity between the shape descriptors of two nodes; the smaller the value, the more similar the shapes of the two nodes.

[0069] To measure the overall shape similarity between the contours of two water bodies, a globally optimal matching relationship needs to be found based on the local differences between all point pairs in their contour point sets. The contour point sets are constructed using the chi-square distance. and Cost matrix between , Point With point Differences in shape context between them:

[0070] in, express and Chi-square distance operator for shape context differences between the two.

[0071] This matrix fully characterizes the local structural differences between any pair of points in the two contour point sets. From the perspective of overall matching, to make the two contours most similar overall, it is necessary to find an optimal point pair correspondence that minimizes the sum of the costs of all matching point pairs. (The point set is then...) and The matching is formalized as finding a from and injective mapping The mapping is Each point in The unique matching point assigned in the middle, the optimal matching is represented as:

[0072] in, Indicates all from arrive A single mapping set, Point Mapped to Point index in; Represents a point set The Middle Points Through mapping In point set Matched points The shape context difference value between them, that is, the minimum cost of a single match for this matching pair. It reflects the degree of difference between the two contours in their overall structure; the smaller the value, the more similar the two contours are in shape.

[0073] Based on the overall matching cost, the similarity metric between two nodes in terms of shape context features is defined as follows:

[0074] in, Represents a point set and This exponential mapping, based on shape context features, ensures similarity. Matching cost As it increases, it monotonically decreases, and its range is... .therefore, The larger the value, the more similar the outlines of the two water bodies are in terms of morphology.

[0075] To more effectively measure the nodes of two water bodies and Regarding the similarity of global shapes, considering the differences in dimensions and value ranges among the components of the Hu invariant moments, and the potential correlation between components, Mahalanobis distance is adopted. Its core idea is to eliminate the correlation between feature components through the covariance matrix and normalize components with different dimensions, thereby reflecting the similarity between feature vectors. Specifically, in a preferred embodiment of this invention, the overall morphological features are introduced using Hu invariant moments. By calculating the second and third central moments of the image region, seven moment invariants that are invariant to translation, rotation, and scaling are constructed. Mahalanobis distance is used to eliminate the correlation between feature components through the covariance matrix and normalize components with different dimensions, thereby reflecting the similarity between feature vectors. For any two water body nodes... and The 7-dimensional eigenvectors of the Hu moments are and The difference is measured as:

[0076] in, The covariance matrix of the Hu moment feature matrices for all water samples is used to characterize the covariance relationship between the feature components of the Hu moment. Water body node and The Mahalanobis distance between the Hu moments is the largest possible value. The smaller the distance, the more similar the two water bodies are in terms of the global shape features described by the Hu moments.

[0077] The similarity measure of Hu moments obtained from Mahalanobis distance is:

[0078] in, Representing water body nodes and The similarity metric is based on the Hu moment feature. The larger the value, the more similar the two water bodies are in terms of the global shape features described by the Hu moment, and vice versa.

[0079] While the two descriptors mentioned above can effectively capture the global and local morphological features of water body nodes, they still have limitations in distinguishing between artificial regular shapes and natural irregular shapes. This is because in coastal scenes, artificial features such as aquaculture ponds often exhibit regular rectangular characteristics, while natural water body shapes (such as bays and natural lakes) show varied features with winding edges and no fixed outline pattern. In a preferred embodiment of this invention, the geometric regularity feature introduces a length-width similarity index (LSI), the core idea of ​​which stems from a fundamental fact in geometry: among all rectangles of equal area, the square has the shortest perimeter. For a water body, the LSI value is:

[0080] in, The circumference of the water body The area of ​​the water body, coefficient This is an empirical scaling factor, typically initialized to 1.0. For an ideal square, its perimeter is exactly... Therefore, the LSI value is 1. Other shapes, due to their longer perimeters, will have LSI values ​​greater than 1. Thus, the closer the LSI value of a body of water is to 1, the closer its shape is to a square, and the higher its regularity; if the value is much greater than 1, the shape is more elongated or irregular.

[0081] Two water body nodes and Similarity in terms of regularity can be measured by their LSI values. and Perform measurement:

[0082] in, This indicates the absolute difference in LSI values ​​between the two water bodies. Represented as water body nodes and The similarity measure is based on geometric regularity features (LSI value), and its value ranges from (0,1]. The larger the value, the closer the geometric regularity of the two water bodies is.

[0083] node and The fusion form weight between Weighted fusion of the similarity of each component feature:

[0084] in, Represents a node and In the The degree of difference in morphological features; its range is The larger the value, the more similar the two nodes are to that feature; These are the corresponding weighting coefficients, used to balance the contributions of different features in the similarity measurement; These represent three types of morphological features, corresponding to shape context, Hu moments, and length-width similarity index, respectively.

[0085] Finally, by weighted fusion of the above three types of morphological features, a fusion morphological weight is constructed between nodes. This metric organically integrates the local contour structure, global shape invariance, and geometric regularity information of water body nodes, effectively characterizing the morphological similarity between nodes. However, the fusion of morphological weights... This approach primarily relies on the morphological characteristics of nodes themselves, without fully considering the overall correlation and structural matching of nodes in spatial distribution. This limitation may lead to biases in similarity measurements in regions with complex shapes or irregular distributions. Therefore, it is necessary to introduce optimal transmission theory and construct structural weights that can globally match the morphological distribution between nodes.

[0086] The effectiveness of traditional Random Walk (RW) methods relies on the assumption of local smoothness, meaning that in graphs with uniform feature distribution, adjacent nodes tend to have similar labels. However, this assumption often fails in graph data with high feature heterogeneity and complex structures. Specifically, when there are extensive non-smooth regions in the graph (such as semantic boundaries with abrupt feature changes or sparsely connected heterogeneous communities), relying solely on the propagation mechanism of local neighborhoods cannot form globally consistent inference paths. This leads to discrepancies in label inferences when crossing different regions, compromising the consistency of global classification results. To address this issue, Optimal Transport (OT) theory is introduced to construct an OT-RW framework. The core motivation lies in OT modeling the optimal coupling of probability quality in geometric space, transforming the label assignment problem into a problem of minimizing the cost of probability quality transport in feature space, thus overcoming the dependence of traditional RW on the smoothness of local neighborhoods. Unlike traditional random walk (RW) methods that rely on local neighborhood relationships for label propagation, OT (Optimal Targeting) takes a global approach, directly establishing semantic correspondences between labeled and unlabeled nodes by calculating the overall optimal solution for all possible transmission paths, thus overcoming the limitations imposed by the local smoothness assumption. Specifically, the problem of inferring the category label of an unlabeled node set based on a set of labeled nodes is modeled as an OT problem. This approach has two significant advantages: (1) The global optimization mechanism of OT can alleviate the inherent local bias of existing RW and improve the robustness of label propagation in non-uniform feature scenarios; (2) By explicitly modeling the optimal transmission cost between nodes based on features and structure, this method can achieve cross-regional semantic consistency reasoning while maintaining the continuity of local context, thereby obtaining detection results superior to existing random walk and its derivative methods.

[0087] To achieve optimal transmission modeling, the source and target distributions need to be clearly defined. A labeled seed point is used to represent a uniform discrete probability measure consisting of all coastal water nodes, and this seed point serves as the target measure. In a preferred embodiment of this invention, the source distribution corresponds to the labeled seed nodes, and the target distribution consists of all coastal water nodes. Assuming the labeled seed points are... ,in, This represents the water body node in the crystallization pool obtained from the salt field index extraction. This indicates that the water body nodes are obtained after certain screening of non-salt paddy water bodies; the remaining unmarked water body nodes are denoted as... And satisfy ,in This represents all water features on the land side within a 15-kilometer radius of the coastline. In the graph model built based on water body nodes, the labeled seed points are regarded as the source distribution, and the probability measure is... Defined as feature space The discrete measure on the vector evenly distributes the probability mass across the feature vector corresponding to each labeled node:

[0088] in, Indicates the number of marked water body nodes. Let be the corresponding uniform probability mass, and satisfy . ; node eigenvectors, and ; It is located at the feature point Dirac measure on; similarly, Treating all coastal water nodes as the target distribution, probability measure Defined as a uniform discrete probability measure consisting of all coastal water nodes:

[0089] in, Indicates the total number of nodes in all water bodies; For nodes Uniform probability mass on, and satisfying ; For located at feature points The Dirac measure on.

[0090] The transmission process between a source measure and a target measure is described using a coupling measure, and the optimal transmission coupling matrix is ​​obtained by solving the feasible solution set of the coupling measure. Specifically, as a preferred embodiment of this invention, describing the transmission process between a source measure and a target measure using a coupling measure includes: Source measurement With target measure The support sets are the labeled seed node sets. and the set of all water body nodes ,in , To mark the number of nodes and all water body nodes; measured from the source. to target measurement The transmission process is governed by coupling measures that satisfy mass conservation. Description: The coupling measure satisfies the edge distribution condition. and ,in, and Coupling measures exist and Edge measure; edge distribution condition guarantees that the total mass flowing out of the source node is equal to the edge measure; The total mass flowing into the target node is equal to All coupling measures that satisfy the edge distribution condition Construct a feasible solution set :

[0091] in, Indicates all and Given the set of joint probability distributions of marginal distributions, the goal of optimal transmission is to find a coupling measure. This minimizes the total transmission cost, described as in Linear programming problem on:

[0092] in, For the source node To the target node The transmission cost between them is defined as a weighted combination of the distances to multiple morphological features:

[0093] in, The total number of feature types, These are the weight coefficients of each feature, satisfying... , For nodes and In the A measure of difference on each feature; Using the squared Euclidean distance as a measure of feature difference, the squared Euclidean distance is expressed as:

[0094] in, and Representing nodes respectively and In the Vector representation on class features, when hour, This represents the shape context feature vector corresponding to the node; hour, This represents the Hu moment eigenvector corresponding to the node; hour, Represents the LSI feature corresponding to the node; symbol The Euclidean norm of a vector; transmission cost, making the optimal transmission problem equivalent to minimizing the source measure in the sense of a weighted feature space. and target measure Between Wasserstein distance, will Wasserstein distance is expressed as:

[0095] in, It is the first Weight coefficients of class features Indicates in middle The lower bound is obtained; the optimal transmission coupling matrix is ​​then solved. Therefore, minimizing the transmission cost is essentially equivalent to reducing the overall morphological difference between the source node set and the target node set in terms of probability distribution, providing the optimal global allocation basis for tag propagation.

[0096] Since directly solving the above optimal transport problem has high computational complexity, entropy regularization is introduced to improve solution efficiency and numerical stability, resulting in the Wasserstein distance with entropy regularization:

[0097] in, For transmission coupling entropy, is the regularization coefficient. The introduction of entropy not only ensures the strict convexity of the objective function, but also facilitates efficient solution using the Sinkhorn algorithm.

[0098] To transform the theoretical model into a discrete form suitable for numerical computation, the probability measure and transmission coupling need to be discretized and modeled. Specifically, in the discrete form, the source and target measures correspond to probability vectors, respectively. And satisfy Transmission cost matrix between nodes Given, its elements Within this discrete framework, all transmission plans that satisfy the mass conservation constraint constitute a feasible solution set. This set corresponds to a specific geometric structure within the framework of optimal transmission theory, namely, a transmission-coupled polytope, which takes the following form:

[0099] in, and Given two vectors of dimensions n and m, both consisting of only 1s, and their coupling matrices. Satisfying the marginal distribution condition and That is, each source node The total mass flowing to all target nodes is equal to their initial mass. Each target node The total mass received from all source nodes equals its target mass. Therefore, the discrete-form optimal transport problem can be described as a problem within a coupled multicellular structure. To minimize the transmission cost, the optimal transmission problem can be formulated as:

[0100] in, For the cost matrix, For Frobenius inner product. Directly solving equation (23) still faces numerical instability and high computational complexity. Therefore, an entropy regularization term is introduced to relax the above equation to:

[0101] in, For regularization functions, This indicates that the source node is represented. Transmitted to the target node The probability mass, The strength of regularization is controlled to regulate the weight of the entropy term in the objective function. For efficient solution, a Lagrange multiplier vector is introduced. and To handle marginal constraints, the problem can be transformed into an unconstrained optimization problem, with the corresponding Lagrangian function being:

[0102] Its unfolded form is as follows:

[0103] This function corresponds to the minimization objective of the primal optimization problem. In the entropy regularization term... Under the influence of the function about It is strictly convex. Therefore, by setting the partial derivative... The stability condition corresponding to the global minimum can be obtained directly:

[0104] The optimal solution can be obtained by rearranging:

[0105] To simplify the expression above, a scaling vector is defined. and kernel matrix Then it is written as:

[0106] When vector and Satisfy edge conditions When this happens, the solution can be efficiently obtained through Sinkhorn iteration as follows:

[0107] After iterative convergence, the optimal transmission coupling matrix is ​​obtained. .

[0108] A coupled morphological weight matrix for a random walk model is constructed, and an absorption Markov random walk model is built by introducing absorption intensity to extract coastal salt fields from optical images. Specifically, in a preferred embodiment of this invention, the coupled morphological weight matrix is ​​a fused morphological weight matrix based on local features. With the global correlation matrix extracted from optimal transmission Combine and couple the morphological weight matrix Represented as:

[0109] in, It is a fusion morphology weight matrix; The global incidence matrix is ​​obtained by optimal transmission coupling calculation, and its elements are... Reflects the nodes and In the global transport plan, the strength of the association between each node and the labeled seed node; a larger value indicates that the two are more likely to belong to the same category in the global distribution; symbol This represents the Hadamard product of matrices. The Hadamard product combines local shape weights with global correlation weights, achieving a complementary effect between the two and fusing shape weights. It can effectively characterize the similarity of neighboring nodes in terms of morphological details, but its perception of the association between distant or morphologically heterogeneous nodes is weak; global weight By capturing semantic relationships at the distribution level through optimal transmission, it can identify nodes that are dissimilar in local features but belong to the same category in the global distribution, thus compensating for the shortcomings of local methods. The combination of the two allows the model to retain sensitivity to local structures while introducing global semantic information that transcends the local neighborhood, enhancing its adaptability to heterogeneous graph structures and alleviating the over-reliance on the traditional assumption of local smoothness.

[0110] To differentiate the strength of the category information received by each node from the labeled seed node during a random walk, a reliable quantitative metric is needed. This can be achieved using the optimal transmission coupling matrix. Define nodes Transmission confidence To mark the sum of coupling quality transmitted from the water body node to this node:

[0111] in, Indicates the first 1 marker seed point; This reflects the prior category information of the labeled node being passed to the node. The larger the overall strength, the closer the relationship between the node and the tag set in the sense of optimal transmission.

[0112] Introducing absorption strength The absorption intensity is defined as a nonlinear mapping of the transmission confidence:

[0113] This function ensures And it is It is a monotonically increasing function. When When very low, A value close to 0 indicates that the node absorbs or retains almost no external label information; when... At higher levels, A value that quickly approaches 1 indicates that the node can effectively absorb and retain category information from the labeled node.

[0114] Based on absorption strength and fusion weight Construct a Markov random walk model with node set Divided into transient sets and absorption state set ,Right now Among them, the salt field absorption end The corresponding marked salt field seed node, i.e. Non-salt field absorption end For the corresponding marked non-saline water body seed node, i.e. Under the absorbed Markov random walk model, the nodes arrive State transition probability Represented as:

[0115] in, It is a node and The boundary weight between; It is a node The degree; It is a node Normalized transport confidence for salt field water body categories, and and They are nodes Transmission confidence levels for salt pan and non-salt pan water body categories. It is a very small constant used to ensure that the denominator is not zero; correspondingly, For nodes Normalized transport confidence for non-saline water body categories; make Represents a node Absorbed end The probability of absorption, based on the principle of total probability in Markov chains, is determined by the nodes. Absorbed end The probability of absorption is determined by both the probability of transfer to other transient nodes and the probability of direct transfer to the foreground absorption end, that is:

[0116] If the absorption probability of a node is written in vector form Then the vector expression of equation (35) is:

[0117] in, Represent a The identity matrix; It is a diagonal matrix, where the diagonal elements are the absorption intensities of each node. Transition matrix Adjacency matrix The row-normalized matrix whose elements ; Given the normalized confidence vector of a node, the absorption probability of a node has a unique solution:

[0118] According to the maximum absorption probability criterion, for any transient node Its categories are:

[0119] in, Represents transient nodes The final classification result (category label) under the maximum absorption probability criterion is a vector of absorption probabilities. In the middle, node The category index corresponding to the maximum absorption probability component is the category to which the node belongs, as determined by the random walk model.

[0120] To quantitatively analyze the extraction performance of the proposed algorithm in salt fields, the Intersection over Union (IOU) ratio and Kappa coefficient were used to evaluate the performance of the salt field extraction rows.

[0121] IOU directly measures the degree of spatial overlap between the predicted results and the actual annotations by calculating the ratio of the intersection to the union of the predicted results and the actual annotations. It can be expressed as:

[0122] Where TP represents the number of pixels correctly classified as salt fields, FP represents the number of pixels misclassified as salt fields, TN represents the number of pixels correctly classified as non-salt fields, and FN represents the number of pixels misclassified as non-salt fields. A higher OA value indicates better overall classification performance of the model. A higher IOU value indicates a greater overlap between the extracted results and the ground truth map, meaning the extracted results better reflect the actual distribution of salt fields. The Kappa coefficient is used to evaluate the consistency between the extracted results and the truth map. Its calculation formula is as follows:

[0123] in, The ratio of the sum of the number of correctly classified samples in each category to the total number of samples can be expressed as:

[0124] The ratio of the sum of the products of the actual sample size and the predicted sample size for each category to the square of the total sample size can be expressed as: .

[0125] Example To verify the effectiveness of the proposed algorithm in the extraction task of dense coastal salt fields, the patent systematically compared the extraction performance of the proposed algorithm with four existing methods, namely ASI-MLSD, NDSI, ZBP and HPAC, in Shandong Province. Figure 3 The results of salt field extraction experiments in Shandong Province are shown in the following figures: (a) is the ground truth map of salt field distribution in the region, where salt field areas are highlighted in red; (b)–(f) show the salt field extraction results of five algorithms in the region, where blue indicates correctly extracted salt field areas, green indicates missed salt field water bodies, and red indicates incorrectly extracted non-salt field water bodies.

[0126] Overall, all five methods can achieve extraction in salt pan areas to some extent. However, as... Figure 3 As shown, various comparative algorithms still exhibit varying degrees of missed detections and false detections. Specifically, the NDSI-based algorithm struggles to effectively extract evaporation ponds, primarily because it cannot adequately distinguish the spectral characteristics of low-salinity evaporation ponds from other seawater (such as aquaculture ponds). In contrast, the ZBP algorithm achieves more complete salt field extraction by fusing spectral information from multiple typical land features; however, due to the spectral similarity between evaporation ponds and aquaculture ponds, it still suffers from numerous missed detections of evaporation ponds. Furthermore, because some dikes and salt fields have relatively small spectral differences, ZBP sometimes misclassifies dikes as salt fields during the extraction process.

[0127] The HPCA method outperforms NDSI and ZBP in salt field extraction, likely because it primarily relies on near-infrared reflectance approaching zero to distinguish evaporation ponds from other water bodies. However, HPCA may fail when the near-infrared spectral value of the evaporation pond is close to zero. Furthermore, since this algorithm mainly considers near-infrared reflectance, it can still produce many false positives if other land features have small spectral differences with the evaporation or crystallization ponds in this band. ASI-MLSD can extract most of the evaporation ponds that NDSI and ZBP failed to identify effectively. Compared to HPCA, ASI-MLSD utilizes the spatial structure features between evaporation and crystallization ponds for extraction, making it more robust than relying solely on spectral differences. However, due to missed detections in the crystallization pond extraction stage, significant missed detections also occur in subsequent evaporation pond extraction based on spatial structure. Experimental results show that the method of this invention can effectively extract salt field water in this area.

[0128] Table 1. Performance Evaluation of Salt Field Extraction in Shandong Province

[0129] To quantitatively evaluate the extraction performance of ASI-MLSD, NDSI, ZBP, HPAC, and OT-RW algorithms in coastal salt fields, Table 1 calculates the extraction performance of these five methods in salt fields in Shandong Province. Analysis revealed that the patented method had higher IOU and Kappa coefficients than the comparative algorithms, with an IOU increase of at least 20% and a Kappa coefficient increase of at least 15%.

[0130] In summary, the patented method outperforms the comparative methods in multiple indicators. The proposed OT-RW method still exhibits a small number of false positives and false negatives during salt field extraction, mainly due to: 1) adhesion between some small water bodies and the evaporation pond during dam reinforcement; and 2) salt precipitation during crystallization pond extraction leading to missed extraction in some crystallization areas, thus affecting the overall extraction accuracy of the salt field. Despite this, from... Figure 3 As shown in Table 1, the experimental results demonstrate that OT-RW can effectively extract salt fields. This method uses initially labeled water seed points to construct the association between coastal water nodes based on optimal transport theory, and completes the automatic extraction of salt fields using a random walk algorithm. While maintaining high extraction integrity, it significantly reduces misclassification and omission errors, demonstrating good robustness and applicability.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for extracting coastal salt fields from optical images based on optimal transport random walk, characterized in that, include: A coastal water mask with complete boundary information is generated by using an improved normalized differential water index water body detection method and a dam enhancement technique based on the Frangi filter. The Frangi filter is used to enhance the linear characteristics of the dam, highlight the continuous linear structure, and suppress speckle-like noise; Typical coastal land cover type samples were selected, and spectral curve analysis was performed on samples of different land cover types. An improved modified salt field index was constructed, and the crystallization pools were identified by combining the temporal spectral characteristics of salt fields. An undirected graph model of the coastal water body is established based on the principle of spatial proximity, and the spatial adjacency relationship between water body segments is transformed into the topological connection between graph nodes. The morphological differences and category associations between water bodies are described by the local structural features, overall morphological features and geometric regularity features of the node's own contour, and the fusion morphological weights between nodes are constructed. The seed points are used to represent the uniform discrete probability measure of all coastal water nodes, and this measure is used as the target measure. The transmission process between the source measure and the target measure is described using a coupling measure, and the optimal transmission coupling matrix is ​​obtained by solving the feasible solution set of the coupling measure. A coupled morphological weight matrix for a random walk model is constructed, and an absorption Markov random walk model is constructed by introducing absorption intensity to extract coastal salt fields from optical images.

2. The method for extracting coastal salt fields from optical images based on optimal transport random walk according to claim 1, characterized in that, The typical coastal land cover type samples include crystallization ponds, aquaculture ponds, tidal flats, buildings, and sediment; sentinel-2 multispectral images were obtained based on the land cover type samples. The reflectance of a certain wavelength band is used as a spectral indicator to distinguish between crystallization ponds and aquaculture ponds. The reflectance of a certain band serves as a spectral indicator for distinguishing crystallization pools, sediment, and tidal flats. The reflectance of a wavelength band serves as a spectral indicator for identifying buildings; The improved modified salt pan index (IASI) is used to maximize the difference between crystallization ponds and other similar water bodies and disturbing features. The improved modified salt pan index is defined as follows: in, These represent the reflectance of bands 3, 4, 8, and 12 in Sentinel-2, respectively. Indicates the whole scene image The arithmetic mean of the band reflectivity.

3. The method for extracting coastal salt fields from optical images based on optimal transport random walk according to claim 1, characterized in that, The construction of the undirected graph specifically includes: Assuming there exists inland areas of the coastal zone An independent water body unit ,in For the first Each water body unit will be divided into several water body units. Treat it as a node This will affect the coastal waters. Using undirected graphs This indicates that the set of nodes is... ; An undirected graph The edge set, where each edge connects two adjacent nodes; the edge set and the weight of each edge are described from the perspective of the neighboring nodes; For any water body node , with geometric center Construct a circular neighborhood using the reference point. The radius is defined as: in, for The length of the longest side of the smallest bounding rectangle. The scale parameter and This is used to control the extent of the circular neighborhood, including the nodes. Neighborhood The collection of water bodies is represented as: in, Represents any and The Euclidean distance between the centroids of the two water bodies; Set constraints to ensure that the neighborhood contains a sufficient number of other nodes: Neighborhood Internal The number of other water body nodes besides itself should not be less than the preset value. That is, satisfying Then the node and The set of edges for other nodes is represented as follows: Undirected graph The edge set is the union of all edges, represented as: 。 4. The method for extracting coastal salt fields from optical images based on optimal transport random walk according to claim 1, characterized in that, The local structural features are constructed based on the shape context contour structure similarity of the nodes. The shape context describes the local structural features of the contour by constructing a spatial distribution histogram of contour points in logarithmic polar coordinates. For any two water body nodes... and Extract its contour sampling point set, denoted as and ,in and The number of sampling points for the outlines of the two water bodies; for Establish a logarithmic-polar coordinate system with any point in the graph as the origin, and divide the surrounding area into... Radial intervals and A total of angle intervals, Each statistical unit, through statistics The rest The number of points falling into each unit is normalized to construct a shape context descriptor; the shape context descriptor is essentially a histogram of the local spatial distribution of contour points, reflecting the local structural features of the contour centered on that point; for any two points , The shape context descriptor differences are measured using chi-square distance: in, For histogram interval indexing, express and The chi-square distance between them express Statistical units Histogram representing the distribution of relative positions ; Construct a set of contour points based on the chi-square distance. and Cost matrix between , Point With point Differences in shape context between them: in, express and Chi-square distance operator for shape context differences between elements; Point set and The matching is formalized as finding a from and injective mapping The mapping is Each point in The unique matching point assigned in the middle, the optimal matching is represented as: in, Indicates all from arrive A single mapping set, Point Mapped to Point index in; Represents a point set Middle Points Through mapping In point set Matched points The shape context difference value between them, i.e., the cost of a single match for a matching pair, is the minimum cost. This reflects the degree of difference between the two contours in their overall structure; Based on the overall matching cost, the similarity metric between two nodes in terms of shape context features is defined as follows: in, Represents a point set and Similarity measure based on shape context features.

5. The method for extracting coastal salt fields from optical images based on optimal transport random walk according to claim 1, characterized in that, The overall morphological features introduce Hu invariant moments. By calculating the second and third central moments of the image region, seven moment invariants that are invariant to translation, rotation, and scaling are constructed. Mahalanobis distance is used to eliminate the correlation between feature components through the covariance matrix, and components with different dimensions are normalized to reflect the similarity between feature vectors. For any two water body nodes... and The 7-dimensional eigenvectors of the Hu moments are and The difference is measured as: in, The covariance matrix of the Hu moment feature matrices for all water samples is used to characterize the covariance relationship between the feature components of the Hu moment. Water body node and Mahalanobis distance, a feature of the Hu moment. The similarity metric for Hu moments obtained from the Mahalanobis distance is as follows: in, Representing water body nodes and Similarity measure based on Hu moment features.

6. The method for extracting coastal salt fields from optical images based on optimal transport random walk according to claim 1, characterized in that, The geometric regularity features introduce a length-width similarity index (LSI). For a body of water, the LSI value is: in, The circumference of the water body The area of ​​the water body, coefficient For the empirical scaling factor, Two water body nodes and Similarity in terms of regularity can be measured by their LSI values. and Perform measurement: in, This indicates the absolute difference in LSI values ​​between the two water bodies. Represented as water body nodes and Similarity measures based on geometric regularity features; node and The fusion form weight between Weighted fusion of the similarity of each component feature: in, Represents a node and In the The degree of difference in morphological features; These are the corresponding weighting coefficients, used to balance the contributions of different features in the similarity measurement; These represent three types of morphological features, corresponding to shape context, Hu moments, and length-width similarity index, respectively.

7. The method for extracting coastal salt fields from optical images based on optimal transport random walk according to claim 1, characterized in that, The source distribution corresponds to the labeled seed nodes, while the target distribution consists of all coastal water nodes; assuming the labeled seed points are... ,in, This represents the water body node in the crystallization pool obtained from the salt field index extraction. This indicates that the water body nodes are obtained after certain screening of non-salt paddy water bodies; the remaining unmarked water body nodes are denoted as... And satisfy ,in This represents all water features on the land side within a 15-kilometer radius of the coastline. In the graph model built based on water body nodes, the labeled seed points are regarded as the source distribution, and the probability measure is... Defined as feature space The discrete measure on the vector evenly distributes the probability mass across the feature vector corresponding to each labeled node: in, Indicates the number of marked water body nodes. Let be the corresponding uniform probability mass, and satisfy . ; node eigenvectors, and ; It is located at the feature point Dirac measure on; similarly, Treating all coastal water nodes as the target distribution, probability measure Defined as a uniform discrete probability measure consisting of all coastal water nodes: in, This indicates the total number of nodes in all water bodies. For nodes Uniform probability mass on, and satisfying ; For located at feature points Dirac measure on.

8. The method for extracting coastal salt fields from optical images based on optimal transport random walk according to claim 7, characterized in that, The description of the transmission process from the source measure to the target measure using coupled measures specifically includes: Source measurement With target measure The support sets are the labeled seed node sets. and the set of all water body nodes ,in , To mark the number of nodes and all water body nodes; measured from the source. to target measurement The transmission process is governed by coupling measures that satisfy mass conservation. The coupling measure is described as satisfying the edge distribution condition. and ,in, and Coupling measures exist and Edge measure; the edge distribution condition guarantees that the total mass flowing out from the source node is equal to the edge measure; The total mass flowing into the target node is equal to All coupling measures that satisfy the aforementioned edge distribution conditions Construct a feasible solution set : in, Indicates all and Given the set of joint probability distributions of marginal distributions, the goal of optimal transmission is to find a coupling measure. This minimizes the total transmission cost, described as in Linear programming problem on: in, For the source node To the target node The transmission cost between them is defined as a weighted combination of the distances to multiple morphological features: in, The total number of feature types, These are the weight coefficients of each feature, satisfying... , For nodes and In the A measure of difference in each feature; Using the squared Euclidean distance as a feature difference measure, the squared Euclidean distance is expressed as: in, and Representing nodes respectively and In the Vector representation on class features, when hour, This represents the shape context feature vector corresponding to the node; hour, This represents the Hu moment eigenvector corresponding to the node; hour, Represents the LSI feature corresponding to the node; symbol The Euclidean norm of the vector is denoted by the transmission cost, which makes the optimal transmission problem equivalent to minimizing the source measure in the sense of the weighted feature space. and target measure Between Wasserstein distance, will the above Wasserstein distance is expressed as: in, It is the first Weight coefficients of class features Indicates in middle The lower bound is obtained; the optimal transmission coupling matrix is ​​obtained by solving.

9. The method for extracting coastal salt fields from optical images based on optimal transport random walk according to claim 1, characterized in that, The coupling morphological weight matrix will be based on the fusion morphological weight matrix of local features. With the global correlation matrix extracted from optimal transmission The coupling morphological weight matrix is ​​combined. Represented as: in, It is a fusion morphology weight matrix; The global incidence matrix is ​​obtained by optimal transmission coupling calculation, and its elements are... Reflects the nodes and In the global transport plan, the strength of the association between each node and the marked seed node; Using the optimal transmission coupling matrix Define nodes Transmission confidence To mark the sum of coupling quality transmitted from the water body node to this node: in, Indicates the first 1 marker seed point; Introducing absorption strength The absorption intensity is defined as a nonlinear mapping of transmission confidence: Based on absorption strength and fusion weight Construct a Markov random walk model with node set Divided into transient sets and absorption state set ,Right now Among them, the salt field absorption end The corresponding marked salt field seed node, i.e. Non-salt field absorption end For the corresponding marked non-saline water body seed node, i.e. Under the absorbed Markov random walk model, the nodes arrive State transition probability Represented as: in, It is a node and The boundary weights between them; It is a node The degree; It is a node Normalized transport confidence for salt field water body categories, and and They are nodes Transmission confidence levels for salt pan and non-salt pan water body categories. It is a very small constant used to ensure that the denominator is not zero; correspondingly, For nodes Normalized transport confidence for non-saline water body categories; make Represents a node Absorbed end The probability of absorption, based on the principle of total probability in Markov chains, is determined by the nodes. Absorbed end The probability of absorption is determined by both the probability of transfer to other transient nodes and the probability of direct transfer to the foreground absorption end, that is: If the absorption probability of a node is written in vector form Then the vector expression of equation (35) is: in, Represent a The identity matrix; It is a diagonal matrix, where the diagonal elements are the absorption intensities of each node. Transition matrix Adjacency matrix The row-normalized matrix whose elements ; Given the normalized confidence vector of a node, the absorption probability of a node has a unique solution: According to the maximum absorption probability criterion, for any transient node Its categories are: in, Represents transient nodes The classification result under the maximum absorption probability criterion is taken as the absorption probability vector. In the middle, node The category index corresponding to the maximum absorption probability component is the category to which the node belongs, as determined by the random walk model.

Images