Forest LiDAR point cloud crown-trunk segmentation method and system based on soft semantic grouping

Abstract

The invention discloses a forest LiDAR point cloud crown-trunk segmentation method and system based on soft semantic grouping, and the method comprises the steps: constructing a soft semantic grouping SSGM module, carrying out the processing of input forest LiDAR point cloud data, and carrying out the soft semantic annotation of points in a crown-trunk transition region through a probability multi-class distribution mechanism; constructing a feature fusion clustering (FCM) module, receiving the point cloud features processed by the SSGM module, and performing instance-level clustering through a density-invariant clustering algorithm; constructing a deformable convolution up-sampling DCUM module, performing up-sampling on the clustered point cloud features, and adapting to a non-rigid geometric structure of the forest point cloud through learnable spatial offset; performing end-to-end joint training and parameter optimization on the SSGM module, the FCM module and the DCUM module to complete panoramic segmentation of the crown and the trunk in the forest LiDAR point cloud; the method adapts to point cloud density and distribution difference of different LiDAR scanning platforms, overcomes dependence of a traditional method on a single data source, and realizes cross-platform and cross-season stability.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of artificial intelligence and forestry information technology, and particularly relates to a forest LiDAR point cloud crown-trunk segmentation method and system based on soft semantic grouping. BACKGROUND

[0002] With the global climate change and the in-depth promotion of sustainable development strategy, high-precision monitoring and efficient management of forest resources have become the core task of ecological environment protection and forestry economic development. At present, the industry has an increasingly urgent demand for real-time acquisition and accurate analysis of forest three-dimensional structure information, especially in key applications such as biomass estimation, carbon sink evaluation, tree species classification and growth state monitoring, which urgently need a technical system that can support large-scale and fine operation.

[0003] In the prior art, LiDAR (Light Detection and Ranging) as an active remote sensing means has been widely used in the collection and processing of forest point cloud data. The current common point cloud segmentation method mainly relies on traditional computer vision algorithms based on geometric features (such as curvature, normal vector, height, etc.) or combines clustering strategies (such as region growing, Euclidean clustering, etc.) to realize single tree segmentation. In addition, some methods use segmentation networks based on deep learning, such as PointNet++, RandLA-Net, etc., trying to learn semantic or instance-level feature expression from the original point cloud to distinguish different structures such as trunk, branches and leaves and ground.

[0004] However, the prior art has a most obvious shortcoming: when dealing with complex forest scenes, it lacks consistent generalization ability for data obtained from different scanning platforms (such as ground, airborne and mobile LiDAR). Due to the significant differences in density, angle of view and coverage range of point clouds collected by different platforms, existing methods often perform well in a certain specific data source, but when migrating across platforms, they have problems such as significant decrease in segmentation accuracy, inconsistency in feature extraction, etc., which seriously limits their practical application value in large-scale, multi-source collaborative monitoring. SUMMARY

[0005] The purpose of the present application is to provide a forest LiDAR point cloud crown-trunk segmentation method based on soft semantic grouping, which can adapt to the characteristics of point clouds from different scanning platforms and significantly improve the cross-platform generalization ability. On the other hand, a forest LiDAR point cloud crown-trunk segmentation system based on soft semantic grouping is provided.

[0006] Technical solution: The forest LiDAR point cloud crown-trunk segmentation method according to the present application comprises the following steps:

[0007] (1) Construct a soft semantic grouping (SSGM) module to process the input forest LiDAR point cloud data, and use a probabilistic multi-class assignment mechanism to perform soft semantic labeling on the points in the crown-trunk transition area;

[0008] (2) Construct a feature fusion clustering (FCM) module to receive the point cloud features processed by the SSGM module, and perform instance-level clustering through a density-invariant clustering algorithm;

[0009] (3) Construct a deformable convolution up-sampling (DCUM) module to up-sample the clustered point cloud features, and adapt to the non-rigid geometric structure of the forest point cloud through a learnable spatial offset;

[0010] (4) Perform end-to-end joint training and parameter optimization on the SSGM module, FCM module and DCUM module to complete the panoramic segmentation of the crown and trunk in the forest LiDAR point cloud.

[0011] The present application uses a probabilistic multi-class assignment mechanism through a soft semantic grouping (SSGM) module to effectively handle the uncertainty in the crown-trunk transition area, achieving more fine and robust semantic labeling. The probabilistic design can adapt to the distribution differences of point clouds from different scanning platforms. The feature fusion clustering (FCM) module combines a density-invariant clustering algorithm to achieve high-precision instance-level clustering in complex forest structure, and its strong robustness to point cloud density changes significantly improves the generalization ability of cross-platform data. The deformable convolution up-sampling (DCUM) module adaptively captures the non-rigid crown geometry through a learnable spatial offset, effectively overcoming the geometric feature deformation problem caused by different scanning angles and platforms. Finally, through end-to-end joint training and optimization, high-precision and seamless panoramic segmentation of the crown and trunk in multi-platform forest LiDAR point cloud is achieved, showing excellent cross-platform adaptability and generalization performance.

[0012] Preferably, the SSGM module calculates the soft semantic label of a point as follows:

[0013] For each point pi in the point cloud, its soft semantic label is generated by a temperature-scaled Softmax function:

[0014]

[0015] where S i,k is the soft semantic score of point p i in class k, τ sem is the temperature parameter, and K is the total number of classes.

[0016] Soft semantic grouping is achieved through a weighted adjacency graph, and the edge weight w i,j is calculated based on spatial distance and semantic similarity:

[0017]

[0018] Among them, σ controls the influence of spatial distance, λ adjusts the weight of semantic similarity, and p i ′ and p j ′ represents the coordinates of the transformed point, S i ′ and S j ′ is a soft semantic tag.

[0019] This soft semantic labeling calculation method significantly smooths the category probability distribution by introducing a temperature scaling mechanism, effectively solving the uncertainty problem of point cloud classification in transitional regions and enhancing the tolerance for blurry and noisy points. The weighted adjacency graph constructed by combining spatial distance and semantic similarity constraints not only maintains the continuity of the spatial topology of the point cloud but also achieves deep fusion of semantic context information. It can accurately capture subtle feature changes at the canopy-trunk junction, suppress classification outliers, and ultimately achieve high-quality soft grouping results with smooth boundaries and semantic consistency in complex forest environments, providing a reliable foundation for subsequent instance segmentation.

[0020] Preferably, the density-invariant clustering algorithm includes:

[0021] Calculate point p using K nearest neighbors. i Local density D i :

[0022]

[0023] Where K is the number of nearest neighbors, and count(·) and volume(·) represent the number of points and volume, respectively;

[0024] Calculate the adaptive clustering radius ri:

[0025] r i =r0·(1-V i )·exp(λ·D i )

[0026] Where r0 is the basic radius, V i =|N i ·(0,0,1)| represents the verticality feature, N i For point p i The normal vector, where λ is the density adjustment coefficient.

[0027] The density-invariant clustering algorithm effectively overcomes the problem of uneven density caused by scanning angle and occlusion in forest point cloud by introducing local density calculation and adaptive radius mechanism. By combining the normal vector analysis of the perpendicularity feature, the ability to distinguish between trunk and crown regions is significantly improved, enabling the algorithm to adaptively adjust the clustering scale in different regions. This design not only ensures the fine segmentation of high-density point clouds, but also avoids over-segmentation in low-density regions, ultimately achieving more accurate and stable instance-level clustering results in complex forest structure, significantly improving the adaptability to different tree species and growth forms.

[0028] Preferably, the FCM module is also used to calculate a geometric consistency score to constrain the clustering results:

[0029]

[0030] where |C k | is the number of points in cluster C k , ∇N i is the normal vector gradient of point p i , ∇N k is the average normal vector gradient of cluster center c k , f i and f k are the point feature and cluster center feature respectively.

[0031] The geometric consistency score calculation mechanism effectively enhances the geometric rationality of the clustering results through dual constraints of normal vector gradient and feature similarity. By evaluating the directional consistency of the normal vector and the feature space consistency between points and cluster centers, it significantly suppresses abnormal clustering caused by point cloud noise and density changes, ensuring smooth geometric transition and consistent feature expression within the same instance. This design not only improves the clustering accuracy of trunk and crown regions, but also enhances the recognition ability of complex tree structure and branching patterns, ultimately achieving instance segmentation results with high geometric consistency and semantic coherence.

[0032] Preferably, the deformable convolution up-sampling module calculates the output feature in the following way:

[0033]

[0034] where y(p0) is the output feature, x(·) is the input feature, w(p n ) is the convolution weight, R is the receptive field, and Δp n is the learned offset.

[0035] The deformable convolution up-sampling module enables the convolution kernel to adaptively adjust the receptive field spatial distribution through the learned spatial offset, effectively overcoming the limitations of traditional fixed grid convolution on irregular crown geometry structure; this dynamic receptive field adjustment mechanism significantly enhances the representation ability of non-rigid crown shape, can accurately capture the subtle bending and complex spatial trend of branches and leaves, while maintaining the continuity and integrity of the geometric structure during feature up-sampling, ultimately realizing high-quality point cloud reconstruction with clear boundaries and rich details, greatly improving the accuracy and visual coherence of crown-trunk segmentation.

[0036] Preferably, the deformable convolution up-sampling module further comprises a PRNet instance optimization network for realizing feature fusion through a multi-scale feature pyramid network, wherein the feature Fl of the lth layer is calculated as:

[0037] F l =Upsample(F l+1 )+Conv 3×3 (C l )+Attention(F l+1 ,C l )

[0038] wherein Upsample(·) is an up-sampling operation, C l is the feature of the lth layer of the backbone network, Conv 3×3 (·) is a 3x3 convolution, and Attention(F l+1 ,C l ) is a cross-layer attention mechanism.

[0039] The PRNet instance optimization network depth-fuses semantic information of different levels through a multi-scale feature pyramid structure, significantly enhancing the richness and diversity of feature representation; through the cross-layer attention mechanism, it adaptively weights and fuses low-level detailed features and high-level semantic features, effectively preserving the fine geometric structure of crown branches and leaves and the structural features of the trunk; combined with the up-sampling operation and convolution optimization, it realizes multi-scale feature reconstruction from coarse to fine, not only improving the accuracy and continuity of the segmentation boundary, but also significantly enhancing the adaptability to different scale crown structures, ultimately producing instance segmentation results with high structural consistency and detail preservation.

[0040] Preferably, the total loss function corresponding to the end-to-end joint training is:

[0041]

[0042] wherein L semantic is a semantic segmentation loss, L instance is an instance segmentation loss, L geometry is a geometric consistency loss, and Lrefinement For instance optimization loss, λ1, λ2, λ3 are balance weights.

[0043] The multi-task joint loss function realizes deep coupling and mutual enhancement between different modules in the end-to-end training process through the synergistic optimization of semantic segmentation, instance segmentation, geometric consistency and instance optimization loss; the semantic segmentation loss ensures the accuracy of point-level classification, the instance segmentation loss strengthens the independence of clustering results, the geometric consistency loss maintains the smooth transition of object surfaces, and the instance optimization loss guarantees the fineness of the final segmentation boundary; such a comprehensive constraint mechanism effectively solves multiple challenges in forest point cloud segmentation, such as mismatch between semantic and instance information, geometric structure rupture and boundary ambiguity, significantly improves the convergence stability and segmentation precision of the overall model, and finally realizes the global optimal solution of tree crown-trunk panoramic segmentation.

[0044] In a second aspect, the forest LiDAR point cloud tree crown-trunk segmentation system provided by the present application comprises:

[0045] A soft semantic grouping module is used to process the input forest LiDAR point cloud data, and a probabilistic multi-class allocation mechanism is used to perform soft semantic labeling on the points in the tree crown-trunk transition area.

[0046] A feature fusion clustering module is connected to the soft semantic grouping module and is used to receive the point cloud features processed by the SSGM module and perform instance-level clustering through a density-invariant clustering algorithm.

[0047] A deformable convolution upsampling module is connected to the feature fusion clustering module and is used to upsample the clustered point cloud features and adapt to the non-rigid geometric structure of forest point clouds through a learnable spatial offset.

[0048] A model training and optimization module is used to perform end-to-end joint training and parameter optimization on the soft semantic grouping module, feature fusion clustering module and deformable convolution upsampling module to complete panoramic segmentation of tree crowns and tree trunks in forest LiDAR point clouds.

[0049] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program capable of being loaded and executed by the processor to implement the soft semantic grouping-based forest LiDAR point cloud tree crown-trunk segmentation method.

[0050] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the soft semantic grouping-based forest LiDAR point cloud tree crown-trunk segmentation method.

[0051] Beneficial effects: compared with the prior art, the present application has the following remarkable advantages: 1, by the density-invariant clustering algorithm and the adaptive radius calculation mechanism, the point cloud density and distribution difference of different LiDAR scanning platforms are effectively adapted, the dependence of the traditional method on a single data source is overcome, and stable performance across platforms and seasons is realized; 2, a soft semantic grouping mechanism is adopted, points in the crown-trunk transition area are flexibly semantically labeled through probability multi-class allocation and temperature scaling function, and the segmentation accuracy of complex areas is significantly improved; 3, by using the spatial offset learning mechanism in the deformable convolution upsampling module, the receptive field and feature extraction method are dynamically adjusted, and the geometric features of irregular shapes such as tree crowns and tree trunks in forest point clouds are better captured; 4, by means of geometric consistency constraint and multi-scale feature fusion mechanism, the spatial structure rationality and semantic consistency are considered in the clustering process, and the integrity and reliability of the instance-level segmentation result are improved. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 The method flowchart of the embodiment of the present application is shown in the figure;

[0053] Figure 2 The current research challenge diagram of the embodiment of the present application is shown in the figure;

[0054] Figure 3 The CTS panoramic segmentation network architecture diagram of the embodiment of the present application is shown in the figure;

[0055] Figure 4 The PRNet instance optimization network structure diagram of the embodiment of the present application is shown in the figure;

[0056] Figure 5 The experimental flowchart description diagram of the embodiment of the present application is shown in the figure;

[0057] Figure 6 The research area diagram of the embodiment of the present application is shown in the figure;

[0058] Figure 7 The forest point cloud panoramic segmentation result diagram of the embodiment of the present application is shown in the figure;

[0059] Figure 8 The instance quality loss and semantic class consistency loss diagram of the embodiment of the present application is shown in the figure;

[0060] Figure 9 The parameter analysis visualization diagram under extreme conditions of the embodiment of the present application is shown in the figure;

[0061] Figure 10 The comparative analysis diagram of the embodiment of the present application and the prior art is shown in the figure;

[0062] Figure 11 The crown volume and projected area relationship regression analysis diagram of the embodiment of the present application is shown in the figure;

[0063] Figure 12 This is a visualization diagram of the spatial distribution of canopy volume under different forest types and scanning platforms in an embodiment of the present invention;

[0064] Figure 13 This is a schematic diagram of the correlation analysis of tree growth parameters in an embodiment of the present invention;

[0065] Figure 14 This is a schematic diagram for verifying the accuracy of morphological parameter extraction in an embodiment of the present invention. Detailed Implementation

[0066] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0067] Example 1: Design and Implementation of CTS Panoramic Segmentation Framework

[0068] This embodiment provides a panoramic segmentation method for forest LiDAR point cloud canopy-trunk based on soft semantic grouping.

[0069] Overall framework architecture: such as Figure 2 As shown in the figure, this illustrates three key challenges in forest point cloud segmentation: (a) dimensionality limitations, (b) multi-platform adaptability, and (c) variations in point cloud density. The CTS framework employs an end-to-end deep learning architecture, primarily comprising the following components: 1. Point Cloud Feature Extractor: Using PointNet++ as the backbone network to extract geometric and semantic features of the point cloud; 2. Soft Semantic Grouping Module (SSGM): Implementing probabilistic multi-class assignment to address semantic ambiguity in the canopy-trunk transition region; 3. Feature Fusion Clustering Module (FCM): Achieving cross-platform instance separation through density-invariant clustering algorithms; 4. Deformable Convolutional Upsampling Module (DCUM): Adapting to the non-rigid geometric structure of forest point clouds; 5. PRNet Instance Optimization Network: Improving instance segmentation accuracy through multi-scale feature fusion; 6. Segmentation Head: Outputting semantic segmentation and instance segmentation results. Figure 1 As shown, the method includes the following steps:

[0070] Step 1: Construct the Soft Semantic Grouping (SSGM) module to process the input forest LiDAR point cloud data and use a probabilistic multi-class assignment mechanism to perform soft semantic annotation on points in the canopy-trunk transition region.

[0071] For each point pi in the point cloud, its soft semantic label is generated using a temperature-scaled Softmax function:

[0072]

[0073] Among them, S i,k For point p i The original semantic score under category k, τ semK is the total number of categories for the temperature parameter;

[0074] Soft semantic grouping is achieved by weighted adjacency graph, combining spatial distance and semantic similarity to calculate edge weight w i,j :

[0075]

[0076] where σ controls the influence of spatial distance, λ adjusts the weight of semantic similarity, and p i ′ and p j ′ are the transformed point coordinates, S i ′ and S j ′ are soft semantic labels.

[0077] Step 2, build a feature fusion clustering FCM module, receive the point cloud features processed by the SSGM module, and perform instance-level clustering through a density-invariant clustering algorithm;

[0078] The local density D i of point p i is calculated by K-nearest neighbors:

[0079]

[0080] where K is the number of neighbors, count(·) and volume(·) represent the point number and volume, respectively;

[0081] The adaptive clustering radius r i is calculated:

[0082] r i =r0·(1-V i )·exp(λ·D i )

[0083] where r0 is the basic radius, V i =|N i ·(0,0,1)| is the verticality feature, N i is the normal vector of point p i , and λ is the density adjustment coefficient.

[0084] The FCM module is also used to calculate a geometric consistency score to constrain the clustering result:

[0085]

[0086] where |C k | is the number of points in cluster C k , ▽N i is the normal vector gradient of point p i , ▽N k is the average normal vector gradient of cluster center c k , and fi and f k are point features and cluster center features, respectively.

[0087] Step 3, a deformable convolution up-sampling DCUM module is constructed to up-sample the clustered point cloud features, and a learnable spatial offset is used to adapt to the non-rigid geometric structure of the forest point cloud;

[0088] The deformable convolution up-sampling module calculates the output features in the following manner:

[0089]

[0090] where y(p0) is the output feature, x(·) is the input feature, w(p n ) is the convolution weight, R is the receptive field, and Δp n is the learned offset.

[0091] As shown in Figure 4 , the deformable convolution up-sampling module further includes a PRNet instance optimization network for realizing feature fusion through a multi-scale feature pyramid network, wherein the feature Fl of the lth layer is calculated as:

[0092] F l =Upsample(F l+1 )+Conv 3×3 (C l )+Attention(F l+1 , C l )

[0093] where Upsample(·) is an up-sampling operation, C l is the feature of the lth layer of the backbone network, Conv 3×3 (·) is a 3x3 convolution, and Attention(F l+1 , C l ) is a cross-layer attention mechanism.

[0094] Step 4, the SSGM module, the FCM module and the DCUM module are trained and optimized in an end-to-end manner, and panoramic segmentation of tree crowns and tree trunks in forest LiDAR point clouds is completed, and a panoramic segmentation network architecture diagram is shown in Figure 3 .

[0095] The total loss function corresponding to the end-to-end joint training is:

[0096]

[0097] where L semantic is a semantic segmentation loss, L instance is an instance segmentation loss, and L geometryL is a geometry consistency loss refinement L is an instance optimization loss, λ1, λ2, λ3 are balance weights; a quality loss and a semantic category consistency loss are shown in a schematic diagram as Figure 8 Figure 9 The parameter sensitivity and robustness analysis of the algorithm under extreme environmental conditions are shown.

[0098] Based on the similar inventive concept, the embodiment of the present application also provides a forest LiDAR point cloud crown-trunk panoramic segmentation system corresponding to the forest LiDAR point cloud crown-trunk panoramic segmentation method, comprising:

[0099] A soft semantic grouping module is used for processing the input forest LiDAR point cloud data, and a probability multi-class allocation mechanism is used for soft semantic labeling of points in the crown-trunk transition area;

[0100] A feature fusion clustering module is connected with the soft semantic grouping module, and is used for receiving the point cloud features processed by the SSGM module and performing instance-level clustering through a density-invariant clustering algorithm;

[0101] A deformable convolution upsampling module is connected with the feature fusion clustering module, and is used for upsampling the clustered point cloud features and adapting the non-rigid geometric structure of the forest point cloud through a learnable spatial offset;

[0102] A model training and optimization module is used for end-to-end joint training and parameter optimization of the soft semantic grouping module, the feature fusion clustering module and the deformable convolution upsampling module, so as to complete the panoramic segmentation of the crown and the trunk in the forest LiDAR point cloud.

[0103] Embodiment two: trunk diameter measurement and crown three-dimensional volume calculation

[0104] Based on the accurate crown-trunk segmentation result, the embodiment realizes the trunk diameter measurement and the crown three-dimensional volume calculation, and shows the practical value of the CTS framework in the forest phenotype analysis, as shown in Figure 6 and Figure 7 The geographical location and environmental characteristics of different forest research areas involved in the experiment and the comparison of segmentation results of each data set are shown, wherein (a) and (b) show the overhead view and the side view of the original data set; (c) and (d) show the overhead view and the side view after panoramic segmentation; (e) shows the crown instance extraction result; (f) shows the trunk instance extraction result; (g) shows the single tree instance extraction result; (h) shows the trunk and crown semantic segmentation result of the single tree.

[0105] ​Diameter at Breast Height (DBH) measurement: DBH is a standard forestry measurement parameter reflecting the development of tree trunk. Based on the accurate segmentation result of CTS framework, the present application extracts the trunk cross section at 1.3 meters above the ground to obtain the trunk point set T 1.3 = {t s |1.2 ≤ z s ≤ 1.4, t s ∈ T}, where T is the trunk point set, z s is the height coordinate of point t s .

[0106] The horizontal projection of the cross section is fitted to a circular profile using the least squares method, and the equation of the fitted circle (x-a) 2 +(y-b) 2 =r 2 produces the diameter at breast height DBH=2r. Through the accurate segmentation of the CTS framework, the accuracy of the diameter measurement is significantly improved, achieving a high-precision measurement of DBH RMSE of 0.021m on mature forest data sets.

[0107] Crown three-dimensional volume calculation: Crown volume is a basic parameter for biomass estimation and growth assessment. Based on the accurate segmentation result of the CTS framework, the present application realizes spatial discretization by dividing the 3D space into regular cubic grids, and the voxel size parameter is v=0.1m.

[0108] The spatial range of each voxel V i,j,k is defined as:

[0109] {(x, y, z)|i·v ≤ x < (i+1)·v, j·v ≤ y < (j+1)·v, k·v ≤ z < (k+1)·v}

[0110] The total volume calculation is V=N v ·v 3 , where N v is the total number of occupied voxels. Through the accurate segmentation of the CTS framework, the accuracy of crown volume calculation is significantly improved, achieving accurate calculation of average volume of 6.354m 3 on multiple data sets.

[0111] Crown projection area calculation: Crown projection area represents the horizontal range covered by the crown and is a key parameter for assessing light competition and resource utilization. The present application adopts an orthogonal projection strategy to project the D crown point cloud onto a 2D horizontal plane. Given the crown point set P, the projection forms a 2D point set Q={q j |q j =(x j ,y j ), p j ∈P c}.

[0112] The area is calculated by constructing the minimum convex polygon using the shoelace formula:

[0113]

[0114] where (x1, y1), (x2, y2),..., (x M ,y M ) represent the convex hull vertex sequence.

[0115] Figure 5 The complete experimental process from data acquisition, preprocessing to segmentation result evaluation is shown.

[0116] Example three: forest growth pattern analysis and morphological parameter relationship research

[0117] Based on the accurate morphological parameters extracted by the CTS framework, this embodiment realizes forest growth pattern analysis and morphological parameter relationship research, and provides a scientific basis for forest ecology research.

[0118] Crown volume-area relationship analysis: By setting the voxel size to 0.1m, the present application analyzes the regression relationship between crown volume and projected area in all six data sets. The results show that there is a significant difference in growth pattern between mature trees and young trees; Figure 11 The regression relationship between crown volume and projected area in six data sets and the difference in development stage are shown, where the X-axis represents the crown volume (m 3 ), and the Y-axis represents the horizontal projected area (m 2 ). Panels (a) to (f) correspond to MA, MM, MT, YA, YM and YT data sets, respectively.

[0119] Mature trees show a more stable linear relationship between crown volume and area, with higher R 2 values, which is attributed to their more regular crown structure and balanced growth pattern. Young trees show lower correlation coefficients due to unbalanced branch development and significant crown shape changes during the rapid growth phase.

[0120] The fitting slope of mature trees is generally steeper than that of young trees, indicating that volume increase leads to more significant crown area expansion in mature trees. This reflects a higher proportion of lateral growth in mature crowns, while young trees prioritize vertical growth for light competition.

[0121] Morphological parameter correlation analysis: To understand the relationship between different growth parameters, the present application constructs a correlation matrix of key tree features, including tree height, diameter at breast height, crown diameter, crown area and crown volume. Correlation analysis reveals a strong positive correlation between crown-related parameters; Figure 13The correlation strength and direction between each growth parameter are demonstrated by heat map, where the heat map adopts cold and hot color marks, red color represents positive correlation, blue color represents negative correlation, and the deeper the color, the stronger the correlation. This analysis reveals the interdependence between different morphological characteristics and growth strategies.

[0122] The correlation coefficient between tree height and crown parameters (area and volume) indicates a moderate positive correlation. Mature trees generally show stronger correlation than young trees. This difference reflects the more stable growth pattern of mature trees, where resource allocation between height and crown development is more balanced.

[0123] Cross-platform data fusion analysis: One of the important advantages of the CTS framework is its cross-platform adaptability. This invention analyzes data from different LiDAR platforms (backpack, ground, and unmanned aerial vehicle), verifying the generalization ability of the framework. Figure 10 The performance differences between the method in this paper and other advanced methods are intuitively demonstrated through top view and side view comparison, where (a) is the real label top view; (b) is the Point2tree top view; (c) is the method in this invention top view; (d) is the ForAINet top view; (e) is the real label side view; (f) is the Point2tree side view; (g) is the method in this invention side view; (h) is the ForAINet side view. Figure 12 The spatial distribution characteristics of crown volume under different platforms and forest types are intuitively demonstrated, where panels (a) to (f) represent MA, MM, MT, YA, YM, and YT datasets, respectively.

[0124] The CTS framework maintains stable segmentation accuracy on different platforms: - Backpack LiDAR (MLS): Point density of about 8,000-9,000 points / m 2 , highest segmentation accuracy - Ground LiDAR (TLS): Point density over 4,000-5,000 points / m 2 , higher segmentation accuracy - Unmanned aerial vehicle LiDAR (ALS): Point density over 1,000 points / m 2 , stable segmentation accuracy

[0125] This cross-platform adaptability provides important technical support for multi-source data fusion and forest monitoring. Figure 14 The extraction accuracy of key parameters such as diameter at breast height and crown projection area is verified through regression analysis, where the regression analysis of predicted values and true values: (a) is the diameter at breast height of the MM dataset; (b) is the crown projection area of the MA dataset; (c) is the diameter at breast height of the YM dataset; (d) is the crown projection area of the YA dataset.

[0126] Example Four: Practical application of forest resource monitoring and management

[0127] This embodiment demonstrates the practical application value of the CTS framework in forest resource monitoring and management, including large-scale forest survey, carbon sink assessment, and biodiversity protection.

[0128] Large-scale forest survey application: Based on the accurate segmentation results of the CTS framework, the invention can automatically extract key morphological indicators for forest growth phenotype research. These parameters provide quantitative measurements for understanding tree architecture, growth patterns, and phenotypic variation at different developmental stages and environmental conditions.

[0129] Automatically extracted DBH, tree height, crown area, and volume support large-scale forest surveys and breeding programs. The established crown volume-area regression model reveals the dynamic growth correlation across developmental stages, supporting the development of growth prediction models and yield estimation systems. These models can provide information for forest management decisions and help predict future forest productivity.

[0130] Carbon sink assessment and biomass calculation: Comprehensive crown mapping and volume quantification can accurately assess forest biomass, carbon storage, and timber resources. Based on the accurate segmentation of the CTS framework, the invention can:

[0131] 1. Calculate the biomass of individual trees: Through the diameter at breast height-biomass relationship model, combined with accurate diameter at breast height measurement results; 2. Estimate forest carbon storage: Based on the accumulation of individual biomass, considering species differences and growth stages; 3. Monitor carbon sink trends: Through time series analysis, achieve monthly-level carbon sink dynamic monitoring.

[0132] These information support sustainable forest management and environmental impact assessment, providing scientific basis for carbon sink trading and ecological compensation.

[0133] Biodiversity protection application: The CTS framework plays an important role in biodiversity monitoring, which can:

[0134] 1. Species identification and classification: Based on automatic identification of morphological features, the accuracy rate reaches 85%; 2. Habitat assessment: Identify key habitat features and assess the integrity of forest structure; 3. Ecological corridor planning: Provide scientific basis for protection planning and support ecological network construction.

[0135] Industrial application prospects: The CTS framework has broad industrial application prospects:

[0136] 1. Forestry management software: Provide intelligent management tools for forestry departments, support forest resource monitoring and planning; 2. Carbon sink trading platform: Provide accurate monitoring data for carbon sink projects, support carbon sink quantification and trading; 3. Ecological protection system: Provide intelligent monitoring solutions for nature reserves, support biodiversity protection; 4. Research and education platform: Provide technical support for forestry research and teaching, promote discipline development.

[0137] Through the above specific application examples, the CTS framework fully demonstrates its practical value in key applications such as forest resource monitoring, carbon sink assessment, and biodiversity protection, providing important technical support for smart forestry and ecological protection.

Claims

1. A forest LiDAR point cloud canopy-trunk segmentation method based on soft semantic grouping, characterized in that, Includes the following steps: (1) Construct a soft semantic grouping (SSGM) module to process the input forest LiDAR point cloud data and use a probabilistic multi-class assignment mechanism to perform soft semantic annotation on the points in the canopy-trunk transition region; (2) Construct a feature fusion clustering FCM module to receive point cloud features processed by the SSGM module and perform instance-level clustering using density-invariant clustering algorithm; (3) Construct a deformable convolutional upsampling DCUM module to upsample the clustered point cloud features and adapt the non-rigid geometric structure of the forest point cloud through learnable spatial offset. (4) Perform end-to-end joint training and parameter optimization on the SSGM module, FCM module and DCUM module to complete the panoramic segmentation of tree canopy and trunk in forest LiDAR point cloud.

2. The forest LiDAR point cloud canopy-trunk segmentation method according to claim 1, characterized in that, The SSGM module calculates the soft semantic labels of points in the following way: For each point pi in the point cloud, its soft semantic label is generated using a temperature-scaled Softmax function: Among them, S i,k For point p i The original semantic score under category k, τ sem Here, K represents the temperature parameter and K represents the total number of categories. Soft semantic grouping is achieved through a weighted adjacency graph, and edge weights w are calculated by combining spatial distance and semantic similarity. i,j : Among them, σ controls the influence of spatial distance, λ adjusts the weight of semantic similarity, and p i ′ and p j ′ represents the coordinates of the transformed point, S i ′ and S j ′ is a soft semantic tag.

3. The forest LiDAR point cloud canopy-trunk segmentation method according to claim 1, characterized in that, The density-invariant clustering algorithm includes: Calculate point p using K nearest neighbors. i Local density D i : Where K is the number of nearest neighbors, and count(·) and volume(·) represent the number of points and volume, respectively; Calculate the adaptive clustering radius ri: r i =r0·(1-V i )·exp(λ·D i ) Where r0 is the basic radius, V i =|N i ·(0,0,1)| represents the verticality feature, N i For point p i The normal vector, where λ is the density adjustment coefficient.

4. The forest LiDAR point cloud canopy-trunk segmentation method according to claim 1, characterized in that, The FCM module is also used to calculate geometric consistency scores to constrain clustering results: Among them, |C k |For clustering C k The number of points in ▽N i For point p i The gradient of the normal vector, ▽N k For cluster center c k The average normal gradient, f i and f k These are point features and cluster center features, respectively.

5. The forest LiDAR point cloud canopy-trunk segmentation method according to claim 1, characterized in that, The deformable convolutional upsampling module calculates the output features in the following manner: Where y(p0) is the output feature, x(·) is the input feature, and w(p) is the input feature. n ) represents the convolution weights, R represents the receptive field, and Δp represents the convolution weights. n This is the learned offset.

6. The forest LiDAR point cloud canopy-trunk segmentation method according to claim 1, characterized in that, The deformable convolutional upsampling module also includes a PRNet instance optimization network for feature fusion via a multi-scale feature pyramid network, where the feature Fl of the l-th layer is calculated as follows: F l =Upsample(F l+1 )+Conv 3×3 (C l )+Attention(F l+1 ,C l ) Where Upsample(·) is the upsampling operation, C l For the features of the l-th layer of the backbone network, Conv 3×3 (·) represents a 3×3 convolution, Attention(F) l+1 C l This is a cross-layer attention mechanism.

7. The forest LiDAR point cloud canopy-trunk segmentation method according to claim 1, characterized in that, The total loss function corresponding to the end-to-end joint training is: Among them, L semantic For semantic segmentation loss, L instance L is the instance segmentation loss. geometry For geometric consistency loss, L refinement The loss is optimized for instances, and λ1, λ2, and λ3 are balancing weights.

8. A forest LiDAR point cloud canopy-trunk segmentation system based on soft semantic grouping, characterized in that, include: The soft semantic grouping module is used to process the input forest LiDAR point cloud data and uses a probabilistic multi-class assignment mechanism to perform soft semantic annotation on points in the canopy-trunk transition region. The feature fusion and clustering module, connected to the soft semantic grouping module, is used to receive point cloud features processed by the SSGM module and perform instance-level clustering using the density-invariant clustering algorithm. The deformable convolutional upsampling module, connected to the feature fusion clustering module, is used to upsample the clustered point cloud features and adapt to the non-rigid geometric structure of the forest point cloud through learnable spatial offset. The model training and optimization module is used to perform end-to-end joint training and parameter optimization of the soft semantic grouping module, feature fusion clustering module and deformable convolution upsampling module to complete the panoramic segmentation of tree canopy and trunk in forest LiDAR point cloud.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the forest LiDAR point cloud canopy-trunk segmentation method based on soft semantic grouping as described in any one of claims 1 to 7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the forest LiDAR point cloud canopy-trunk segmentation method based on soft semantic grouping according to any one of claims 1 to 7.

Images