A method and system for intelligent extraction of topographic feature lines based on deep learning

Abstract

The application relates to a topographic feature line intelligent extraction method and system based on deep learning. The method comprises the following steps: constructing a space graph structure based on the morphological channel initial features and the process channel initial features of each grid point, and iteratively updating the morphological channel initial features and the process channel initial features of each node in the space graph structure through a terrain-hydrology collaborative evolution layer until converging to a fixed point to obtain morphological channel convergent features and process channel convergent features; classifying and identifying the morphological channel convergent features to obtain a topographic feature type label; calculating the geodesic distance of adjacent node pairs in the feature space in the space graph structure based on the morphological channel convergent features of each node, and performing virtual node interpolation and connection edge establishment on the adjacent node pairs with topological discontinuity to obtain a topographic structure line network. The method can realize automatic reconstruction from discrete feature points to a complete topological network by combining geometric morphological attributes and hydrological process attributes.

Description

Technical Field

[0001] This invention belongs to the field of terrain feature line technology, and in particular relates to a method and system for intelligent extraction of terrain feature lines based on deep learning. Background Technology

[0002] With the development of digital terrain analysis technology, digital elevation models (DEMs) have become the core data source for obtaining terrain feature information. Automatically and accurately extracting terrain feature lines from DEMs is a core prerequisite for achieving accurate digital terrain analysis.

[0003] In existing topographic feature line extraction techniques, traditional methods rely on local differential geometric analysis (LDG). This involves calculating geometric properties such as profile curvature, planar curvature, and slope variability of DEM grid points, selecting curvature extrema as candidate feature points, and then using tracking or region growing algorithms to connect these candidate points to form ridgelines and valleys. However, LDD methods are susceptible to local data fluctuations, resulting in numerous noisy points and missed features. To address these shortcomings, one approach incorporates physical simulations of surface water flow, calculating physical quantities such as surface water flow direction and catchment area to aid in identifying valleys and ridgelines, thus mitigating noise and feature omission issues to some extent. Another approach employs deep learning techniques such as conditional generative adversarial networks (GANs) to transform feature line extraction into a pixel-by-pixel image segmentation or generation problem, effectively improving local recognition accuracy.

[0004] However, the aforementioned methods, whether local differential geometric analysis or deep learning, simplify feature line extraction into a purely geometric pattern recognition problem, losing crucial geomorphological information. Valley lines are not merely sets of local lowest points; they are natural paths for surface runoff convergence, their morphology and direction strictly constrained by hydrological processes. Ridge lines, on the other hand, directly determine surface runoff bifurcation patterns and watershed boundaries. Furthermore, even when hydrological simulation is incorporated, some methods operate independently of feature line extraction, lacking bidirectional information interaction and constraints, and failing to reflect the interaction between valley lines and water flow paths. Real DEMs are often sparse and discontinuous due to factors such as vegetation obstruction, acquisition errors, and flat terrain. Traditional methods rely on distance threshold rules or manual intervention to repair breaks, and erroneous connections are prone to occur in complex terrain areas. Deep learning methods have not fundamentally solved the problem of topological generation of feature line networks, lacking the ability to infer reasonable connectivity relationships in sparse data regions. Summary of the Invention

[0005] Therefore, it is necessary to provide a deep learning-based intelligent extraction method and system for terrain feature lines, which addresses the above-mentioned technical problems by having geometric morphological features and hydrological process features mutually constrain and converge collaboratively through an alternating message passing mechanism.

[0006] Firstly, this application provides a method for intelligent extraction of terrain feature lines based on deep learning, including:

[0007] Based on the digital elevation model of the target area, the geometric feature vector and process feature vector of each grid point are calculated, and the geometric feature vector and process feature vector are used as the initial features of the morphological channel and the initial features of the process channel, respectively.

[0008] Based on the initial morphological and process channel features of each grid point, a spatial graph structure with grid points as nodes is constructed. A pre-trained topographic-hydrological co-evolution layer iteratively updates the initial morphological and process channel features of each node in the spatial graph structure until convergence to a fixed point, yielding the convergent morphological and process channel features of each node. The topographic-hydrological co-evolution layer includes morphological and process channels, which exchange and mutually correct features through bidirectional coupling gating. The topographic-hydrological co-evolution layer iteratively updates the initial morphological and process channel features by performing spatial message passing within the morphological and process channels respectively.

[0009] The morphological channel convergence features of each node are classified and identified to obtain the terrain feature type label corresponding to each node;

[0010] Based on the morphological channel convergence characteristics of each node, the geodesic distance between adjacent node pairs in the feature space is calculated. Virtual node interpolation and connection edge establishment are performed for adjacent node pairs with topological breaks to obtain the terrain structure line network. Adjacent node pairs with topological breaks satisfy the following conditions: the geodesic distance is less than the preset feature distance threshold, and the terrain feature type labels corresponding to the two nodes in the adjacent node pair are the same.

[0011] Secondly, this application also provides a deep learning-based intelligent terrain feature line extraction system, comprising:

[0012] The feature calculation module is used to calculate the geometric feature vector and process feature vector of each grid point based on the acquired digital elevation model of the target area, and use the geometric feature vector and process feature vector as the initial features of the morphology channel and the initial features of the process channel, respectively.

[0013] The co-evolution module is used to construct a spatial graph structure with grid points as nodes based on the initial features of the morphological and process channels of each grid point. A pre-trained topographic-hydrological co-evolution layer iteratively updates the initial features of the morphological and process channels of each node in the spatial graph structure until convergence to a fixed point, obtaining the convergent features of the morphological and process channels for each node. The topographic-hydrological co-evolution layer includes morphological and process channels, which exchange and mutually correct features through bidirectional coupling gating. The topographic-hydrological co-evolution layer achieves iterative updates of the initial features of the morphological and process channels by performing spatial message passing within the morphological and process channels respectively.

[0014] The classification and recognition module is used to classify and recognize the morphological channel convergence features of each node, and obtain the terrain feature type label corresponding to each node.

[0015] The topology reconstruction module is used to calculate the geodesic distance between adjacent node pairs in the feature space based on the morphological channel convergence features of each node. For adjacent node pairs with topological breaks, virtual node interpolation and connection edge establishment are performed to obtain the terrain structure line network. Adjacent node pairs with topological breaks meet the following conditions: the geodesic distance is less than the preset feature distance threshold, and the terrain feature type labels corresponding to the two nodes in the adjacent node pair are the same.

[0016] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-described deep learning-based intelligent extraction methods for terrain feature lines.

[0017] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the above-described deep learning-based intelligent extraction methods for terrain feature lines.

[0018] The aforementioned deep learning-based intelligent extraction method and system for terrain feature lines calculates geometric feature vectors representing geometric morphology and process feature vectors representing hydrological confluence information for each grid point of the digital elevation model of the target area, providing a dual information foundation of morphology and process for co-evolution. The terrain-hydrology co-evolution layer achieves local aggregation of neighborhood information through spatial message passing within the morphological and process channels. Simultaneously, it utilizes bidirectional coupling gating to exchange and mutually correct features between the two channels, allowing morphological features to be adjusted under the physical constraints of process information and process features to be corrected under the geometric constraints of morphological information. After iterative updates and convergence to a fixed point, the converged features of the morphological channels simultaneously contain both the geometric morphological attributes of the terrain and the hydrological process attributes, thus solving the problem that purely geometric methods cannot understand the semantics of landform formation. This enables classification and recognition to be based on richer information. By using rich and physically reasonable information to determine whether each node belongs to a ridgeline, valleyline, or slope break, the model's ability to identify complex landform units is improved. Based on the convergence characteristics of morphological channels, the geodesic distance between adjacent node pairs in the feature space is calculated, reflecting the true connectivity cost of two nodes on the morphological feature manifold. Combined with the condition that the two nodes have the same terrain feature type label, topological breaks are determined. This makes break detection no longer dependent on fixed distance thresholds or human experience, but based on the continuity and semantic consistency in the feature space for intelligent inference. Virtual nodes are inserted at the location of the determined break to establish connecting edges. The resulting terrain structure line network not only maintains the spatial continuity and smoothness of the feature lines in the feature space, but also fundamentally repairs the break defects in the original data caused by vegetation shading, collection errors, or flat terrain, realizing the automatic reconstruction from discrete feature points to a complete topological network. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the intelligent extraction method for terrain feature lines based on deep learning according to the present invention.

[0021] Figure 2 This is a structural diagram of the intelligent terrain feature line extraction system based on deep learning according to the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] In one embodiment, such as Figure 1 As shown, a deep learning-based intelligent terrain feature line extraction method is provided. This embodiment illustrates the application of this method to an intelligent terrain feature line extraction terminal. It is understood that this method can also be applied to an intelligent terrain feature line extraction server, and further to an intelligent terrain feature line extraction system comprising an intelligent terrain feature line extraction terminal and an intelligent terrain feature line extraction server, and is implemented through the interaction between the intelligent terrain feature line extraction terminal and the intelligent terrain feature line extraction server. In this embodiment, the method includes the following steps:

[0024] S101. Based on the obtained digital elevation model of the target area, calculate the geometric feature vector and process feature vector of each grid point, and use the geometric feature vector and process feature vector as the initial features of the morphological channel and the initial features of the process channel, respectively.

[0025] As an illustration, after acquiring the digital elevation model of the target area, the intelligent terrain feature line extraction terminal performs a grid-by-grid scan of the digital elevation model, calculating a geometric feature vector for each grid point. (Grid point) The geometric feature vector, i.e., the initial feature of the morphological channel, is Among them, elevation value Reflects the elevation and slope of the land surface at the grid points. The slope aspect value reflects the degree of surface inclination at the grid points. Reflects the orientation and curvature of the grid points on the Earth's surface. Planar curvature reflects the degree of surface curvature along the slope direction. The maximum curvature reflects the degree of surface curvature perpendicular to the slope direction. and minimum curvature The second-order Laplace elevation model reflects the maximum and minimum curvature of the surface at grid points in different directions. This reflects the local smoothness of the grid point elevation field. Furthermore, the intelligent terrain feature line extraction terminal also calculates the elevation of each grid point using a built-in differentiable merge solver based on the digital elevation model. The process feature vector is the initial feature of the process channel. ,in, The upstream catchment area, For grid points The probability of flow to eight neighboring grid points. This represents the cumulative energy index of the water flow along the course.

[0026] S102. Based on the initial features of the morphological channel and the initial features of the process channel of each grid point, a spatial graph structure with grid points as nodes is constructed. The initial features of the morphological channel and the initial features of the process channel of each node in the spatial graph structure are iteratively updated through a pre-trained topographic-hydrological co-evolution layer until convergence to a fixed point, thus obtaining the convergent features of the morphological channel and the convergent features of the process channel of each node. The topographic-hydrological co-evolution layer includes morphological channels and process channels. The morphological channels and process channels exchange features and correct each other through bidirectional coupling gating. The topographic-hydrological co-evolution layer realizes the iterative update of the initial features of the morphological channel and the initial features of the process channel by performing spatial message passing in the morphological channel and the process channel respectively.

[0027] Furthermore, in the spatial graph structure, each grid point is treated as a node, and the initial feature vector of each node is composed of the initial features of the grid point's morphological channel and the initial features of its process channel. The intelligent terrain feature line extraction terminal establishes adjacency relationships between nodes in the spatial graph structure. The adjacency relationship is established by connecting each grid point to all grid points within a preset radius, or by connecting the grid point to its K nearest neighbors.

[0028] Specifically, the intelligent terrain feature line extraction terminal inputs the spatial map structure into a pre-trained terrain-hydrological co-evolution layer for iterative updates. This layer includes two parallel processing channels: a morphological channel and a process channel. In each iteration, the morphological channel performs intra-channel spatial message passing on the morphological channel features of each node it maintains, and the process channel also performs intra-channel spatial message passing on the process channel features of each node it maintains. After the spatial message passing within both the morphological and process channels is completed, the intelligent terrain feature line extraction terminal performs feature exchange and mutual correction between the morphological and process channels through bidirectional coupling gating. For example, the bidirectional coupling gating includes correction from the morphological channel to the process channel and correction from the process channel to the morphological channel. In the correction from the morphological channel to the process channel, the intelligent terrain feature line extraction terminal inputs the aggregated features of the morphological channel into a multilayer perceptron and outputs a gating vector after passing through an activation function. This gating vector controls the extent to which the correction information from the morphological channel is injected into the various feature dimensions of the process channel. The correction from the process channel to the morphological channel adopts a symmetrical design. The intelligent terrain feature line extraction terminal calculates the gating vector and correction amount from process to morphology based on the aggregated features of the process channel, and corrects the aggregated features of the morphological channel. The bidirectional coupling gating realizes the information interaction and mutual constraint between the two channels by correcting process features with morphological features and morphological features with process features. The intelligent terrain feature line extraction terminal repeats the above iterative update process until the sum of the norms of the feature changes of all nodes in two adjacent iterations is less than a preset threshold. It is determined that the terrain-hydrological co-evolution layer has converged to a fixed point. The morphological channel features of each node are the morphological channel convergence features, and the process channel features of each node are the process channel convergence features.

[0029] S103. Classify and identify the morphological channel convergence features of each node to obtain the terrain feature type label corresponding to each node.

[0030] Optionally, the classification and recognition process is implemented using a classification multilayer perceptron. This classification multilayer perceptron takes the morphological channel convergent feature vector of each node as input, passes it through several fully connected layers and nonlinear activation functions, and finally outputs the probability distribution of a node belonging to each terrain feature type. The classification multilayer perceptron structure includes an input layer, multiple intermediate hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimension of the morphological channel convergent feature vector. The number of neurons in the intermediate hidden layers can be progressively decreased to extract high-level semantic features. The number of neurons in the output layer is equal to the total number of terrain feature types. The intermediate hidden layers use nonlinear activation functions to introduce nonlinear combination capabilities between features. The output layer uses a softmax activation function to convert the output values ​​into a probability distribution with a sum of 1, where each component corresponds to a terrain feature type. Terrain feature types include four categories: ridgelines, valley lines, slope break lines, and non-feature points. During the training of the classification multilayer perceptron, the intelligent terrain feature line extraction terminal uses a cross-entropy loss function as a supervision signal. During the inference phase, it directly takes the category with the highest probability as the terrain feature type label for the node.

[0031] S104. Based on the morphological channel convergence characteristics of each node, calculate the geodesic distance between adjacent node pairs in the feature space of the spatial graph structure, and perform virtual node interpolation and connection edge establishment for adjacent node pairs with topological breaks to obtain the terrain structure line network; adjacent node pairs with topological breaks satisfy the condition that the geodesic distance is less than the preset feature distance threshold, and the terrain feature type labels corresponding to the two nodes in the adjacent node pair are the same.

[0032] Furthermore, for any node in the spatial graph structure The intelligent terrain feature line extraction terminal searches for nodes in space. Search for candidate nodes within the preset spatial neighborhood radius The intelligent terrain feature line extraction terminal will extract nodes. With nodes The spatial path between them is parameterized, using parameters Indicates from node To the node The normalized position on the path, The range of values ​​is At each location along the spatial path At this location, the intelligent terrain feature line extraction terminal obtains the position through spatial interpolation. The corresponding morphological channel feature vector values ​​can be spatially interpolated using linear or spline interpolation based on the grid point elevations and feature values ​​along the path. The intelligent terrain feature line extraction terminal calculates the morphological channel feature vector values ​​along the path with respect to parameters. The derivative of the position is obtained. The gradient of the morphological channel feature vector along the path direction is used, and the L2 norm of the gradient is taken as the local rate of change at that location. The local rate of change reflects the drastic change of the morphological channel features as it moves along the spatial path; the smaller the local rate of change, the more gradual the change of morphological features along the path near that location. The intelligent terrain feature line extraction terminal uses this local rate of change along the path from... arrive Integrate to obtain the integral value corresponding to the spatial path. Optionally, due to the nodes... and nodes There may be multiple different spatial paths between them. The intelligent terrain feature line extraction terminal selects the path that minimizes the integral value from all possible spatial paths. The path with the minimum integral value is the node. and nodes The geodesic distance between them in the characteristic space.

[0033] Furthermore, the condition for determining topological breaks is that two conditions must be met simultaneously: first, the geodesic distance between adjacent node pairs is less than a preset feature distance threshold; second, the terrain feature type labels corresponding to the two nodes in the adjacent node pair are the same. For adjacent node pairs determined to have topological breaks, the intelligent terrain feature line extraction terminal extracts data from the nodes... and nodes The insertion point is the location where the rate of change of the convergence feature of the morphological channel along the spatial path is the minimum. The insertion point is determined by comparing the local rate of change of each point on the path and taking the minimum value. At the insertion point, the terrain feature line intelligent extraction terminal inserts a virtual node. The spatial coordinates of the virtual node are based on the parameters of the insertion point on the path. The value is obtained through spatial interpolation. The morphological channel feature vector of the virtual node is assigned based on the interpolated feature vector at the insertion position, and the process channel feature vector of the virtual node is assigned through interpolation at the same position. The terrain feature line intelligent extraction terminal assigns the virtual node with the node's feature value. and nodes Same terrain feature type label, and in virtual nodes and nodes Establish a connection edge between the virtual nodes and the node. A connecting edge is also established between them, thereby connecting the originally broken terrain feature lines through this virtual node.

[0034] In the aforementioned deep learning-based intelligent terrain feature line extraction method, the geometric and process feature vectors of each grid point are calculated based on the digital elevation model of the target area. These vectors serve as the initial features for the morphological and process channels, respectively, providing dual-dimensional information on terrain morphology and hydrological processes for feature line extraction. This avoids the problem of losing the semantic meaning of landform formation due to a single geometric representation. A spatial graph structure is constructed using grid points as nodes. Through a pre-trained terrain-hydrological co-evolution layer, spatial message passing is performed within the morphological and process channels. Bidirectional coupling gating is used to achieve feature exchange and mutual correction between the two channels. The two types of initial features are iteratively updated until convergence to a fixed point, thereby achieving bidirectional coupling co-evolution of morphology and process. This ensures that the final features simultaneously satisfy both geometric and hydrological constraints, fundamentally addressing the issue of... It bridges the semantic gap between geometry and process; based on the converged morphological channel features of each node, it performs classification and identification to obtain corresponding terrain feature type labels. While ensuring the accuracy of feature point classification and identification, it enables the output feature points to have both morphological attributes and hydrological process semantics, which can directly support downstream hydrological analysis and other geomorphological applications; based on the converged morphological channel features, it calculates the geodesic distance between adjacent node pairs in the feature space of the spatial graph structure. For adjacent node pairs of topological fractures that meet the same terrain feature type label and whose geodesic distance is less than a preset threshold, it performs virtual node interpolation and establishes connecting edges to obtain a complete terrain structure line network. It naturally completes feature line fracture completion by relying on the continuous manifold structure embedded in the feature space, without the need to set up an independent fracture repair module. At the same time, it has stronger topological inference capabilities and robustness in sparse data areas.

[0035] In one embodiment, the process feature vector is calculated through the following steps:

[0036] S21. For each grid point, calculate the directional slope from the grid point to its eight neighboring grid points, and use a differentiable softmax function to convert each directional slope into the flow probability from the eight neighboring grid points to the grid point, thus obtaining the eight-directional flow probability of each grid point.

[0037] Optionally, the expression for the flow probability is:

[0038] ;

[0039] in, For neighborhood grid points To grid point directional slope, For temperature parameters, For grid points The set of eight neighboring grid points;

[0040] Schematic illustration: A terrain feature line intelligent extraction terminal determines grid points. Spatial location index in the digital elevation model, and read from the digital elevation model. Elevation values ​​and surrounding areas Furthermore, the elevation values ​​of the eight directly adjacent grid points are calculated, with these eight grid points located directly above, below, to the left, to the right, and diagonally opposite the current grid point. The terrain feature line intelligent extraction terminal calculates the elevation values ​​of each neighboring grid point relative to the current grid point. directional slope ,in, For grid points Elevation value, For neighborhood grid points Elevation value, For grid points To neighboring grid points Planar distance, It can be equal to the grid size of the digital elevation model in orthogonal directions. It can be positive or negative; a positive value indicates along The elevation decreases along the direction; a negative value indicates that the elevation decreases along the direction. Direction and elevation increase. Optional, temperature parameter. It is a positive real number whose physical function is to control the shape characteristics of the probability distribution of flow direction, and the temperature parameter. The larger the value, the more concentrated the flow probability is in the direction with the steepest slope. Temperature parameter The smaller the value, the more evenly the flow probability is distributed in the eight directions. For each neighboring grid point According to its relative to grid points The magnitude of the elevation difference and the direction of the slope are assigned a probability share for the flow to the current grid point, and this probability share is related to the neighboring grid points. The direction slope is directly proportional to the exponential value. The greater the direction slope, that is, the greater the elevation difference, the higher the probability of flow direction for the neighboring grid points. As a normalization factor, it ensures that the sum of the probabilities of the eight flows is strictly equal to 1, thus satisfying the basic mathematical constraints of the probability distribution.

[0041] S22. Based on the eight-directional flow probabilities of each grid point, the conjugate gradient method is used to iteratively solve the confluence equations to obtain the upstream catchment area of ​​each grid point.

[0042] Optionally, the expression for the confluence equations is:

[0043] ;

[0044] In the formula, For grid points The upstream catchment area, To obtain from neighboring grid points Flow to grid points The probability of the flow direction, For neighborhood grid points The upstream catchment area.

[0045] Furthermore, the intelligent terrain feature line extraction terminal constructs a set of confluence equations based on the eight-directional flow probabilities of all grid points. This set of equations is a linear system where the unknown is the upstream catchment area of ​​each grid point, and the number of equations equals the total number of grid points. In the confluence equations, for any given grid point… Its upstream catchment area Equals one unit area plus grid points Each of the eight neighboring grid points Merging into grid points The sum of the catchment areas, of which from a certain neighboring grid point Merging into grid points The catchment area is equal to the neighboring grid points upstream catchment area Multiply by the neighborhood grid points Flow to grid points Flow probability Specifically, For surface water to flow into the grid points from the upstream area The total number of grid cells traversed during the time period, The larger the grid size, the more grid points there are. The stronger the cumulative effect in hydrological runoff, the better. For dimensionless grid counting, To obtain from neighboring grid points Flow to the current grid point The probability of the flow direction, The range of values ​​is And the sum of the flow probabilities of all eight neighboring grid points is 1.

[0046] Optionally, the intelligent terrain feature line extraction terminal can use the conjugate gradient method to solve the confluence equations. The conjugate gradient method is an iterative numerical method for solving symmetric positive definite linear equations. In each iteration, it gradually approximates the exact solution of the equations by constructing a set of search directions conjugate to the coefficient matrix. When using the conjugate gradient method to solve the confluence equations, the intelligent terrain feature line extraction terminal rewrites the equations in matrix form. In this matrix form, the diagonal elements of the coefficient matrix are all 1s, and the off-diagonal elements are the negative values ​​of the flow probability from the corresponding grid point in the column to the corresponding grid point in the row. The intelligent terrain feature line extraction terminal initializes the upstream catchment area of ​​all grid points as a zero vector as the initial value for iteration. Then, it calculates the current residual vector, which is equal to the constant vector of the right-hand side minus the coefficient matrix multiplied by the current solution vector. The intelligent terrain feature line extraction terminal constructs search directions in the direction of the residual vector and determines the optimal step size by performing a one-dimensional exact search along the search direction. The optimal step size minimizes the objective function along the search direction. The intelligent terrain feature line extraction terminal updates the current solution vector according to the optimal step size and calculates the updated residual vector. Then, it calculates the conjugate parameters based on the updated residual vector and corrects the search direction, making the new search direction conjugate with all historical search directions about the coefficient matrix. The intelligent terrain feature line extraction terminal repeats the above iterative process until the norm of the current residual vector is less than the preset convergence threshold. At this point, the solution vector is the numerical solution of the upstream catchment area of ​​all grid points.

[0047] S23. Based on the upstream catchment area and local slope of each grid point, calculate the cumulative flow energy index along the route, and generate the process feature vector of each grid point based on the eight-directional flow probability, upstream catchment area and cumulative flow energy index along the route.

[0048] For example, after obtaining the upstream catchment area of ​​each grid point, the intelligent terrain feature line extraction terminal calculates the local slope value of each grid point from the digital elevation model. The local slope value is calculated by taking the maximum directional slope between the grid point and its eight neighboring grid points as the local slope of the grid point, or by using the least squares method to perform plane fitting on the elevation values ​​of the grid point and all grid points within a preset window around it, and taking the maximum tilt angle of the fitted plane as the local slope value of the grid point.

[0049] Furthermore, the intelligent terrain feature line extraction terminal calculates the cumulative water flow energy index along the grid points based on the upstream catchment area and local slope values. ,in, From the source of the watershed to the grid points The set of all grid points along the water flow path, Grid points along the water flow path The elevation difference along the flow direction, For grid points The square root of the upstream catchment area, A quantity with the dimension of length. It has the dimension of energy and, after being accumulated along the path, represents the total mechanical energy carried by the water flow when it reaches the current grid point. Optionally, the terrain feature line intelligent extraction terminal extracts the data based on the grid points. The eight-directional flow probability determines the main flow direction, which is the direction with the highest probability value among the eight flow directions. The terrain feature line intelligent extraction terminal starts from the source grid point of the watershed and tracks the main flow direction grid by grid until it reaches the outlet grid point of the watershed, forming a water flow path from the source to the outlet.

[0050] In one embodiment, the initial features of the morphological channels and the initial features of the process channels of each node in the spatial map structure are iteratively updated through a pre-trained topography-hydrology co-evolution layer until convergence to a fixed point, resulting in the convergent features of the morphological channels and the convergent features of the process channels of each node, including:

[0051] S31. For each node in the spatial graph structure, based on the morphological channel features and process channel features of the node and its corresponding neighboring nodes in the previous iteration step, the morphological channel message passing coefficient and process channel message passing coefficient of the current iteration step are calculated by a multilayer perceptron; wherein, when the current iteration step is the first step, the morphological channel features and process channel features in the previous iteration step are the initial features of the morphological channel and the initial features of the process channel, respectively.

[0052] Optionally, the expressions for the morphological channel message passing coefficient and the process channel message passing coefficient of the current iteration step are as follows:

[0053] ;

[0054] ;

[0055] In the formula, For the current iteration step The morphological channel message transmission coefficient, For the current iteration step Process channel message transmission coefficient, For the node in the previous iteration step The morphological channel characteristics, For the node in the previous iteration step Process channel characteristics, For the node in the previous iteration step Neighboring nodes The morphological channel characteristics, For the node in the previous iteration step Neighboring nodes Process channel characteristics, This is a vector concatenation operation;

[0056] As an illustration, the intelligent terrain feature line extraction terminal will extract nodes. In the previous iteration, the morphological channel features and process channel features were respectively associated with each neighboring node. In the previous iteration, the morphological channel features and process channel features are concatenated end-to-end into a single concatenated vector according to a preset order. The terrain feature line intelligent extraction terminal inputs the concatenated vector into a morphological channel message passing coefficient calculation multilayer perceptron. The morphological channel message passing coefficient calculation multilayer perceptron contains at least one fully connected layer and a nonlinear activation function layer to obtain the neighborhood nodes. Relative to node The unnormalized morphological channel message passing coefficients. The unnormalized morphological channel message passing coefficients reflect the message passing performance in the neighborhood nodes. To the node Neighboring nodes when transmitting morphological channel information The original weights that should be assigned, taking into account the nodes and neighboring nodes Their respective morphological and hydrological conditions. The intelligent terrain feature line extraction terminal extracts data from nodes. All unnormalized morphological channel message passing coefficients corresponding to neighboring nodes are subjected to flexible maximum value normalization to obtain the neighboring node-to-node ratio in the current iteration step. The morphological channel message passing coefficient. The morphological channel message passing coefficient reflects the morphological channel message passing coefficient at the node. During the morphological feature update process, each neighboring node The importance of the morphological and process information it carries.

[0057] Specifically, Calculate the multilayer perceptron for the morphological channel message passing coefficients. Calculate the multilayer perceptron (MLP) for the process channel message passing coefficients. All MLPs used in the topography-hydrology co-evolution layer, including... , , , , , Both classification multilayer perceptrons and multilayer perceptrons can adopt a unified two-layer structure, with the hidden layer dimension of each multilayer perceptron set to 256. The hidden layer can use Parametric ReLU (PReLU). PReLU preserves the non-saturating properties of ReLU while allowing a learnable slope in the negative range, thus avoiding the "death" problem of neurons and better adapting to the complex nonlinear features in terrain data. For the output layer, due to the gating mechanism, the output value must be between 0 and 1, using a gated generative multilayer perceptron. and The output layer of the multilayer perceptron uses the Sigmoid activation function, while other multilayer perceptrons... , , , The output layer does not require an activation function and directly outputs the linear transformation result.

[0058] S32. Based on the morphological channel message passing coefficient of the current iteration step, the morphological channel features of the neighboring nodes in the previous iteration step are weighted and aggregated to obtain the morphological channel aggregated features of the current iteration step. Based on the process channel message passing coefficient of the current iteration step, the process channel features of the neighboring nodes in the previous iteration step are weighted and aggregated to obtain the process channel aggregated features of the current iteration step.

[0059] For example, the calculation expression for morphological channel aggregation features is as follows: ,in, For nodes The morphological channel spatial neighborhood node set, where each node in the neighborhood set is spatially distant from the node. No more than all nodes within a preset spatial radius. Neighboring nodes To the node The morphological channel message transmission coefficient, Given a learnable linear transformation weight matrix, The dimension is the product of the morphological channel feature dimension and the morphological channel feature dimension. Its function is to perform a linear projection transformation on the morphological channel features of neighboring nodes to enhance the expressive power of the features. For neighboring nodes The morphological channel features from the previous iteration step. The calculation expression for the process channel aggregation features is: ,in, For nodes The process channel space neighborhood node set, the neighborhood set not only includes nodes Spatially adjacent nodes also include nodes located along the direction of water flow. Upstream and downstream nodes are used to ensure that the propagation of hydrological signals strictly follows the physical flow direction of the terrain. For this neighboring node To the node Process channel message transmission coefficient, It is a learnable linear transformation weight matrix, whose dimension is the process channel feature dimension multiplied by the process channel feature dimension. For neighboring nodes Process channel features in the previous iteration step.

[0060] S33. Based on the morphological channel aggregation features, the gating vector of the morphological channel to the process channel is calculated by a multilayer perceptron to obtain the first gating vector. The process channel aggregation features are then corrected based on the first gating vector to obtain the process channel update features.

[0061] Optionally, the expressions for the first gating vector and the process channel update features are as follows:

[0062] ;

[0063] ;

[0064] In the formula, This is the first gating vector. For the current iteration step Middle node Process channel characteristics, For the current iteration step Middle node The morphological channel aggregation characteristics, For the current iteration step Middle node Process channel aggregation characteristics, It is the sigmoid activation function. For element-wise product;

[0065] Specifically, for each node, the terrain feature line intelligent extraction terminal inputs morphological channel aggregation features into the gating system to generate a multilayer perceptron. The vector output by the first gating multilayer perceptron is processed by an activation function, which maps its input element by element to... This yields the first gating vector from the morphological channel to the process channel. The first gating generates a multilayer perceptron containing at least one fully connected layer and at least one nonlinear activation function layer. The expression for the activation function is: ,in, Generate any element value from the multilayer perceptron output for the first gating control. The function value approaches 1 as the independent variable approaches positive infinity, approaches 0 as the independent variable approaches negative infinity, and equals 0.5 when the independent variable is 0. Each element of the first gating vector independently controls the injection intensity of the morphological channel's correction information into the corresponding feature dimension of the process channel. When the value of an element is close to one, it indicates that the morphological information in that dimension should be fully injected into the process channel; when the value of an element is close to zero, it indicates that the morphological information in that dimension should not be injected into the process channel.

[0066] Optionally, the terrain feature line intelligent extraction terminal simultaneously inputs the morphological channel aggregation features into a correction value to generate a multilayer perceptron. The output vector is the correction vector of the morphological channel to the process channel. The network structure of the first correction vector generation multilayer perceptron can be the same as or different from that of the first gated generation multilayer perceptron. The correction vector represents the correction information generated in the process channel feature space after the morphological channel features are nonlinearly mapped. The correction information includes the physical constraints of the morphological features on the process features. The terrain feature line intelligent extraction terminal multiplies each element of the first gated vector with each element of the corresponding position in the correction vector to obtain an element-wise gated correction vector. The multiplication operation is the Hadamard product operation. The terrain feature line intelligent extraction terminal adds the element-wise gated correction vector to the process channel aggregated features element-wise to obtain the updated process channel features, which are the process channel features corrected by the morphological channel information.

[0067] S34. Based on the process channel aggregation features, the gating vector of the process channel to the morphology channel is calculated by a multilayer perceptron to obtain the second gating vector. The morphology channel aggregation features are then corrected based on the second gating vector to obtain the morphology channel update features.

[0068] Optionally, the expressions for the second gating vector and the morphological channel update features are as follows:

[0069] ;

[0070] ;

[0071] In the formula, This is the second gating vector. Update features for morphological channels.

[0072] Specifically, each element of the second gating vector independently controls the injection intensity of the process channel's correction information into the corresponding feature dimension of the morphological channel, thereby realizing the physical guidance of morphological information from process information. This represents the correction information generated in the morphological channel feature space after the process channel features are nonlinearly mapped, which includes the physical constraints of the water catchment process on the terrain morphology.

[0073] S35. Perform residual connection and layer normalization on the process channel update features and morphology channel update features respectively to obtain the morphology channel features and process channel features of the current iteration step; where the current iteration step is the last step, the morphology channel features and process channel features are the morphology channel convergence features and process channel convergence features respectively.

[0074] Optionally, the terrain feature line intelligent extraction terminal will update the morphological channel features of the current iteration step. Morphological channel features compared to the previous iteration Element-by-element summation yields the morphological channel residual features. After residual concatenation, the intelligent terrain feature line extraction terminal performs layer normalization, with each node's morphological channel residual feature serving as an independent normalization unit. The intelligent terrain feature line extraction terminal then applies the same residual concatenation and layer normalization processing to the process channel update features, i.e., the process channel update features of the current iteration step. Process channel features compared to the previous iteration step After adding each element, layer normalization is performed to obtain the final process channel features of the current iteration step.

[0075] In one embodiment, based on the morphological channel convergence characteristics of each node, the geodesic distance between adjacent node pairs in the feature space of the spatial graph structure is calculated, and virtual node interpolation and connection edge establishment are performed for adjacent node pairs with topological breaks to obtain a terrain structure line network, including:

[0076] S41. For each node, search for neighboring nodes within a preset spatial neighborhood radius to obtain the corresponding neighboring node pairs, and calculate the geodesic distance between each neighboring node pair in the feature space.

[0077] Optionally, the expression for geodesic distance is:

[0078] ;

[0079] in, For adjacent node pairs geodesic distance, For the node to adjacent nodes Spatial path, Location on the node spatial path The morphological channel convergence characteristics at the location.

[0080] This is illustrative of the node currently being processed. The intelligent terrain feature line extraction terminal uses nodes Using the spatial coordinates as the center, a circular search area is constructed with the lower limit of the preset spatial neighborhood radius as the inner radius and the upper limit of the preset spatial neighborhood radius as the outer radius. The terrain feature line intelligent extraction terminal searches for all nodes that meet the spatial distance conditions within the circular search area. , will node With the current node Pairing up yields the current node. This corresponds to several pairs of adjacent nodes. Further, the adjacent node pairs... The spatial path between two nodes is parameterized using parameters. This represents the normalized process position along the spatial path from the starting node to the ending node, and the range of values ​​for s is... ,in, Corresponding to the spatial position of the starting node, The spatial location corresponding to the termination node. The intelligent terrain feature line extraction terminal uses parameters... Within the range of values, the spatial path is discretely sampled with a preset sampling step size to obtain a series of discrete path points. Each path point corresponds to a spatial coordinate and a morphological channel convergence feature value at that spatial coordinate. This morphological channel convergence feature value is obtained by spatial interpolation of the morphological channel convergence features of known nodes in the spatial graph structure. The spatial interpolation method can be inverse distance weighted interpolation, kriging interpolation, or spline interpolation. For example, For nodes and nodes Geodesic distance between them The morphological channel convergent feature vector The derivative of the parameter s along the path, It is the L2 norm. Indicates all possible paths Find the path that minimizes the integral value.

[0081] S42. If the geodesic distance between adjacent node pairs is less than the preset feature distance threshold, and the terrain feature type labels corresponding to the two nodes in the adjacent node pair are the same, then it is determined that there is a topological break in the adjacent node pair.

[0082] Optionally, the preset feature distance threshold is a value determined statistically based on the distribution of geodesic distances between adjacent nodes on the complete terrain feature line network in the training dataset within the geodesic space. Specifically, it can be the average of the geodesic distance distribution plus or minus a certain number of standard deviations, or a value determined empirically. The preset feature distance threshold represents the upper limit to which the geodesic distance between adjacent nodes on the same continuous terrain feature line should not exceed in the feature space. When the geodesic distance between two nodes exceeds this threshold, it indicates that the distance between these two nodes in the feature space is too large, and they are likely to belong to different terrain feature lines, and therefore should not be connected. The causes of topological breaks may include missing elevation data in the original digital elevation model due to vegetation cover, noise generated during data acquisition obscuring the curvature signal in the area, or the terrain being too flat, preventing the area from possessing obvious curvature extrema and thus failing to be selected as a candidate node.

[0083] S43. On the node spatial path of adjacent node pairs where topological breaks are determined, the position with the smallest rate of change of morphological channel features along the node spatial path is searched as the insertion position, and a virtual node is inserted at the insertion position; the terrain feature type label of the virtual node is the same as that of the adjacent node pair.

[0084] The intelligent terrain feature line extraction terminal traverses all sampling locations along the node spatial path and compares the rate of change values ​​at each location. The insertion point is determined by finding the position with the smallest rate of change. This insertion point represents the location where the morphological channel features change most gently as the terrain feature line moves along the node spatial path. It is precisely the position where the terrain feature line is most likely to be truncated after a break, and also the most reasonable connection point to link the two ends of the break through a virtual node. The intelligent terrain feature line extraction terminal generates a virtual node at this insertion point. The spatial coordinates of this virtual node are determined by linear interpolation of the path endpoint coordinates based on the parameter position of the insertion point on the path. The morphological channel convergence feature vector and the process channel convergence feature vector of the virtual node are assigned eigenvalues ​​obtained through spatial interpolation, ensuring that the virtual node is precisely located on the continuous path of the two nodes at the two ends of the topological break in the feature space. The virtual node acts as a bridge between the nodes at the two ends of the break in the spatial graph structure, allowing terrain feature lines that are spatially separated but continuous in the feature space to be connected through this virtual node.

[0085] S44. Connect the virtual nodes to the two nodes in the adjacent node pairs to obtain the terrain structure line network.

[0086] For example, the connection edges are established by adding two new undirected or directed edges to the spatial graph structure maintained by the terrain feature line intelligent extraction terminal, connecting the two nodes in adjacent node pairs respectively. After all break points are repaired, the node set in the spatial graph structure obtained by the terrain feature line intelligent extraction terminal includes the nodes corresponding to the original grid points and the newly inserted virtual nodes. The edge set includes the existing connecting edges in the original spatial graph structure and the newly established repaired connecting edges. All nodes in the entire graph structure form a connected linear network through the connecting edges. Each connected path in the linear network corresponds to a complete terrain feature line, and the set of all paths constitutes the terrain structure line network.

[0087] In one embodiment, the topography-hydrology co-evolution layer is trained through the following steps:

[0088] S51. Obtain the sample digital elevation model and the labeled terrain structure line network corresponding to the sample digital elevation model, and calculate the initial features of the sample morphology channel and the initial features of the sample process channel based on the sample digital elevation model to construct the sample space map structure.

[0089] In a schematic representation, the intelligent terrain feature line extraction terminal acquires a training dataset containing multiple sample digital elevation models (DEMs) and a corresponding labeled terrain structure line network for each DEM. The labeled terrain structure line network is manually drawn on the sample DEMs by geographic information experts, or extracted from high-resolution DEMs generated from high-precision LiDAR point cloud data after manual verification. Each position of every line segment in the labeled terrain structure line network is labeled with a category label indicating whether that position belongs to a ridgeline, valley line, or slope break. The intelligent terrain feature line extraction terminal calculates the geometric feature vector of each sample grid point based on the sample DEMs grid-by-grid, and calculates the process feature vector of each sample grid point using a differentiable merge solver. The geometric feature vector and process feature vector are used as the initial features of the sample morphology channel and the initial features of the sample process channel for that sample grid point, respectively. Furthermore, the intelligent terrain feature line extraction terminal establishes a sample spatial map structure based on the spatial locations of all sample grid points, with each sample grid point in the sample spatial map structure serving as a sample node.

[0090] S52. The initial features of the sample morphology channel and the initial features of the sample process channel of each sample node in the sample spatial map structure are iteratively updated through the topography-hydrology co-evolution layer until they converge to a fixed point, thus obtaining the convergence features of the sample morphology channel and the convergence features of the sample process channel.

[0091] Similarly, the intelligent extraction terminal for terrain feature lines inputs the initial features of the sample morphology channel and the initial features of the sample process channel of each sample node in the sample spatial map structure into the current terrain-hydrology co-evolution layer to be trained. The current trainable parameters of the terrain-hydrology co-evolution layer are the initial values ​​or the values ​​updated in the previous iteration. The intelligent terrain feature line extraction terminal iteratively updates the initial features of the sample morphology channel and the initial features of the sample process channel of each sample node in the sample spatial map structure through the terrain-hydrology co-evolution layer. In each iteration, the morphology channel and the process channel perform intra-channel spatial message passing. After each iteration, the intelligent terrain feature line extraction terminal compares the sum of the norms of the changes in the sample morphology channel features of all sample nodes in the current iteration step with those in the previous iteration step. When the sum of the norms is less than the preset convergence threshold, it is determined that the iteration of the terrain-hydrology co-evolution layer on the sample spatial map structure has converged to a fixed point. At this time, the iteration stops and the sample morphology channel features and sample process channel features of the current iteration step are used as the sample morphology channel convergence features and sample process channel convergence features of each sample node in the sample spatial map structure, respectively.

[0092] S53. In the feature space, perform local principal direction analysis on the convergence features of the sample morphology channels of each sample node to obtain the feature line direction of each sample node, and construct a topological consistency loss function based on the consistency measure of the feature line direction between adjacent sample nodes.

[0093] Furthermore, for the currently processed sample node, the intelligent terrain feature line extraction terminal determines the set of spatial neighboring nodes of the sample node in the sample space graph structure. The intelligent terrain feature line extraction terminal collects the convergent feature vectors of the sample morphology channels of all sample nodes in the sample node and its spatial neighboring node set. These convergent feature vectors are arranged into a matrix by rows. The intelligent terrain feature line extraction terminal then centers the matrix, calculating the mean of each column and subtracting the mean from each element in that column to obtain the centered feature matrix. The intelligent terrain feature line extraction terminal calculates the covariance matrix of the centered feature matrix and performs eigenvalue decomposition on the covariance matrix to obtain all eigenvalues ​​and corresponding eigenvectors. The eigenvector corresponding to the largest eigenvalue is extracted as the principal gradient direction in the feature space at the sample node. The intelligent terrain feature line extraction terminal repeats the above operations for each sample node in the sample space graph structure to obtain the principal gradient direction in the feature space at each sample node.

[0094] Optionally, for any pair of adjacent sample nodes in the sample space graph structure, the intelligent terrain feature line extraction terminal calculates the morphological channel category probability distribution and process channel category distribution of the adjacent sample node pair based on the principal gradient direction of each sample node in its feature space and the spatial coordinates of the two sample nodes, and then calculates the feature line direction consistency metric. The closer the consistency metric value is to 1, the more likely the adjacent sample node pair is arranged along the same direction in the feature space and belongs to the same continuous terrain feature line. The intelligent terrain feature line extraction terminal constructs a topological consistency loss function based on the feature line direction consistency metrics of all adjacent sample node pairs. The topological consistency loss function is the sum of the negative logarithms of the consistency metrics of all adjacent sample node pairs in all iteration steps. The smaller the value of the topological consistency loss function, the more continuous and consistent the morphological channel convergence features of the adjacent sample nodes are arranged in the feature space.

[0095] S54. Construct a total loss function based on the topology consistency loss function, the preset node classification loss function, the channel consistency loss function, and the feature sparsity loss function. Iterate the trainable parameters of the topography-hydrology co-evolution layer through backpropagation until the total loss function converges to obtain the pre-trained topography-hydrology co-evolution layer. The backpropagation method uses the implicit function differentiation method to calculate the gradient of the total loss function with respect to the trainable parameters and updates the trainable parameters based on the gradient.

[0096] Optionally, the total loss function can be obtained using the following formula:

[0097] ;

[0098] ;

[0099] in, For node classification loss function, Let the topology consistency loss function be... Let the channel consistency loss function be... The characteristic sparsity loss function is... , , For weight hyperparameters, For the set of sample nodes, Let KL divergence be the KL divergence. The probability distribution of morphological channel categories for the convergence features of morphological channels in the samples. The probability distribution of process channel categories for the convergence features of the sample process channels;

[0100] Optionally, the expression corresponding to the implicit function differentiation method can be obtained through the following formula:

[0101]

[0102] in, Let be the fixed point after convergence. It is the identity matrix. The joint state vector, composed of the sample morphology channel convergence features and the sample process channel convergence features of all sample nodes, satisfies the fixed-point equation. , These are trainable parameters for the topographic-hydrological co-evolution layer.

[0103] For example, the node classification loss function is calculated based on the cross-entropy between the class probability distribution predicted by the classification multilayer perceptron and the true class labels provided by the labeled terrain structure line network, using the sample morphology channel convergence features of each sample node in the sample space graph structure. This node classification loss function drives the model to correctly identify the terrain feature type of each sample node. The channel consistency loss function is calculated by the terrain feature line intelligent extraction terminal inputting the sample morphology channel convergence features of each sample node into a first classification head to obtain the morphology channel class probability distribution. The convergence features of the sample process channels of the sample nodes are input into a second classification head to obtain the probability distribution of process channel categories. Through calculation and The channel consistency loss function is calculated by averaging the KL divergence of all sample nodes. This channel consistency loss function drives the classification predictions of the morphological and process channels for the same sample node to converge, thus ensuring that the two channels reach a unified classification judgment after co-evolution. The feature sparsity loss function is calculated based on the sum of the L1 norms of the convergence features of the sample morphological channels of all sample nodes. The feature sparsity loss function constrains the activation values ​​of the convergence features of the sample morphological channels of sample nodes in non-feature regions to approach zero, thereby reducing the risk of overfitting.

[0104] Furthermore, during the backpropagation process, the intelligent terrain feature line extraction terminal treats the multiple iterative updates of the terrain-hydrological co-evolution layer as solving a fixed-point equation. The iterative process, in which, This is the joint state vector composed of the sample morphology channel features and sample process channel features of all sample nodes. This is the set of all trainable parameters in the topographic-hydrological co-evolution layer. The intelligent topographic feature line extraction terminal calculates the total loss function for converged fixed points. The gradient is calculated using the implicit function differentiation method to determine the total loss function with respect to the trainable parameters. gradient , Let be the Jacobian matrix of the iterative function with respect to the joint state vector at the fixed point. The intelligent terrain feature line extraction terminal updates the trainable parameters of the terrain-hydrology co-evolution layer based on gradients through an optimizer. Once one training iteration is completed, the terrain feature line intelligent extraction terminal repeats the above forward propagation, loss calculation, gradient calculation, and parameter update process until the total loss function value no longer decreases on the validation set or reaches the preset maximum number of training rounds. At this point, the terrain-hydrology co-evolution layer is considered to have been trained, and the trainable parameters are solidified and saved for inference.

[0105] In one embodiment, in the feature space, local principal direction analysis is performed on the convergence features of the sample morphology channels of each sample node to obtain the feature line directions of each sample node, and a topological consistency loss function is constructed based on the consistency measure of the feature line directions between adjacent sample nodes, including:

[0106] S61. Based on the convergence characteristics of the sample morphology channels of each sample node, in the feature space, principal component analysis is performed on the set of adjacent sample nodes in the spatial neighborhood of the sample node and the first principal component direction is extracted to obtain the principal gradient direction of the feature space of each sample node.

[0107] To illustrate, for the currently processed sample node, the terrain feature line intelligent extraction terminal determines the set of neighboring nodes in space based on the sample spatial graph structure. The neighboring node set is determined by establishing a circular spatial neighborhood window with the spatial coordinates of the sample node as the center and a preset neighborhood search radius. All other sample nodes falling within this circular spatial neighborhood window are included in the spatial neighborhood node set. The terrain feature line intelligent extraction terminal obtains the sample morphological channel convergence feature vector corresponding to each sample node in the spatial neighborhood node set, and simultaneously obtains the sample morphological channel convergence feature vector of the current sample node itself. These sample morphological channel convergence feature vectors are combined into a two-dimensional matrix in a row-stacked manner. The number of rows in the two-dimensional matrix equals the number of nodes in the spatial neighborhood node set plus the current sample node itself, and the number of columns equals the dimension of the sample morphological channel convergence feature vectors. Further, the terrain feature line intelligent extraction terminal performs a centering process on the two-dimensional matrix, that is, it calculates the arithmetic mean of all elements in each column of the two-dimensional matrix, and then subtracts the arithmetic mean of the column elements from each element in the column to obtain the centered feature matrix. The intelligent terrain feature line extraction terminal calculates the covariance matrix of the centralized feature matrix, where the first... Line number The elements of the column are equal to the elements of all sample nodes at the 1st position. Dimensional features and the first The covariance of each sample node on the dimensional feature, where the covariance is the variance of each sample node on the dimensional feature. The centered value on the dimensional feature and the dimensional feature on the th dimension The sum of the products of the centered values ​​on the feature dimension is divided by the number of sample nodes minus one. Optionally, the intelligent terrain feature line extraction terminal performs eigenvalue decomposition on the covariance matrix. Eigenvalue decomposition decomposes the covariance matrix into a set of eigenvalues ​​and an eigenvector corresponding to each eigenvalue. The eigenvalues ​​are scalars representing the magnitude of the variance along the eigenvector direction, and the eigenvectors are unit vectors with the same dimension as the eigenvectors, representing the direction corresponding to the variance. The intelligent terrain feature line extraction terminal finds the eigenvalue with the largest value from all eigenvalues ​​and extracts the eigenvector corresponding to the largest eigenvalue as the principal gradient direction in the feature space at the sample node. The principal gradient direction in the feature space represents the direction in which the morphological channel convergence feature changes the most in the feature space within the local spatial neighborhood of the sample node.

[0108] S62. Based on the principal gradient direction and spatial coordinates of the feature space corresponding to the two sample nodes in each adjacent sample node pair, the feature line consistency measure of each adjacent sample node pair is calculated.

[0109] Optionally, the expression for the feature line consistency measure is:

[0110] ;

[0111] In the formula, For adjacent sample node pairs A measure of consistency of feature lines between them. For sample nodes The principal gradient direction in the feature space at that location, For sample nodes The principal gradient direction in the feature space at that location, For sample nodes spatial coordinates, For sample nodes spatial coordinates, This refers to the spatial bandwidth parameter.

[0112] Furthermore, for a pair of adjacent sample nodes, the terrain feature line intelligent extraction terminal obtains the principal gradient direction vectors of the feature space at the first sample node and the second sample node. The intelligent terrain feature line extraction terminal calculates the cosine of the angle between these two principal gradient direction vectors and takes their absolute value. The intelligent terrain feature line extraction terminal simultaneously acquires the spatial coordinates of the first and second sample nodes, and calculates the Euclidean distance between the two spatial coordinates. Specifically, the spatial bandwidth parameter... For a pre-defined positive real number, To control the scaling parameter of the rate at which spatial distance decays as a measure of consistency, The larger the value, the more gradual the attenuation effect of spatial distance. The smaller the value, the more severe the attenuation effect of spatial distance. It is a spatial distance attenuation factor between zero and one, when the spatial distance between two sample nodes is zero. When the spatial distance between two sample nodes approaches infinity It approaches zero.

[0113] S63. Construct a topological consistency loss function based on the feature line consistency measure of each pair of adjacent sample nodes.

[0114] Optionally, the expression for the topology consistency loss function is:

[0115]

[0116] In the formula, Let the topology consistency loss function be... This represents the total number of iterations. For sample nodes space in front Adjacent sample nodes are adjacent sample nodes in the set.

[0117] Optionally, for each iteration step from the first iteration to the Tth iteration, the intelligent terrain feature line extraction terminal obtains the feature line consistency metric values ​​corresponding to all adjacent sample node pairs at each iteration step, where T is the pre-set maximum number of iteration steps for the terrain-hydrological co-evolution layer. For each adjacent sample node pair, the intelligent terrain feature line extraction terminal obtains the feature line consistency metric values ​​corresponding to all adjacent sample node pairs at each iteration step. The characteristic line consistency metric is taken as the logarithm with the natural constant as the base. Since the consistency metric value is in the range of zero to one, when the consistency metric approaches one... When the consistency metric approaches zero It tends towards negative infinity. The intelligent terrain feature line extraction terminal... Taking the negative value yields a non-negative numerical value, which indicates that the consistency measure is close to one. Approaching zero, when consistency metrics are far from zero. It is a large positive value. The topology consistency loss function penalizes the discontinuity of all adjacent sample node pairs in the feature space in all iteration steps. The smaller the value of the topology consistency loss function, the more consistent the principal gradient direction of adjacent sample nodes in the feature space, that is, the sample nodes on the same terrain feature line form a continuous low-dimensional manifold structure in the feature space.

[0118] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0119] Based on the same inventive concept, this application also provides a deep learning-based intelligent terrain feature line extraction system for implementing the aforementioned deep learning-based intelligent terrain feature line extraction method. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more deep learning-based intelligent terrain feature line extraction system embodiments provided below can be found in the limitations of the deep learning-based intelligent terrain feature line extraction method described above, and will not be repeated here.

[0120] In one exemplary embodiment, such as Figure 2 As shown, a deep learning-based intelligent terrain feature line extraction system 200 is provided, comprising:

[0121] The feature calculation module 201 is used to calculate the geometric feature vector and process feature vector of each grid point based on the acquired digital elevation model of the target area, and use the geometric feature vector and process feature vector as the initial features of the morphological channel and the initial features of the process channel, respectively.

[0122] The co-evolution module 202 is used to construct a spatial graph structure with grid points as nodes based on the initial features of the morphological and process channels of each grid point. A pre-trained topographic-hydrological co-evolution layer iteratively updates the initial features of the morphological and process channels of each node in the spatial graph structure until convergence to a fixed point, obtaining the convergent features of the morphological and process channels for each node. The topographic-hydrological co-evolution layer includes morphological and process channels, which exchange and mutually correct features through bidirectional coupling gating. The topographic-hydrological co-evolution layer achieves iterative updates of the initial features of the morphological and process channels by performing spatial message passing within the morphological and process channels respectively.

[0123] The classification and recognition module 203 is used to classify and recognize the morphological channel convergence features of each node to obtain the terrain feature type label corresponding to each node.

[0124] The topology reconstruction module 204 is used to calculate the geodesic distance between adjacent node pairs in the feature space based on the morphological channel convergence features of each node, and to perform virtual node interpolation and connection edge establishment for adjacent node pairs with topological breaks to obtain the terrain structure line network. The adjacent node pairs with topological breaks satisfy the following conditions: the geodesic distance is less than the preset feature distance threshold, and the terrain feature type labels corresponding to the two nodes in the adjacent node pairs are the same.

[0125] In one embodiment, the feature calculation module 201 is further configured to:

[0126] For each grid point, calculate the directional slope from the grid point to its eight neighboring grid points, and use a differentiable softmax function to convert each directional slope into the flow probability from the eight neighboring grid points to the grid point, thus obtaining the eight-directional flow probability of each grid point.

[0127] Based on the eight-directional flow probabilities of each grid point, the conjugate gradient method is used to iteratively solve the confluence equations to obtain the upstream catchment area of ​​each grid point.

[0128] Based on the upstream catchment area and local slope of each grid point, the cumulative flow energy index along the route is calculated, and the process feature vector of each grid point is generated based on the eight-directional flow probability, upstream catchment area and cumulative flow energy index along the route.

[0129] In one embodiment, the cooperative evolution module 202 is further configured to:

[0130] For each node in the spatial graph structure, based on the morphological channel features and process channel features of the node and its corresponding neighboring nodes in the previous iteration step, the morphological channel message passing coefficients and process channel message passing coefficients of the current iteration step are calculated by a multilayer perceptron; where, when the current iteration step is the first step, the morphological channel features and process channel features in the previous iteration step are the initial features of the morphological channel and the initial features of the process channel, respectively.

[0131] Based on the morphological channel message passing coefficient of the current iteration step, the morphological channel features of the neighboring nodes in the previous iteration step are weighted and aggregated to obtain the morphological channel aggregated features of the current iteration step. Based on the process channel message passing coefficient of the current iteration step, the process channel features of the neighboring nodes in the previous iteration step are weighted and aggregated to obtain the process channel aggregated features of the current iteration step.

[0132] Based on the morphological channel aggregation features, the gating vector of the morphological channel to the process channel is calculated by a multilayer perceptron to obtain the first gating vector. The process channel aggregation features are then corrected based on the first gating vector to obtain the updated process channel features.

[0133] Based on the process channel aggregation features, the gating vector of the process channel to the morphology channel is calculated by a multilayer perceptron to obtain the second gating vector. The morphology channel aggregation features are then corrected based on the second gating vector to obtain the morphology channel update features.

[0134] The process channel update features and morphology channel update features are subjected to residual connection and layer normalization to obtain the morphology channel features and process channel features of the current iteration step. When the current iteration step is the last step, the morphology channel features and process channel features are the morphology channel convergence features and process channel convergence features, respectively.

[0135] In one embodiment, the topology reconstruction module 204 is further configured to:

[0136] For each node, search for neighboring nodes within a preset spatial neighborhood radius to obtain the corresponding neighboring node pairs, and calculate the geodesic distance between each neighboring node pair in the feature space.

[0137] If the geodesic distance between adjacent node pairs is less than the preset feature distance threshold, and the terrain feature type labels corresponding to the two nodes in the adjacent node pair are the same, then it is determined that there is a topological break in the adjacent node pair.

[0138] On the node spatial path of adjacent node pairs where topological breaks are determined, the position with the smallest rate of change of morphological channel features along the node spatial path is searched as the insertion position, and a virtual node is inserted at the insertion position; the terrain feature type label of the virtual node is the same as that of the adjacent node pair.

[0139] By establishing connection edges between virtual nodes and the two nodes in each adjacent node pair, a terrain structure line network is obtained.

[0140] In one embodiment, the deep learning-based intelligent terrain feature line extraction system 200 is further used for:

[0141] Obtain the sample digital elevation model and the corresponding labeled terrain structure line network, and calculate the initial features of the sample morphology channel and the initial features of the sample process channel based on the sample digital elevation model to construct the sample spatial map structure;

[0142] The initial features of the sample morphology channel and the initial features of the sample process channel of each sample node in the sample spatial map structure are iteratively updated through the topography-hydrology co-evolution layer until they converge to a fixed point, thus obtaining the convergence features of the sample morphology channel and the convergence features of the sample process channel.

[0143] In the feature space, the local principal direction analysis is performed on the convergence features of the sample morphology channels of each sample node to obtain the feature line direction of each sample node, and a topological consistency loss function is constructed based on the consistency measure of the feature line direction between adjacent sample nodes.

[0144] A total loss function is constructed based on the topology consistency loss function, the preset node classification loss function, the channel consistency loss function, and the feature sparsity loss function. The trainable parameters of the topography-hydrology co-evolution layer are iteratively trained using the backpropagation method until the total loss function converges, thus obtaining the pre-trained topography-hydrology co-evolution layer. The backpropagation method uses the implicit function differentiation method to calculate the gradient of the total loss function with respect to the trainable parameters, and updates the trainable parameters based on the gradient.

[0145] In one embodiment, the deep learning-based intelligent terrain feature line extraction system 200 is further used for:

[0146] Based on the convergence characteristics of the sample morphology channels of each sample node, principal component analysis is performed on the set of adjacent sample nodes in the spatial neighborhood of the sample node in the feature space, and the first principal component direction is extracted to obtain the principal gradient direction of the feature space of each sample node.

[0147] Based on the principal gradient direction and spatial coordinates of the feature space corresponding to the two sample nodes in each adjacent sample node pair, the feature line consistency measure of each adjacent sample node pair is calculated.

[0148] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0149] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0150] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0151] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for intelligent extraction of terrain feature lines based on deep learning, characterized in that, The method includes: Based on the digital elevation model of the target area, the geometric feature vector and process feature vector of each grid point are calculated, and the geometric feature vector and the process feature vector are used as the initial features of the morphological channel and the initial features of the process channel, respectively. Based on the initial features of the morphological channels and the initial features of the process channels of each grid point, a spatial graph structure with grid points as nodes is constructed. A pre-trained topographic-hydrological co-evolution layer iteratively updates the initial features of the morphological channels and the initial features of the process channels of each node in the spatial graph structure until convergence to a fixed point, obtaining the convergent features of the morphological channels and the convergent features of the process channels for each node. The topographic-hydrological co-evolution layer includes morphological channels and process channels, which exchange features and mutually correct each other through bidirectional coupling gating. The topographic-hydrological co-evolution layer achieves iterative updates of the initial features of the morphological channels and the initial features of the process channels by performing spatial message passing within the morphological channels and the process channels respectively. The morphological channel convergence features of each node are classified and identified to obtain the terrain feature type label corresponding to each node; Based on the morphological channel convergence characteristics of each node, the geodesic distance between adjacent node pairs in the feature space is calculated, and virtual node interpolation and connection edge establishment are performed for adjacent node pairs with topological breaks to obtain a terrain structure line network; the adjacent node pairs with topological breaks satisfy the condition that the geodesic distance is less than a preset feature distance threshold, and the terrain feature type labels corresponding to the two nodes in the adjacent node pairs are the same.

2. The method according to claim 1, characterized in that, The process feature vector is calculated through the following steps: For each grid point, the directional slope from the grid point to its eight neighboring grid points is calculated, and the directional slope is converted into the flow probability from the eight neighboring grid points to the grid point by a differentiable softmax function, thus obtaining the eight-directional flow probability of each grid point. Based on the eight-directional flow probabilities of each grid point, the conjugate gradient method is used to iteratively solve the confluence equations to obtain the upstream catchment area of ​​each grid point. Based on the upstream catchment area and local slope of each grid point, the cumulative flow energy index along the route is calculated, and based on the eight-directional flow probability, the upstream catchment area and the cumulative flow energy index along the route, the process feature vector of each grid point is generated. The expression for the flow direction probability is: ； In the formula, For neighborhood grid points To grid point directional slope, For temperature parameters, For grid points The set of eight neighboring grid points; The expression for the confluence equations is: ； In the formula, For grid points The upstream catchment area, To obtain from neighboring grid points Flow to grid points The probability of the flow direction, For neighborhood grid points The upstream catchment area.

3. The method according to claim 1, characterized in that, The process iteratively updates the initial features of the morphological channels and the initial features of the process channels of each node in the spatial graph structure through a pre-trained topography-hydrology co-evolution layer until it converges to a fixed point, obtaining the convergent features of the morphological channels and the convergent features of the process channels of each node, including: For each node in the spatial graph structure, based on the morphological channel features and process channel features of the node and its corresponding neighboring nodes in the previous iteration step, the morphological channel message passing coefficient and process channel message passing coefficient of the current iteration step are calculated by a multilayer perceptron; wherein, when the current iteration step is the first step, the morphological channel features and the process channel features in the previous iteration step are respectively the initial features of the morphological channel and the initial features of the process channel. Based on the morphological channel message passing coefficient of the current iteration step, the morphological channel features of the neighboring nodes in the previous iteration step are weighted and aggregated to obtain the morphological channel aggregated features of the current iteration step. Based on the process channel message passing coefficient of the current iteration step, the process channel features of the neighboring nodes in the previous iteration step are weighted and aggregated to obtain the process channel aggregated features of the current iteration step. Based on the morphological channel aggregation features, the gating vector of the morphological channel to the process channel is calculated by a multilayer perceptron to obtain a first gating vector, and the process channel aggregation features are corrected based on the first gating vector to obtain the process channel update features. Based on the process channel aggregation features, the gating vector of the process channel to the morphology channel is calculated by a multilayer perceptron to obtain a second gating vector, and the morphology channel aggregation features are corrected based on the second gating vector to obtain the morphology channel update features. The process channel update features and the morphology channel update features are subjected to residual connection and layer normalization processing respectively to obtain the morphology channel features and the process channel features of the current iteration step; wherein, when the current iteration step is the last step, the morphology channel features and the process channel features are respectively the morphology channel convergence features and the process channel convergence features; The expressions for the morphological channel message passing coefficient and the process channel message passing coefficient in the current iteration step are as follows: ； ； In the formula, For the current iteration step The morphological channel message transmission coefficient, For the current iteration step Process channel message transmission coefficient, For the node in the previous iteration step The morphological channel characteristics, For the node in the previous iteration step Process channel characteristics, For the node in the previous iteration step Neighboring nodes The morphological channel characteristics, For the node in the previous iteration step Neighboring nodes Process channel characteristics, This is a vector concatenation operation; The expressions for the first gating vector and the process channel update feature are as follows: ； ； In the formula, Let be the first gating vector. For the current iteration step Middle node Process channel characteristics, For the current iteration step Middle node The morphological channel aggregation characteristics, For the current iteration step Middle node Process channel aggregation characteristics, It is the sigmoid activation function. For element-wise product; The expressions for the second gating vector and the morphological channel update feature are as follows: ； ； In the formula, This is the second gating vector. Update features for morphological channels.

4. The method according to claim 1, characterized in that, Based on the convergence characteristics of the morphological channels of each node, the geodesic distance between adjacent node pairs in the feature space of the spatial graph structure is calculated, and virtual node interpolation and connection edge establishment are performed on adjacent node pairs with topological breaks to obtain a terrain structure line network, including: For each node, neighboring nodes are searched within a preset spatial neighborhood radius to obtain the corresponding neighboring node pairs, and the geodesic distance between each neighboring node pair in the feature space is calculated. If the geodesic distance between adjacent node pairs is less than a preset feature distance threshold, and the terrain feature type labels corresponding to the two nodes in the adjacent node pair are the same, then it is determined that there is a topological break in the adjacent node pair. On the node spatial path of the adjacent node pair where a topological break is determined to exist, the position with the smallest rate of change of the morphological channel feature along the node spatial path is searched as the insertion position, and a virtual node is inserted at the insertion position; the terrain feature type label of the virtual node is the same as that of the adjacent node pair. By establishing connection edges between the virtual nodes and the two nodes in the adjacent node pair, the terrain structure line network is obtained. The expression for the geodesic distance is: ； In the formula, For adjacent node pairs geodesic distance, For the node to adjacent nodes Spatial path, Location on the node spatial path The morphological channel convergence characteristics at the location.

5. The method according to claim 1, characterized in that, The topographic-hydrological co-evolution layer is obtained through the following steps: Obtain the sample digital elevation model and the labeled terrain structure line network corresponding to the sample digital elevation model, and calculate the initial features of the sample morphology channel and the initial features of the sample process channel based on the sample digital elevation model to construct the sample space map structure; The initial features of the sample morphology channel and the initial features of the sample process channel of each sample node in the sample spatial map structure are iteratively updated through the topography-hydrology co-evolution layer until they converge to a fixed point, thereby obtaining the convergence features of the sample morphology channel and the convergence features of the sample process channel. In the feature space, the local principal direction analysis is performed on the convergence features of the sample morphology channels of each sample node to obtain the feature line direction of each sample node, and a topological consistency loss function is constructed based on the consistency metric of the feature line direction between adjacent sample nodes. A total loss function is constructed based on the topology consistency loss function, the preset node classification loss function, the channel consistency loss function, and the feature sparsity loss function. The trainable parameters of the topography-hydrology co-evolution layer are iteratively trained using the backpropagation method until the total loss function converges, thus obtaining the pre-trained topography-hydrology co-evolution layer. The backpropagation method uses the implicit function differentiation method to calculate the gradient of the total loss function with respect to the trainable parameters, and updates the trainable parameters based on the gradient.

6. The method according to claim 5, characterized in that, In the feature space, local principal direction analysis is performed on the convergence features of the sample morphology channels of each sample node to obtain the feature line directions of each sample node. A topological consistency loss function is then constructed based on the consistency metric of the feature line directions between adjacent sample nodes, including: Based on the convergence characteristics of the sample morphology channels of each sample node, in the feature space, principal component analysis is performed on the set of adjacent sample nodes in the spatial neighborhood of the sample node and the first principal component direction is extracted to obtain the principal gradient direction of the feature space of each sample node. Based on the principal gradient direction and spatial coordinates of the feature space corresponding to the two sample nodes in each adjacent sample node pair, the feature line consistency measure of each adjacent sample node pair is calculated. The topology consistency loss function is constructed based on the feature line consistency metric of each pair of adjacent sample nodes; The expression for the feature line consistency measure is as follows: ； In the formula, For adjacent sample node pairs A measure of consistency of feature lines between them. For sample nodes The principal gradient direction in the feature space at that location, For sample nodes The principal gradient direction in the feature space at that location, For sample nodes spatial coordinates, For sample nodes spatial coordinates, This refers to the spatial bandwidth parameter; The expression for the topology consistency loss function is: ； In the formula, Let the topology consistency loss function be... This represents the total number of iterations. For sample nodes space in front Adjacent sample nodes are adjacent sample nodes in the set.

7. The method according to claim 5, characterized in that, The process of iteratively training the trainable parameters of the topography-hydrology co-evolution layer using backpropagation until the total loss function converges, to obtain the pre-trained topography-hydrology co-evolution layer, includes: The total loss function is obtained using the following formula: ； ； in, For node classification loss function, Let the topology consistency loss function be... Let be the channel consistency loss function. The characteristic sparsity loss function is... , , For weight hyperparameters, For the set of sample nodes, Let KL divergence be the KL divergence. The probability distribution of morphological channel categories for the convergence features of morphological channels in the samples. The probability distribution of process channel categories for the convergence features of the sample process channels; The expression corresponding to the implicit function differentiation method can be obtained through the following formula: ； in, Let this be the fixed point after convergence. It is the identity matrix. The joint state vector, composed of the sample morphology channel convergence features and the sample process channel convergence features of all sample nodes, satisfies the fixed-point equation. , These are the trainable parameters of the topography-hydrology co-evolution layer.

8. A deep learning-based intelligent terrain feature line extraction system, characterized in that, The system includes: The feature calculation module is used to calculate the geometric feature vector and process feature vector of each grid point based on the acquired digital elevation model of the target area, and to use the geometric feature vector and the process feature vector as the initial features of the morphological channel and the initial features of the process channel, respectively. A co-evolution module is used to construct a spatial graph structure with grid points as nodes based on the initial features of the morphological channels and the initial features of the process channels of each grid point. A pre-trained topographic-hydrological co-evolution layer iteratively updates the initial features of the morphological channels and the initial features of the process channels of each node in the spatial graph structure until convergence to a fixed point, obtaining the convergent features of the morphological channels and the convergent features of the process channels for each node. The topographic-hydrological co-evolution layer includes morphological channels and process channels, which exchange and mutually correct features through bidirectional coupling gating. The topographic-hydrological co-evolution layer achieves iterative updates of the initial features of the morphological channels and the initial features of the process channels by performing spatial message passing within the morphological channels and the process channels respectively. The classification and recognition module is used to classify and recognize the morphological channel convergence features of each node to obtain the terrain feature type label corresponding to each node. The topology reconstruction module is used to calculate the geodesic distance between adjacent node pairs in the feature space based on the morphological channel convergence features of each node, and to perform virtual node interpolation and connection edge establishment for adjacent node pairs with topological breaks to obtain a terrain structure line network; the adjacent node pairs with topological breaks satisfy the condition that the geodesic distance is less than a preset feature distance threshold, and the terrain feature type labels corresponding to the two nodes in the adjacent node pairs are the same.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images