Landscape architecture digital earthwork construction method based on remote sensing technology

Abstract

The application discloses a landscape digital earthwork construction method based on remote sensing technology, comprising the following steps: acquiring design terrain data and construction control data, collecting multi-temporal remote sensing images, remote sensing point clouds and camera poses, and generating remote sensing terrain data; constructing a construction terrain RegSDF model; based on the remote sensing terrain data, constructing a permanent ground SDF and a temporary cover SDF; constructing a design ground SDF according to the design terrain data, determining a current ground position, and generating a current construction ground SDF; generating a directional curvature component based on the current construction ground SDF, and correcting to obtain a corrected construction ground SDF; generating a digital construction scheme and executing according to the corrected construction ground SDF, the design ground SDF and the construction control data; collecting re-measurement data, generating a re-measurement permanent ground SDF and a construction deviation, and updating the digital construction scheme. The application improves the ground recognition, excavation and filling calculation and scheme updating precision.

Description

Technical Field

[0001] This invention relates to the field of landscape architecture construction technology, and in particular to a digital earthwork construction method for landscape architecture based on remote sensing technology. Background Technology

[0002] As landscape architecture projects move towards digital construction, earthwork construction needs to simultaneously meet requirements for design elevation, slope continuity, drainage organization, garden path connection, and phased construction control. Existing construction methods rely heavily on manual surveying, point cloud elevation comparison, and two-dimensional cross-section verification. These methods lack the ability to identify temporary soil piles, vegetation obstructions, and construction equipment obstructions within the construction area, and are prone to mistaking non-realistic surfaces for workable surfaces.

[0003] In complex garden terrain, the relationship between the designed terrain and the existing terrain is not a simple elevation difference. Existing methods usually only rely on local elevation differences to determine the excavation and filling volume, which makes it difficult to express the spatial relationship between the existing surface, the designed surface, and the surface under construction at different stages. This results in discontinuous excavation and filling profiles and insufficient accuracy in calculating excavation boundaries, filling boundaries, and remaining construction thickness.

[0004] In addition, existing digital earthwork construction schemes have weak constraints on slope top lines, slope toe lines, drainage boundaries, and garden path boundaries. The retest results are usually only used for acceptance records, making it difficult to write back the construction model and revise subsequent construction schemes in a timely manner. This results in insufficient connection between earthwork transportation, layered filling, slope leveling, and compaction, affecting the refined control of earthwork construction in landscape gardens.

[0005] Therefore, how to provide a digital earthwork construction method for landscape architecture based on remote sensing technology is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a digital earthwork construction method for landscape architecture based on remote sensing technology. This invention constructs a RegSDF model of the construction terrain by acquiring multi-temporal remote sensing images, remote sensing point clouds, and camera poses. It then uses surface diversion units to distinguish between permanent surfaces and temporary covers, generates cut-fill profiles through dual zero-point profile units, corrects the current construction surface SDF through curvature transfer units, and generates a digital construction plan that is iteratively updated based on re-measurement results. This method has the advantages of accurate surface identification, continuous cut-fill calculation, dynamic correction of the construction plan, and fine boundary control.

[0007] The digital earthwork construction method for landscape architecture based on remote sensing technology according to an embodiment of the present invention includes the following steps:

[0008] Acquire design terrain data and construction control data, collect multi-temporal remote sensing images, remote sensing point clouds and camera poses, unify spatial coordinates and divide planar grids to generate remote sensing terrain data;

[0009] Construct a RegSDF model of the construction terrain, including an SDF geometric network, surface diversion units, dual zero-point profile units, and curvature transfer units;

[0010] Based on the SDF geometric network, remote sensing terrain data is processed to generate permanent surface attribution values ​​and temporary cover attribution values, and surface diversion units are invoked to construct permanent surface SDF and temporary cover SDF;

[0011] In the dual zero-point profile unit, the design surface SDF is constructed based on the design terrain data. Combined with the permanent surface SDF, the current zero point and the design zero point are generated, and the current surface location is determined to generate the current construction surface SDF.

[0012] Based on the current construction surface SDF, directional curvature components are generated. In the curvature transfer unit, the components are transferred according to the design terrain data. The current construction surface SDF is corrected by combining the temporary cover assignment value, the temporary cover SDF, and the transferred directional curvature components, and the corrected construction surface SDF is generated.

[0013] Based on the corrected construction surface SDF, design surface SDF, and construction control data, a digital construction plan is generated, and construction equipment is controlled to perform construction.

[0014] Collect remeasured remote sensing images, remeasured remote sensing point clouds, and remeasured camera poses to generate remeasured permanent surface SDF and construction deviations, and update the digital construction plan based on the construction deviations.

[0015] Optionally, the step of acquiring design terrain data and construction control data, and generating remote sensing terrain data, includes:

[0016] Obtain the design elevation, design contour lines, slope top line, slope toe line, drainage boundary, and garden path boundary, and combine them to generate design terrain data;

[0017] The construction stage, completed depth, operating width of construction equipment, single-layer filling thickness, and allowable elevation deviation are obtained and combined to generate construction control data.

[0018] Collect multi-temporal remote sensing images, remote sensing point clouds, and camera poses of the landscape construction area, unify spatial coordinates, and divide the construction area into a planar grid;

[0019] The locations of remote sensing points falling into the same planar grid in different acquisition phases are correlated. The normal is calculated based on the distribution of neighboring points of each remote sensing point. The phase duration and the number of repeated observations are determined according to the continuous occurrence of the correlated remote sensing points in different acquisition phases.

[0020] Remote sensing topographic data is generated by combining multi-temporal remote sensing images, remote sensing point clouds, camera pose, planar grid, normal, temporal persistence, and number of repeated observations.

[0021] Optionally, the construction of the RegSDF model of the construction terrain includes:

[0022] The SDF geometric network and surface optical field network in the RegSDF model are retained;

[0023] The spherical tracing, differentiable surface intersection calculation, Hessian regularization, and minimum surface regularization in the RegSDF model are retained and configured as spherical tracing structure, differentiable surface intersection structure, Hessian matrix calculation structure, and minimum surface regularization structure;

[0024] Following the SDF geometric network, a new surface divergence unit is added, with permanent surface branches and temporary cover branches set in parallel. The permanent surface branches and temporary cover branches generate distance fields respectively and are connected to the surface light field network to form a surface divergence path.

[0025] After the spherical tracking structure and the intersection structure of the differentiable surface, a double zero-point profile unit is added, a design surface branch is set, the design terrain data is used as the processing object of the design surface branch, and a double zero-point profile path is constructed to generate the current zero point and the design zero point in sequence along the gravity direction.

[0026] Between the Hessian matrix calculation structure and the minimum surface regularization structure, a curvature transfer unit is added to construct a curvature transfer path for the directional curvature components to be transferred between adjacent planar meshes.

[0027] Improve the connection relationships between various structures and construct a RegSDF model of the construction terrain.

[0028] Optionally, the construction of the permanent surface SDF and the temporary cover SDF includes:

[0029] The coordinates of remote sensing points in the remote sensing terrain data are input into the SDF geometric network for spatial encoding, generating the distance value and spatial location features corresponding to each remote sensing point.

[0030] In the surface subdivision unit, the elevation difference and elevation continuity of adjacent time-phase remote sensing points within the same planar grid are calculated, the normal consistency of the normal angle between associated remote sensing points is calculated, the same remote sensing point is projected onto multi-temporal remote sensing images according to the camera pose, and the color reprojection difference is calculated.

[0031] The duration of the time phase and the number of repeated observations were normalized.

[0032] Based on elevation continuity, normal consistency, color reprojection difference, normalized value of temporal persistence, and normalized value of repeated observations, permanent surface attribution value and temporary cover attribution value are calculated and generated.

[0033] The permanent surface attribution value and the temporary cover attribution value are used as the update weights for the permanent surface branch and the temporary cover branch, respectively, to update the distance field of the permanent surface branch and the distance field of the temporary cover branch.

[0034] The spherical tracking structure is invoked to search for candidate intersection points in the distance fields of the permanent surface branch and the temporary cover branch along the observation ray defined by the camera pose. The corrected intersection points are determined and the corrected intersection points and their corresponding spatial location features are input into the surface light field network for remote sensing image fitting.

[0035] The fitted permanent surface branch distance field is determined as the permanent surface SDF, and the fitted temporary cover branch distance field is determined as the temporary cover SDF.

[0036] Optionally, generating the current construction surface SDF includes:

[0037] In the design surface branch, the design elevation is used as the elevation control point, the design contour line is discretized into elevation points, the elevation control points and elevation points are connected to generate the design surface, the linear boundary is preserved, and the design surface SDF is constructed.

[0038] In a unified spatial coordinate system, the direction of gravity is determined. In the dual zero-point profile unit, the plane coordinates of the center of each plane grid are used as the starting point. The spherical tracking structure is called to query the candidate zero intersection points of the permanent surface SDF and the design surface SDF along the direction of gravity. The candidate zero intersection point with the smallest distance from the plane center of the plane grid is selected. The differentiable surface intersection structure is called to generate the current zero point and the design zero point.

[0039] Calculate the vertical coordinate difference between the design zero point and the existing zero point, determine the cut-fill mark, determine the absolute value of the vertical coordinate difference as the construction thickness, and combine the plane grid, the existing zero point, the design zero point, the cut-fill mark and the construction thickness to generate the cut-fill profile;

[0040] Compare the completed depth with the construction thickness to determine the current surface location and generate the current construction surface SDF.

[0041] Optionally, generating the modified construction surface SDF includes:

[0042] The Hessian matrix is ​​called to calculate the current construction surface SDF and generate the Hessian matrix corresponding to each planar grid.

[0043] Based on the design surface SDF, calculate the surface normal, the direction of maximum slope, and the contour line direction;

[0044] Using the contour line direction, the maximum slope direction, and the surface normal as the projection directions of the Hessian matrix, respectively, calculate the curvature components of the contour line direction, the curvature components of the maximum slope direction, and the curvature components of the surface normal.

[0045] Based on the temporary cover attribution value and the permanent surface attribution value, the plane grid to be completed is determined, and the plane grid from the curvature source is selected;

[0046] Based on the center line connecting the curvature source planar grid and the to be completed full planar grid, the contour line direction, the maximum slope direction, the center distance, and the permanent surface attribution value, calculate and generate the contour line transfer weight and the slope transfer weight;

[0047] Based on the cut and fill markings, the cut area and fill area are determined, and the curvature component in the direction of maximum slope is transferred according to the slope transfer weight.

[0048] According to the contour line transfer weight, the contour line direction curvature component is transferred along the contour line direction, the surface normal curvature component is retained, and the direction curvature difference between the transferred direction curvature component and the corresponding direction curvature component of the to be filled full-plane grid is calculated.

[0049] The minimum surface regularization structure is invoked to perform minimum surface regularization in the permanent surface branch. The corresponding parameters of the current construction surface SDF are updated based on the directional curvature difference and the minimum surface regularization result, and the corrected construction surface SDF is generated.

[0050] Optionally, generating a digital construction plan and controlling construction equipment to perform construction includes:

[0051] Using the center of each planar grid as the search location, the spherical tracing structure is invoked to query the distance values ​​of the corrected construction surface SDF and the design surface SDF along the gravity direction to determine the construction zero intersection point and the design zero intersection point;

[0052] Based on the construction zero intersection point and the design zero intersection point, determine the remaining construction thickness and calculate the total remaining excavation volume and the total remaining fill volume;

[0053] Select one planar grid from the undivided planar grid to construct the continuous working range and determine the excavation boundary and the filling boundary;

[0054] The remaining earthwork volume of each continuous operation range is accumulated to obtain the remaining excavation volume and the remaining fill volume of the range, and the earthwork transportation volume is determined.

[0055] Calculate the number of filling layers and the thickness of each layer based on the remaining construction thickness and the thickness of the single-layer filling;

[0056] The excavation depth is determined based on the remaining construction thickness, and the vertical coordinates of the zero intersection points of each planar grid are determined as the overall slope elevation to determine the compaction sequence.

[0057] By combining excavation boundaries, excavation depths, fill boundaries, layered fill thicknesses, earthwork transport volumes, slope elevations, and compaction sequences, a digital construction plan is generated.

[0058] According to the digital construction plan, the data is sent to the construction equipment, and the equipment is controlled to carry out earthwork construction.

[0059] Optionally, the generation of remeasured permanent surface SDF and construction deviation, and the updating of the digital construction scheme based on the construction deviation, includes:

[0060] After completing the current construction phase, collect the remeasured remote sensing images, remeasured remote sensing point clouds, and remeasured camera poses. Recalculate the normal, elevation continuity, normal consistency, color reprojection difference, temporal continuity, and number of repeated observations of the remeasured remote sensing points, and update the remote sensing topographic data.

[0061] The SDF geometric network is called to process the coordinates of the remeasured remote sensing points. Based on the surface subtraction unit, the remeasured permanent surface attribution value and the remeasured temporary cover attribution value are generated. Combined with the spherical tracking structure, the differentiable surface intersection structure and the surface light field network, the remeasured permanent surface SDF is generated.

[0062] The spherical tracking structure is invoked to generate the re-measurement zero point and the stage target zero point along the gravity direction, and the construction deviation is calculated and generated.

[0063] When the construction deviation is greater than the allowable elevation deviation, the original values ​​of the same plane grid are replaced by the re-measured permanent surface value and the re-measured temporary cover value, and the vertical distance from the current zero point to the re-measured zero point is updated to the completed depth.

[0064] Based on the updated permanent surface attribution values, temporary cover attribution values, and completed depth, the cut-fill profile and the current construction surface SDF are regenerated, and the digital construction plan is regenerated.

[0065] The beneficial effects of this invention are:

[0066] First, this invention separates permanent surface branches and temporary cover branches by using surface diversion units in the construction terrain RegSDF model. This reduces the interference of vegetation, temporary soil mounds, and construction equipment on the identification of the real construction surface, avoids mistaking temporary cover surfaces as the basis for excavation, filling, and slope trimming, improves the accuracy of permanent surface SDF generation, and makes subsequent excavation and filling judgments more stable.

[0067] Secondly, this invention searches for the zero intersection points of the permanent surface SDF and the design surface SDF respectively through dual zero-point profile units, and generates cut-fill profiles based on the vertical positional relationship between the current zero point and the design zero point. This can unify the design terrain, the current terrain and the stage construction progress into the same spatial expression, reducing the breakage of cut-fill boundaries, construction thickness errors and earthwork volume calculation deviations caused by simply relying on local elevation differences.

[0068] Furthermore, this invention corrects the current construction surface SDF by using curvature transfer units and minimum surface regularization structures, and generates a digital construction plan based on the corrected construction surface SDF. This allows the invention to maintain terrain constraints at drainage boundaries, slope top lines, slope toe lines, and garden path boundaries, improve the continuity of construction surface replenishment below temporary coverage areas, and form a repeatable and updatable closed-loop control between excavation, earthwork transportation, layered filling, slope leveling, and compaction. Attached Figure Description

[0069] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0070] Figure 1 This is an overall flowchart of a digital earthwork construction method for landscape architecture based on remote sensing technology proposed in this invention;

[0071] Figure 2 This is a schematic diagram of the structure and processing flow of the RegSDF model for construction terrain in this invention. Detailed Implementation

[0072] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0073] refer to Figures 1-2 The digital earthwork construction method for landscape architecture based on remote sensing technology includes the following steps:

[0074] The system acquires design elevations, design contour lines, slope top lines, slope toe lines, drainage boundaries, and garden path boundaries to generate design topographic data; it acquires construction stage, completed depth, construction equipment operating width, single-layer filling thickness, and allowable elevation deviation to generate construction control data; it collects multi-temporal remote sensing images, remote sensing point clouds, and camera poses, unifies spatial coordinates and divides them into planar grids, associates temporal remote sensing points, calculates normals, temporal persistence, and the number of repeated observations to generate remote sensing topographic data.

[0075] A construction terrain RegSDF model is constructed based on the RegSDF model, including surface diversion units, double zero-point profile units, and curvature transfer units. The surface diversion units include permanent surface branches and temporary cover branches, and the double zero-point profile units include design surface branches. The SDF geometric network, surface light field network, spherical tracing structure, differentiable surface intersection structure, Hessian matrix calculation structure, and minimum surface regularization structure in the RegSDF model are retained.

[0076] The SDF geometric network processes remote sensing terrain data and generates distance values; the surface distribution unit generates permanent surface attribution values ​​and temporary cover attribution values ​​based on the remote sensing terrain data, and updates the distance fields of the permanent surface branch and the temporary cover branch; the spherical tracking structure, the differentiable surface intersection structure, and the surface light field network are invoked to generate permanent surface SDF and temporary cover SDF;

[0077] The design surface branch generates the design surface SDF based on the design terrain data; the dual zero-point profile unit searches for the zero intersection of the permanent surface SDF and the design surface SDF along the gravity direction, and generates the current zero point and the design zero point after being corrected by the differential surface intersection structure. It generates cut and fill marks and construction thickness based on the vertical position, and generates the cut and fill profile and the current construction surface SDF in combination with the completed depth.

[0078] The Hessian matrix calculation structure processes the current construction surface SDF, generates a Hessian matrix, and generates directional curvature components based on the design surface SDF and design contour lines. The curvature transfer unit transfers directional curvature components based on temporary cover assignment values, cut-fill marks, drainage boundaries, slope top lines, slope toe lines, and garden path boundaries, and interrupts cross-boundary transfer. The minimum surface regularization structure combines the directional curvature components to correct the current construction surface SDF, generating a corrected construction surface SDF.

[0079] The zero intersection point of the construction surface SDF and the design surface SDF is searched and corrected along the gravity direction to determine the remaining construction thickness, remaining excavation volume, and remaining filling volume; the continuous operation range is divided according to the operating width of the construction equipment, and a digital construction plan is generated based on the remaining excavation volume, remaining filling volume, and single-layer filling thickness to control the construction equipment to perform excavation, earthwork transportation, layered filling, slope trimming, and compaction.

[0080] Collect remeasured remote sensing images, remeasured remote sensing point clouds, and remeasured camera poses to generate a remeasured permanent surface SDF; search for the zero intersection point between the remeasured permanent surface SDF and the current construction surface SDF to determine the construction deviation; when the construction deviation is greater than the allowable elevation deviation, update the permanent surface attribution value, temporary cover attribution value, and completed depth, and regenerate the cut-fill profile, correct the construction surface SDF, and digital construction plan; when the construction deviation is not greater than the allowable elevation deviation, update the construction stage until the vertical distance between the remeasured zero point and the design zero point is not greater than the allowable elevation deviation.

[0081] In this embodiment, the step of acquiring design terrain data and construction control data, and generating remote sensing terrain data includes:

[0082] Obtain the design elevation, design contour lines, slope top line, slope toe line, drainage boundary, and garden path boundary, and combine them to generate design terrain data;

[0083] The construction stage, completed depth, operating width of construction equipment, single-layer filling thickness, and allowable elevation deviation are obtained and combined to generate construction control data.

[0084] Collect multi-temporal remote sensing images, remote sensing point clouds, and camera poses of the landscape construction area; unify the spatial coordinates of the multi-temporal remote sensing images, remote sensing point clouds, camera poses, and design terrain data; and divide the construction area into planar grids according to the unified planar coordinates.

[0085] Remote sensing points falling into the same planar grid in different acquisition phases are correlated in terms of location. The normal vector is calculated based on the distribution of neighboring points of each remote sensing point. The temporal persistence and the number of repeated observations are determined according to the continuous occurrence of correlated remote sensing points in different acquisition phases. Specifically, several nearest neighboring points within the same acquisition phase are selected with each remote sensing point as the center. The mean coordinates of the neighboring points are calculated. A covariance matrix is ​​constructed based on the deviation of each neighboring point from the mean coordinates. The covariance matrix is ​​decomposed into eigenvalues, and the eigenvector corresponding to the smallest eigenvalue is determined as the normal vector of the remote sensing point. The orientation of the normal vector is then standardized. The occurrence of correlated remote sensing points at the same location is counted according to the order of acquisition phases. The number of acquisition phases with correlated remote sensing points is determined as the number of repeated observations, and the longest number of acquisition phases with consecutively correlated remote sensing points is determined as the temporal persistence.

[0086] Remote sensing topographic data is generated by combining multi-temporal remote sensing images, remote sensing point clouds, camera pose, planar grid, normal, temporal persistence, and number of repeated observations.

[0087] In this embodiment, constructing the construction terrain RegSDF model includes:

[0088] The RegSDF model refers to a 3D surface reconstruction model based on neural implicit representation. It utilizes directed point clouds and multi-view images to jointly optimize the signed distance field and surface light field, and smooths the propagation of distance information through Hessian regularization and completes missing geometry through minimum surface regularization. It is used to improve the accuracy and stability of surface reconstruction under complex scenes, open scenes and incomplete point cloud conditions.

[0089] The RegSDF model retains the SDF geometric network used for geometric representation and the surface light field network used for appearance representation. The SDF geometric network uses an 8-layer multilayer perceptron with 512 neurons in each hidden layer and skip connections in the middle of the network. It performs position encoding on the 3D spatial coordinates at 6 frequency levels and outputs distance values ​​and 256 channels of spatial position features. The surface light field network uses a 4-layer multilayer perceptron and receives corrected intersection points, surface normals, viewing directions after 4 frequency level position encoding, and spatial position features.

[0090] The spherical tracking, differentiable surface intersection calculation, Hessian regularization, and minimum surface regularization in the RegSDF model are retained. Spherical tracking is configured as a spherical tracking structure, repeatedly querying distance values ​​and moving the query position according to the ray direction until the intersection position of the zero horizontal plane is obtained. Differentiable surface intersection calculation is configured as a differentiable surface intersection structure, and the intersection position is corrected according to the distance value, SDF gradient, and ray direction at the intersection position. The part of Hessian regularization that calculates the analytical Hessian matrix based on the second-order partial derivative of SDF is configured as a Hessian matrix calculation structure, and minimum surface regularization is configured as a minimum surface regularization structure.

[0091] A surface diversion unit is added after the SDF geometric network. The surface diversion unit includes a permanent surface branch and a temporary cover branch set in parallel. The distance values ​​and spatial location features generated by the SDF geometric network are used as common processing objects of the two branches. The permanent surface branch is used to represent the continuous surface involved in excavation, filling and slope shaping, and the temporary cover branch is used to represent the visible surface corresponding to vegetation, temporary soil mounds and construction equipment. A surface diversion path is constructed in which the permanent surface branch and the temporary cover branch generate distance fields respectively and are connected to the surface light field network.

[0092] Following the spherical tracking structure and the differentiable surface intersection structure, a new dual-zero-point profile unit is added. A design surface branch is set within the dual-zero-point profile unit, and the design terrain data is used as the processing object of the design surface branch. The permanent surface SDF generated by the permanent surface branch and the design surface SDF generated by the design surface branch are also used as processing objects of the dual-zero-point profile unit. The permanent surface branch and the design surface branch are connected to the spherical tracking structure respectively. The candidate zero-intersection points generated by the spherical tracking structure are passed to the differentiable surface intersection structure for position correction. A dual-zero-point profile path is constructed, sequentially generating the current zero point and the design zero point along the gravity direction. This path is used to generate cut-fill profiles based on the vertical positional relationship between the current zero point and the design zero point.

[0093] A curvature transfer unit is added between the Hessian matrix calculation structure and the minimum surface regularization structure. The Hessian matrix generated by the Hessian matrix calculation structure based on the current construction surface SDF is used as the processing object of the curvature transfer unit. The contour direction, maximum slope direction, and surface normal determined by the design surface SDF and design contour lines are used as the projection direction of the Hessian matrix to generate directional curvature components. The permanent surface attribution value, temporary cover attribution value, cut and fill mark, drainage boundary, slope top line, slope toe line, and garden road boundary are used as the basis for the transfer of directional curvature components. A curvature transfer path is constructed to transfer directional curvature components between adjacent planar grids and stop the transfer at the drainage boundary, slope top line, slope toe line, and garden road boundary. The transferred directional curvature components are input into the minimum surface regularization structure to complete the construction surface below the temporary cover area and retain the slope break and boundary.

[0094] Connect the output of the SDF geometric network to the surface splitting unit; connect the outputs of the permanent surface branch and the temporary cover branch to the surface light field network through the spherical tracking structure and the differentiable surface intersection structure in sequence; connect the outputs of the permanent surface branch and the designed surface branch to the dual zero-point profile unit through the spherical tracking structure and the differentiable surface intersection structure in sequence.

[0095] Based on SDF geometric network, surface light field network, spherical tracking structure, differentiable surface intersection structure, Hessian matrix calculation structure, minimum surface regularization structure, surface diversion unit, double zero point profile unit and curvature transfer unit, the connection relationship between various structures is improved, and a construction terrain RegSDF model is constructed.

[0096] During the training of the RegSDF model for the construction terrain, training samples were composed of multi-temporal remote sensing images, remote sensing point clouds, camera pose, normals, permanent surface attribution values, and temporary cover attribution values. The SDF geometric network, surface light field network, permanent surface branch, and temporary cover branch participated in parameter updates; the design surface branch, dual-zero profile units, and curvature transfer units operated according to defined calculation rules, without any trainable parameters. Four acquisition phases were selected for each batch, and 2048 observation rays were randomly selected from each acquisition phase. Additionally, 8192 remote sensing points and 4096 spatial locations were extracted from the remote sensing point cloud and the construction area, respectively.

[0097] When calculating the distance loss, the absolute values ​​of the distances output by the permanent surface branch and the temporary cover branch at the remote sensing point are taken, multiplied by the permanent surface attribution value and the temporary cover attribution value respectively, and then averaged. When calculating the normal loss, the SDF gradients of the two branches are normalized, and the normalized result is subtracted from the dot product of the remote sensing point normal by 1, and then averaged according to the corresponding attribution value. When calculating the color loss, the predicted color generated by the surface light field network is compared with the actual color of the remote sensing image, and the average of the absolute differences of the three color channels is calculated. When calculating the Eikonal loss, the absolute difference between the SDF gradient magnitude at each spatial location and 1 is calculated. When calculating the Hessian loss, the absolute values ​​of each element of the Hessian matrix are averaged. When calculating the minimum surface loss, the regularized Dirac function values ​​corresponding to random spatial locations are averaged.

[0098] When calculating the directional curvature loss, the current directional curvature component of each planar mesh is compared with the directional curvature component output by the curvature transfer unit. The absolute value of the difference is taken and averaged according to the transfer weight. Planar meshes on both sides of the drainage boundary, slope crest line, slope toe line, and garden path boundary are not included in this calculation. The weights of distance loss, normal loss, and color loss are set to 1, the weights of Eikonal loss and directional curvature loss are set to 0.1, and the weights of Hessian loss and minimum surface loss are set to 0.01. The total loss is generated by weighted summation.

[0099] The Adam optimizer is used to update model parameters. The initial learning rate is set to 0.0005, and the training epochs are set to 800. In the first 100 epochs, only the distance loss, normal loss, and Eikonal loss are calculated. Subsequently, the calculation of the intersection of differentiable surfaces and other losses are enabled. The learning rate is halved every 200 epochs. Training is stopped when the total loss decreases by less than 0.0001 for 30 consecutive epochs, and the model parameters corresponding to the minimum total loss are saved.

[0100] In this embodiment, the construction of permanent surface SDF and temporary cover SDF includes:

[0101] The coordinates of remote sensing points in the remote sensing terrain data are input into the SDF geometric network. The SDF geometric network performs spatial encoding on the coordinates of the remote sensing points, outputs the distance value and spatial location features corresponding to each remote sensing point, and inputs the distance value and spatial location features into the surface diversion unit.

[0102] In the surface subdivision unit, the elevation difference between adjacent remote sensing points acquired at different time phases within the same planar grid is calculated. The elevation difference is normalized by minimum-maximum value and then subtracted from the normalization result by 1 to generate elevation continuity with a value range of 0 to 1. The cosine value of the normal angle between the associated remote sensing points is calculated and converted to 0 to 1 to generate normal consistency. Based on the camera pose, the same remote sensing point is projected onto multi-temporal remote sensing images acquired at different time phases. The average absolute difference of the corresponding pixel in the three color channels is calculated and divided by 255 to generate color reprojection difference.

[0103] The temporal continuity and the number of repeated observations were normalized. The elevation continuity, normal consistency, 1 minus the color reprojection difference, the normalized value of temporal continuity and the normalized value of repeated observations were multiplied by 0.2 and then summed to generate the permanent surface attribution value. The result of 1 minus the permanent surface attribution value was determined as the temporary cover attribution value.

[0104] The permanent surface attribution value and the temporary cover attribution value are used as the update weights for the permanent surface branch and the temporary cover branch, respectively. The distance values ​​output by the permanent surface branch and the temporary cover branch at each remote sensing point are extracted, and the absolute difference between each distance value and 0 is calculated to generate permanent distance value difference and temporary distance value difference. The partial derivatives of the distance values ​​output by the two branches with respect to the three spatial coordinates of the remote sensing point are calculated, and the three partial derivatives are combined to generate the SDF gradient. The SDF gradient is then normalized, and the dot product of the normalized result and the remote sensing point normal is calculated. The permanent normal difference and the temporary normal difference are generated by subtracting the dot product from 1. Normal difference; Multiply the permanent surface attribution value by the permanent distance difference and the permanent normal difference respectively, sum and average the results for all remote sensing points, update the permanent surface branch parameters in reverse based on the average result, and recalculate the distance value based on the updated parameters to update the distance field of the permanent surface branch; Multiply the temporary cover attribution value by the temporary distance difference and the temporary normal difference respectively, sum and average the results for all remote sensing points, update the temporary cover branch parameters in reverse based on the average result, and recalculate the distance value based on the updated parameters to update the distance field of the temporary cover branch;

[0105] The spherical tracking structure is invoked, with the camera center corresponding to the camera pose as the starting point of the observation ray. The distance values ​​of the permanent surface branch and the temporary cover branch are queried along the direction of the observation ray. The query position is moved according to the distance value obtained each time. When the absolute value of the distance value corresponding to the query position is not greater than 0.005 meters, the movement stops, and the query position is determined as the permanent candidate intersection point and the temporary candidate intersection point, respectively. The differentiable surface intersection structure is invoked to obtain the distance value and SDF gradient at the permanent candidate intersection point and the temporary candidate intersection point, respectively. The dot product of the observation ray direction and the SDF gradient is calculated. The adjustment amount of the intersection point position along the direction of the observation ray is determined based on the distance value and the dot product. The permanent candidate intersection point and the temporary candidate intersection point are moved along the direction of the observation ray by the corresponding intersection point position adjustment amount, respectively, to generate the permanent corrected intersection point and the temporary corrected intersection point.

[0106] When both permanent and temporary correction intersections exist, the Euclidean distances between the permanent and temporary correction intersections and the camera pose are compared. The correction intersection with the smaller Euclidean distance, its corresponding spatial location features, remote sensing point normal, and observation ray direction are input into the surface light field network to generate the predicted color. The actual color of the pixel corresponding to the observation ray in the multi-temporal remote sensing image is extracted, the color difference between the predicted color and the actual color is calculated, and the branch parameters of the surface light field network and the correction intersection with the smaller Euclidean distance are updated according to the color difference.

[0107] When the Euclidean distance between the corrected intersection of the temporary cover branch and the camera pose is less than the Euclidean distance between the corrected intersection of the permanent surface branch and the camera pose, the temporary cover branch is updated based on the color difference of the current acquisition phase, and the permanent surface branch is not updated using the color difference of the current acquisition phase; in other acquisition phases, when the corrected intersection of the permanent surface branch is located in front of the corrected intersection of the temporary cover branch, or when the temporary cover branch has not generated a corrected intersection, the permanent surface branch is updated based on the color difference of the corresponding acquisition phase, and the permanent surface branch parameters are further updated by combining the permanent distance value difference and permanent normal difference of remote sensing points in adjacent planar grids whose permanent surface belonging value is greater than the temporary cover belonging value;

[0108] The distance field is recalculated based on the updated branch parameters. The distance field of the permanent surface branch is determined as the permanent surface SDF, and the distance field of the temporary cover branch is determined as the temporary cover SDF.

[0109] In this embodiment, generating the current construction surface SDF includes:

[0110] In the design surface branch, the design elevation is used as the elevation control point, and the design contour lines are discretized into elevation points according to the corresponding elevations. The elevation control points and elevation points are connected to generate the design surface. Linear boundaries are retained at the top of the slope, the toe of the slope, the drainage boundary, and the boundary of the garden road. Cross-boundary interpolation is not performed on the elevations on both sides of the linear boundaries. The directed distance from each spatial location to the nearest location on the design surface is calculated to generate the design surface SDF.

[0111] In a unified spatial coordinate system, the direction of decreasing vertical coordinates is determined as the gravity direction. In a dual-zero-point profile unit, the plane coordinates of the center of each plane grid are used as the starting point. The spherical tracking structure is called to query the distance values ​​of the permanent surface SDF and the design surface SDF along the gravity direction. The query position is moved according to the obtained distance value. When the absolute value of the distance value is not greater than 0.005 meters, the query position is determined as a candidate zero intersection point.

[0112] When multiple candidate zero intersections exist within the same planar grid, delete the candidate zero intersections whose surface normal is oriented towards the gravity direction, and select the candidate zero intersection with the smallest distance from the center plane of the planar grid from the remaining candidate zero intersections. Then, call the differentiable surface intersection structure, determine the position adjustment amount along the gravity direction based on the distance value, SDF gradient and gravity direction at the candidate zero intersection, move the candidate zero intersection, and generate the current zero point and the design zero point respectively.

[0113] Calculate the vertical coordinate difference between the design zero point and the existing zero point. When the vertical coordinate difference is less than -0.005 meters, set the cut-fill mark of the corresponding plane grid to a cut mark; when the vertical coordinate difference is greater than 0.005 meters, set the cut-fill mark of the corresponding plane grid to a fill mark; when the vertical coordinate difference is between -0.005 meters and 0.005 meters, set the cut-fill mark of the corresponding plane grid to a no-construction mark. Determine the absolute value of the vertical coordinate difference as the construction thickness. Combine the plane grid, existing zero point, design zero point, cut-fill mark, and construction thickness to generate a cut-fill profile.

[0114] Compare the completed depth with the construction thickness. When the completed depth is less than the construction thickness, move the completed depth from the current zero point in the direction pointing to the design zero point. When the completed depth is not less than the construction thickness, take the design zero point as the endpoint of the movement, determine the current ground surface position, and generate the current construction ground surface SDF based on the current ground surface position corresponding to each plane grid.

[0115] In this embodiment, generating the modified construction surface SDF includes:

[0116] The Hessian matrix calculation structure is called to process the current construction surface SDF and generate the Hessian matrix corresponding to each planar grid. Specifically, the Hessian matrix calculation structure performs second-order differentiation on the current construction surface SDF with respect to three spatial coordinates, generates nine second-order partial derivatives by combining the three spatial coordinates in pairs, and arranges the nine second-order partial derivatives into the Hessian matrix corresponding to each planar grid.

[0117] Calculate and normalize the gradient of the design surface SDF at each design zero point to generate the surface normal; project the gravity direction onto the tangent plane corresponding to the surface normal and normalize it to generate the direction of maximum slope; calculate and normalize the cross product of the surface normal and the direction of maximum slope to generate the contour line direction.

[0118] The contour line direction, the maximum slope direction, and the surface normal direction are used as the projection directions of the Hessian matrix, respectively. Each projection direction is multiplied by the Hessian matrix in turn, and the dot product of the multiplication result and the corresponding projection direction is calculated to generate the curvature components of the contour line direction, the curvature component of the maximum slope direction, and the curvature component of the surface normal direction.

[0119] Plane grids with temporary cover attribution values ​​greater than permanent surface attribution values ​​are identified as plane grids to be completed. Among the four adjacent plane grids sharing the same edge, the plane grid with permanent surface attribution values ​​greater than temporary cover attribution values ​​is selected as the curvature source plane grid.

[0120] Based on the center line connecting the curvature source planar grid and the to-be-completed full-planar grid, the contour line direction, the maximum slope direction, the center distance, and the permanent surface attribution value, contour line transfer weights and slope transfer weights are generated, and the contour line transfer weights and slope transfer weights corresponding to the same to-be-completed full-planar grid are normalized respectively; specifically, the unit direction vector from the center of the curvature source planar grid to the center of the to-be-completed full-planar grid is calculated, and the dot product of the unit direction vector with the contour line direction and the maximum slope direction is calculated respectively, and the dot product less than 0 is adjusted to 0; the contour lines are then... The contour transfer weight is generated by multiplying the dot product corresponding to the direction, the permanent surface attribution value of the curvature source plane grid, and the reciprocal of the center distance between the two plane grids. The slope transfer weight is generated by multiplying the dot product corresponding to the direction of maximum slope, the permanent surface attribution value, and the reciprocal of the center distance. The sum of all contour transfer weights and the sum of all slope transfer weights corresponding to the same plane grid to be completed are calculated separately. Each transfer weight is divided by the sum of the corresponding weights to complete the normalization. When the sum of the weights corresponding to any direction is 0, the directional curvature component is not transferred along the corresponding direction.

[0121] When the center line connecting the curvature source planar grid and the planar grid to be completed intersects any one of the following boundaries: drainage boundary, slope top line, slope toe line, and garden path boundary, the corresponding contour transfer weight and slope transfer weight are set to 0, and the remaining contour transfer weight and remaining slope transfer weight are renormalized; planar grids corresponding to cut markings and sharing edges are connected to form cut areas, and planar grids corresponding to fill markings and sharing edges are connected to form fill areas. The adjacent positions of cut areas and non-cut areas are determined as the cut area boundary. Within the excavation area, starting from the curvature source plane grid near the boundary of the excavation area, the curvature component in the direction of maximum slope is transferred along the direction of maximum slope according to the slope transfer weight. Within the fill area, the curvature component in the direction of maximum slope is transferred from the curvature source plane grid with a higher design zero elevation to the plane grid to be filled with a lower design zero elevation according to the slope transfer weight. The curvature component in the direction of contour lines is transferred along the contour line direction to the plane grid to be filled with a design zero elevation difference not greater than the allowable elevation deviation according to the contour line transfer weight.

[0122] The product of the contour curvature component and the contour transfer weight of the curvature source plane grid corresponding to the same full-plane grid to be filled is summed to generate the transferred contour curvature component; the product of the maximum slope direction curvature component and the slope transfer weight of the curvature source plane grid is summed to generate the transferred maximum slope direction curvature component; the surface normal curvature component generated by the full-plane grid to be filled according to the Hessian matrix is ​​retained, and the absolute difference between the transferred contour curvature component, the transferred maximum slope direction curvature component and the corresponding direction curvature component of the full-plane grid to be filled is calculated to generate the contour curvature difference and the maximum slope direction curvature difference;

[0123] The minimum surface regularization structure is invoked, and the distance value of the temporary cover SDF is substituted into the regularization Dirac function of the RegSDF model. The function value corresponding to the sampling spatial location where the temporary cover attribution value is greater than the permanent surface attribution value is averaged, and the visible surface in the temporary cover branch is closed. In the permanent surface branch, minimum surface regularization is performed only on the plane grid to be completed within the same cut area or the same fill area. When the line connecting the centers of adjacent plane grids intersects with any of the boundaries of drainage boundary, slope top line, slope toe line, and garden road boundary, cross-boundary surface completion is stopped. The corresponding parameters of the current construction surface SDF are updated according to the curvature difference in the contour direction, the curvature difference in the maximum slope direction, and the minimum surface regularization result. The distance value and Hessian matrix of the plane grid to be completed are recalculated until the curvature difference in both directions is no greater than 0.001, and the corrected construction surface SDF is generated.

[0124] In this embodiment, generating a digital construction plan and controlling the construction equipment to perform construction includes:

[0125] Using the center of each planar grid as the search position, the spherical tracking structure is called to query the distance values ​​of the corrected construction surface SDF and the design surface SDF along the gravity direction, and the search position is moved according to the distance value. When the absolute value of the distance value is not greater than 0.005 meters, the construction zero intersection point and the design zero intersection point are determined respectively.

[0126] When the construction zero intersection point is higher than the design zero intersection point, the vertical distance between the two zero intersection points is determined as the remaining excavation thickness of the corresponding planar grid; when the construction zero intersection point is lower than the design zero intersection point, the vertical distance between the two zero intersection points is determined as the remaining fill thickness of the corresponding planar grid; the remaining excavation thickness or the remaining fill thickness is determined as the remaining construction thickness, the area of ​​the planar grid is multiplied by the remaining construction thickness to generate the remaining earthwork volume of the corresponding planar grid, and the remaining earthwork volumes of all planar grids are accumulated according to the excavation and fill markings to generate the total remaining excavation volume and the total remaining fill volume;

[0127] Select one planar grid from the undivided planar grid, and merge the adjacent planar grids that share the same edge, have the same cut and fill markings, and whose center distance is not greater than the working width of the construction equipment, until there are no adjacent planar grids that meet the conditions, thus generating a continuous working range; extract the grid edges outside the continuous working range that are not shared with other planar grids in the same continuous working range, and connect the outer grid edges in sequence to generate the excavation boundary or the filling boundary respectively.

[0128] Accumulate the remaining earthwork volume of the planar grid within each continuous operation range to generate the remaining excavation volume and the remaining fill volume for each continuous excavation operation range. Calculate the distance between the center of the continuous excavation operation range and the center of the continuous fill operation range. Select the two continuous operation ranges with the smallest center distance. Determine the smaller value between the remaining excavation volume and the remaining fill volume of the corresponding range as the earthwork transfer volume. Subtract the earthwork transfer volume from the corresponding remaining excavation volume and the remaining fill volume of the range. Repeat this process until the remaining excavation volume or the remaining fill volume of the range is 0.

[0129] Divide the remaining fill thickness of the plane grid corresponding to the fill mark by the single-layer fill thickness and round up to determine the number of fill layers; when there is a remainder in the above division, set the layer fill thickness of the first few layers to the single-layer fill thickness, and determine the layer fill thickness of the last layer as the layer fill thickness; when there is no remainder in the above division, set the layer fill thickness of each fill layer to the single-layer fill thickness.

[0130] The remaining excavation thickness of each plane grid within the continuous excavation operation range is determined as the excavation depth at the corresponding location. The number of filling layers and the thickness of each layer are written into the filling boundary within the continuous filling operation range. The vertical coordinates of the zero intersection points of each plane grid are determined as the whole slope elevation. The compaction sequence is determined according to the order of the filling layers from bottom to top.

[0131] By combining excavation boundaries, excavation depths, fill boundaries, layered fill thicknesses, earthwork transport volumes, slope elevations, and compaction sequences, a digital construction plan is generated.

[0132] According to the digital construction plan, the continuous operation range of excavation and filling, as well as the corresponding construction parameters, are sent to the construction equipment. The construction equipment is then controlled to perform excavation, earthwork transportation, layered filling, layer-by-layer compaction, and slope leveling in sequence.

[0133] In this embodiment, the step of generating the remeasured permanent surface SDF and construction deviation, and updating the digital construction plan based on the construction deviation, includes:

[0134] After completing the current construction phase, collect remeasured remote sensing images, remeasured remote sensing point clouds, and remeasured camera poses. Unify the spatial coordinates of the remeasured remote sensing images, remeasured remote sensing point clouds, and remeasured camera poses and divide them into corresponding planar grids. Associate the remeasured remote sensing points with historical remote sensing points in the same planar grid according to their planar positions. Recalculate the normal, elevation continuity, normal consistency, color reprojection difference, temporal continuity, and number of repeated observations of the remeasured remote sensing points. Incorporate the remeasured remote sensing images, remeasured remote sensing point clouds, remeasured camera poses, and the normal, elevation continuity, normal consistency, color reprojection difference, temporal continuity, and number of repeated observations of the remeasured remote sensing points into the remote sensing topographic data.

[0135] The SDF geometric network processes the coordinates of the remeasured remote sensing points to generate remeasured distance values ​​and remeasured spatial location characteristics. The surface dispersal unit combines the recalculated elevation continuity, normal consistency, color reprojection difference, temporal continuity, and number of repeated observations to generate remeasured permanent surface attribution values ​​and remeasured temporary cover attribution values. Based on the two attribution values, the permanent surface branch and the temporary cover branch are updated. Combining the spherical tracking structure, the differentiable surface intersection structure, and the surface light field network, the remeasured permanent surface SDF is generated.

[0136] The spherical tracking structure is invoked to query the SDF of the remeasured permanent surface and the SDF of the corrected construction surface along the gravity direction. When the absolute value of the distance value is not greater than 0.005 meters, the remeasurement zero point and the stage target zero point are generated respectively. Only when the remeasurement permanent surface value is greater than the remeasurement temporary cover value is the corresponding remeasurement zero point retained. The absolute value of the vertical coordinate difference between the remeasurement zero point and the stage target zero point is calculated to generate the construction deviation.

[0137] When the construction deviation exceeds the allowable elevation deviation, the original values ​​of the same plane grid are replaced by the re-measured permanent surface value and the re-measured temporary cover value. The direction from the current zero point to the re-measured zero point is calculated to see if it is consistent with the direction from the current zero point to the design zero point. If the directions are consistent, the vertical distance from the current zero point to the re-measured zero point is updated to the completed depth, and the part exceeding the construction thickness is adjusted to the construction thickness. When the re-measured permanent surface value is not greater than the re-measured temporary cover value, the re-measured zero points corresponding to construction machinery, temporary soil piles, and vegetation are not included in the completed depth update.

[0138] Based on the updated permanent surface attribution value, temporary cover attribution value, and completed depth, the cut-fill profile and the current construction surface SDF are regenerated. The Hessian matrix, directional curvature components and their transmission results are recalculated to generate the corrected construction surface SDF. The digital construction plan is then regenerated based on the corrected construction surface SDF.

[0139] When the construction deviations of all plane grids with excavation or filling marks in the current construction phase are not greater than the allowable elevation deviation, the construction phase is updated and the next construction phase is executed. After re-measurement in each construction phase, the vertical distance between the re-measured zero point and the design zero point in the plane grid with excavation or filling marks is calculated. Construction ends when all vertical distances are not greater than the allowable elevation deviation.

[0140] Example 1: To verify the feasibility of this invention in practice, it was applied to a landscape earthwork construction scenario. The construction area was a landscape terrain with gentle slopes, terraces, winding paths, and drainage ditches. The effective construction area included in the calculation was 3816.7 square meters. After dividing the area into planar grids of 0.5 meters × 0.5 meters, the outer edges of the paths and non-construction corner areas were removed, leaving 14892 effective planar grids. The design terrain data included design elevation, design contour lines, slope top lines, slope toe lines, drainage boundaries, and path boundaries. The design elevation was between 24.38 meters and 28.16 meters. The total length of the slope top line was 173.4 meters, the total length of the slope toe line was 196.8 meters, the total length of the drainage boundary was 241.6 meters, and the total length of the path boundary was 318.9 meters. The construction control data records the construction stage, completed depth, operating width of construction equipment, single-layer fill thickness, and allowable elevation deviation. The operating width of construction equipment is 2.2 meters, the single-layer fill thickness is 0.28 meters, and the allowable elevation deviation is 0.04 meters. The completed depth of the current construction stage is mainly distributed between 0.12 meters and 0.47 meters. This scenario includes localized vegetation cover, temporary soil piles, brief stops of construction equipment, and abrupt changes in slope elevation. Traditional manual measurement and single-point cloud comparison can easily mistake temporary covered surfaces for the actual construction surface, affecting the updating of cut-fill profiles and construction plans.

[0141] In this scenario, remote sensing images, point clouds, and camera poses were collected at nine different time points before construction. A total of 326 remote sensing images were acquired, with a single image resolution of 5472×3648 pixels. The effective point cloud count was 1.746 million points, with an average point density of approximately 457 points / square meter. 326 sets of camera poses were recorded. After unifying the multi-temporal remote sensing images, point clouds, and camera poses to the same spatial coordinates, the remote sensing points were merged according to a planar grid. The normal direction, elevation continuity, normal consistency, temporal continuity, number of repeated observations, and color reprojection difference of the remote sensing points were calculated. On-site verification revealed that the grids affected by vegetation cover and temporary soil mounds were mainly distributed in the inner turning area of ​​the garden path, the soil mound area at the toe of the slope, and near the drainage ditch, affecting 1687 grids, which constitute a portion of the effective planar grid. After inputting remote sensing topographic data into the construction topographic RegSDF model, the SDF geometric network extracts the surface spatial representation. The surface subdivision unit distinguishes between permanent surface and temporary cover based on elevation continuity, normal consistency, temporal continuity, number of repeated observations, and color reprojection difference, and constructs permanent surface SDF and temporary cover SDF respectively.

[0142] During construction, the dual-zero-point profile unit determines the current zero point and design zero point based on the permanent surface SDF and design surface SDF, transforming the vertical relationship between them into a cut-fill profile, and generating the current construction surface SDF in conjunction with the completed depth. The excavation areas identified in this stage are mainly concentrated on the edges of higher-elevation terraces and the transition zone at the top of the slope, while the filling areas are mainly concentrated in the low-lying areas at the toe of the slope and the connecting area outside the garden path. The curvature transfer unit corrects the current construction surface SDF based on the directional curvature components, ensuring the construction surface below the temporary cover area maintains slope continuity, and stopping cross-boundary surface completion at drainage boundaries, the top line of the slope, the toe line of the slope, and the boundary of the garden path. The digital construction scheme is generated based on the corrected construction surface SDF, the design surface SDF, and construction control data. After completing the current construction phase, re-measurement remote sensing images, re-measurement remote sensing point clouds, and re-measurement camera poses will be collected. Each re-measurement will collect remote sensing images from 3 to 4 time phases. The effective number of points in the re-measurement remote sensing point cloud will be maintained between 493,000 and 668,000 points. The re-measurement permanent surface SDF and construction deviation will be generated, and the subsequent digital construction plan will be updated accordingly.

[0143] Table 1 Comparison of different digital processing methods for earthwork construction

[0144] Permanent surface identification accuracy 84.6% 88.1% 91.7% Cut-fill boundary matching rate 78.9% 83.5% 89.2% Average error in earthwork volume calculation 14.8% 10.6% 6.9% Temporary coverage misjudgment rate 18.5% 13.2% 7.6% Time spent adjusting the plan after retesting 46 minutes 31 minutes 19 minutes Stage construction deviation pass rate 74.5% 81.3% 90.6%

[0145] As shown in Table 1, the permanent surface identification accuracy of the method of this invention is 91.7%, which is higher than the 84.6% of the traditional manual measurement method and the 88.1% of the point cloud elevation comparison method. This result mainly comes from the differentiated processing of permanent surface attribution values ​​and temporary cover attribution values ​​by the surface diversion unit, so that the visible surfaces formed by vegetation, temporary mounds, and construction equipment do not directly participate in the construction of permanent surface SDF, reducing the interference of non-real construction surfaces on subsequent calculations.

[0146] Regarding the cut-fill boundary matching rate, the method of this invention achieves 89.2%, which is higher than the 78.9% of the traditional manual measurement method and the 83.5% of the point cloud elevation comparison method. This is because the present invention processes both the permanent surface SDF and the design surface SDF simultaneously through dual zero-point profile units, directly generating cut-fill profiles based on the vertical relationship between the current zero point and the design zero point. Compared with simply using the elevation difference of measuring points for judgment, this method can more stably express the continuous changes of the slope and the position of the design boundary.

[0147] Regarding the average error in earthwork volume calculation, the method of this invention has an error rate of 6.9%, lower than the 14.8% of traditional manual measurement methods and the 10.6% of point cloud elevation comparison methods. This result indicates that the present invention has better spatial consistency in calculating remaining construction thickness, remaining excavation volume, and remaining fill volume. Especially in areas with local obstruction and undulating slopes, the present invention reduces volume deviations caused by local point cloud omissions and misjudgments of temporary coverage by correcting the SDF of the construction surface in subsequent earthwork volume calculations.

[0148] Regarding the misclassification rate of temporary cover, the method of this invention has a rate of 7.6%, lower than the 18.5% of traditional manual measurement methods and the 13.2% of point cloud elevation comparison methods. This indicator can directly reflect the effectiveness of surface dispersal units. Traditional manual measurement methods rely on on-site judgment for temporary soil mounds and areas obscured by construction equipment, while point cloud elevation comparison methods easily mistake areas with abrupt elevation changes for terrain changes. In contrast, this invention combines elevation continuity, normal consistency, temporal continuity, number of repeated observations, and color reprojection difference for attribution judgment, thus reducing misclassification.

[0149] Regarding the time required for adjusting the plan after re-survey, the method of this invention takes 19 minutes, shorter than the 46 minutes of the traditional manual measurement method and the 31 minutes of the point cloud elevation comparison method. This result comes from the continuous update relationship between the re-surveyed permanent surface SDF, construction deviation, and digital construction plan. After the re-survey, there is no need to rebuild the complete construction judgment process; it is only necessary to update the relevant planar grid based on the re-surveyed permanent surface attribution value, the re-surveyed temporary cover attribution value, and the construction deviation, and regenerate the corrected construction surface SDF and digital construction plan.

[0150] Regarding the pass rate of stage construction deviations, the method of this invention achieved 90.6%, which is higher than the 74.5% of the traditional manual measurement method and the 81.3% of the point cloud elevation comparison method. This improvement mainly comes from the correction processing of the current construction surface SDF by the curvature transfer unit and the minimum surface regularization structure, making the slope completion below the temporary cover area more continuous, and maintaining construction constraints at the drainage boundary, slope top line, slope toe line, and garden path boundary.

[0151] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A digital earthwork construction method for landscape architecture based on remote sensing technology, characterized in that: Includes the following steps: Acquire design terrain data and construction control data, collect multi-temporal remote sensing images, remote sensing point clouds and camera poses, unify spatial coordinates and divide planar grids to generate remote sensing terrain data; Construct a RegSDF model of the construction terrain, including an SDF geometric network, surface diversion units, dual zero-point profile units, and curvature transfer units; Based on the SDF geometric network, remote sensing terrain data is processed to generate permanent surface attribution values ​​and temporary cover attribution values, and surface diversion units are invoked to construct permanent surface SDF and temporary cover SDF; In the dual zero-point profile unit, the design surface SDF is constructed based on the design terrain data. Combined with the permanent surface SDF, the current zero point and the design zero point are generated, and the current surface location is determined to generate the current construction surface SDF. Based on the current construction surface SDF, directional curvature components are generated. In the curvature transfer unit, the components are transferred according to the design terrain data. The current construction surface SDF is corrected by combining the temporary cover assignment value, the temporary cover SDF, and the transferred directional curvature components, and the corrected construction surface SDF is generated. Based on the corrected construction surface SDF, design surface SDF, and construction control data, a digital construction plan is generated, and construction equipment is controlled to perform construction. Collect remeasured remote sensing images, remeasured remote sensing point clouds, and remeasured camera poses to generate remeasured permanent surface SDF and construction deviations, and update the digital construction plan based on the construction deviations.

2. The digital earthwork construction method for landscape architecture based on remote sensing technology according to claim 1, characterized in that, The process of acquiring design terrain data and construction control data, and generating remote sensing terrain data, includes: Obtain the design elevation, design contour lines, slope top line, slope toe line, drainage boundary, and garden path boundary, and combine them to generate design terrain data; The construction stage, completed depth, operating width of construction equipment, single-layer filling thickness, and allowable elevation deviation are obtained and combined to generate construction control data. Collect multi-temporal remote sensing images, remote sensing point clouds, and camera poses of the landscape construction area, unify spatial coordinates, and divide the construction area into a planar grid; The locations of remote sensing points falling into the same planar grid in different acquisition phases are correlated. The normal is calculated based on the distribution of neighboring points of each remote sensing point. The phase duration and the number of repeated observations are determined according to the continuous occurrence of the correlated remote sensing points in different acquisition phases. Remote sensing topographic data is generated by combining multi-temporal remote sensing images, remote sensing point clouds, camera pose, planar grid, normal, temporal persistence, and number of repeated observations.

3. The digital earthwork construction method for landscape architecture based on remote sensing technology according to claim 1, characterized in that, The construction terrain RegSDF model includes: The SDF geometric network and surface optical field network in the RegSDF model are retained; The spherical tracing, differentiable surface intersection calculation, Hessian regularization, and minimum surface regularization in the RegSDF model are retained and configured as spherical tracing structure, differentiable surface intersection structure, Hessian matrix calculation structure, and minimum surface regularization structure; Following the SDF geometric network, a new surface divergence unit is added, with permanent surface branches and temporary cover branches set in parallel. The permanent surface branches and temporary cover branches generate distance fields respectively and are connected to the surface light field network to form a surface divergence path. After the spherical tracking structure and the intersection structure of the differentiable surface, a double zero-point profile unit is added, a design surface branch is set, the design terrain data is used as the processing object of the design surface branch, and a double zero-point profile path is constructed to generate the current zero point and the design zero point in sequence along the gravity direction. Between the Hessian matrix calculation structure and the minimum surface regularization structure, a curvature transfer unit is added to construct a curvature transfer path for the directional curvature components to be transferred between adjacent planar meshes. Improve the connection relationships between various structures and construct a RegSDF model of the construction terrain.

4. The digital earthwork construction method for landscape architecture based on remote sensing technology according to claim 1, characterized in that, The construction of permanent surface SDF and temporary cover SDF includes: The coordinates of remote sensing points in the remote sensing terrain data are input into the SDF geometric network for spatial encoding, generating the distance value and spatial location features corresponding to each remote sensing point. In the surface subdivision unit, the elevation difference and elevation continuity of adjacent time-phase remote sensing points within the same planar grid are calculated, the normal consistency of the normal angle between associated remote sensing points is calculated, the same remote sensing point is projected onto multi-temporal remote sensing images according to the camera pose, and the color reprojection difference is calculated. The duration of the time phase and the number of repeated observations were normalized. Based on elevation continuity, normal consistency, color reprojection difference, normalized value of temporal persistence, and normalized value of repeated observations, permanent surface attribution value and temporary cover attribution value are calculated and generated. The permanent surface attribution value and the temporary cover attribution value are used as the update weights for the permanent surface branch and the temporary cover branch, respectively, to update the distance field of the permanent surface branch and the distance field of the temporary cover branch. The spherical tracking structure is invoked to search for candidate intersection points in the distance fields of the permanent surface branch and the temporary cover branch along the observation ray defined by the camera pose. The corrected intersection points are determined and the corrected intersection points and their corresponding spatial location features are input into the surface light field network for remote sensing image fitting. The fitted permanent surface branch distance field is determined as the permanent surface SDF, and the fitted temporary cover branch distance field is determined as the temporary cover SDF.

5. The digital earthwork construction method for landscape architecture based on remote sensing technology according to claim 1, characterized in that, The generation of the current construction surface SDF includes: In the design surface branch, the design elevation is used as the elevation control point, the design contour line is discretized into elevation points, the elevation control points and elevation points are connected to generate the design surface, the linear boundary is preserved, and the design surface SDF is constructed. In a unified spatial coordinate system, the direction of gravity is determined. In the dual zero-point profile unit, the plane coordinates of the center of each plane grid are used as the starting point. The spherical tracking structure is called to query the candidate zero intersection points of the permanent surface SDF and the design surface SDF along the direction of gravity. The candidate zero intersection point with the smallest distance from the plane center of the plane grid is selected. The differentiable surface intersection structure is called to generate the current zero point and the design zero point. Calculate the vertical coordinate difference between the design zero point and the existing zero point, determine the cut-fill mark, determine the absolute value of the vertical coordinate difference as the construction thickness, and combine the plane grid, the existing zero point, the design zero point, the cut-fill mark and the construction thickness to generate the cut-fill profile; Compare the completed depth with the construction thickness to determine the current surface location and generate the current construction surface SDF.

6. The digital earthwork construction method for landscape architecture based on remote sensing technology according to claim 1, characterized in that, The generation of the modified construction surface SDF includes: The Hessian matrix is ​​called to calculate the current construction surface SDF and generate the Hessian matrix corresponding to each planar grid. Based on the design surface SDF, calculate the surface normal, the direction of maximum slope, and the contour line direction; Using the contour line direction, the maximum slope direction, and the surface normal as the projection directions of the Hessian matrix, respectively, calculate the curvature components of the contour line direction, the curvature components of the maximum slope direction, and the curvature components of the surface normal. Based on the temporary cover attribution value and the permanent surface attribution value, the plane grid to be completed is determined, and the plane grid from the curvature source is selected; Based on the center line connecting the curvature source planar grid and the to be completed full planar grid, the contour line direction, the maximum slope direction, the center distance, and the permanent surface attribution value, calculate and generate the contour line transfer weight and the slope transfer weight; Based on the cut and fill markings, the cut area and fill area are determined, and the curvature component in the direction of maximum slope is transferred according to the slope transfer weight. According to the contour line transfer weight, the contour line direction curvature component is transferred along the contour line direction, the surface normal curvature component is retained, and the direction curvature difference between the transferred direction curvature component and the corresponding direction curvature component of the to be filled full-plane grid is calculated. The minimum surface regularization structure is invoked to perform minimum surface regularization in the permanent surface branch. The corresponding parameters of the current construction surface SDF are updated based on the directional curvature difference and the minimum surface regularization result, and the corrected construction surface SDF is generated.

7. The digital earthwork construction method for landscape architecture based on remote sensing technology according to claim 1, characterized in that, The process of generating a digital construction plan and controlling construction equipment to perform construction includes: Using the center of each planar grid as the search location, the spherical tracing structure is invoked to query the distance values ​​of the corrected construction surface SDF and the design surface SDF along the gravity direction to determine the construction zero intersection point and the design zero intersection point; Based on the construction zero intersection point and the design zero intersection point, determine the remaining construction thickness and calculate the total remaining excavation volume and the total remaining fill volume; Select one planar grid from the undivided planar grid to construct the continuous working range and determine the excavation boundary and the filling boundary; The remaining earthwork volume of each continuous operation range is accumulated to obtain the remaining excavation volume and the remaining fill volume of the range, and the earthwork transportation volume is determined. Calculate the number of filling layers and the thickness of each layer based on the remaining construction thickness and the thickness of the single-layer filling; The excavation depth is determined based on the remaining construction thickness, and the vertical coordinates of the zero intersection points of each planar grid are determined as the overall slope elevation to determine the compaction sequence. By combining excavation boundaries, excavation depths, fill boundaries, layered fill thicknesses, earthwork transport volumes, slope elevations, and compaction sequences, a digital construction plan is generated. According to the digital construction plan, the data is sent to the construction equipment, and the equipment is controlled to carry out earthwork construction.

8. The digital earthwork construction method for landscape architecture based on remote sensing technology according to claim 1, characterized in that, The process of generating and re-measuring permanent surface SDF and construction deviations, and updating the digital construction plan based on the construction deviations, includes: After completing the current construction phase, collect the remeasured remote sensing images, remeasured remote sensing point clouds, and remeasured camera poses. Recalculate the normal, elevation continuity, normal consistency, color reprojection difference, temporal continuity, and number of repeated observations of the remeasured remote sensing points, and update the remote sensing topographic data. The SDF geometric network is called to process the coordinates of the remeasured remote sensing points. Based on the surface subtraction unit, the remeasured permanent surface attribution value and the remeasured temporary cover attribution value are generated. Combined with the spherical tracking structure, the differentiable surface intersection structure and the surface light field network, the remeasured permanent surface SDF is generated. The spherical tracking structure is invoked to generate the re-measurement zero point and the stage target zero point along the gravity direction, and the construction deviation is calculated and generated. When the construction deviation is greater than the allowable elevation deviation, the original values ​​of the same plane grid are replaced by the re-measured permanent surface value and the re-measured temporary cover value, and the vertical distance from the current zero point to the re-measured zero point is updated to the completed depth. Based on the updated permanent surface attribution values, temporary cover attribution values, and completed depth, the cut-fill profile and the current construction surface SDF are regenerated, and the digital construction plan is regenerated.

Images