A method, medium and system for extracting a coastal water area based on a LiDAR point cloud TIN triangle form

Abstract

The present application provides a kind of based on LiDAR point cloud TIN triangle form coastal zone water area extraction method, medium and system, belong to coastal zone water area technical field, the present application is by to target area airborne LiDAR point cloud ground point extraction and irregular triangle net construction, calculate the long short side parameter of triangle envelope rectangle, introduce local space point array density adaptive correction operator to dynamic correction is carried out to discriminant threshold, then adopt the artificial intelligence model of collaborative embedding based on multi-resolution pyramid tensor, long span dense residual network, adaptive activation operator and anisotropic diffusion tensor cooperatively embedded output still water area confidence, in combination with the long length discriminant threshold, identify still water area triangle and eliminate isolated long and narrow triangle noise, finally, the continuous still water area vector surface layer of generation is combined, solves the technical problem that airborne LiDAR point cloud is in land and water junction area due to point cloud uneven density and leads to water area boundary discriminant threshold invalid, still water area extraction precision unstable.

Description

Technical Field

[0001] This invention belongs to the field of coastal water area technology, and more specifically, relates to a method, medium, and system for extracting coastal water areas based on the triangular shape of LiDAR point cloud TIN. Background Technology

[0002] Coastal water boundary extraction is a fundamental technical step in coastal wetland mapping, tidal flat geomorphology monitoring, and coastline dynamic analysis. Existing technologies, water boundary extraction methods based on airborne LiDAR point clouds, typically rely on irregular triangular meshes. They identify elongated triangles in still water areas by calculating the ratio of the longer side to the shorter side, thus achieving a geometric description of the water boundary. This method is applicable to areas with uniform point cloud density and is widely used for vector surface extraction of inland lakes and regular rivers. However, in typical coastal geomorphic areas such as tidal channels and tidal flats, the point cloud density at the land-water interface exhibits significant spatial heterogeneity due to dramatic topographic relief, uneven vegetation cover, and multiple echo superpositions. In existing technologies, when there is a large deviation between the local point cloud density at the land-water interface and the global average density, the fixed longer-to-shorter-side ratio discrimination threshold cannot adapt to local point cloud density variations. This leads to the omission of small, elongated triangles in areas with higher density and the misidentification of non-water triangles in areas with lower density, resulting in a systematic bias in the water boundary extraction results. In other words, existing technologies have technical problems such as the failure of water boundary discrimination thresholds in land-water interface areas due to uneven density of point clouds, and unstable extraction accuracy in still water areas by airborne LiDAR point clouds. Summary of the Invention

[0003] In view of this, the present invention provides a method, medium and system for extracting coastal water areas based on the triangular shape of LiDAR point cloud TIN, which can solve the technical problems in the prior art where the water boundary discrimination threshold fails due to the uneven density of airborne LiDAR point clouds in the land-water interface area, and the extraction accuracy in still water areas is unstable.

[0004] The present invention is implemented as follows: The first aspect of the present invention provides a method for extracting coastal water areas based on the triangular morphology of LiDAR point cloud TINs, comprising the following steps:

[0005] Acquire airborne LiDAR point cloud data of the target area, perform classification processing to extract the ground point cloud set, and the still water area in the ground point cloud set does not contain ground points;

[0006] Based on the three-dimensional spatial grid block streaming incremental partitioning mechanism, an irregular triangular network is constructed on the ground point cloud set to obtain a triangular set covering all ground points.

[0007] Calculate the length of the long side and the length of the short side of the envelope rectangle of each triangle in the triangle set, and introduce a local spatial lattice density adaptive correction operator to correct the ratio of the long side to the short side;

[0008] A high-order gradient flow and continuous state information transmission model is used to process the corrected triangle geometric parameters and output the confidence score of the triangle in the still water region.

[0009] Based on the confidence score of triangles in still water areas and combined with the narrow length discrimination threshold, triangles in still water areas are identified, and isolated narrow triangle noise points are removed.

[0010] The identified triangles in the still water area are merged to generate a continuous vector surface layer of the still water area.

[0011] Specifically, the three-dimensional spatial mesh block streaming incremental partitioning mechanism divides the ground point cloud into multiple mesh blocks according to the three-dimensional spatial location. Each block is loaded into memory in batches in a streaming manner and incremental triangulation is performed. A dynamic circular buffer stack is set up inside the block to temporarily store the point cloud data to be processed. Lock-free atomic queues are used between blocks to achieve parallel multi-threaded independent network construction. After the network construction of each block is completed, boundary stitching is performed.

[0012] Specifically, the local spatial lattice density adaptive correction operator calculates the median distance between nearest neighbors within the local region containing the triangle. Correction factor The calculation formula is ,in The average nearest neighbor distance of the regional point cloud. The calibration coefficients are used, and the corrected narrow-length discriminant is: And simultaneously satisfy ,in For narrow length discrimination threshold, This is the lower limit value for the length of the longer side.

[0013] The high-order gradient flow and continuous state information transmission model is equipped with a pre-processing module at its input end, which converts the irregular triangular mesh into a multi-resolution pyramid tensor. Each layer of the tensor corresponds to a triangle geometric parameter matrix under different spatial resolutions. The parameter matrix includes four types of feature channels: the length of the long side of the envelope rectangle, the length of the short side, the correction coefficient, and the deflection angle of the direction vector.

[0014] The backbone of the high-order gradient flow and continuous state information transmission model consists of multiple stacked feature extraction layers. Each layer contains multiple adaptive activation operators. The parameters of the adaptive activation operators are reconstructed in real time by the state entropy of the output feature map of the preceding layer. Each resolution feature layer is connected through a long-span dense residual network, and the cross-layer connection weights are dynamically adjusted by the inter-layer gradient flow optimization operator.

[0015] The high-order gradient flow and continuous state information transmission model is equipped with a dynamic gated routing module, which selects the signal transmission path based on the edge pixel ambiguity score of the current feature map. When the edge pixel ambiguity score exceeds the ambiguity threshold, it jumps to the high-resolution local backtracking sub-network to perform secondary correction on the features of the region before merging into the backbone network.

[0016] The higher-order gradient flow and continuous-state information transmission model is embedded with a tidal flat elevation field spatial smoothing algorithm based on anisotropic diffusion equations, and the anisotropic diffusion tensor... The construction formula is ,in For local elevation gradient, Let be the diffusion intensity function that monotonically decreases as the gradient increases. The weighting coefficients of the model convolution kernel are dynamically adjusted using a unit tensor and diffusion intensity.

[0017] The loss function of the high-order gradient flow and continuous state information transmission model is composed of a weighted sum of a classification cross-entropy term and a boundary perimeter-area ratio anisotropy penalty term. The boundary perimeter-area ratio anisotropy penalty term is used to constrain the geometric regularity of the predicted water boundary.

[0018] The high-order gradient flow and continuous state information transmission model nests a minimum spanning tree connectivity discrimination algorithm. Triangles with confidence scores exceeding the confidence score threshold are abstracted as graph nodes, and adjacent triangles with shared edges are abstracted as graph edges. The edge weight is taken as the difference in confidence scores between adjacent triangles. After constructing the minimum spanning tree, the size of the connected region is determined based on the number of nodes in the connected components. Connected components with a size lower than the number threshold are determined as isolated misjudged regions and are removed.

[0019] Specifically, the removal of isolated, narrow triangular noise points involves an anisotropic filtering mechanism based on topological connectivity and spatial area accumulation constraints. By iteratively calculating the geometric centroid distance and intersecting side length of the connected triangular domains, connected domains with accumulated area values ​​below an area threshold and centroid distances exceeding a distance threshold are identified as isolated noise points and removed.

[0020] The higher-order gradient flow and continuous-state information transmission model is configured with a confidence adjustment function, which calculates the adjustment value based on four data points: the ratio of the longer side to the shorter side of the triangle, the correction coefficient, the state entropy, and the local point cloud density. ,when When the response slope parameter of the adaptive activation operator is lowered below the lower limit adjustment threshold, the response slope parameter of the adaptive activation operator is reduced; when When the response is between the lower and upper adjustment thresholds, the response slope parameter remains constant; when... When the response slope parameter is higher than the upper limit adjustment threshold, increase the response slope parameter.

[0021] In the scenario where the triangle crossing the water area at the bend of the tidal channel presents an asymmetrical distortion, the principal component decomposition of the triangle vertex coordinate matrix is ​​performed to obtain the direction vector of the triangle's major axis. Combined with the coefficient of variation of the triangle's major axis direction along the tidal channel, a nonlinear judgment surface is constructed with the combined effect of three parameters: the ratio of the long side to the short side, the deflection angle of the direction vector, and the coefficient of variation, to replace the single threshold judgment method based on the ratio of the long side to the short side.

[0022] The calculation of triangle geometric parameters is implemented using a parallel geometric computing kernel based on the graphics processor's vectorized instruction set. The coordinates of the three-dimensional vertices of the triangle are mapped into a single instruction multiple data stream form, and the long side length, short side length, and envelope rectangle parameters are extracted synchronously through a multi-channel parallel method.

[0023] The training dataset for the high-order gradient flow and continuous state information transmission model is obtained by collecting airborne LiDAR point cloud data under different tidal conditions in multiple typical coastal areas. Irregular triangular networks are constructed for the ground point clouds in each area and geometric morphology parameters are calculated. Simultaneously, measured waterline data of the corresponding time phases are collected as the labeling basis. The triangular geometric morphology parameter matrix and the measured waterline labeling results are matched according to spatial location to form training sample pairs, which are divided into training set, validation set and test set according to a preset ratio.

[0024] Among them, calibration coefficient The value range is 0.3 to 0.8; the lower limit adjustment threshold is 0.3, and the upper limit adjustment threshold is 0.7. , , , The value was determined by collecting multiple sets of on-site measured data from the target area, combining multiple iterative experiments under different tide levels and point cloud densities, and comparing the degree of agreement between the identification results and the measured waterline.

[0025] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, are used to perform the above-described method for extracting coastal water areas based on the triangular morphology of LiDAR point cloud TIN.

[0026] A third aspect of the present invention provides a coastal water area extraction system based on the triangular shape of LiDAR point cloud TIN, comprising the aforementioned computer-readable storage medium, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0027] This invention introduces a local spatial lattice density adaptive correction operator. It dynamically calculates a correction coefficient based on the ratio between the median nearest distance to the nearest neighbor in the local region of a triangle and the global average nearest neighbor distance, and applies this correction to the long-short side ratio discriminant. This allows the discrimination threshold to adaptively adjust with changes in local point cloud density, thus solving the technical problems of discrimination threshold failure and unstable extraction accuracy in still water areas caused by uneven point cloud density in land-water interface regions. Furthermore, this invention employs a high-order gradient flow and continuous-state information transmission model to perform multi-scale deep fusion of the corrected geometric parameters and embeds anisotropic diffusion tensors to enhance feature contrast at land-water interfaces, enabling the model to maintain stable discrimination capability even in complex boundary regions. In summary, this invention solves the technical problems mentioned in the background art, such as the failure of the water boundary discrimination threshold and unstable extraction accuracy in still water areas caused by uneven point cloud density in airborne LiDAR point clouds in land-water interface regions. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method of the present invention.

[0029] Figure 2 Comparison of the spatial distribution of candidate triangles before and after removing isolated noise points.

[0030] Figure 3 This is a diagram showing the results of vector surface extraction in a still water region.

[0031] Figure 4 This is a display diagram showing an irregular triangular mesh overlaid on top of each other. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0033] like Figure 1 The diagram shown is a flowchart of a method for extracting coastal water areas based on the triangular shape of LiDAR point cloud TIN, provided by the first aspect of this invention. This method includes the following steps:

[0034] S01. Acquire airborne LiDAR point cloud data of the target area, perform classification processing to extract a ground point cloud set, wherein the still water area in the ground point cloud set does not contain ground points.

[0035] S02. Based on the three-dimensional spatial grid block-based incremental partitioning mechanism, an irregular triangular network is constructed on the ground point cloud set to obtain a triangular set covering all ground points.

[0036] S03. Calculate the length of the long side and the length of the short side of the envelope rectangle of each triangle in the triangle set, and introduce a local spatial lattice density adaptive correction operator to correct the ratio of the long side to the short side.

[0037] S04. The corrected triangle geometric parameters are processed using a high-order gradient flow and continuous state information transmission model, and the confidence score of the triangle in the still water region is output.

[0038] S05. Based on the confidence level of the triangle in the still water area and combined with the narrow length discrimination threshold, identify the triangle in the still water area and remove isolated narrow triangle noise points.

[0039] S06. Merge the identified triangles in the still water area to generate a continuous vector surface layer of the still water area.

[0040] The non-publicly known concepts and technical features in the above steps are explained as follows.

[0041] The aforementioned 3D spatial mesh block streaming incremental triangulation mechanism refers to dividing the ground point cloud into multiple mesh blocks according to the 3D spatial location during the construction of irregular triangulation. Each block is loaded into memory in batches in a streaming manner and incremental triangulation is performed. A dynamic circular buffer stack is set up within each block to temporarily store the point cloud data to be processed. Lock-free atomic queues are used between blocks to achieve parallel multi-threaded independent network construction. After the network construction of each block is completed, boundary stitching is performed, thereby controlling the memory usage to the size of a single block, avoiding memory overflow caused by loading ultra-large-scale vertices at the same time, and eliminating concurrent contention for locks between threads in global triangulation.

[0042] The local spatial lattice density adaptive correction operator is used to correct the discrimination threshold failure caused by uneven point cloud density at the land-water interface before calculating the ratio of the long side length to the short side length of the envelope rectangle. The correction coefficient is calibrated by calculating the median distance of the nearest neighbor points within the local region where the triangle is located, denoted as . The correction factor is denoted as Its calculation formula is ,in The average nearest neighbor distance of the regional point cloud. The calibration coefficient ranges from 0.3 to 0.8. The corrected narrow length discriminant is: the length of the longer side of the enclosing rectangle. Length of the shorter side of the enclosing rectangle The ratio multiplied by the correction factor ,Right now ,in This is the threshold for narrow length discrimination. The above discriminant formula also requires the length of the longer side of the enclosing rectangle. It is used to filter out extremely small, narrow triangles generated in areas of high point cloud density, where , , , The value was determined by collecting multiple sets of on-site measured data from the target area, combining multiple iterative experiments under different tide levels and point cloud densities, and comparing the degree of agreement between the identification results and the measured waterline.

[0043] The calculation of the geometric parameters of the triangles, which are in the tens of millions, is implemented by a parallel geometric computing kernel based on the vectorized instruction set of the graphics processor. The coordinates of the three-dimensional vertices of the triangle are mapped into a single instruction multiple data stream form. The long side length, short side length and envelope rectangle parameters are extracted synchronously through a multi-channel parallel method to alleviate the problem of excessive consumption of single-core computing resources and the bottleneck of computing efficiency.

[0044] The removal of isolated, narrow triangular noise points is achieved using an anisotropic filtering mechanism based on topological connectivity and spatial area accumulation constraints. By iteratively calculating the geometric centroid distance and intersecting side length of the connected domains of the triangle, connected domains with a cumulative area value lower than a set area threshold and a centroid distance exceeding a set distance threshold are identified as isolated noise points and removed. The area threshold and distance threshold are determined by statistical analysis of multiple sets of measured point cloud data from the shallow water area of ​​the tidal flats.

[0045] The principal component analysis method extracts the direction vector of the triangle. It is used in scenarios where the triangles crossing the water area in the bend of the tidal channel exhibit asymmetrical distortion. By performing principal component decomposition on the coordinate matrix of the triangle vertex, the direction vector of the major axis of the triangle is obtained. Combined with the coefficient of variation of the direction of the major axis of the triangle along the direction of the tidal channel, a nonlinear judgment surface is constructed with the combined effect of three parameters: the ratio of the long side to the short side, the deflection angle of the direction vector, and the coefficient of variation. This surface is used to replace the single threshold judgment method based on the ratio of the long side to the short side, thereby improving the segmentation accuracy of the bend area.

[0046] The high-order gradient flow and continuous-state information transmission model has the following structure: A pre-processing module is set at the model input to convert the irregular triangular mesh into a multi-resolution pyramid tensor. Each layer of the tensor corresponds to a matrix of geometric parameters of triangles at different spatial resolutions, including four types of feature channels: the length of the long side of the envelope rectangle, the length of the short side, the correction coefficient, and the deflection angle of the direction vector. The model backbone consists of multiple stacked feature extraction layers. Each layer contains multiple adaptive activation operators. The parameters of these adaptive activation operators are reconstructed in real time from the state entropy of the output feature maps of the preceding layers. The higher the state entropy value, the greater the adjustment range of the response slope of the activation operator. The feature layers at different resolutions are connected through a long-span dense residual network to fuse shallow high-resolution detail features with deep low-resolution semantic features across scales. The cross-layer connection weights are dynamically adjusted by an inter-layer gradient flow optimization operator, which calculates the fusion weights based on the gradient similarity between adjacent feature maps. The model incorporates a dynamic gating routing module to select the signal transmission path based on the edge pixel ambiguity score of the current feature map. When the edge pixel ambiguity score exceeds a set threshold, the conditional branch jumps to a high-resolution local backtracking subnetwork for secondary correction of the region's features before rejoining the backbone network. A confidence regression layer at the model output outputs the confidence score of each triangle belonging to the still water region. The model's loss function is a weighted sum of a classification cross-entropy term and a boundary perimeter-to-area ratio anisotropy penalty term. This penalty term constrains the geometric regularity of the predicted water boundary, suppressing misclassifications of jagged boundaries.

[0047] The steps for establishing the training dataset for the high-order gradient flow and continuous state information transmission model specifically include: selecting multiple typical coastal areas, covering various landform types such as tidal channels, tidal flats, and coastal wetland waterways; collecting airborne LiDAR point cloud data under different tidal conditions; constructing irregular triangular networks for the ground point clouds in each area and calculating their geometric morphological parameters; simultaneously collecting measured waterline data of the corresponding time phase as a labeling basis; matching the triangular geometric morphological parameter matrix with the measured waterline labeling results according to spatial location to form training sample pairs; and dividing the sets into training set, validation set, and test set according to a preset ratio.

[0048] The training steps of the high-order gradient flow and continuous state information transmission model specifically include: inputting the multi-resolution pyramid tensor from the training set into the model; calculating the confidence of each triangular still water region through forward propagation; calculating the deviation between the prediction result and the labeled result based on the loss function weighted by the classification cross-entropy term and the anisotropy penalty term of the boundary perimeter-area ratio; updating the parameters of each layer of the model using the backpropagation algorithm, wherein the fusion weight of the inter-layer gradient flow optimization operator and the response slope parameter of the adaptive activation operator are updated synchronously and iteratively; monitoring the model convergence using a validation set during training; stopping training after reaching the preset number of iterations or the convergence of the loss function; and obtaining the trained model.

[0049] The model embeds a spatial smoothing algorithm for the tidal flat elevation field based on anisotropic diffusion equations. This algorithm works in deep collaboration with the feature transfer process of the model's hidden layer: during grid elevation feature transfer, the algorithm constructs anisotropic diffusion tensor based on local elevation gradients, denoted as... Its construction formula is ,in For local elevation gradient, Let be the diffusion intensity function that monotonically decreases as the gradient increases. It is a unit tensor. Intensity diffusion is implemented along the low gradient direction inside the water area, and diffusion is suppressed along the high gradient direction at the water-land boundary. The numerical value of this diffusion intensity dynamically adjusts the weighting coefficients of the model convolution kernel to enhance the feature contrast at the water-land interface, enabling the model to accurately capture the dramatic changes in the location of the tidal channel bank.

[0050] The resource allocation method of the model is as follows: For the feature maps of each layer of the multi-resolution pyramid tensor, a corresponding proportion of GPU memory is allocated based on the resolution. High-resolution layers, due to their large number of parameters, receive more GPU memory, while low-resolution layers receive less. For the gradient flow optimization operator computation task between layers in the long-span dense residual network, the number of computation threads is dynamically allocated based on the distribution density of triangles, with regions having higher point cloud density receiving more computation threads. When the dynamic gated routing module jumps to a high-resolution local backtracking sub-network, it triggers the memory allocation module to temporarily expand the memory space corresponding to that sub-network. After the sub-network computation is completed, this memory space is released. During data looping, training samples are loaded from disk into a memory buffer in batches. The buffer capacity has a functional relationship with the batch size and the dimension of the triangle geometric parameter matrix to avoid memory overflow.

[0051] The graph theory algorithm nested in the model is a minimum spanning tree connectivity discrimination algorithm, which is used to post-process the confidence matrix of triangles in the still water area output by the model: triangles with confidence exceeding a preset threshold are abstracted as graph nodes, and triangles with shared edges are abstracted as graph edges. The edge weight is taken as the difference in confidence of adjacent triangles. After constructing the minimum spanning tree, the scale of the connected region is determined based on the number of nodes in the connected components in the tree. Connected components with a scale lower than a preset number are determined as isolated misjudged regions and are removed, which is used to improve the continuity and integrity of the water area vector surface.

[0052] The designed adjustment function is a confidence adjustment function, used to adjust the response slope parameter of the model's adaptive activation operator. This function is calculated based on four data points: the ratio of the longer to shorter side of the triangle, the correction coefficient, the state entropy, and the local point cloud density, to obtain an adjustment value, denoted as [value missing]. .when When the response slope is reduced, the parameters of the adaptive activation operator are adjusted to suppress noise interference; when When, keep the response slope parameter constant; when At the same time, the adaptive activation operator parameters are adjusted by increasing the response slope to enhance the contrast of boundary features.

[0053] The principle and technical effects of the high-order gradient flow and continuous state information transmission model are as follows: This model uses a multi-resolution pyramid tensor to uniformly represent the geometric features of triangles at different scales. It achieves deep fusion of cross-scale features through interlayer gradient flow optimization operators and long-span dense residual networks, enabling the model to consider both the local morphological details of minute tidal channels and the semantic information of the overall water area distribution. An adaptive activation operator dynamically reconstructs response parameters based on state entropy, ensuring stable discrimination capabilities even in complex boundary regions with uneven point cloud density and irregular noise distribution. A dynamic gating routing mechanism triggers local backtracking corrections for blurred edge regions, improving the positioning accuracy of narrow, winding waterway boundaries. The collaborative embedding of anisotropic diffusion tensors and the model's feature transmission process enhances the feature contrast at the land-water interface, comprehensively improving the geometric accuracy and boundary continuity of water area vector surface extraction.

[0054] The specific implementation of step S01 is as follows: First, an airborne LiDAR system is used to conduct aerial surveys of the target coastal area to acquire raw point cloud data containing three-dimensional coordinates and echo intensity information. The raw point cloud typically includes various categories such as ground points, vegetation points, building points, and areas of water without echoes. A progressive irregular triangulation filtering algorithm is applied to the raw point cloud to extract ground points. This algorithm iteratively determines the distance and angular deviation between a point and the current triangulation, incorporating points that meet the ground point criteria into the ground point set. Since the laser pulse is reflected by a specular surface in still water, it produces almost no effective echoes. Therefore, the still water area in the ground point cloud set naturally does not contain ground points, forming a point cloud void between the still water and the surrounding land surface points, providing a physical basis for subsequent water area identification.

[0055] The specific implementation of step S02 is as follows: Given that coastal point cloud data typically reaches tens of millions to hundreds of millions of vertices, loading all ground points into memory simultaneously for overall triangulation would lead to memory overflow and lock contention among multiple threads. This step employs a three-dimensional spatial grid block-based streaming incremental triangulation mechanism to address these issues. Specifically, the ground point cloud is divided into regular grid blocks according to three-dimensional spatial coordinates. Each block is configured with a dynamic circular buffer stack, and the block data is loaded into memory in batches using a streaming method. Parallel multi-threaded independent execution of Delaunay triangulation is achieved between blocks through lock-free atomic queues, eliminating lock contention among threads. After a single block completes triangulation, the corresponding memory is immediately released, maintaining memory usage at the data size of a single block. After all blocks are triangulated, the boundary triangles between adjacent blocks are stitched together to ensure the topological continuity of the irregular triangulation network, ultimately resulting in a set of triangles covering all ground points.

[0056] The specific implementation of step S03 is as follows: For each triangle in the triangle set, calculate its axis-aligned envelope rectangle and extract the length of the longer side of the envelope rectangle. With the length of the shorter side Due to the significant spatial heterogeneity of point cloud density at the land-water interface along the coast, directly using a fixed threshold for judgment is not feasible. Determining whether a triangle belongs to a narrow, elongated body of water can lead to numerous false positives in high-density areas and false negatives in low-density areas. Therefore, a local spatial lattice density adaptive correction operator is introduced: for each triangle, the median distance from each point in its local neighborhood to its nearest neighbor is calculated, denoted as [the median distance]. ; using the average nearest neighbor distance of global ground points as ,according to Calculate the correction factor ,in The calibration coefficient has a reference range of 0.3 to 0.8. The corrected narrow length discrimination condition is as follows: and Lower limit of the length of the long side Used to filter out extremely small, narrow triangles generated in high-density areas. , , , The parameters of the envelope rectangle were determined through iterative experiments using multiple sets of field-measured data from the target area. The extraction of these parameters employs a parallel geometric computing kernel based on a GPU vectorized instruction set, mapping the 3D vertex coordinates of triangles into a single-instruction, multiple-data-stream format. This multi-channel parallel processing simultaneously handles a large set of triangles, effectively alleviating the bottleneck of single-core computing.

[0057] The specific implementation of step S04 is as follows: The corrected triangular geometric parameters are input into the high-order gradient flow and continuous state information transmission model. The preprocessing module converts the irregular triangular mesh into a multi-resolution pyramid tensor. Each layer of the tensor contains four types of feature channels: the length of the long side of the envelope rectangle, the length of the short side, the correction coefficient, and the deflection angle of the direction vector. Different layers correspond to different spatial resolutions. The model backbone is composed of multiple stacked feature extraction layers. The adaptive activation operator within each layer reconstructs the response slope parameters in real time based on the state entropy of the feature map output from the preceding layer. The higher the state entropy, the greater the slope adjustment, enabling the model to maintain a stronger perception of complex boundary regions. The resolution layers are connected across scales through a long-span dense residual network. The inter-layer gradient flow optimization operator dynamically adjusts the fusion weights based on the gradient similarity between adjacent feature maps, effectively fusing shallow high-resolution detail features with deep low-resolution semantic features. The dynamic gated routing module selects the transmission path based on the edge pixel ambiguity score. When the score exceeds the ambiguity threshold, a high-resolution local backtracking sub-network is triggered to perform secondary correction on the features of that region. The model embeds an anisotropic diffusion tensor. Strong diffusion is applied in the low-gradient direction within the water area, while diffusion is suppressed in the high-gradient direction at the water-land boundary. The weighting coefficients of the convolution kernel are dynamically adjusted to enhance the feature contrast at the water-land interface. The confidence adjustment function calculates the adjustment value based on the ratio of the longer to shorter sides of the triangle, the correction coefficient, the state entropy, and the local point cloud density. ,when When the response slope is reduced, Time remains unchanged, when The response slope is increased to ensure the model maintains stable discrimination capability under different point cloud conditions. The confidence regression layer at the output end outputs the confidence value of each triangle belonging to the still water region.

[0058] The specific implementation of step S05 is as follows: Based on the confidence level of the still water region triangle output in step S04, and combined with the modified narrow length discrimination condition, the confidence level exceeding the confidence threshold and satisfying the condition is... and The triangles under the given conditions were initially identified as triangles in still water areas. For asymmetric twisted triangles appearing at the bends of tidal channels, principal component decomposition was performed on the vertex coordinate matrix of the triangles to extract the major axis direction vector. A nonlinear judgment surface was constructed by combining the ratio of the long side to the short side, the deflection angle of the direction vector, and the coefficient of variation, replacing the single threshold discrimination. Subsequently, an anisotropic filtering mechanism based on topological connectivity and spatial area accumulation constraints was used to remove isolated narrow triangle noise points: the geometric centroid distance and intersecting side length of the connected domains were calculated iteratively. Connected domains with a cumulative area value lower than the area threshold and a centroid distance exceeding the distance threshold were identified as isolated noise points and removed. The area threshold and distance threshold reference values ​​were determined by statistical analysis of multiple sets of measured point cloud data in the shallow water area of ​​the tidal flat. At the same time, the minimum spanning tree connectivity discrimination algorithm was used to post-process the confidence matrix, identifying connected components with fewer nodes than the number threshold as isolated misjudged regions and removing them, further ensuring the continuity and integrity of the water area vector surface.

[0059] The specific implementation of step S06 is as follows: Perform topological merging on the still water region triangles identified in step S05. Specifically, merge adjacent still water region triangles sharing edges into connected polygonal regions, eliminate shared boundaries within the triangles, retain the outer contour edges to form water boundary polylines, and then perform geometric simplification using the Douglas-Puk algorithm on the boundary polylines to reduce node redundancy in the output vector surface. Finally, output a continuous still water region vector surface layer. Each surface feature in the layer corresponds to an independent connected water body, and the surface feature attributes record statistical information such as the area, perimeter, and average confidence score of that water body for subsequent tidal flat geomorphological analysis.

[0060] It should be noted that the key technologies of this invention include: a local spatial lattice density adaptive correction operator, which dynamically generates correction coefficients by sensing the deviation between the median distance of the nearest neighbor points in the local neighborhood of a triangle and the global mean, and applies them to the long-short side ratio discriminant, so that the discrimination boundary adaptively drifts with the local point cloud density, fundamentally overcoming the systematic failure defect of the globally fixed threshold in regions of density heterogeneity; a high-order gradient flow and continuous state information transmission model, which uniformly represents the geometric features of different scales through a multi-resolution pyramid tensor, and achieves cross-scale deep fusion by combining a long-span dense residual network and an inter-layer gradient flow optimization operator, and an adaptive activation operator dynamically reconstructs response parameters based on state entropy, so that the model maintains stable discrimination capability in complex boundary regions; and a collaborative embedding of anisotropic diffusion tensor and model feature transmission process, which strengthens diffusion inside water bodies and suppresses diffusion at the water-land boundary, dynamically adjusts the convolution kernel weights, and enhances the feature contrast at the water-land interface. The three key technologies work together to ensure the density adaptability of input features through adaptive correction, enhance the semantic understanding depth through model fusion, and strengthen the contrast of boundary features through anisotropic diffusion. The combined effect of these three technologies enables water boundary extraction to maintain geometric accuracy and boundary continuity even in complex coastal landforms with significant point cloud density spatial heterogeneity.

[0061] It should be noted that in processing point cloud data in coastal tidal channel bends, the meandering direction of the tidal channels results in significant directional variations in the land-water boundary at the bends. Triangles spanning the water area often exhibit asymmetrical, distorted geometry in this region, with their major axes significantly deviating from the overall direction of the tidal channel. Under these conditions, traditional methods rely solely on the ratio of the long side to the short side for threshold discrimination, failing to perceive the spatial relationship between the triangle's major axis and the tidal channel's direction. This leads to numerous asymmetrically distorted triangles at the bends being misclassified due to their long-to-short side ratios approaching the critical value, resulting in a systematic segmentation bias in the water boundary at the bends. The reason for this technical problem is that the asymmetrical distortion of the water triangles at the bends causes their envelope rectangles to have a smaller long-to-short side ratio, overlapping with the ratio range of land-based, non-water triangles. A single threshold discrimination surface cannot effectively distinguish between water and non-water triangles within this overlapping range; that is, the discrimination feature space exhibits nonlinear aliasing in the bend region, which linear threshold classification methods cannot structurally handle. The usual solution to the aforementioned technical problems is to manually set multiple threshold combinations to adapt to tidal channel sections with different degrees of curvature. However, the curvature of tidal channels varies continuously in space, and there are blind spots between the preset discrete threshold combinations. Furthermore, the curvature of tidal channels differs significantly in different regions and under different tidal conditions, making it difficult for manual threshold combinations to generalize and resulting in classification errors in the transition areas of bends. This invention effectively solves this technical problem by performing principal component decomposition on the coordinate matrix of triangle vertices, extracting the direction vector of the major axis of the triangle, and combining it with the coefficient of variation of the direction of the major axis of the triangle along the tidal channel. The ratio of the long side to the short side, the deflection angle of the direction vector, and the coefficient of variation are used to construct a nonlinear decision surface, replacing the single threshold discrimination method based on the ratio of the long side to the short side. This nonlinear decision surface can describe the geometric distribution of water triangles in a three-dimensional feature space, and explicitly model the direction deflection information of asymmetric twisted triangles at bends, making water and non-water triangles that were originally mixed in the single-parameter threshold space separable again in the three-dimensional feature space. Meanwhile, the high-order gradient flow and continuous-state information transmission model fuses the local morphological features of the bend region with the semantic features of the overall tidal channel orientation across scales using a multi-resolution pyramid tensor. This allows the model to utilize the contextual information of the surrounding tidal channel orientation to assist in determining the attribution of triangles at the bend, further compensating for the insufficient discriminative power of local geometric features in the bend region. The synergistic effect of these two approaches systematically improves the segmentation accuracy of the water boundary at the bend, avoiding the systematic classification bias inherent in single-threshold discrimination methods in the bend region.

[0062] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, are used to perform the above-described method for extracting coastal water areas based on the triangular morphology of LiDAR point cloud TIN.

[0063] A third aspect of the present invention provides a coastal water area extraction system based on the triangular shape of LiDAR point cloud TIN, comprising the aforementioned computer-readable storage medium. The system can be any one of a computer, a server, or a microcontroller. The computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0064] Specifically, the principle of this invention is as follows: The reason this invention can solve the above-mentioned technical problems is that the local spatial lattice density adaptive correction operator fundamentally changes the way the discrimination threshold works. Traditional methods use a globally fixed threshold to determine whether a triangle belongs to a still water region. This threshold is reasonable when the global point cloud density is uniform, but in the land-water interface region where the point cloud density has significant spatial heterogeneity, there is a structural mismatch between the fixed threshold and the actual local density. This invention calculates the median distance between the nearest neighbors of the triangle in its local region. Distance from global average nearest neighbor The ratio is used to construct the correction coefficient. This correction coefficient can sense the local point cloud sparsity, appropriately relaxing the discrimination threshold in low-density regions and appropriately tightening the discrimination threshold in high-density regions, thus making the discriminant... It maintains a reasonable discrimination boundary under different density conditions. Simultaneously, a lower limit on the length of the long side of the envelope rectangle is introduced. The constraints filter out extremely small, narrow triangles generated in high-density areas, further eliminating systematic misjudgments introduced by density changes. Based on the adaptive correction of geometric parameters, this invention uses a high-order gradient flow and continuous-state information transmission model to process the corrected parameters. This model converts irregular triangular networks into multi-resolution pyramid tensors and achieves cross-scale feature fusion through a long-span dense residual network, enabling the model to simultaneously consider the local morphological details of subtle tidal channels and the semantic information of the overall water area distribution. The adaptive activation operator dynamically reconstructs response parameters based on state entropy, maintaining stable discrimination capabilities in complex boundary areas with uneven point cloud density and irregular noise distribution. The dynamic gated routing module triggers a local backtracking subnetwork for secondary correction in blurred edge areas, improving the positioning accuracy of narrow and winding waterway boundaries. The embedded anisotropic diffusion tensor distinguishes between the interior of the water area and the water-land boundary based on the local elevation gradient, implementing strong diffusion within the water area and suppressing diffusion at the water-land boundary, dynamically adjusting the weighting coefficients of the convolution kernel to enhance the feature contrast at the water-land interface. Finally, the geometric continuity of the output water vector surface is ensured through noise removal from isolated elongated triangles and post-processing based on minimum spanning tree connectivity. The above steps have a strict logical progression: adaptive correction precedes this, ensuring that the geometric parameters input to the model reflect the true local density conditions; model fusion follows, utilizing multi-scale semantic information to compensate for the shortcomings of simple geometric discrimination in complex terrains; and post-processing removes isolated noise, ensuring the topological integrity of the final result. These three steps work synergistically to constitute the complete technical logic of this invention, enabling stable extraction of still water areas under complex coastal terrain conditions.

[0065] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0066] The specific implementation of step S01 is as follows: acquire airborne lidar point cloud data of the target area, use a progressively encrypted irregular triangular mesh filtering algorithm to classify ground points in the original point cloud, iteratively remove non-ground points such as vegetation and buildings, and obtain a ground point cloud set. In still water areas, the laser pulse is absorbed by the water or undergoes specular reflection, resulting in missing echo signals. Therefore, the still water areas in the ground point cloud set do not contain effective surface points, forming a point cloud blank area.

[0067] The specific implementation of step S02 is as follows: the ground point cloud set is divided into several grid blocks according to the three-dimensional spatial location, let the first... The point set of each block is ,in For the first The number of points within each block , , The first The three-dimensional coordinates of each point are determined. Each block is configured with a dynamic circular buffer stack, which loads data into memory in batches using a streaming method and performs incremental Delaunay triangulation. Lock-free atomic queues are used between blocks to achieve parallel, multi-threaded, and independent network construction. After triangulation within a block, triangles at the boundaries of adjacent blocks are stitched together to eliminate topological discontinuities at these boundaries, ultimately resulting in a set of triangles covering all ground points. , This represents the total number of triangles.

[0068] The specific implementation of step S03 is as follows: for the triangle set Each triangle in Calculate the length of the longer side of its axis-aligned envelope rectangle. With the length of the shorter side The units are all in meters, and a local spatial lattice density adaptive correction operator is introduced to correct the ratio of the long side to the short side. It is a triangle The median distance between nearest neighbors within a local area is obtained by sorting all ground points within the local area and taking the median distance between them, and the unit is meters (m). The average nearest neighbor distance of the point cloud in the target area is obtained by averaging the nearest neighbor distances of the ground point cloud across the entire area, and the unit is meters (m). Correction coefficient. The formula for calculation is:

[0069] ;

[0070] In the formula, The calibration coefficient, ranging from 0.3 to 0.8, was determined through iterative experiments using multiple sets of field-measured data from the target area. This is a dimensionless correction coefficient. The corrected narrow-length discriminant is:

[0071] ;

[0072] In the formula, The threshold for narrow-length discrimination is dimensionless and determined by comparing the degree of agreement between the identification results and the measured waterline. The discriminant formula also requires... , This is the filtering threshold for the longest side, in meters, used to filter out extremely small, narrow triangular noise points. The value was determined through multiple iterative experiments.

[0073] The specific implementation of step S04 is as follows: A high-order gradient flow and continuous-state information transmission model is used to process the corrected triangular geometric parameters, outputting the confidence score of each triangle belonging to the still water region. The geometric parameters corresponding to the set of triangles are organized into a multi-resolution pyramid tensor. Layer tensor corresponding to resolution layer The characteristic matrix below ,in For the first The number of triangles in the layer, This represents the number of feature channels, with each channel corresponding sequentially. , , and direction vector deflection angle , The unit is rad. The response slope of each adaptive activation operator in the model backbone. The state entropy of the feature map output by the preceding layer Real-time reconstruction, state entropy is achieved by normalizing the pixel values ​​of the output feature map of the preceding layer and then statistically analyzing the discrete probability distribution. Then, taking the Shannon entropy, we get:

[0074] ;

[0075] In the formula, For the first The state entropy of the layer feature map is dimensionless. For the first The first layer feature map The normalized probability of a discrete state, dimensionless. It is the natural logarithm. The higher the value, the greater the adjustment range of the response slope of the adaptive activation operator. The confidence adjustment function calculates the adjustment value based on four data points: the ratio of the longer side to the shorter side of the triangle, the correction coefficient, the state entropy, and the local point cloud density. :

[0076] ;

[0077] In the formula, It is a triangle Point cloud density of the local area, in units of This is obtained by statistically analyzing the ratio of the number of points to the area within a local region. It is a non-linear mapping function, obtained by fitting the training data, and the output is... It is a dimensionless scalar. When When, reduce the response slope ;when At that time, keep Unchanged; when At that time, improve The interlayer gradient flow optimization operator calculates the fusion weights based on the gradient similarity between the feature maps of two adjacent layers. :

[0078] ;

[0079] In the formula, For the first The gradient of the layer feature matrix, For the first The gradient of the layer feature matrix, For inner product operations, It is the Euclidean norm. These are dimensionless fusion weights; larger values ​​indicate greater similarity in features between adjacent layers and a higher degree of fusion. Anisotropic diffusion tensor. The construction formula for embedding hidden layer feature propagation is as follows:

[0080] ;

[0081] In the formula, It is a triangle The elevation gradient vector of the local area, in units of That is, dimensionless Its Euclidean norm Let be the diffusion intensity function, which monotonically decreases with increasing gradient, and its expression is:

[0082] ;

[0083] In the formula, The diffusion control coefficient is dimensionless and is determined through statistical analysis of measured elevation data of tidal flats. It is typically taken as 1 to 3 times the standard deviation of the regional elevation gradient. for identity matrix It is a dimensionless diffusion tensor; in the low gradient region inside the water body. When the value approaches 1, strong diffusion occurs; in high gradient regions at the land-water boundary. A value approaching 0 suppresses diffusion and enhances boundary contrast. For triangles... vertex coordinate matrix Principal component decomposition is performed to obtain the major axis direction vector. Direction vector deflection angle Defined as The angle between the reference direction and the direction vector. Combined with the coefficient of variation of the direction vector along the tidal channel direction. (Dimensionless, the ratio of standard deviation to mean), construct a nonlinear decision surface. :

[0084] ;

[0085] In the formula, It is a three-parameter nonlinear decision function, obtained by fitting the training samples. This is a dimensionless decision value, used to replace the single long-to-short side ratio threshold for discrimination, improving the accuracy of curve region segmentation. Model loss function. From the classification cross entropy term Anisotropy penalty term compared to boundary perimeter-area ratio Weighted composition:

[0086] ;

[0087] In the formula, This is the weighting coefficient for the penalty term, dimensionless, typically ranging from 0.1 to 0.5, and determined by parameter tuning of the validation set. and All are dimensionless loss values A penalty is applied to the perimeter-to-area ratio of the predicted water boundary to constrain the geometric regularity of the boundary and suppress jagged false positives. The model output, via a confidence regression layer, outputs the confidence score for each triangle belonging to the still water region. The training dataset is dimensionless. It selects multiple typical coastal areas, collects airborne lidar point clouds at different tidal levels, constructs an irregular triangular network, calculates the geometric morphological parameter matrix, and simultaneously collects corresponding temporal measured waterline data as annotation basis. Training sample pairs are formed by matching spatial locations, and the dataset is divided into training, validation, and test sets according to a preset ratio. The training process uses the backpropagation algorithm to update the parameters of each layer of the model. and Synchronous iterative updates are performed to monitor convergence on the validation set. Training stops once the preset number of iterations or the loss function converges.

[0088] The specific implementation of step S05 is as follows: based on the confidence level Combined with narrow length discrimination threshold The method identifies triangles in still water regions. For the identified triangles, an anisotropic filtering mechanism based on topological connectivity and spatial area accumulation constraints is used to remove isolated, narrow, noisy triangles. Triangles that meet the confidence threshold are abstracted as graph nodes, and shared edges between adjacent triangles are abstracted as graph edges. The edge weight is taken as the difference in confidence scores between adjacent triangles. After constructing a minimum spanning tree, the number of nodes, geometric centroid distance, and area accumulation of each connected component are calculated. The area threshold is set as follows: Distance threshold is The cumulative area of ​​connected components is lower than And the distance between the centroids exceeds Connected components are identified as isolated noise and are removed. and This was determined through statistical analysis of multiple sets of measured point cloud data from the shallow water area of ​​the tidal flats.

[0089] The specific implementation of step S06 is as follows: spatial merging processing is performed on the identified still water area triangles, the shared edges of adjacent triangles are dissolved, a continuous still water area vector surface layer is generated, and the output is a vector surface format file as the coastal water area extraction result.

[0090] To better understand and implement the present invention, the following is a specific application scenario of the present invention, Example 2: This example selects a typical tidal flat area in a coastal zone as the test scenario. The landform types in this area include tidal channels, tidal flat water bodies and coastal wetland waterways. The terrain is complex and the point cloud density at the land-water interface has significant spatial heterogeneity, which is a typical working condition for verifying the applicability of the present method.

[0091] Airborne LiDAR aerial surveying employed a line-scan method to achieve full coverage acquisition of the target area. The flight altitude was approximately 500 m, using a laser pulse repetition frequency of 400 kHz, a scan angle of ±20°, and a flight strip overlap rate of 30%. After flight strip stitching and coordinate transformation, the original point cloud contained approximately 120 million points in the target area, with an average point density of approximately 12 points / ... However, in the area where tidal creek slopes intersect with vegetation cover, the local density variation ranges from 3 to 18 points / year. .

[0092] In step S01, a progressive irregular triangular mesh filtering algorithm was used to extract ground points from the original point cloud, resulting in approximately 78 million ground points. Due to the lack of effective echoes from the laser specular reflection in the tidal channels and tidal flats, obvious point cloud holes were formed in the ground point cloud set, with the total area of ​​point cloud holes accounting for approximately 28% of the survey area.

[0093] In step S02, the ground point cloud is divided into regular grid blocks with sides of 100 m × 100 m according to 3D spatial coordinates, generating approximately 960 blocks, each containing an average of about 81,000 ground points. A streaming incremental partitioning mechanism for the 3D spatial grid blocks is employed, with each block undergoing Delaunay triangulation in parallel via a lock-free atomic queue. The peak memory usage per block is approximately 480 MB, and overall memory usage is controlled within the size of a single block, without memory overflow. After boundary stitching, a set of approximately 155 million triangles is obtained.

[0094] Execute step S03 to calculate the length of the longer side of the envelope rectangle for all triangles. With the length of the shorter side The calculation process employs a parallel geometry computation kernel based on the GPU's vectorized instruction set, using a single-instruction multiple-data (SMD) parallel processing method. Extracting the geometric parameters of 155 million triangles took approximately 42 seconds. Subsequently, the median distance to the local nearest neighbors of each triangle was calculated. Global average nearest neighbor distance Take 0.29 m as the calibration coefficient. After iterative experiments, the value was set to 0.55, and then... Calculate the correction coefficients for each triangle. In high-density areas of tidal channel banks, Approximately 0.17m, corresponding to Approximately 0.78; in low-density areas of tidal flats, Approximately 0.48 m, corresponding to Approximately 1.37. After correction, the effective discrimination boundaries of different density regions tend to be consistent spatially, no longer exhibiting systematic shifts due to changes in point cloud density. Narrow-length discrimination threshold. The experiment yielded a value of 4.5, which is the lower limit for the length of the longer side. The value is 0.35 m. The distribution of correction coefficients under different local point cloud density conditions is shown in Table 1:

[0095] Table 1. Statistical table of correction coefficients corresponding to different median local nearest neighbor distances.

[0096]

[0097] As shown in Table 1, as the median local nearest neighbor distance increases from 0.12 m to 0.48 m, the correction factor... The monotonically increasing value from 0.67 to 1.37 reflects the adaptive response mechanism of the local spatial lattice density adaptive correction operator to the spatial heterogeneity of point cloud density: in dense point cloud regions, the correction coefficient is less than 1, which tightens the overall effective discrimination threshold and suppresses the misjudgment of extremely small and narrow triangles; in sparse point cloud regions, the correction coefficient is greater than 1, which moderately relaxes the effective discrimination threshold and avoids the under-judgment of real water triangles due to low density, thus solving the technical defect of the fixed threshold systematically failing in density heterogeneous regions.

[0098] Step S04 involves inputting the corrected triangular geometric parameters into the high-order gradient flow and continuous-state information transmission model. The preprocessing module converts the parameter matrix into a 3-layer multi-resolution pyramid tensor, with each layer's feature map resolution ratio at 1:2:4, including four feature channels: long side length, short side length, correction coefficient, and direction vector deflection angle. The model backbone consists of 12 stacked feature extraction layers, and the adaptive activation operator dynamically adjusts the response slope parameter based on the state entropy of each layer. The inter-layer gradient flow optimization operator dynamically adjusts the cross-layer fusion weights of the long-span dense residual network based on the gradient similarity of adjacent layer feature maps. The dynamic gated routing module triggers a high-resolution local backtracking sub-network in areas where the edge pixel ambiguity score exceeds the ambiguity threshold, performing secondary correction on the features of blurred boundary areas such as tidal channel slopes. Anisotropic diffusion tensor... Strong diffusion is implemented in low-gradient regions within the water area, while diffusion is suppressed in high-gradient regions at the water-land boundary, thus enhancing the feature contrast of the model's convolutional kernel at the water-land interface. The confidence regression layer outputs the confidence score value for each triangle belonging to the still water region. The distribution of the number of triangles in each confidence interval of the model output is shown in Table 2.

[0099] Table 2. Statistical distribution of confidence intervals in the triangle of still water region from the model output.

[0100]

[0101] As shown in Table 2, the confidence distribution of the model output exhibits a clear polarization trend. Triangles with confidence levels in the range [0, 0.2] account for 70.9% of the total, triangles with confidence levels in the range [0.8, 1.0] account for 14.4%, and triangles in the intermediate range account for approximately 14.7%. This distribution trend indicates that the synergistic effect of the anisotropic diffusion tensor and the adaptive activation operator effectively enhances the feature contrast at the water-land interface, enabling the model to make a deterministic classification judgment for the vast majority of triangles. Only a small number of triangles near the true boundary have confidence levels in the intermediate range, indicating that the overall feature discrimination ability of the model is at a reasonable level.

[0102] Step S05 is executed, using a confidence threshold of 0.65 as the discrimination criterion, combined with the corrected narrow-length discrimination criterion, to initially identify triangles in the still water area, resulting in approximately 23.1 million candidate triangles. For asymmetric twisted triangles in the tidal channel bends, principal component decomposition is used to extract the major axis direction vector, and a three-parameter nonlinear decision surface is constructed using the coefficient of variation for correction, adjusting the classification results of approximately 180,000 triangles at bends. Subsequently, an anisotropic filtering mechanism is used to remove noise from isolated narrow-length triangles, with an area threshold set to... With a distance threshold of 8 meters, approximately 3700 isolated noise nodes and connected components were removed, involving about 410,000 triangles. For example... Figure 2 As shown, the spatial distribution of candidate triangles changes significantly before and after the removal of isolated noise points. After removal, the boundaries of the connected domains in the tidal channel and tidal flat water areas are more regular, and the discrete noise points are significantly reduced.

[0103] Step S06 involves topological merging of approximately 22.69 million triangles in the still water area to eliminate shared internal edges. After extracting the outer contour edges, the Douglas-Pock algorithm is used for simplification, with a simplification distance threshold of 0.15 m. This results in a final vector surface layer of the still water area, containing 1823 water surface features. The total water area represents approximately 27.6% of the survey area, which is roughly consistent with the 28% point cloud void area ratio in step S01, with a difference of approximately 1.4 percentage points. This difference stems from the presence of a small number of valid echo ground points in some extremely shallow water areas, representing an inherent error at the methodological level, within acceptable engineering limits. Figure 3 and Figure 4 As shown in the figure, the vector surface extraction results of the still water area and the overlay display of the irregular triangular network show that the boundary positioning matches the measured water edge line on the ground well, and no systematic offset is observed.

[0104] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for extracting coastal water zones based on the triangular morphology of LiDAR point cloud TIN, characterized in that, Includes the following steps: Acquire airborne LiDAR point cloud data of the target area, perform classification processing to extract the ground point cloud set, and the still water area in the ground point cloud set does not contain ground points; Based on the three-dimensional spatial grid block streaming incremental partitioning mechanism, an irregular triangular network is constructed on the ground point cloud set to obtain a triangular set covering all ground points. Calculate the length of the long side and the length of the short side of the envelope rectangle of each triangle in the triangle set, and introduce a local spatial lattice density adaptive correction operator to correct the ratio of the long side to the short side; A high-order gradient flow and continuous state information transmission model is used to process the corrected triangle geometric parameters and output the confidence score of the triangle in the still water region. Based on the confidence score of triangles in still water areas and combined with the narrow length discrimination threshold, triangles in still water areas are identified, and isolated narrow triangle noise points are removed. The identified triangles in the still water area are merged to generate a continuous vector surface layer of the still water area.

2. The method for extracting coastal water areas based on the triangular morphology of LiDAR point cloud TIN according to claim 1, characterized in that, The three-dimensional spatial mesh block streaming incremental partitioning mechanism specifically divides the ground point cloud into multiple mesh blocks according to the three-dimensional spatial location. Each block is loaded into memory in batches in a streaming manner and incremental triangulation is performed. A dynamic circular buffer stack is set up inside the block to temporarily store the point cloud data to be processed. Lock-free atomic queues are used between blocks to achieve parallel multi-threaded independent network construction. After the network construction of each block is completed, boundary stitching is performed.

3. The method for extracting coastal water areas based on the triangular morphology of LiDAR point cloud TIN according to claim 2, characterized in that, The method for eliminating isolated, narrow triangular noise points specifically employs an anisotropic filtering mechanism based on topological connectivity and spatial area accumulation constraints. By iteratively calculating the geometric centroid distance and intersecting side length of the connected triangular regions, connected regions with a cumulative area value lower than the area threshold and a centroid distance exceeding the distance threshold are identified as isolated noise points and eliminated.

4. The method for extracting coastal water areas based on the triangular morphology of LiDAR point cloud TIN according to claim 3, characterized in that, The input end of the high-order gradient flow and continuous state information transmission model is equipped with a pre-processing module, which converts the irregular triangular mesh into a multi-resolution pyramid tensor. Each layer of the tensor corresponds to a triangle geometric parameter matrix under different spatial resolutions. The parameter matrix includes four types of feature channels: the length of the long side of the envelope rectangle, the length of the short side, the correction coefficient, and the deflection angle of the direction vector.

5. The method for extracting coastal water areas based on the triangular morphology of LiDAR point cloud TIN according to claim 4, characterized in that, The backbone of the high-order gradient flow and continuous state information transmission model consists of multiple stacked feature extraction layers. Each layer contains multiple adaptive activation operators. The parameters of the adaptive activation operators are reconstructed in real time by the state entropy of the output feature map of the preceding layer. Each resolution feature layer is connected through a long-span dense residual network, and the cross-layer connection weights are dynamically adjusted by the inter-layer gradient flow optimization operator.

6. The method for extracting coastal water areas based on the triangular morphology of LiDAR point cloud TIN according to claim 5, characterized in that, The high-order gradient flow and continuous state information transmission model is equipped with a dynamic gated routing module, which selects the signal transmission path based on the edge pixel ambiguity score of the current feature map. When the edge pixel ambiguity score exceeds the ambiguity threshold, it jumps to the high-resolution local backtracking sub-network to perform secondary correction of the features in that area before merging into the backbone network.

7. The method for extracting coastal water areas based on the triangular morphology of LiDAR point cloud TIN according to claim 6, characterized in that, The loss function of the high-order gradient flow and continuous state information transmission model is composed of a weighted sum of a classification cross-entropy term and a boundary perimeter-area ratio anisotropy penalty term. The boundary perimeter-area ratio anisotropy penalty term is used to constrain the geometric regularity of the predicted water boundary.

8. The method for extracting coastal water areas based on the triangular morphology of LiDAR point cloud TIN according to claim 7, characterized in that, The high-order gradient flow and continuous state information transmission model nests a minimum spanning tree connectivity discrimination algorithm. Triangles with confidence scores exceeding the confidence score threshold are abstracted as graph nodes, and adjacent triangles with shared edges are abstracted as graph edges. The edge weight is taken as the difference in confidence scores between adjacent triangles. After constructing the minimum spanning tree, the size of the connected component is determined based on the number of nodes in the connected component. Connected components with a size lower than the number threshold are determined as isolated misjudged regions and are removed.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform the coastal water area extraction method based on the LiDAR point cloud TIN triangle morphology as described in any one of claims 1-8.

10. A coastal water zone extraction system based on the triangular morphology of LiDAR point cloud TIN, characterized in that, The system comprises the computer-readable storage medium of claim 9, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

Images