Land high-precision three-dimensional modeling optimization method and system based on unmanned aerial vehicle remote sensing

Abstract

The invention discloses a land high-precision three-dimensional modeling optimization method and system based on unmanned aerial vehicle remote sensing, and relates to the technical field of remote sensing surveying and mapping, and the method comprises the steps: carrying out the local multi-view geometric solution and increment optimization of a multi-source view block sequence, and obtaining a view block geometric parameter set; performing online training on the neural elevation coding network by using the view block geometric parameter set to generate a gridding result; carrying out semantic projection and grid subdivision on the gridding result, carrying out clustering to obtain a plurality of improvement operation blocks, and carrying out graph construction on each improvement operation block by using the preprocessed topographic data to obtain a plot instance scene graph; performing visual interaction on the plot instance scene graph to generate a renovation index, and updating the plot instance scene graph to obtain an updated plot instance scene graph; according to the method, the neural elevation field is constructed into the plot instance scene graph, so that the problem of weak semantic association of plot improvement is solved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing and mapping technology, and in particular to a method and system for optimizing high-precision 3D land modeling based on UAV remote sensing. Background Technology

[0002] 3D modeling technology based on UAV remote sensing has become the mainstream method for obtaining high-precision terrain data in land consolidation projects. Most existing 3D modeling methods rely on structured light, LiDAR, or stereo image matching, generating digital surface models through multi-view geometry of aerial imagery, and then combining digital orthophotos and digital surface models to achieve terrain mapping and feature identification. To improve the accuracy and timeliness of the models, researchers have proposed multi-view stereo reconstruction algorithms, deep learning semantic segmentation models, and point cloud fusion technology, enabling UAVs to achieve high-precision surface modeling in large-scale and complex terrains.

[0003] Existing methods still have room for improvement. First, they typically employ offline reconstruction and step-by-step optimization, resulting in low accuracy in multi-source data fusion. Second, the lack of spatiotemporal constraints and semantic relationships between multi-source data makes it difficult to form an interactive and updatable digital terrain representation structure. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a high-precision 3D land modeling optimization method based on UAV remote sensing to solve the problems of low accuracy of multi-source data fusion and weak semantic association of land parcel remediation.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for optimizing high-precision 3D land modeling based on UAV remote sensing, comprising, Collect and preprocess terrain data, improvement planning data, and UAV remote sensing data of the target area to construct a multi-source view block sequence; Local multi-view geometry solution and incremental optimization are performed on the multi-source view block sequence to obtain the view block geometry parameter set. The view block geometry parameter set is then used to train the neural elevation coding network online to generate gridded results. Semantic projection and grid subdivision are performed on the gridded results, and clustering is carried out to obtain multiple remediation operation blocks. The preprocessed terrain data is used to construct a graph for each remediation operation block to obtain a scene graph of the land parcel instance. The system provides interactive visualization of the land parcel instance scene map, generates remediation indicators, and updates the land parcel instance scene map to obtain the updated land parcel instance scene map.

[0007] As a preferred embodiment of the high-precision 3D land modeling optimization method based on UAV remote sensing described in this invention, the specific steps for constructing a multi-source view block sequence using preprocessed terrain data, land reclamation planning data, and UAV remote sensing data are as follows. The importance of each plot is calculated using the preprocessed remediation planning data, and the distribution of plot importance is obtained. Based on the distribution of land parcel importance, preprocessed terrain data, and UAV remote sensing data, the flight path of the UAV is planned to obtain the UAV flight path and the set of photo points. Control the drone to perform aerial photography based on the drone's flight path and the set of photo points to obtain an aerial image sequence. Then, aggregate the aerial image sequence to obtain a multi-source view block sequence.

[0008] As a preferred embodiment of the high-precision 3D land modeling optimization method based on UAV remote sensing described in this invention, the specific steps for performing local multi-view geometric solution and incremental optimization on the multi-source view block sequence to obtain the view block geometric parameter set are as follows. Local feature extraction and matching are performed on the multi-source view block sequence, and local multi-view geometry is solved to obtain the absolute camera exterior orientation parameters and pixel ray directions; Incremental optimization of the absolute camera exterior orientation parameters and pixel ray directions yields a set of view block geometric parameters.

[0009] As a preferred embodiment of the high-precision 3D land modeling optimization method based on UAV remote sensing described in this invention, the specific steps for training the neural elevation coding network online using a set of view block geometric parameters to generate a gridded result are as follows. The neural elevation field is obtained by using the view block geometry parameter set to train the local region of the neural elevation coding network through online backpropagation. The neural elevation field is sampled in a gridded manner to obtain the gridded results.

[0010] As a preferred embodiment of the high-precision 3D land modeling optimization method based on UAV remote sensing described in this invention, the specific steps for semantic projection and mesh subdivision of the meshed results are as follows: Semantic projection is performed on the meshing results using a multi-source view block sequence to obtain semantically meshed results. The semantic meshing results are extended to three-dimensional space using a neural elevation field to obtain a three-dimensional surface mesh.

[0011] As a preferred embodiment of the high-precision 3D land modeling optimization method based on UAV remote sensing described in this invention, the specific steps for constructing maps of each remediation operation block using preprocessed terrain data to obtain instance scene maps of the blocks are as follows. The mapping information between land parcels and remediation work areas is determined based on the preprocessed terrain data and remediation planning data; Using mapping information, land parcels and remediation work areas are abstracted into nodes, and edge relationships are established between each node to obtain a land parcel relationship graph; The parcel relationship graph is extended into a parcel instance scene graph using neural elevation fields and semantically meshed results.

[0012] As a preferred embodiment of the high-precision 3D land modeling and optimization method based on UAV remote sensing described in this invention, the specific steps for visualizing and interacting with the plot instance scene map to generate remediation indicators are as follows. Load the plot instance scene map into the 3D visualization platform, and respond to the user's drawing operations in the visualization interface to obtain the geometric shape drawn by the user; The user-drawn geometry is bound to the corresponding node, and the node is updated to obtain the governance indicators.

[0013] As a preferred embodiment of the high-precision 3D land modeling optimization method based on UAV remote sensing described in this invention, the specific steps for updating the plot instance scene map to obtain the updated plot instance scene map are as follows. The target terrain is obtained by locally constraining and adjusting the neural elevation field using the remediation index; The target terrain is used to adjust the plot instance scene map to obtain the updated plot instance scene map.

[0014] As a preferred embodiment of the high-precision 3D land modeling optimization method based on UAV remote sensing described in this invention, the specific steps for adjusting the plot instance scene map using the target terrain are as follows: Based on the target terrain and remediation indicators, a local gridded result is generated, and the three-dimensional surface grid is locally replaced to obtain an updated three-dimensional surface grid. The updated 3D surface grid is used to adjust the plot relationship map and then expanded to obtain the updated plot instance scene map.

[0015] Secondly, this invention provides a high-precision 3D land modeling and optimization system based on UAV remote sensing, comprising: The module collects and preprocesses terrain data, improvement planning data, and UAV remote sensing data of the target area to construct a multi-source view block sequence; The training module performs local multi-view geometry solving and incremental optimization on the multi-source view block sequence to obtain a set of view block geometry parameters. The set of view block geometry parameters is then used to train the neural elevation coding network online to generate a gridded result. The instance module performs semantic projection and grid subdivision on the gridded results, and performs clustering to obtain multiple remediation operation blocks. The preprocessed terrain data is used to construct a graph for each remediation operation block to obtain a plot instance scene graph. The update module provides interactive visualization of the land parcel instance scene map, generates remediation indicators, and updates the land parcel instance scene map to obtain the updated land parcel instance scene map.

[0016] The beneficial effects of this invention are as follows: By establishing a land 3D modeling optimization process based on neural elevation coding, dynamic reconstruction of terrain elevation can be achieved on the basis of multi-source remote sensing data fusion, realizing high-density acquisition and high-precision 3D reconstruction of high-value remediation areas, thus solving the problem of low accuracy of multi-source data fusion; by constructing the neural elevation field as a plot instance scene map, incremental updates of terrain and remediation indicators and real-time adjustment of scene map attributes are realized, thus solving the problem of weak semantic association of plot remediation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart of an optimization method for high-precision 3D land modeling based on UAV remote sensing; Figure 2 A schematic diagram of a high-precision 3D land modeling and optimization system based on UAV remote sensing; Figure 3 A schematic diagram for constructing a multi-source view block sequence; Figure 4 This is a schematic diagram of geometric solution and incremental optimization. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0022] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for optimizing high-precision 3D land modeling based on UAV remote sensing, including the following steps: S1: Collect and preprocess terrain data, improvement planning data, and UAV remote sensing data of the target area to construct a multi-source view block sequence.

[0023] S1.1: Perform coordinate projection and elevation unification on the terrain data, perform coordinate projection on the improvement planning data, and perform time alignment on the UAV remote sensing data to obtain preprocessed terrain data, preprocessed improvement planning data, and preprocessed UAV remote sensing data.

[0024] Furthermore, the elevation values ​​in the terrain data are standardized to a numerical format in meters, and the coordinate system of the terrain data is converted to the Gauss-Kruger projection plane coordinate system. The coordinate system of the remediation planning data is converted to the Gauss-Kruger projection plane coordinate system, and the coordinate system of the UAV remote sensing data is converted to the Gauss-Kruger projection plane coordinate system. Outlier detection and removal are performed on the UAV remote sensing data, and linear interpolation is performed on the UAV remote sensing data according to the timestamp to obtain the preprocessed UAV remote sensing data.

[0025] S1.2: Assign values ​​to each plot in the preprocessed remediation planning data according to whether it is farmland, within the remediation scope, or a sensitive area. Farmland is assigned a value of 1, non-farmland is assigned a value of 0, within the remediation scope is assigned a value of 1, non-remediation is assigned a value of 0, and sensitive area is assigned a value of 1, non-sensitive area is assigned a value of 0. Whether it is farmland, within the remediation scope, or a sensitive area is an inherent attribute of the remediation planning data and does not need to be classified. The importance of each plot of land is calculated using the preprocessed remediation planning data, expressed as follows: ; in, For the first The importance of each plot of land, For land parcel indexing, Indicates the first Is each plot of land a farmland area? Indicates the first Is each plot of land within the scope of the remediation? Indicates the first Is each plot of land a sensitive area? This is the weight for farmland areas, with a value of 0.4. The weight for the scope of the rectification is set to 0.4. The weight for the sensitive area is 0.2. It should also be noted that the weights for farmland areas and remediation areas are set at 0.4 because farmland areas are priority areas for protection and key improvement in land remediation. Plots within the remediation area directly participate in the current construction and planning implementation. A weight higher than 0.4 would result in plots that are farmland areas but not within the remediation area, or plots that are not farmland areas but within the remediation area, contributing more, leading to an overly extreme result in plot importance, making it difficult to balance accuracy and cost in flight route planning. A weight lower than 0.4 would weaken the importance of farmland areas and plots within the remediation area, potentially allowing some non-critical plots to receive similar aerial photography configurations as key farmland areas, failing to highlight the policy orientation of "key protection." The weight for sensitive areas is set at 0.2 because sensitive areas require monitoring but are not necessarily directly remediated. A weight higher than 0.2 would over-inflate the importance of plots in sensitive areas, interfering with the priority ranking of farmland and remediation areas. A weight lower than 0.2 would weaken the identification of plots in sensitive areas in terms of plot importance, failing to reflect the necessary safety and monitoring in flight route planning.

[0026] The importance of each land parcel is interpolated and rasterized according to its spatial location to obtain the distribution of land parcel importance.

[0027] S1.3: As can be seen from step S1.2, the importance of each plot can only be one of 0, 0.2, 0.4, 0.6, 0.8 and 1.0. Because a plot with an importance of 0.8 or 1.0 means that it is both farmland and within the remediation area, it is classified as a first-level important plot. A plot with an importance of 0.4 or 0.6 means that it is either farmland or within the remediation area, so it is classified as a second-level important plot. A plot with an importance of 0 or 0.2 means that it is neither farmland nor within the remediation area, so it is classified as a third-level important plot.

[0028] For Tier 1 important plots, a flight altitude of 80 meters, a forward overlap of 80%, and a lateral overlap of 70% were set; for Tier 2 important plots, a flight altitude of 100 meters, a forward overlap of 75%, and a lateral overlap of 65% were set; and for Tier 3 important plots, a flight altitude of 120 meters, a forward overlap of 70%, and a lateral overlap of 60% were set. This yielded a set of flight altitude and overlap parameters corresponding to each importance level. The flight altitudes were set to 80 meters, 100 meters, and 120 meters respectively because, given the focal length and pixel size configuration of the UAV camera, a flight altitude of 80 meters allows for finer ground resolution for primary important plots, a flight altitude of 100 meters allows for medium-precision ground resolution for secondary important plots, and a flight altitude of 120 meters allows for a more relaxed resolution requirement for tertiary important plots while ensuring they can be used for 3D modeling. Using 80% forward overlap and 70% lateral overlap for primary important plots can achieve denser multi-view constraints. Using 75% forward overlap and 65% lateral overlap for secondary important plots can reduce some redundant acquisitions while ensuring matching quality. Using 70% forward overlap and 60% lateral overlap for tertiary important plots can reduce the number of aerial images and operating costs while meeting the basic requirements of 3D modeling.

[0029] The flight altitude parameter and overlap parameter are used to calculate the route spacing and trigger spacing, as expressed by: ; in, For the distance between routes, For flight altitude parameters, For the camera's horizontal field of view, Lateral overlap; ; in, For trigger spacing, For flight altitude parameters, For the camera's longitudinal field of view, For heading overlap; Substitute the flight altitude and overlap parameters of the primary, secondary, and tertiary important plots into the expressions for calculating the flight path spacing and trigger spacing to obtain the UAV flight path and photo point set, including the primary flight path spacing, secondary flight path spacing, tertiary flight path spacing, primary trigger spacing, secondary trigger spacing, and tertiary trigger spacing.

[0030] S1.4: Control the drone to carry out aerial photography operations according to the drone flight path and the set of shooting points, and continuously record the drone's spatial position, attitude angle and shooting time during the flight to obtain a sequence of aerial images; Image sharpness evaluation methods from computer vision are used to filter and remove substandard aerial images from the aerial image sequence: the aerial images in the sequence are converted into grayscale images, and the sharpness evaluation value of each grayscale image is calculated using the Tenengrad gradient function, expressed as: ; in, The sharpness rating is... For the horizontal Sobel operator, This represents the convolution operation. For grayscale images at pixels grayscale value at that location The Sobel operator is for the vertical direction.

[0031] Based on the statistical distribution of historical sharpness evaluation values, a percentile method is used to set the sharpness threshold. For example, historical sharpness evaluation values ​​are arranged from smallest to largest, and the 5th percentile historical sharpness evaluation value is used as the sharpness threshold. The 5th percentile historical sharpness evaluation value is chosen because the top 5% of historical sharpness evaluation values ​​usually include unusable aerial images due to various factors. If the historical sharpness evaluation value exceeds 5%, the selection criteria will be too strict, which may cause aerial images of acceptable quality to be misjudged as unqualified. If the historical sharpness evaluation value is below 5%, the selection criteria will be too lenient, which may cause poor-quality aerial images to be retained, resulting in a decrease in the accuracy of the generated neural elevation field and 3D surface grid. Aerial images with a sharpness rating below the sharpness threshold are removed from the aerial image sequence to obtain a valid set of aerial images.

[0032] Using the capture time of each aerial image as an index, the aerial image, spatial location, attitude angle, and the corresponding land plot are bound together to form a multi-source view block. All multi-source view blocks are arranged in chronological order of capture time to form a multi-source view block sequence.

[0033] It should also be noted that by planning the drone flight paths based on the importance distribution of land parcels, high-density data collection of high-value remediation areas was achieved, providing a data foundation for subsequent high-precision 3D modeling and improving the accuracy and efficiency of multi-source data fusion.

[0034] S2: Perform local multi-view geometry solution and incremental optimization on the multi-source view block sequence to obtain the view block geometry parameter set. Use the view block geometry parameter set to train the neural elevation coding network online and generate gridded results.

[0035] S2.1: Using scale-invariant feature transform in computer vision methods to extract feature points and descriptors for each aerial image in a multi-source view block sequence: Gaussian blurring is performed on the aerial images according to the Gaussian standard deviation in multiples to obtain a set of Gaussian images at different scales. Multiple Gaussian images at different scales are stacked to form an image scale space. Gray-level difference operation is performed on Gaussian images at adjacent scales in the image scale space to obtain a Gaussian difference pyramid containing the Gaussian difference value of each pixel. Each pixel in the Gaussian difference pyramid is compared with the Gaussian difference value of all its neighboring pixels. When the Gaussian difference value of a pixel is less than or greater than the Gaussian difference value of all its neighboring pixels, the pixel is regarded as a candidate feature point. The gradient magnitude and direction of each pixel within the neighborhood of a candidate feature point are calculated using the following expression: ; in, For pixels gradient magnitude, For pixels The horizontal gradient, For pixels The vertical gradient, horizontal gradient and vertical gradient are obtained by convolving the Sobel operator with the Gaussian blurred image at the scale of the feature point; ; in, For pixels The direction, with a value range of [0°, 360°), It is a two-parameter arctangent function. For pixels The horizontal gradient, For pixels The vertical gradient; The orientation of each pixel is mapped to 36 bins according to its value range. For example, a pixel with an orientation of 15° corresponds to the 2nd bin, and a pixel with an orientation of 5° corresponds to the 1st bin. The gradient magnitude of each pixel in the neighborhood of the candidate feature point is accumulated to the corresponding bin to form an orientation gradient histogram. It should also be noted that bin represents the bar of the histogram of oriented gradients. The orientation of a pixel is mapped to 36 bins according to its value range because 36 bins can achieve a good balance between the discriminative power of features and the stability against noise. If the number of bins is too small, the orientation division is too coarse, and gradients in different directions may be assigned to the same bin, reducing the discriminative power of the descriptor and making it easier to make mistakes during feature matching. If the number of bins is too large, the orientation division is too fine. Small image noise, viewpoint changes, or candidate feature point localization errors may cause the gradient direction to jump from one bin to an adjacent bin, making the histogram peak unstable and thus reducing the stability of the descriptor against noise. The direction corresponding to the bin with the largest accumulated gradient magnitude in the directional gradient histogram is taken as the primary direction of the candidate feature point, and the direction corresponding to the bin with an accumulated gradient magnitude of not less than 80% of the primary direction is taken as the secondary direction of the candidate feature point. Since there are bins with accumulated gradient magnitudes of not less than 80% of the primary direction, it means that the candidate feature point has gradient information in multiple directions. Creating multiple different directions for the candidate feature point can ensure that at least one direction can match well when the image is rotated. If 90% is selected, it will be too strict and may miss some candidate feature points that do have stable multi-directionality, reducing the success rate of feature matching. If 60% is selected, it will be too lenient, causing many feature points to be assigned multiple directions. On the one hand, it increases the burden of calculation and matching, and on the other hand, it may introduce noise or unstable directions, which will reduce the accuracy of matching. The neighborhood of a candidate feature point is divided into multiple sub-blocks. For each sub-block, the gradient magnitudes of all pixels within the sub-block are accumulated into histograms in multiple directions. The histograms of all sub-blocks are then connected to obtain the descriptor. For each descriptor in an aerial image, the nearest Euclidean distance descriptor in another aerial image is searched as a candidate matching point. When the ratio of the nearest Euclidean distance to the second nearest Euclidean distance is less than 0.8, each descriptor is associated with its corresponding matching point to form a matching relationship between aerial images. 0.8 is chosen because it provides an optimal engineering balance between recall and precision. If 0.95 is chosen, the filtering criteria become very lenient, which may lead to a slight increase in recall but a sharp decrease in precision. If 0.7 is chosen, the filtering criteria become very strict, which may lead to a high precision but a sharp decrease in recall.

[0036] S2.2: Using the spatial position and attitude angle in the multi-source view block sequence as initial values, construct a linear equation system using multiple sets of descriptors and corresponding matching points, and solve the fundamental matrix of the linear equation system using the least squares method. Perform singular value decomposition on the fundamental matrix to obtain the camera relative rotation matrix and translation vector, i.e., the relative pose between adjacent view blocks.

[0037] The relative poses between adjacent view blocks are used to triangulate the descriptor and the corresponding matching point to obtain the 3D point coordinates. The camera exterior orientation parameters and the 3D point coordinates are then optimized by clustering adjustment with the goal of minimizing the reprojection error, using the 3D coordinates of the descriptor and the corresponding matching point to obtain the absolute camera exterior orientation parameters and pixel ray directions.

[0038] S2.3: Using the absolute camera exterior orientation parameters, pixel ray directions, and 3D point coordinates as initial values, construct the overall optimization objective function, expressed as: ; in, To optimize the objective function value overall, The total number of aerial images in a multi-source view block sequence. This represents the total number of three-dimensional point coordinates. For robust kernel functions, For camera projection function, For the first The camera relative rotation matrix of Zhang's aerial images For the first Translation vector of Zhang's aerial image For the first Three-dimensional point coordinates, For the first The coordinates of the 3D point at the th... The coordinates of two-dimensional image points observed on Zhang's aerial image. Represents the L2 norm; The overall optimization objective function is optimized nonlinearly using the bundle adjustment method: the camera exterior orientation parameters and 3D point coordinates are taken as parameters to be estimated, and the projection error of each 3D point coordinate is taken as the observation residual. The Gauss-Newton iteration method is used to linearize the overall optimization objective function step by step. The incremental vector of the parameters to be estimated is obtained by solving the normal equation. The incremental vector of the parameters to be estimated is superimposed on the parameters to be estimated to obtain the high-precision camera exterior orientation parameters and pixel ray direction set, i.e., the view block geometric parameter set. The expression for the normal equation is: ; in, For Jacobian matrices, Let be the increment vector of the parameter to be estimated. Indicates transpose. This is the increment vector of the observation residuals.

[0039] S2.4: For each pixel observation, query the elevation value and color prediction value of the sampling point in the neural elevation coding network along the corresponding pixel ray direction, read the real color value of the aerial image from the multi-source view block sequence, use the Softmax function to normalize the elevation value of each sampling point, use the normalized elevation value of each sampling point as the weight, and perform a weighted summation of the color prediction values ​​of all sampling points to obtain the composite color of the pixel; Pixel observation refers to observation data that includes pixel location and color value; The mean square error between the synthesized color and the real pixel color value is calculated. The weight parameters of the neural elevation coding network are updated using the stochastic gradient descent method with the goal of minimizing the mean square error, and the neural elevation field is obtained.

[0040] S2.5: Generate a geographic grid within the target area according to the specified grid resolution. Query the elevation values ​​of the neural elevation field at each vertex of the geographic grid. Calculate the standard deviation of the elevation values ​​at each vertex. Normalize the reciprocal of the standard deviation using min-max normalization to obtain the confidence score of the elevation values ​​at each vertex. Fill the confidence scores into the corresponding vertices to obtain the gridded result.

[0041] It should also be noted that the accuracy of the camera's exterior orientation parameters and 3D point coordinates was improved through geometric solutions and incremental optimization. By training the neural elevation field online, the dynamic construction of the neural elevation field was achieved, solving the problem of low accuracy in multi-source data fusion.

[0042] S3: Perform semantic projection and grid subdivision on the gridded results, and cluster them to obtain multiple remediation operation blocks. Use the preprocessed terrain data to construct a graph for each remediation operation block to obtain a plot instance scene graph.

[0043] S3.1: Using the high-precision camera exterior orientation parameters in the view block geometry parameter set, the aerial images of the multi-source view block sequence in S1.4 are projected onto the geographic grid corresponding to the gridded result to obtain the projected aerial images.

[0044] Historical aerial images of land reclamation areas with labeled land cover categories were collected as training samples. The number of iterations for the semantic segmentation model was set to 150, and the batch size was set to 16. This was because 150 iterations were sufficient for the loss curve of the semantic segmentation model to converge fully on the training set, while avoiding overfitting. If the number of iterations was higher than 150, it might increase the risk of overfitting, while if it was lower than 150, it might lead to underfitting. A batch size of 16 could achieve a good balance between gradient update stability and training efficiency. If the batch size was higher than 16, it would place higher demands on hardware and might affect the generalization ability of the semantic segmentation model. If the batch size was lower than 16, it might lead to unstable training and lower efficiency.

[0045] The training samples are input into the semantic segmentation model. The encoder performs convolution and pooling operations on the training samples to obtain training semantic feature maps. The decoder performs upsampling and convolution operations on the training semantic features to obtain prediction probability feature maps for each land cover category. The Softmax function is used to map the prediction probability feature maps of each land cover category to the prediction probability distribution of each land cover category. The prediction probabilities of all pixels in the same grid are averaged to obtain the prediction probability of the land cover category for each grid. The cross-entropy loss value between the prediction probability of each land cover category and the labeled land cover category is calculated. The partial derivative of the cross-entropy loss value with respect to the parameters of the semantic segmentation model is calculated using the backpropagation method. The parameters of the semantic segmentation model are iteratively updated using the stochastic gradient descent method with the goal of minimizing the cross-entropy loss value until the required number of iterations is reached, resulting in a trained semantic segmentation model.

[0046] The projected aerial imagery is input into the trained semantic segmentation model, and the forward propagation operation is performed during the training process of the semantic segmentation model to obtain the probability of each land cover category in the grid. The land cover category with the highest probability among the land cover categories in the grid is used as the semantic label of the grid. The semantic label is written into the corresponding grid to obtain the semantically meshed result.

[0047] S3.2: Read the planar coordinates of each grid vertex in the semantically meshed result, take the elevation value of the neural elevation field at the planar coordinates of the corresponding grid vertex as the height, merge it with the planar coordinates to form the three-dimensional spatial coordinates of the grid vertex, and connect all grid vertices into a triangular mesh according to the topological relationship of the semantic labels to form a three-dimensional surface mesh.

[0048] S3.3: Calculate the slope and undulation of each triangular facet in the three-dimensional surface grid, use the slope and undulation as clustering features, and use the mean-shift clustering algorithm to cluster all triangular facets to obtain the remediation work blocks.

[0049] Furthermore, the cosine of the angle between the normal vector and the perpendicular vector of each triangular facet is calculated. The cosine of the angle is then substituted into the arcsine function, and the complementary angle is taken to obtain the slope of the triangular facet. The perpendicular vector is a unit vector perpendicular to the sea level.

[0050] Calculate the standard deviation of the difference between the elevation values ​​of the triangular patch and all neighboring triangular patches, and use it as the undulation of the triangular patch; Max-min normalization was used to normalize the slope and undulation of the triangular facets respectively. The bandwidth of the mean-drift clustering algorithm was set to 0.2 and the convergence threshold to 0.001. This is because a bandwidth of 0.2 means that the influence radius of each cluster center point covers 20% of the feature space span, which can achieve a good balance between avoiding the generation of too many small clusters and preventing different terrain features from being inappropriately merged into one cluster. A bandwidth greater than 0.2 may lead to over-smoothing and loss of important details, while a bandwidth less than 0.2 may be too sensitive to noise and may divide continuous terrain areas into a large number of fragmented small clusters, destroying the integrity of the remediation blocks. A convergence threshold of 0.001 can balance computational accuracy and efficiency. A convergence threshold greater than 0.001 may cause the mean-drift clustering algorithm to terminate prematurely and the clustering results to be inaccurate. A convergence threshold less than 0.001 may increase the number of unnecessary iterations and increase the computational cost, while the improvement in the accuracy of the clustering results is limited. The slope and undulation of each normalized triangular facet are combined as cluster feature vectors. Each cluster feature vector is regarded as an initial cluster center. For each initial cluster center, the mean of all feature vectors within the bandwidth is calculated, and the initial cluster center is moved to the position of the mean of the feature vectors. The moving step is iterated until the Euclidean distance moved by all cluster centers during the iteration is less than the convergence threshold. When all cluster centers are determined to have converged to the density peak, all triangular faces that have converged to the same density peak are classified into the same category. All triangular faces of the same category are merged into a continuous region to form a remediation block.

[0051] S3.4: Perform spatial overlay analysis on the boundary of the remediation operation block and the land parcel boundary in the preprocessed remediation planning data. If the remediation operation block is completely located within the boundary of a certain land parcel, establish the attribution mapping from the remediation operation block to the corresponding land parcel. If the remediation operation block spans multiple land parcels, divide the remediation operation block according to the land parcel boundary to obtain sub-blocks, and establish the mapping from the divided sub-blocks to the corresponding land parcels to obtain the mapping information between land parcels and remediation operation blocks.

[0052] S3.5: Abstract the land parcels as parent nodes and the remediation work blocks as child nodes. Establish adjacency edges between adjacent child nodes and establish belonging edges between child nodes and their corresponding parent nodes according to the mapping information to form a land parcel relationship graph.

[0053] S3.6: Bind attributes to each parent node and child node in the land parcel relationship diagram: Calculate the average elevation, average slope, and average undulation of the area represented by each parent node and child node as attribute indicators, bind the attribute indicators to the corresponding parent node and child node, and form a land parcel instance scene diagram.

[0054] It should also be noted that the semantic information of the 3D mesh is enhanced by semantic projection, and the plots are associated with the remediation work blocks by the mean-shift clustering algorithm and the construction of a plot relationship graph, forming a plot instance scene graph, which solves the problem of weak semantic association of plot remediation.

[0055] S4: Visualize and interact with the plot instance scene map, generate remediation indicators, and update the plot instance scene map to obtain the updated plot instance scene map.

[0056] S4.1: Load the plot instance scene map into the 3D visualization platform for rendering and display. Receive and respond to the geometric shapes drawn by the user through the input device in the visualization interface, and record the vertex coordinates of the geometric shapes. It should also be noted that the geometry includes the target elevation value at the vertex coordinates.

[0057] The geometry is overlaid onto the scene graph of the land parcel instance, and the vertex coordinates are written to the parent and child nodes representing the corresponding land parcel and the remediation operation block to obtain the remediation indicators; The difference between the target elevation value at the treatment index and the corresponding elevation value in the neural elevation field is used as the elevation adjustment amount. The stochastic gradient descent method is used to minimize the elevation adjustment amount as the objective. The weight parameters of the neural elevation coding network are updated within the geometric coverage area to obtain the target terrain.

[0058] S4.2: Query the elevation values ​​of the target terrain at each grid vertex within the geometric coverage area, and generate the local meshing results according to the steps in S2.5. Use the local meshing results to generate the updated 3D surface mesh according to the steps in S3.1 and S3.2. Using the updated 3D surface mesh, rebind the attributes of the parent and child nodes according to the steps in S3.6 to obtain the updated plot instance scene map.

[0059] It should also be noted that: the intuitive generation and real-time adjustment of remediation indicators are achieved through visual interaction, and the neural elevation field and plot instance scene map are updated based on the remediation indicators, realizing incremental updates of terrain and remediation indicators, and improving the interactivity and flexibility of land remediation planning.

[0060] This embodiment also provides a high-precision 3D land modeling and optimization system based on UAV remote sensing, including: The module collects and preprocesses terrain data, improvement planning data, and UAV remote sensing data of the target area to construct a multi-source view block sequence; The training module performs local multi-view geometry solving and incremental optimization on the multi-source view block sequence to obtain a set of view block geometry parameters. The set of view block geometry parameters is then used to train the neural elevation coding network online to generate a gridded result. The instance module performs semantic projection and grid subdivision on the gridded results, and performs clustering to obtain multiple remediation operation blocks. The preprocessed terrain data is used to construct a graph for each remediation operation block to obtain a plot instance scene graph. The update module provides interactive visualization of the land parcel instance scene map, generates remediation indicators, and updates the land parcel instance scene map to obtain the updated land parcel instance scene map.

[0061] This embodiment also provides a computer device applicable to the land high-precision 3D modeling optimization method based on UAV remote sensing, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the land high-precision 3D modeling optimization method based on UAV remote sensing proposed in the above embodiment.

[0062] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0063] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the method for optimizing high-precision 3D land modeling based on UAV remote sensing, as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0064] In summary, this invention addresses the problem of low accuracy in multi-source data fusion by establishing a land 3D modeling optimization process based on neural elevation coding, enabling dynamic reconstruction of terrain elevation on the basis of multi-source remote sensing data fusion, achieving high-density acquisition and high-precision 3D reconstruction of high-value remediation areas; and by constructing the neural elevation field as a plot instance scene map, it realizes incremental updates of terrain and remediation indicators and real-time adjustment of scene map attributes, thus solving the problem of weak semantic association in plot remediation.

[0065] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for optimizing high-precision 3D land modeling based on UAV remote sensing, characterized by: include, Collect and preprocess terrain data, improvement planning data, and UAV remote sensing data of the target area to construct a multi-source view block sequence; Local multi-view geometry solution and incremental optimization are performed on the multi-source view block sequence to obtain the view block geometry parameter set. The view block geometry parameter set is then used to train the neural elevation coding network online to generate gridded results. Semantic projection and grid subdivision are performed on the gridded results, and clustering is carried out to obtain multiple remediation operation blocks. The preprocessed terrain data is used to construct a graph for each remediation operation block to obtain a scene graph of the land parcel instance. The system provides interactive visualization of the land parcel instance scene map, generates remediation indicators, and updates the land parcel instance scene map to obtain the updated land parcel instance scene map.

2. The method for optimizing high-precision 3D land modeling based on UAV remote sensing according to claim 1, characterized in that: The process of constructing a multi-source view block sequence using preprocessed terrain data, remediation planning data, and UAV remote sensing data involves the following specific steps. The importance of each plot is calculated using the preprocessed remediation planning data, and the distribution of plot importance is obtained. Based on the distribution of land parcel importance, preprocessed terrain data, and UAV remote sensing data, the flight path of the UAV is planned to obtain the UAV flight path and the set of photo points. Control the drone to perform aerial photography based on the drone's flight path and the set of photo points to obtain an aerial image sequence. Then, aggregate the aerial image sequence to obtain a multi-source view block sequence.

3. The method for optimizing high-precision 3D land modeling based on UAV remote sensing according to claim 1, characterized in that: The process of performing local multi-view geometry solving and incremental optimization on the multi-source view block sequence to obtain the view block geometry parameter set is as follows: Local feature extraction and matching are performed on the multi-source view block sequence, and local multi-view geometry is solved to obtain the absolute camera exterior orientation parameters and pixel ray directions; Incremental optimization of the absolute camera exterior orientation parameters and pixel ray directions yields a set of view block geometric parameters.

4. The method for optimizing high-precision 3D land modeling based on UAV remote sensing according to claim 1, characterized in that: The online training of the neural elevation coding network using a set of view block geometric parameters to generate a gridded result follows these steps. The neural elevation field is obtained by using the view block geometry parameter set to train the local region of the neural elevation coding network through online backpropagation. The neural elevation field is sampled in a gridded manner to obtain the gridded results.

5. The method for optimizing high-precision 3D land modeling based on UAV remote sensing according to claim 1, characterized in that: The specific steps for semantic projection and grid subdivision of the gridded results are as follows. Semantic projection is performed on the meshing results using a multi-source view block sequence to obtain semantically meshed results. The semantic meshing results are extended to three-dimensional space using a neural elevation field to obtain a three-dimensional surface mesh.

6. The method for optimizing high-precision 3D land modeling based on UAV remote sensing according to claim 1, characterized in that: The preprocessed terrain data is used to construct maps for each remediation operation block, resulting in instance scene maps of the land parcels. The specific steps are as follows. The mapping information between land parcels and remediation work areas is determined based on the preprocessed terrain data and remediation planning data; Using mapping information, land parcels and remediation work areas are abstracted into nodes, and edge relationships are established between each node to obtain a land parcel relationship graph; The parcel relationship graph is extended into a parcel instance scene graph using neural elevation fields and semantically meshed results.

7. The method for optimizing high-precision 3D land modeling based on UAV remote sensing according to claim 1, characterized in that: The specific steps for visualizing and interacting with the plot instance scene map to generate remediation indicators are as follows. Load the plot instance scene map into the 3D visualization platform, and respond to the user's drawing operations in the visualization interface to obtain the geometric shape drawn by the user; The user-drawn geometry is bound to the corresponding node, and the node is updated to obtain the governance indicators.

8. The method for optimizing high-precision 3D land modeling based on UAV remote sensing according to claim 1, characterized in that: The steps for updating the plot instance scene map to obtain the updated plot instance scene map are as follows. The target terrain is obtained by locally constraining and adjusting the neural elevation field using the remediation index; The target terrain is used to adjust the plot instance scene map to obtain the updated plot instance scene map.

9. The method for optimizing high-precision 3D land modeling based on UAV remote sensing according to claim 8, characterized in that: The specific steps for adjusting the plot instance scene map using the target terrain are as follows. Based on the target terrain and remediation indicators, a local gridded result is generated, and the three-dimensional surface grid is locally replaced to obtain an updated three-dimensional surface grid. The updated 3D surface grid is used to adjust the plot relationship map and then expanded to obtain the updated plot instance scene map.

10. A high-precision 3D land modeling and optimization system based on UAV remote sensing, based on the high-precision 3D land modeling and optimization method based on UAV remote sensing as described in any one of claims 1 to 9, characterized in that: include, The module collects and preprocesses terrain data, improvement planning data, and UAV remote sensing data of the target area to construct a multi-source view block sequence; The training module performs local multi-view geometry solving and incremental optimization on the multi-source view block sequence to obtain a set of view block geometry parameters. The set of view block geometry parameters is then used to train the neural elevation coding network online to generate a gridded result. The instance module performs semantic projection and grid subdivision on the gridded results, and performs clustering to obtain multiple remediation operation blocks. The preprocessed terrain data is used to construct a graph for each remediation operation block to obtain a plot instance scene graph. The update module provides interactive visualization of the land parcel instance scene map, generates remediation indicators, and updates the land parcel instance scene map to obtain the updated land parcel instance scene map.

Images