A forest structure reconstruction system based on three-dimensional laser scanning point cloud modeling

Abstract

The present application relates to the technical field of three-dimensional reconstruction, and proposes a forest structure reconstruction system based on three-dimensional laser scanning point cloud modeling. The system acquires forest point cloud data through multi-source laser scanning, and completes data preprocessing in combination with auxiliary perception information; on this basis, a hierarchical feature expression mode is constructed, a three-constraint joint mechanism and a multi-source feature balance mechanism are introduced, and the accurate separation of tree trunks, branches and canopies is realized; further, the cross-scale structure expression capability is enhanced through structure perception feature propagation; in the three-dimensional reconstruction stage, based on skeleton topology and parameter-driven modeling, in combination with complex area compensation and adaptive refinement strategy, high-fidelity reconstruction of forest structure is realized; finally, single-tree structure model and forest structure parameters are output. The present application improves the point cloud segmentation precision and reconstruction quality, and is suitable for forest resource investigation, carbon storage evaluation and ecological monitoring scenes.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional reconstruction technology, and in particular to a forest structure reconstruction system based on three-dimensional laser scanning point cloud modeling. Background Technology

[0002] In the field of forest structure modeling based on 3D laser scanning point cloud data, existing intelligent technologies have gradually introduced deep learning methods (such as PointNet, PointNet++, SparseConv, Mask3D, etc.) to perform semantic segmentation or instance segmentation on point cloud data, and combined geometric fitting, graph clustering, and surface reconstruction algorithms to achieve 3D structure representation. However, due to the significant multi-level structural features and strong spatial coupling relationships in forest scenes, existing methods still have several key technical bottlenecks in practical applications: First, existing point cloud instance segmentation models mostly rely on local geometric features or sparse convolutional features, lacking a unified modeling capability for cross-scale structural continuity and hierarchical topological relationships, resulting in problems in the trunk-branch-canopy transition zone. First, the domain is prone to instance sticking or over-segmentation problems. Second, multi-source features (such as geometric structural features, spatial layer features, and echo attribute features) usually adopt simple splicing or fixed weight mechanisms during the fusion process, lacking the ability to adaptively adjust the information contribution of different structural regions, which easily leads to single feature dominance, thereby reducing the recognition accuracy of fine branch structures, canopy boundaries, and sparse regions. Third, in the process of constructing regional features, existing methods are mostly based on static adjacency relationships or local receptive fields for feature propagation, which is difficult to effectively characterize the cross-level propagation path of point clouds in complex spatial structures and global contextual dependencies, limiting the overall representation ability of occluded areas and structural fracture areas. In addition, in the 3D reconstruction stage, the traditional combination method based on cylinder fitting and Alpha Shape lacks a collaborative modeling mechanism for local surface continuity and multi-scale detail restoration when dealing with trunk and branch connection areas and canopy boundary areas, which easily produces surface fractures, over-smoothing, or local geometric distortion, making it difficult to achieve a unified expression of structural continuity and detail realism. Summary of the Invention

[0003] This invention addresses the shortcomings of existing intelligent forest point cloud processing and 3D reconstruction technologies, such as insufficient multi-level structural representation, imbalanced feature fusion, and limited accuracy in reconstructing complex surfaces. It proposes a forest structure reconstruction system based on 3D laser scanning point cloud modeling. The core of this system lies in constructing a unified structure-aware modeling mechanism that spans "point cloud representation—instance segmentation—structure reconstruction." In the point cloud instance segmentation stage, a hierarchical feature representation framework of "local point blocks—regional aggregation" is constructed. A three-constraint joint mechanism, including statistical structural consistency, principal direction coupling consistency, and adjacency topological continuity, is introduced to collaboratively constrain the structural relationships between features at different levels. Simultaneously, a multi-source feature mutual information balancing mechanism is combined to adaptively adjust the fusion contributions of geometric structural features, spatial layer features, and echo attribute features. This maintains the continuous evolutionary relationship between the trunk, branches, and canopy in the feature space and suppresses single-source features. The problem is primarily driven by a feature-based approach. Furthermore, by introducing a structure-aware random walk-driven cross-point block feature propagation mechanism, the co-occurrence relationship of paths and the stability of cross-layer propagation are modeled, enabling adaptive aggregation of regional contextual features to enhance the overall expressive power for occluded regions, sparse regions, and complex overlapping structures. In the 3D reconstruction stage, a layered reconstruction strategy based on skeleton topology and parameter-driven approaches is employed. For complex surface regions such as trunk-branch connection areas and canopy boundaries, a nonlinear mapping function integrating local coordinates, structural orientation, and geometric features is constructed to jointly compensate for local surface continuity and normal variations. Simultaneously, an adaptive refinement mechanism driven by frequency domain errors is combined to enhance the reconstructed surface with multi-scale details, thereby improving the recovery capability of local geometric details while ensuring overall topological continuity. Ultimately, this achieves high-precision and robust reconstruction of the forest 3D structure model.

[0004] This invention proposes a forest structure reconstruction system based on three-dimensional laser scanning point cloud modeling. The system includes: a forest point cloud acquisition module, an auxiliary perception acquisition module, an edge preprocessing module, a point cloud instance segmentation module, a single tree clustering segmentation module, a skeleton extraction and cylinder fitting module, a surface reconstruction and inversion module, and a result management and application output module.

[0005] The forest point cloud acquisition module obtains the original forest point cloud dataset.

[0006] The auxiliary sensing and acquisition module acquires multi-source auxiliary observation datasets;

[0007] The edge preprocessing module performs time synchronization, coordinate unification, noise point removal, outlier removal, drift compensation, occlusion marking, spatial pruning, and quality assessment on the original forest point cloud dataset and the multi-source auxiliary observation dataset, and outputs a preprocessed point cloud dataset.

[0008] The point cloud instance segmentation module constructs a 3D instance segmentation model based on Mask3D; it performs hierarchical block encoding on the preprocessed point cloud dataset to form local point block feature representations and region-level aggregated feature representations; it inputs the local point block feature representations and region-level aggregated feature representations into the 3D instance segmentation model, and introduces a three-constraint joint mechanism and a multi-source feature mutual information balance constraint mechanism to collaboratively optimize the feature representation during the training process of the 3D instance segmentation model, generating forest point cloud segmentation results with instance identifiers and category labels, and outputting a set of instance-labeled point clouds;

[0009] The single-tree clustering and segmentation module constructs a point cloud similarity map based on the trunk, branch, and canopy instances in the instance-annotated point cloud set. It then performs spectral clustering based on spatial proximity, structural continuity, and canopy connectivity to aggregate multiple types of point cloud instances belonging to the same tree into single-tree candidate objects and outputs a single-tree level point cloud object set.

[0010] The skeleton extraction and cylinder fitting module performs structural analysis and parametric modeling on a single tree-level point cloud object set, and outputs a tree skeleton topology model and a cylinder parametric model.

[0011] The surface reconstruction and inversion module performs parameter-driven geometric reconstruction of trunks and branches based on the tree skeleton topology model and cylindrical parameter model. For the canopy point cloud subset within the single-tree level point cloud object set, it employs the Alpha Shape reconstruction algorithm to perform boundary extraction and surface generation, identifying complex surface regions. Local continuous surface compensation and frequency domain adaptive refinement are then performed on these complex surface regions. Based on this, unified spatial registration and organization of the trunk, branches, and canopy structure are performed to construct single-tree and scene-level 3D forest structure models. These models are then structurally encapsulated and formatted to output a forest structure reconstruction model. Based on the tree skeleton topology model, cylindrical parameter model, and forest structure reconstruction model, the module calculates single-tree height, diameter at breast height (DBH), crown width, canopy projected area, canopy volume, branch height, trunk inclination angle, branch angle distribution, canopy closure, and stand density, forming a forest structure parameter dataset.

[0012] The results management and application output module performs model storage and version management on the forest structure parameter dataset and conducts statistical analysis of indicators. Based on this, it performs 3D visualization and report generation on the forest structure parameter dataset and corresponding 3D structure model, and publishes and shares it through external system interfaces. It outputs forest structure reconstruction results, individual tree attribute files and sample plot structure statistical results, as well as application results for forest resource surveys, carbon storage accounting, management assessment and disaster loss analysis.

[0013] Furthermore, the local point block feature representation and the region-level aggregated feature representation are input into the 3D instance segmentation model. During the training process of the 3D instance segmentation model, a three-constraint joint mechanism and a multi-source feature mutual information balance constraint mechanism are introduced to collaboratively optimize the feature representation, and the process of outputting the instance annotation point cloud includes the following steps:

[0014] Step S1: Perform hierarchical block encoding on the preprocessed point cloud dataset according to the preset spatial partitioning rules; specifically, based on the spatial coordinate distribution of the point cloud, the point density variation and the vertical height layer information, perform local clustering partitioning on the discrete points of the point cloud in the preprocessed point cloud dataset to form multiple local point block units; then, according to the hierarchical organization method from the fine-grained local point block layer to the coarse-grained regional aggregation layer, perform hierarchical indexing and association organization on each local point block unit to form a point block representation set with hierarchical topological relationship;

[0015] Step S2: Perform local feature extraction processing on each local point block in the point block representation set to generate local point block feature representation; specifically, extract geometric structure features, spatial layer features and echo attribute features from the point cloud in each local point block, and perform joint encoding on the extracted multi-class features to obtain a local point block feature representation that characterizes the current local point block morphology, layer and echo response.

[0016] Step S3: Based on the local point block feature representation, construct the point block association structure by combining the relationship between each local point block; on this basis, perform high-order topological association modeling on multiple local point blocks that meet the requirements of spatial adjacency, morphological continuity and hierarchical correlation to form high-order association units; and perform structure-aware random walk sampling on the point block association structure and high-order association units to form regional aggregated feature representation.

[0017] Step S4: Input the local point block feature representation and the region-level aggregated feature representation into the 3D instance segmentation model, perform candidate instance generation and instance mask prediction processing, and in the process of candidate instance generation, perform aggregation processing on multiple local point block feature representations belonging to the same candidate instance to generate candidate instance aggregated feature representation;

[0018] Step S5: During the parameter training phase of the 3D instance segmentation model, a three-constraint joint mechanism is introduced to perform collaborative consistency constraints on the statistical structure, main direction coupling relationship, and adjacency topology relationship between the local point block feature representation and the candidate instance aggregate feature representation.

[0019] Step S6: Under the collaborative constraint framework of the three-constraint joint mechanism, a multi-source feature mutual information balance constraint mechanism is introduced during the training process of the 3D instance segmentation model to perform adaptive balance adjustment on the information contribution of geometric structure features, spatial layer features and echo attribute features in the process of forming candidate instance aggregation feature representation.

[0020] Step S7: Based on the 3D instance segmentation model jointly optimized by the three-constraint joint mechanism and the multi-source feature mutual information balance constraint mechanism, perform 3D instance recognition, instance attribution determination and semantic separation processing on tree trunk instances, branch instances, canopy instances, understory vegetation instances and ground surface instances in the forest point cloud, generate forest point cloud segmentation results with instance identifiers and category labels, and output instance labeled point cloud set according to the preset point cloud organization format.

[0021] Furthermore, step S3 specifically includes the following steps:

[0022] Step S31: Based on the feature representation of local points and blocks, perform joint analysis on the spatial coordinate position, geometric extension features and vertical height layer information of each local point and block to determine the spatial proximity relationship, structural continuity relationship and inter-layer progression relationship between local points and blocks; and perform association organization on local points and blocks according to the spatial proximity relationship, structural continuity relationship and inter-layer progression relationship to construct the point and block association structure.

[0023] Step S32: Based on the point block association structure, perform high-order topological association modeling on multiple local point blocks that meet the requirements of spatial adjacency, morphological continuity and hierarchical correlation, and organize multiple local point blocks with multiple common affiliations in the same tree structure chain, the same branch extension area, the same canopy area or the same continuous surface area into high-order association units.

[0024] Step S33: On the point-block association structure and higher-order association units, take each local point block as the starting point, perform structure-aware random walk sampling according to the path selection criteria of spatial proximity priority, structural continuity priority and inter-layer progression priority, and generate the corresponding random walk sequence; and construct cross-point block feature propagation weights based on the spatial proximity, structural continuity and layer association between local point blocks.

[0025] Step S34: Based on the random walk sequence, perform statistical analysis on the path co-occurrence frequency, continuous arrival intensity and cross-layer propagation stability between local points and blocks, and adaptively adjust the cross-point block feature propagation weights according to the statistical results to obtain the adjusted cross-point block feature propagation weights.

[0026] Step S35: Combining the multi-dimensional structural coupling relationship represented by the higher-order correlation unit and the adjusted cross-point block feature propagation weight, perform regional-level joint aggregation processing on the local point block features belonging to the same candidate tree, the same canopy area, or the same continuous surface area. This allows the local point blocks to further integrate multi-step structural path context information while maintaining local geometric morphology information, vertical layer distribution information, and echo response difference information, thus forming a regional-level aggregated feature representation.

[0027] Furthermore, obtaining the forest structure reconstruction model specifically includes the following steps:

[0028] Step B1: Based on the tree skeleton topology model and cylindrical parameter model, perform parameter-driven geometric reconstruction on the trunk segment and branch segment to form the initial surface of the trunk and the initial surface of the branches; at the same time, for the canopy point cloud subset in the single tree level point cloud object set, use the Alpha Shape reconstruction algorithm to perform canopy boundary extraction and initial surface generation to obtain the initial surface of the canopy.

[0029] Step B2: Based on the connection relationship between the trunk node and branch node in the tree skeleton topology model, and combined with the local geometric continuity characteristics of the initial surface of the trunk and the initial surface of the branches, determine the connection region between the trunk and branches; and based on the initial surface of the canopy, combined with the local point density variation, normal dispersion, boundary curvature fluctuation and spatial distribution unevenness of the corresponding canopy point cloud subset, determine the canopy boundary variation region; the trunk and branch connection region and the canopy boundary variation region are combined as the complex surface region;

[0030] Step B3: For each complex surface region, construct a local coordinate domain and map the point cloud data and corresponding structural orientation information within the region to the local coordinate domain; within the local coordinate domain, construct a nonlinear mapping function, and use local coordinates, structural orientation information, and local geometric features as inputs to jointly estimate the local surface offset, normal correction, and surface undulation trend in the trunk-branch connection area and the canopy boundary change area; perform local continuous surface compensation based on the estimation results;

[0031] Step B4: After completing the local continuous surface compensation, the compensated local reconstructed surface is obtained; frequency domain consistency analysis is performed on the compensated local reconstructed surface and the corresponding original point cloud region to obtain the frequency domain error distribution; based on the frequency domain error distribution, adaptive thinning processing is performed on the compensated local reconstructed surface to form an enhanced surface structure set.

[0032] Step B5: Based on the enhanced surface structure set, perform unified spatial registration, topological connection, and hierarchical organization processing on the trunk surface, branch surface, and canopy surface; specifically, based on the structural region marking information, apply continuity constraints and connection consistency correction to the trunk-branch connection region and the canopy boundary region to ensure that each structural unit maintains geometric continuity and topological consistency in the same single-tree coordinate system; further, perform scene-level organization and integration on multiple single-tree structural models according to the spatial distribution relationship of the sample plots to construct single-tree and scene-level forest three-dimensional structural models, and output the forest structure reconstruction model.

[0033] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0034] This invention constructs a hierarchical feature representation framework of "local point-block - region aggregation" and introduces a three-constraint joint mechanism to achieve unified modeling and collaborative expression of multi-level structural relationships such as tree trunks, branches, and canopies in forest point clouds. This improves the continuity and separability of different structural units in the feature space and solves the problems of structural adhesion, blurred boundaries, and inaccurate segmentation that existing instance segmentation methods are prone to in complex forest stand environments. This invention enhances the instance recognition accuracy and segmentation stability of high-density forest stands, structurally intersecting areas, and occlusion scenarios, thereby providing high-quality and structurally consistent input data support for subsequent single-tree-level structural reconstruction.

[0035] This invention introduces a multi-source feature mutual information balancing mechanism and a structure-aware random walk-driven feature propagation mechanism to achieve adaptive fusion and dynamic weight adjustment of geometric structural features, spatial layer features, and echo attribute features in the instance modeling process. This enhances the collaborative expression ability of multi-source information in different structural regions and solves the problem of insufficient fine-grained structure recognition ability caused by single feature dominance in the prior art. It also enhances the overall representation ability and robustness of fine branch structures, canopy boundaries, understory vegetation interlacing areas, and sparse point cloud areas, thus enabling the invention to obtain more stable and complete structural expression results in complex forest scenarios.

[0036] This invention constructs a hierarchical reconstruction mechanism based on skeleton topology and parameter-driven principles, and combines nonlinear mapping compensation for complex surface regions with an adaptive refinement strategy driven by frequency domain errors. This enables high-fidelity modeling of key structures such as trunk-branch connection regions and canopy boundaries, improving the coordination between local detail restoration and overall structural continuity in the 3D reconstruction model. It solves the problems of surface breakage, over-smoothing, and local morphological distortion that easily occur in traditional reconstruction methods, and enhances the accuracy and reliability of forest structure models in single-tree parameter inversion and scene-level modeling. Thus, it significantly improves the practical value and engineering applicability of this invention in applications such as forest resource surveys, carbon storage accounting, and management assessment. Attached Figure Description

[0037] Figure 1 This invention proposes a forest structure reconstruction system based on three-dimensional laser scanning point cloud modeling.

[0038] Figure 2This diagram illustrates the comparison of point cloud instance segmentation results between Example 1 and Example 2. The left image shows the segmentation result obtained using a 3D instance segmentation model optimized with a three-constraint joint mechanism and a multi-source feature mutual information balance constraint mechanism. This model can accurately distinguish the trunk, branches, and canopy in the tree structure. In the segmentation result, the boundaries of each structural region are clear, the hierarchical relationship is well-defined, and the trunk, branches, and canopy maintain good structural continuity and spatial consistency, without obvious structural overlap. In contrast, the right image shows the segmentation result obtained using a conventional optimization method based on instance segmentation cross-entropy loss function and mask prediction loss function. The canopy region exhibits a blurred canopy with low outline clarity; structural overlap exists in the trunk and branch regions, the overall structural separation ability is weak, and it is difficult to accurately reflect the true spatial relationship between the various components of the tree. Detailed Implementation

[0039] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0040] Example 1, according to Figure 1 , Figure 2 This invention proposes a forest structure reconstruction system based on three-dimensional laser scanning point cloud modeling. The system includes: a forest point cloud acquisition module, an auxiliary perception acquisition module, an edge preprocessing module, a point cloud instance segmentation module, a single tree clustering segmentation module, a skeleton extraction and cylinder fitting module, a surface reconstruction and inversion module, and a result management and application output module.

[0041] In this embodiment, area A, a mixed artificial and natural regeneration forest plot, was selected as the test area. The plot area was set at 100 m × 100 m, totaling 10,000 m², with an average slope of 18.6° and an altitude range of 1268 m to 1315 m. The main tree species in the plot included Chinese fir, Masson pine, oak broad-leaved trees, and a small number of shrubs, with an understory vegetation coverage of approximately 34.7%. A total of 428 valid sample trees were planted in the plot, of which 391 were trees with a diameter at breast height (DBH) greater than 5 cm, with an average DBH of 18.4 cm, an average tree height of 13.7 m, and a crown width range of 1.2 m to 6.8 m.

[0042] The forest point cloud acquisition module consists of multiple types of laser scanning acquisition units, including ground-based fixed laser scanning equipment, vehicle-mounted mobile laser scanning equipment, and UAV-borne laser scanning equipment. Each acquisition unit works collaboratively for different application scenarios. The ground-based fixed laser scanning equipment is used to acquire high-precision, omnidirectional static point cloud data. The vehicle-mounted mobile laser scanning equipment is installed on the roof of the inspection vehicle and performs continuous scanning operations along the forest patrol route to acquire information on the spatial structure of the lower and middle layers of the forest stands on both sides of the road. The UAV-borne laser scanning equipment is mounted under a multi-rotor UAV to perform a strip-like scan of the canopy area. This module acquires data on the upper structure of forest areas to build a forest point cloud data acquisition system covering different vertical levels. During the data acquisition process, based on the preset acquisition path planning results and station layout scheme, the module performs multi-point, multi-observation angle, and cross-time period laser scanning operations on tree trunks, branches, canopies, understory vegetation, and surface undulations within the target forest area. Through the above acquisition process, raw point cloud data containing spatial three-dimensional coordinate information, echo intensity information, scanning direction information, timestamp information, and device pose parameters are acquired. The data is then structured and encapsulated to form a raw forest point cloud dataset.

[0043] In this embodiment, a pulsed terrestrial 3D laser scanner was used, with a total of 16 fixed stations set up. The station spacing was 20 m to 30 m. The horizontal scanning angle of a single station was 360°, the vertical scanning range was -45° to +135°, the angular resolution was set to 0.035°, the ranging accuracy was ±2 mm, and the average number of points acquired per station was approximately 18.5 million. The scanning time for each station was approximately 6 minutes, and a total of approximately 296 million static point clouds of the ground were acquired. These points were mainly used to acquire the lower and middle parts of tree trunks and branches, as well as the spatial structure of the understory.

[0044] The mobile laser scanning device was mounted on the top of the inspection vehicle at a height of 2.8 m. It was used to collect data along the central forest road and surrounding passageways of the sample plot. The driving speed was controlled at 8 km / h, the scanning frequency was 320 kHz, the line scanning frequency was 100 Hz, and the average point spacing was approximately 2.5 cm. A total of 4 paths were scanned, with a cumulative driving length of 2.7 km, and approximately 142 million points of spatial structure of the middle and lower layers of the forest stand were obtained on both sides of the road.

[0045] A multi-rotor UAV equipped with an airborne laser scanning device was used. The flight altitude was set at 55 m, the flight strip overlap rate was set at 35%, the flight speed was 4.5 m / s, the pulse repetition frequency was 240 kHz, and the scanning field of view was 70°. A total of 8 flight strips were scanned, and approximately 118 million points of the upper canopy point cloud were obtained. The average ground point density was 168 pts / m², and the average point density in the canopy area was 124 pts / m².

[0046] The auxiliary sensing and acquisition module is set up in conjunction with the forest point cloud acquisition module and communicates with the edge preprocessing module. It synchronously acquires GNSS positioning data, IMU attitude data and environmental meteorological data, provides positioning calibration, attitude compensation and environmental calibration support for the raw point cloud data, and outputs a multi-source auxiliary observation dataset.

[0047] In this embodiment, the GNSS uses a dual-frequency RTK positioning device with a horizontal positioning accuracy of ±1.5 cm and an elevation positioning accuracy of ±2.5 cm. The IMU sampling frequency is set to 200 Hz, and the attitude angle measurement error does not exceed 0.08°. The meteorological acquisition unit records the ambient temperature, humidity, wind speed, and light intensity every 10 seconds. During the acquisition period, the average air temperature was 24.3 ℃, the average relative humidity was 68.5%, and the average wind speed was 1.6 m / s.

[0048] The edge preprocessing module performs time synchronization, coordinate unification, noise point removal, outlier removal, drift compensation, occlusion marking, spatial pruning, and quality assessment on the original forest point cloud dataset and the multi-source auxiliary observation dataset, and outputs a preprocessed point cloud dataset.

[0049] The point cloud instance segmentation module is deployed in the graphics processing node of the cloud computing platform and communicates with the edge preprocessing module and the single tree clustering segmentation module to construct a 3D instance segmentation model based on Mask3D. Hierarchical block encoding is performed on the preprocessed point cloud dataset to form local point block feature representations and region-level aggregated feature representations. These local point block feature representations and region-level aggregated feature representations are input into the 3D instance segmentation model. During the training process of the 3D instance segmentation model, a three-constraint joint mechanism and a multi-source feature mutual information balance constraint mechanism are introduced to collaboratively optimize the feature representations. This maintains the continuity between the local geometric structure and the instance-level aggregated structure and suppresses the dominant role of a single feature source in the instance segmentation results, thereby improving the instance separation accuracy of the trunk, branches, and canopy regions. The module generates forest point cloud segmentation results with instance identifiers and category labels, and outputs a set of instance-labeled point clouds.

[0050] The point cloud instance segmentation module is deployed on a cloud GPU server, which is configured with dual CPUs, 256 GB of memory, and four NVIDIA A100 graphics cards.

[0051] Model training parameters: 15 manually labeled sample areas were used as the training set, 5 sample areas as the validation set, and 5 sample areas as the test set. The manually labeled objects included tree trunk instances, branch instances, canopy instances, understory vegetation instances, and ground surface instances, with a total of 12,864 valid labeled instances. The training parameters are as follows: Optimizer: AdamW; Initial learning rate: 1e-4; Weight decay: 0.01; Batch size: 8; Number of training epochs: 180 epochs; Learning rate decay strategy: decayed to 0.1 of the original value in epochs 90 and 140 respectively; Local patch feature dimension: 128 dimensions; Region-level aggregated feature dimension: 256 dimensions; Candidate instance aggregated feature dimension: 256 dimensions.

[0052] The following instance segmentation results were obtained on the test set: trunk instance mAP: 91.3%; branch instance mAP: 84.7%; canopy instance mAP: 88.9%; understory vegetation instance mAP: 79.6%; ground surface instance mAP: 94.1%; the overall instance segmentation mAP was 87.7%, which is an overall improvement of 6.4 percentage points compared to the basic Mask3D model without the introduction of the three-constraint joint mechanism and the multi-source feature mutual information balance constraint mechanism, with an improvement of 8.1 percentage points in the fine structure region of branches and 7.3 percentage points in the overlapping region of the canopy; the final output instance annotation point cloud identified a total of 10,532 forest object instances.

[0053] The single-tree clustering and segmentation module is deployed on a cloud computing platform and communicates with the skeleton extraction and cylinder fitting module. It constructs a point cloud similarity map based on the trunk, branch, and canopy instances in the instance-annotated point cloud set, and performs spectral clustering processing based on spatial proximity, structural continuity, and canopy connectivity to aggregate multiple types of point cloud instances belonging to the same tree into single-tree candidate objects and output a single-tree level point cloud object set.

[0054] The skeleton extraction and cylinder fitting module is deployed on a high-performance computing node in a cloud computing platform and communicates with the surface reconstruction and inversion module. It performs structural analysis and parametric modeling on a single-tree level point cloud object set. Specifically, it includes: first, extracting the trunk skeleton based on the spatial distribution characteristics and connectivity of the single-tree level point cloud to generate the trunk centerline skeleton; second, performing topological extension analysis and branch tracing along the trunk centerline skeleton to construct the trunk-branch hierarchical connection relationship, forming the tree skeleton topology; third, for the point cloud subsets corresponding to the trunk centerline skeleton and each branch path, using a cylinder model fitting method based on Random Sample Consensus (RANSAC), performing segmented cylinder parameter estimation on the trunk and branch segments to obtain the radius parameter, axial direction vector, and length parameter of each segment; finally, unifying the skeleton topology with the cylinder parameters of each segment to output the tree skeleton topology model and cylinder parameter model.

[0055] The surface reconstruction and inversion module, deployed in the 3D modeling node of the cloud computing platform, performs parameter-driven geometric reconstruction of the trunk and branches based on the tree skeleton topology model and cylindrical parameter model. For the canopy point cloud subset within the single-tree level point cloud object set, it uses the Alpha Shape reconstruction algorithm to perform boundary extraction and surface generation, identifying complex surface regions. Local continuous surface compensation and frequency domain adaptive refinement are performed on these complex surface regions. Based on this, unified spatial registration and organization of the trunk, branches, and canopy structure are performed to construct single-tree and scene-level 3D forest structure models. These models are then structured and formatted to output a forest structure reconstruction model. Based on the tree skeleton topology model, cylindrical parameter model, and forest structure reconstruction model, the module calculates single-tree height, diameter at breast height (DBH), crown width, canopy projected area, canopy volume, branch height, trunk inclination angle, branch angle distribution, canopy closure, and stand density, forming a forest structure parameter dataset.

[0056] The statistical results of the whole plot in this embodiment are as follows: average tree height: 13.92 m; average diameter at breast height: 18.11 cm; average crown width: 3.47 m; average crown projection area: 8.94 m²; average crown volume: 24.63 m³; average height below the branch: 4.26 m; average trunk angle: 5.1°; average first-order branch angle: 39.8°; plot canopy closure: 0.73; stand density: 402 trees / ha.

[0057] Fifty sample trees were selected and compared with manually measured data. The results are as follows: tree height RMSE: 0.37 m; diameter at breast height RMSE: 0.82 cm; crown width RMSE: 0.29 m; branch height RMSE: 0.34 m; relative error of crown projection area: 5.6%. It can be concluded that the system has high accuracy in forest structure parameter inversion.

[0058] The results management and application output module is deployed on a visualization terminal. It is used to perform model storage and version management on the forest structure parameter dataset and to conduct statistical analysis of indicators. On this basis, it performs 3D visualization and report generation on the forest structure parameter dataset and corresponding 3D structure model, and publishes and shares it through external system interfaces. It outputs forest structure reconstruction results, individual tree attribute files and sample plot structure statistical results, as well as application results for forest resource surveys, carbon storage accounting, management assessment and disaster loss analysis.

[0059] Example 2, according to Figure 2The difference between this embodiment and Embodiment 1 lies in the feature optimization mechanism of the point cloud instance segmentation module in the 3D instance segmentation model. In this embodiment, the point cloud instance segmentation module still constructs a 3D instance segmentation model based on Mask3D and performs hierarchical block encoding on the preprocessed point cloud dataset to form local point block feature representations and regional aggregated feature representations. The local point block feature representations and regional aggregated feature representations are then input into the 3D instance segmentation model to generate candidate instance aggregated feature representations. Unlike Embodiment 1, which introduces a three-constraint joint mechanism and a multi-source feature mutual information balance constraint mechanism to collaboratively optimize feature representations during model training, this embodiment uses the instance segmentation cross-entropy loss function and the mask prediction loss function to optimize and train the model parameters. After the model training is completed, based on the trained 3D instance segmentation model, 3D instance recognition, instance attribution determination, and semantic separation processing are performed on the input point cloud data to generate forest point cloud segmentation results with instance identifiers and category labels, and an instance-annotated point cloud set is output.

[0060] Example 3, based on Example 1, involves inputting local point block feature representations and region-level aggregated feature representations into a 3D instance segmentation model. During the training process, a three-constraint joint mechanism and a multi-source feature mutual information balance constraint mechanism are introduced to collaboratively optimize the feature representation, resulting in the output of an instance annotation point cloud. The specific steps include:

[0061] Step S1: Perform hierarchical block encoding on the preprocessed point cloud dataset according to the preset spatial partitioning rules; specifically, based on the spatial coordinate distribution of the point cloud, the point density variation and vertical height layer information, perform local clustering partitioning on the discrete points of the point cloud in the preprocessed point cloud dataset to form multiple local point block units; then, according to the hierarchical organization method from fine-grained local point block layer to coarse-grained regional aggregation layer, perform hierarchical indexing and association organization on each local point block unit to form a point block representation set with hierarchical topological relationship, thereby transforming the original disordered forest point cloud into a hierarchical structured input that can be used for subsequent instance modeling;

[0062] Step S2: Perform local feature extraction processing on each local point block in the point block representation set to generate local point block feature representations. Specifically, extract geometric structure features, spatial layer features, and echo attribute features from the point cloud within each local point block, and perform joint encoding on the extracted multi-class features to obtain local point block feature representations that characterize the current local point block morphology, layer, and echo response. Among them, geometric structure features are used to characterize the slender shape of tree trunks, branching shape, canopy clustering shape, discrete shape of understory vegetation, and continuous shape of the ground surface; spatial layer features are used to characterize the relative position of the current point block in the vertical structure of the forest; and echo attribute features are used to characterize the intensity of laser echoes and their changing response, thereby providing a fine-grained discrimination basis for subsequent instance separation of different forest objects.

[0063] Step S3: Based on the local point block feature representation, construct a point block association structure by combining the relationships between the local point blocks; on this basis, perform high-order topological association modeling on multiple local point blocks that satisfy spatial adjacency, morphological continuity and hierarchical correlation to form high-order association units; and perform structure-aware random walk sampling on the point block association structure and high-order association units. Based on the co-occurrence relationship and path continuity of each local point block during the random walk, adaptively adjust the cross-point block feature propagation weights to complete the regional feature aggregation processing and form a regional aggregated feature representation.

[0064] Step S4: Input the local point patch feature representation and the regional aggregated feature representation into the 3D instance segmentation model, perform candidate instance generation and instance mask prediction processing, and in the process of candidate instance generation, perform aggregation processing on multiple local point patch feature representations belonging to the same candidate instance to generate candidate instance aggregated feature representation; wherein, the 3D instance segmentation model uses local point patch feature representation to represent fine-grained boundary changes, and uses regional aggregated feature representation to represent mesoscale structural continuity information and large-scale contextual semantic information, and in the process of multi-layer feature interaction, local geometric details and regional contextual information jointly participate in candidate instance modeling, instance attribution determination and category prediction, thereby improving the separability between trunk instances, branch instances, canopy instances, understory vegetation instances and ground surface instances;

[0065] Step S5: In the parameter training stage of the 3D instance segmentation model, a three-constraint joint mechanism is introduced to perform cooperative consistency constraints on the statistical structure, main direction coupling relationship and adjacency topology relationship between the local point block feature representation and the candidate instance aggregate feature representation, so as to maintain the continuous evolution relationship of the trunk region, branch region and canopy region within the same tree in the feature space, and maintain the structural separability between different trees.

[0066] The three-constraint joint mechanism includes statistical structure consistency constraints, principal direction coupling consistency constraints, and adjacency topology continuity constraints. By coordinating the statistical distribution structure, principal axis extension relationship, and hierarchical adjacency topology relationship between local point block feature representations and candidate instance aggregated feature representations, it achieves a consistent expression of cross-hierarchical structural associations while maintaining the distribution characteristics of different levels of feature representations. Specifically, the statistical structure consistency constraint constrains the structural correlation between the local point block layer and the candidate instance layer in a second-order statistical sense; the principal direction coupling consistency constraint constrains the principal axis continuity of the trunk, branches, and canopy in the feature space; and the adjacency topology continuity constraint constrains the graph connectivity and feature propagation smoothness between structural units within the same tree. This maintains the continuous topological relationship of structural components within the same tree in the feature space and achieves implicit structural separation between different trees without explicit separation constraints. The three-constraint joint mechanism participates in the training and optimization process of the 3D instance segmentation model in the form of a joint loss function and updates the model parameters through backpropagation. Its expression is:

[0067] ;

[0068] in, This represents the total structural consistency loss function corresponding to the three-constraint joint mechanism; This indicates a statistical structural consistency term. Indicates the weighting coefficients of the statistical structure consistency constraint; This indicates the main direction coupling consistency term. The weighting coefficients represent the consistency constraints of the main direction coupling. Represents the adjacency topological continuity term. Represents the weighting coefficients of the adjacency topology continuity constraint;

[0069] Statistical structural consistency term:

[0070] ;

[0071] in, The normalized cross-covariance matrix represents the relationship between local point features and candidate instance aggregate features, and is used to characterize the correlation between the two types of features in terms of statistical change trends. The normalized covariance matrix represents the local point block feature itself, used to characterize the internal statistical structure of the local point block feature; The normalized covariance matrix represents the aggregated features of the candidate instance itself, used to characterize the overall statistical structure of the candidate instance features; This represents the Frobenius norm, used to measure the overall energy of a matrix.

[0072] Main direction coupling consistency term:

[0073] ;

[0074] in, Indicates the number of main feature directions selected; Indicates a local point-block hierarchy. Indicates the candidate instance hierarchy; The first part represents the result of decomposing the local point block feature statistical matrix. One principal direction vector; The first element is obtained by decomposing the aggregated feature statistical matrix of candidate instances. One principal direction vector; Represents the dot product of vectors, used to measure the angular relationship between two principal directions;

[0075] When the main directions of features at different levels are more consistent, the value of this item is smaller, thus ensuring the continuity of the trunk extension direction, branch forking direction and canopy expansion direction in the feature space.

[0076] Adjacency topological continuity term:

[0077] ;

[0078] in, The embedding representation matrix of local point patches in the feature space; Indicates matrix transpose; The graph Laplacian matrix represents the hierarchical relationship graph. This matrix is ​​constructed based on the spatial proximity, structural continuity, and vertical hierarchical relationships between nodes. The trace operation of the matrix is ​​used to calculate the overall graph smoothing energy;

[0079] The smaller the value of this term is when adjacent nodes belonging to the same tree are represented more closely in the feature space, thus ensuring the topological continuity and stability of the internal structure of the tree during feature propagation.

[0080] Step S6: Under the collaborative constraint framework of the three-constraint joint mechanism, a multi-source feature mutual information balance constraint mechanism is introduced during the training process of the 3D instance segmentation model. Adaptive balance adjustment is performed on the information contribution of geometric structure features, spatial layer features, and echo attribute features in the formation of candidate instance aggregated feature representation. Specifically, the information correlation strength between candidate instance aggregated feature representation and each source feature representation is measured respectively. Based on the degree of information retention of each source feature in candidate instance aggregated feature representation, dynamic weighted constraints are performed on the multi-source feature fusion process. This ensures that candidate instance aggregated feature representation maintains the collaborative expression of local geometric morphology information, vertical layer distribution information, and echo response difference information in the feature space. This suppresses the dominant role of a single feature source in the instance segmentation result and improves the stability and robustness of instance segmentation in fine branch structure region, canopy edge region, understory vegetation interleaved region, and ground transition region.

[0081] Step S7: Based on the 3D instance segmentation model jointly optimized by the three-constraint joint mechanism and the multi-source feature mutual information balance constraint mechanism, perform 3D instance recognition, instance attribution determination and semantic separation processing on tree trunk instances, branch instances, canopy instances, understory vegetation instances and ground surface instances in the forest point cloud, generate forest point cloud segmentation results with instance identifiers and category labels, and output instance labeled point cloud set according to the preset point cloud organization format for subsequent single tree clustering segmentation, skeleton extraction, cylinder fitting and forest structure reconstruction processing.

[0082] Example 4, this example is based on Example 3. In this example, step S3 specifically includes the following steps:

[0083] Step S31: Based on the feature representation of local points and blocks, perform joint analysis on the spatial coordinates, geometric extension features, and vertical height layer information of each local point and block to determine the spatial proximity, structural continuity, and interlayer progression relationships between local points and blocks; and based on the spatial proximity, structural continuity, and interlayer progression relationships, perform association organization on the local points and blocks to construct a point and block association structure for subsequent regional aggregation, thereby forming a basic connection skeleton for path propagation and relationship expansion between local points and blocks;

[0084] Step S32: Based on the point-block association structure, perform high-order topological association modeling on multiple local point blocks that meet the requirements of spatial adjacency, morphological continuity and hierarchical correlation. Organize multiple local point blocks with multi-fold co-ownership relationships in the same tree structure chain, the same branch extension area, the same canopy area or the same continuous surface area into high-order association units. This expands the structural expression ability between local point blocks from pairwise adjacency relationships to multi-point collaborative relationships, thereby improving the association relationship between local point blocks from a single adjacency connection to a high-order structural coupling relationship oriented towards continuous transmission from trunk to branch to canopy.

[0085] Step S33: On the point-block association structure and higher-order association units, taking each local point block as the starting point, perform structure-aware random walk sampling according to the path selection criteria of spatial proximity priority, structural continuity priority, and inter-layer progression priority. This allows the walk path to visit multiple times along the effective structural association direction between local point blocks, generating the corresponding random walk sequence. Based on the spatial proximity, structural continuity, and layer association degree between local point blocks, construct cross-point block feature propagation weights to characterize the basic feature transfer capability between local point blocks. The random walk sequence is used to record the co-occurrence relationship, cross-layer arrival relationship, and path continuity information of local point blocks during the multi-step propagation process, thereby transforming the static point block adjacency relationship into a dynamic path relationship that can characterize the reachability of the regional structural context.

[0086] Step S34: Based on the random walk sequence, perform statistical analysis on the path co-occurrence frequency, continuous arrival intensity, and cross-layer propagation stability between local points and blocks, and adaptively adjust the cross-point block feature propagation weights according to the statistical results. This gives higher propagation weights to local points and blocks that maintain stable co-occurrence, continuous structural paths, and consistent cross-layer propagation during multi-step random walks, while assigning lower propagation weights to local points and blocks that are unstable in co-occurrence, have frequent cross-structure jumps, or obvious path breaks. This is to suppress erroneous aggregation caused by overlapping branches and leaves, local occlusion, sparse point clouds, and boundary aliasing, and obtain the adjusted cross-point block feature propagation weights.

[0087] Step S35: Combining the multi-structure coupling relationship represented by the higher-order correlation unit and the adjusted cross-point block feature propagation weight, perform regional-level joint aggregation processing on the local point block features belonging to the same candidate tree, the same canopy area, or the same continuous surface area. This allows the local point blocks to further integrate multi-step structural path context information while maintaining local geometric morphology information, vertical layer distribution information, and echo response difference information, forming a regional-level aggregated feature representation. This enhances the overall representation ability of occluded areas, sparse areas, overlapping areas, and cross-scale continuous structures.

[0088] Example 5, based on Example 4, describes the acquisition of the forest structure reconstruction model, specifically including the following steps:

[0089] Step B1: Based on the tree skeleton topology model and cylindrical parameter model, perform parameter-driven geometric reconstruction on the trunk segment and branch segment to form the initial surface of the trunk and the initial surface of the branches; at the same time, for the canopy point cloud subset in the single tree level point cloud object set, use the Alpha Shape reconstruction algorithm to perform canopy boundary extraction and initial surface generation to obtain the initial surface of the canopy.

[0090] Step B2: Based on the connection relationship between the trunk node and branch node in the tree skeleton topology model, and combined with the local geometric continuity characteristics of the initial surface of the trunk and the initial surface of the branches, determine the connection region between the trunk and branches; and based on the initial surface of the canopy, combined with the local point density variation, normal dispersion, boundary curvature fluctuation and spatial distribution unevenness of the corresponding canopy point cloud subset, determine the canopy boundary variation region; the trunk and branch connection region and the canopy boundary variation region are combined as the complex surface region;

[0091] Step B3: For each complex surface region, a local coordinate domain is constructed, and the point cloud data and corresponding structural orientation information within the region are mapped to the local coordinate domain. The structural orientation information includes the tree skeleton orientation, the cylinder axial orientation, and the local normal orientation. Within the local coordinate domain, a nonlinear mapping function is constructed to characterize the continuous change relationship of the local surface. Using local coordinates, structural orientation information, and local geometric features as inputs, the local surface offset, normal correction, and surface undulation trend of the trunk-branch connection region and the canopy boundary change region are jointly estimated. Based on the estimation results, local continuous surface compensation is performed on the surface transition morphology of the trunk-branch connection, the surface extension morphology of the branch fork, and the local envelope morphology of the canopy boundary, thereby suppressing surface breakage, abrupt changes, and excessive smoothing caused by cylinder segment splicing, local point cloud sparsity, and boundary overlap.

[0092] The nonlinear mapping function can be expressed as:

[0093] ;

[0094] in, Indicates the number of a localized complex surface region. Indicates the first Surface compensation vectors for local regions; Indicates the coordinate position within the local coordinate domain; Indicates structural orientation information; Indicates local curvature characteristics; Represents local point density characteristics; This represents a nonlinear mapping function used to describe the continuous variation relationship of a local surface;

[0095] Step B4: After completing the local continuous surface compensation, the compensated local reconstructed surface is obtained; frequency domain consistency analysis is performed on the compensated local reconstructed surface and the corresponding original point cloud region to obtain the frequency domain error distribution characterizing the degree of detail recovery in each local region; based on the frequency domain error distribution, adaptive thinning processing is performed on the compensated local reconstructed surface, and regional attribute labels reflecting the complexity of the regional structure and the thinning state are generated simultaneously, thereby forming an enhanced surface structure set containing local thinning results and regional attribute labels;

[0096] The frequency domain error distribution can be expressed as:

[0097] ;

[0098] in, Indicates the first The frequency domain error value of a local region is used to characterize the degree of detail recovery in that region. Indicates the first Original point cloud region data corresponding to each local region; Indicates the first Compensated local reconstructed surface data corresponding to each local region; Represents the frequency domain transform operator; Represents the inverse frequency domain transform operator; This indicates a frequency band truncation operator, used to extract frequency components within a preset frequency range; This represents the local mean smoothing operator; This represents the L1 norm, used to measure the overall difference between two types of frequency band response results;

[0099] Step B5: Based on the enhanced surface structure set, perform unified spatial registration, topological connection, and hierarchical organization processing on the trunk surface, branch surface, and canopy surface; specifically, based on the structural region marking information, apply continuity constraints and connection consistency correction to the trunk-branch connection region and the canopy boundary region to ensure that each structural unit maintains geometric continuity and topological consistency in the same single-tree coordinate system; further, perform scene-level organization and integration on multiple single-tree structural models according to the spatial distribution relationship of the sample plots to construct single-tree and scene-level forest three-dimensional structural models, and output the forest structure reconstruction model.

[0100] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A forest structure reconstruction system based on 3D laser scanning point cloud modeling, characterized in that, include: The edge preprocessing module acquires a preprocessed point cloud dataset; The point cloud instance segmentation module constructs a 3D instance segmentation model based on Mask3D. Hierarchical block encoding is performed on the preprocessed point cloud dataset to form local point block feature representations and region-level aggregated feature representations; The local point block feature representation and the region-level aggregated feature representation are input into the 3D instance segmentation model. During the training process of the 3D instance segmentation model, a three-constraint joint mechanism and a multi-source feature mutual information balance constraint mechanism are introduced for collaborative optimization, and finally the instance annotation point cloud is output. The single-tree clustering and segmentation module constructs a point cloud similarity map based on the trunk instances, branch instances, and canopy instances in the instance-annotated point cloud set, performs spectral clustering processing, and outputs a single-tree level point cloud object set; The skeleton extraction and cylinder fitting module performs structural analysis and parametric modeling on a single tree-level point cloud object set, and outputs a tree skeleton topology model and a cylinder parametric model. The surface reconstruction and inversion module drives geometric reconstruction based on the tree skeleton topology model and cylindrical parameter model. For the canopy point cloud subset in the single-tree level point cloud object set, the Alpha Shape reconstruction algorithm is used to perform boundary extraction and surface generation to determine complex surface regions. Local continuous surface compensation and frequency domain adaptive refinement processing are performed on the complex surface regions to output the forest structure reconstruction model.

2. The forest structure reconstruction system based on three-dimensional laser scanning point cloud modeling according to claim 1, characterized in that: The process of obtaining the instance annotation point cloud includes the following steps: Step S1: Perform hierarchical and block-based encoding on the preprocessed point cloud dataset according to the preset spatial partitioning rules; specifically, perform local clustering partitioning on the discrete points of the point cloud in the preprocessed point cloud dataset to form local point block units; then, according to the hierarchical organization method from fine-grained local point block layer to coarse-grained regional aggregation layer, perform hierarchical indexing and association organization on each local point block unit to form a point block representation set. Step S2: Extract geometric structure features, spatial layer features, and echo attribute features from the point cloud within each local point block in the point block representation set, and perform joint encoding on the extracted multi-class features to obtain the local point block feature representation; Step S3: Based on the local point block feature representation, construct the point block association structure by combining the relationship between each local point block; on this basis, perform high-order topological association modeling to form high-order association units; and perform structure-aware random walk sampling on the point block association structure and high-order association units to form regional aggregated feature representation. Step S4: Input the local point block feature representation and the region-level aggregated feature representation into the 3D instance segmentation model to generate candidate instance aggregated feature representations; Step S5: During the parameter training phase of the 3D instance segmentation model, a three-constraint joint mechanism is introduced to perform collaborative consistency constraints on the statistical structure, main direction coupling relationship, and adjacency topology relationship between the local point block feature representation and the candidate instance aggregate feature representation. Step S6: Under the collaborative constraint framework of the three-constraint joint mechanism, a multi-source feature mutual information balance constraint mechanism is introduced to perform adaptive balance adjustment on the information contribution of geometric structure features, spatial layer features and echo attribute features in the process of candidate instance aggregation feature representation formation. Step S7: Based on the 3D instance segmentation model jointly optimized by the three-constraint joint mechanism and the multi-source feature mutual information balance constraint mechanism, perform 3D instance recognition, instance attribution determination and semantic separation processing to generate forest point cloud segmentation results with instance identifiers and category labels, and output instance labeled point cloud set according to the preset point cloud organization format.

3. The forest structure reconstruction system based on three-dimensional laser scanning point cloud modeling according to claim 2, characterized in that: Step S3 specifically includes the following steps: Step S31: Based on the feature representation of local points and blocks, perform joint analysis on the spatial coordinate position, geometric extension features and vertical height layer information of each local point and block to determine the spatial proximity relationship, structural continuity relationship and inter-layer progression relationship between local points and blocks, and construct the point and block association structure. Step S32: Based on the point-block association structure, perform high-order topological association modeling on spatially adjacent, morphologically continuous and hierarchically related local point blocks, and organize them into high-order association units; Step S33: On the point-block association structure and higher-order association unit, take each local point block as the starting point, perform structure-aware random walk sampling according to the path selection criteria of spatial proximity priority, structural continuity priority and inter-layer progression priority, generate random walk sequence, and construct cross-point block feature propagation weights. Step S34: Based on the random walk sequence, perform statistical analysis on the path co-occurrence frequency, continuous arrival intensity and cross-layer propagation stability between local points and blocks, and adaptively adjust the cross-point block feature propagation weights according to the statistical results to obtain the adjusted cross-point block feature propagation weights. Step S35: Combine the multi-structure coupling relationship represented by the higher-order correlation unit and the adjusted cross-point block feature propagation weight to form a regional aggregated feature representation.

4. The forest structure reconstruction system based on three-dimensional laser scanning point cloud modeling according to claim 3, characterized in that: Random walk sequences record the co-occurrence relationships, cross-layer arrival relationships, and path continuity information of local point blocks during multi-step propagation.

5. A forest structure reconstruction system based on three-dimensional laser scanning point cloud modeling according to claim 2, characterized in that: The three-constraint joint mechanism includes statistical structure consistency constraint, main direction coupling consistency constraint, and adjacency topology continuity constraint.

6. The forest structure reconstruction system based on three-dimensional laser scanning point cloud modeling according to claim 1, characterized in that: Obtaining the forest structure reconstruction model includes the following steps: Step B1: For the canopy point cloud subset in the single-tree level point cloud object set, the Alpha Shape reconstruction algorithm is used to extract the canopy boundary and generate the initial surface to obtain the initial canopy surface; Step B2: Based on the connection relationship between the trunk node and the branch node in the tree skeleton topology model, determine the connection area between the trunk and the branch; and based on the initial surface of the canopy, combined with the local point density change, normal dispersion, boundary curvature fluctuation and spatial distribution unevenness of the canopy point cloud subset, determine the canopy boundary change area; and take the trunk and branch connection area and the canopy boundary change area together as the complex surface area; Step B3: For each complex surface region, construct a local coordinate domain; within the local coordinate domain, construct a nonlinear mapping function to jointly estimate the local surface offset, normal correction, and surface undulation trend in the trunk-branch connection area and the canopy boundary change area; perform local continuous surface compensation based on the estimation results. Step B4: After completing the local continuous surface compensation, the compensated local reconstructed surface is obtained; frequency domain consistency analysis is performed on the compensated local reconstructed surface and the corresponding original point cloud region to obtain the frequency domain error distribution; combined with the frequency domain error distribution, the compensated local reconstructed surface is refined to form an enhanced surface structure set. Step B5: Based on the enhanced surface structure set, perform unified spatial registration, topological connection and hierarchical organization processing to output the forest structure reconstruction model.

Images