A method for extracting wood structure for low-density forest point cloud

Abstract

This invention relates to the fields of 3D point cloud processing and digital forestry technology, specifically a method for extracting wood structures from low-density forest point clouds. Addressing the challenges of mixed distribution of wood structures (trunks and branches) and leaf structures in forest point clouds, low point cloud density, and insufficient local neighborhood points leading to instability of traditional fixed-scale features and difficulty in reliably separating wood and leaves, this invention proposes a complete workflow of "density-adaptive multi-scale geometric feature extraction—random forest supervised classification—post-processing optimization such as spatial consistency." This achieves automatic identification and extraction of wood structure points and outputs confidence and spatial consistency results that can be used for engineering diagnostics. Therefore, it possesses stable and usable engineering application capabilities even in scenarios where conventional methods are inadequate, such as low density, weak neighborhood support, and lack of real-world annotations.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional point cloud processing and digital forestry technology, specifically a method for extracting wood structure from low-density forest point clouds. Background Technology

[0002] Forest point clouds are an important data foundation for individual tree structure analysis, biomass inversion, and forestry parameter measurement. However, actual forest point clouds often suffer from low point density, severe occlusion, and uneven distribution, resulting in insufficient local neighborhood points. This makes the estimation of geometric features such as principal direction, normal, and curvature based on the covariance matrix unstable, leading to speckle and structural breaks in the classification results.

[0003] Wood structures typically exhibit linear extension and stable main direction characteristics; leaf structures often appear as clusters or flakes, with localized dispersion and susceptibility to occlusion. Existing methods rely on intensity or color information for threshold segmentation, resulting in insufficient robustness across devices; deep learning methods require large amounts of labeled data and are sensitive to density changes; traditional machine learning methods depend on fixed radii or fixed K-nearest neighbor parameters, making them prone to failure in low-density scenarios.

[0004] Furthermore, in low-density scenes, clustered leaves and sparse branches often coexist, resulting in significant overlap between the wood structure point cloud and the leaf point cloud in the feature space. Using only fixed physical thresholds (such as linearity threshold, curvature threshold, and verticality threshold) often fails to address both misclassification and omission. Therefore, a wood structure extraction method is needed that can adapt to point cloud density, integrate multi-scale interpretable geometric features, and possess spatial consistency diagnosis and optimization mechanisms. Summary of the Invention

[0005] To address the aforementioned problems and shortcomings, and to resolve the issues of scale sensitivity, unstable geometric features, and significant speckle in the extraction of wood structures from low-density forest point clouds, this invention provides a method for extracting wood structures from low-density forest point clouds. The core idea of ​​this invention is: to characterize the sparsity of the point cloud using nearest neighbor distance statistics, automatically generating a multi-scale neighborhood radius sequence matching the density; to extract geometric features at multiple radius scales and multiple K-nearest neighbor scales, and to synchronously sort principal component analysis (PCA) eigenvalues ​​and eigenvectors to ensure consistency in the physical semantics of the vector features; to introduce principal direction verticality features to enhance the identification ability of tree trunks and branches; and to employ random forest learning for multi-feature nonlinear discrimination boundaries, combined with post-processing such as spatial consistency filtering to suppress speckle noise, resulting in stable wood structure point cloud extraction results.

[0006] A method for extracting the wood structure from point clouds in low-density forests includes the following steps:

[0007] Step 1: Training Data Input and Label Construction. Read the point clouds of the wood structure and leaves from the training data, extract the 3D coordinates (x, y, z) of all points to form the training sample point set P. Assign label 1 to the wood structure points and label 0 to the leaf points, then concatenate them to form labeled training samples.

[0008] Step 2: Point Cloud Density Statistics and Adaptive Scale Determination. A KD tree is constructed for the training sample point set P from Step 1, and several points are randomly sampled to calculate the nearest neighbor distance, obtaining the mean nearest neighbor distance mu (unit: m). mu is used to characterize the sparsity of the point cloud. A multi-scale radius sequence r_i = mu * m_i is generated based on mu, where m_i is the scaling factor. A minimum radius lower limit r_min (unit: m) is set to ensure that the number of neighboring points meets the minimum requirement under low-density conditions in the target forest area.

[0009] Step 3: Multi-scale geometric feature extraction. For each point in the training sample point set P, perform a spherical neighborhood query at the corresponding radius r_i to obtain a neighborhood point set N_i. Calculate the neighborhood covariance matrix and perform eigenvalue decomposition on the neighborhood point set N_i that meets the neighborhood point number requirement, obtaining eigenvalues ​​lambda1>=lambda2>=lambda3 and corresponding eigenvectors v1, v2, v3. Then, simultaneously sort the eigenvalues ​​and eigenvectors in descending order to ensure that v1 corresponds to the largest eigenvalue lambda1 and v3 corresponds to the smallest eigenvalue lambda3.

[0010] Based on eigenvalues, linearity, flatness, sphericity, curvature, anisotropy, and eigenenthalpy features are constructed, and local point density is calculated. Perpendicularity features are further introduced: normal perpendicularity V3 = |v3_z|, principal direction perpendicularity V1 = |v1_z|. The principle of principal direction perpendicularity feature discrimination is described in [link to relevant documentation]. Figure 3 Then, at the K-nearest neighbor (KNN) scale, simplified PCA features and the reciprocal of the average nearest neighbor distance are extracted, and global location features are added to form a multi-scale feature vector for each point. Finally, the multi-scale feature vectors of all points in the training sample point set P are combined and concatenated row by row to construct a complete feature matrix.

[0011] Step 4: Random Forest Classifier Training. The feature matrix obtained in Step 3 is standardized, and the training and test sets are stratified according to labels. The random forest classifier is trained, and the trained random forest classifier and standardized parameters are saved.

[0012] The output of a trained random forest classifier is: predicted label. and the probability of belonging to the category of wooden structure Predicted Labels It belongs to {0,1}, where 1 represents woody structure and 0 represents leaf.

[0013] Step 5: Classification and Result Output of the Forest Point Cloud to be Classified. The random forest classifier trained in Step 4 is used for each point in the forest point cloud to be classified, and the predicted label is output. and the probability of belonging to the category of wooden structure The predicted labels and class probabilities from these outputs together constitute the classification result for each point. Based on the predicted labels from the outputs... The set of points representing the wood structure is obtained by filtering, and the output is wood_only.las; at the same time, the complete classification point cloud classified_new_forest.las is also output, which supports visualization and subsequent processing.

[0014] Furthermore, in step 2, when r_i is less than r_min, r_min is taken to ensure that the number of neighborhood points meets the minimum requirement under the low-density conditions of the target forest area.

[0015] Furthermore, in step 3, when the number of neighborhood points is less than the minimum neighborhood point threshold min_neighbors, the scale feature is set to 0 to avoid instability in covariance estimation.

[0016] Furthermore, the complete classification point cloud, classified_new_forest.las, output in step 5 is used to search for neighboring points (i.e., points whose Euclidean distance from the point does not exceed r_filter) within a spatial range centered on each point. The proportion of points predicted as wood structures within this neighborhood is counted, and this proportion is used as the neighborhood consistency index s. When s≥tau, the label of the point is corrected to wood structure 1; when s≤1-tau, the label of the point is corrected to leaf 0; otherwise, the initial predicted label of the point remains unchanged, thereby suppressing sporadic misclassified points and improving spatial connectivity and stability.

[0017] In summary, this invention addresses the challenges of mixed distribution of woody structures (trunks and branches) and leaf structures in forest point clouds, low point cloud density, and insufficient local neighborhood points, which lead to instability of traditional fixed-scale features and difficulty in reliably separating wood and leaves. It proposes a complete workflow of "density-adaptive multi-scale geometric feature extraction—random forest supervised classification—post-processing optimization such as spatial consistency," enabling automatic identification and extraction of woody structure points and outputting confidence and spatial consistency results suitable for engineering diagnostics. Furthermore, this invention does not limit training data to simulated point clouds; as long as reliable point-level labels are provided, model training can be completed. These labels can originate from simulated point clouds, manually labeled point clouds, or high-confidence pseudo-labeled point clouds. The trained model can be directly transferred to inference and woody structure extraction in real unlabeled forest point clouds, and the extraction effect is evaluated using spatial consistency diagnostic indicators. Therefore, it possesses stable and usable engineering application capabilities even in scenarios where conventional methods struggle with low density, weak neighborhood support, and a lack of real labels. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0019] Figure 2 This is a schematic diagram illustrating point cloud density statistics and adaptive scaling.

[0020] Figure 3 A schematic diagram illustrating the distinction between the principal direction perpendicularity feature and the normal direction perpendicularity feature.

[0021] Figure 4 This is a schematic diagram of the wood structure extraction results for the labeled test set.

[0022] Figure 5 A schematic diagram illustrating the wood structure extraction effect of real, unlabeled forest point clouds. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings and examples.

[0024] The following example uses point cloud data from a low-density forest. For ease of explanation, this example provides two types of data: (1) labeled training data, used for supervised training and quantitative verification; and (2) real unlabeled point cloud data, used to demonstrate the engineering application effect and spatial consistency diagnosis. The labeled data consists of two LAS format point cloud files, wood.las and leaf.las, representing the wood structure point cloud and leaf point cloud, respectively. The real unlabeled data is new_forest.las, a mixed point cloud of wood structure and leaves. All point cloud data contain the three-dimensional coordinate information (x, y, z) of the points, where x, y, and z are in meters (m). When the point cloud file contains color attributes, each point also contains RGB color information.

[0025] A method for extracting wood structure from point clouds in low-density forests, such as Figure 1 As shown, it includes the following steps:

[0026] Step 1: Training Data Input and Label Construction. Read the point clouds of the wood structure and leaves from the training data, extract the 3D coordinates (x, y, z), and construct labels. Assign a value of 1 to the wood structure points and a value of 0 to the leaf points, forming the training sample point set P and its corresponding label Y.

[0027] Step 2: Point Cloud Density Statistics and Adaptive Scale Determination. A KD tree is constructed for the training sample point set P, and the nearest neighbor distance is calculated by randomly sampling points to obtain the mean nearest neighbor distance mu (unit: m). mu is used to characterize the average sparsity of the point cloud. A multi-scale radius sequence r_i = mu * m_i is generated based on mu, where m_i is the radius multiplier coefficient. A minimum radius lower limit r_min (unit: m) is set; when r_i is less than r_min, r_min is used to ensure that the number of neighboring points meets the minimum requirement under low-density conditions. The adaptive scale determination process is described in [link to documentation]. Figure 2 .

[0028] Step 3: Multi-scale geometric feature extraction. For a point p in the training sample point set P, perform a spherical neighborhood query at each radius r_i to obtain the neighborhood point set N_i(p).

[0029] When the number of neighboring points is less than the minimum neighboring point threshold min_neighbors, the scale feature is set to 0 to avoid instability in covariance estimation.

[0030] For a neighborhood point set that meets the required number of neighborhood points, the covariance matrix is ​​calculated and eigenvalues ​​and eigenvectors are decomposed to obtain eigenvalues ​​and eigenvectors. The eigenvalues ​​and their corresponding eigenvectors are then simultaneously sorted in descending order to ensure that the principal direction corresponding to the largest eigenvalue and the normal corresponding to the smallest eigenvalue have consistent physical semantics. Linearity, flatness, sphericity, curvature, anisotropy, and eigenentropy are calculated based on the eigenvalues. Based on the eigenvectors, the perpendicularity of the normal V3 = |v3_z| and the perpendicularity of the principal direction V1 = |v1_z| are introduced; the discrimination principle is described in [link to relevant documentation]. Figure 3 Furthermore, local point density is calculated based on the number of neighborhood points and neighborhood volume to characterize the sparsity of neighborhood sampling. Further, the K nearest neighbors of each point are retrieved at multiple K-nearest neighbor scales to form a KNN neighborhood. Based on the eigenvalues ​​of the neighborhood covariance matrix, three simplified PCA features—linearity, flatness, and sphericity—are calculated. The reciprocal of the average nearest neighbor distance within the neighborhood is used as a local density index. Global location features are then added to form a multi-scale feature vector for each point. Finally, the multi-scale feature vectors of all points in the point set are concatenated row-wise to construct the complete feature matrix.

[0031] Step 4: Random Forest Model Training. The feature matrix obtained in Step 3 is standardized; the labeled point set is divided into training and test sets by hierarchical classification based on labels, and a random forest classifier is trained to establish a supervised discrimination model for point-level woody / leaf structure.

[0032] In this embodiment, the results of extracting the wooden structure from the labeled test set are shown after training. Figure 4 Save the random forest model and its standardized parameters for subsequent classification of forest point clouds. The output of the trained random forest classifier is: predicted label. and the probability of belonging to the category of wooden structure Predicted Labels It belongs to {0,1}, where 1 represents woody structure and 0 represents leaf.

[0033] Step 5: Classification and Result Output of the Forest Point Cloud to be Classified. Read the real unlabeled forest point cloud new_forest.las (the forest point cloud to be classified), and treat each point in this point cloud as a point to be classified. Then, call the random forest model saved in Step 4 to perform point-level inference. The trained random forest classifier will perform neighborhood construction and multi-scale geometric feature extraction according to Steps 2 and 3, and use the standardized parameters saved in Step 4 to perform a uniform transformation on the features; output the preliminary predicted label for each point to be classified. And the probability P(y=1|x) of the wooden structure category, based on the output predicted label. The set of points representing the wood structure is obtained through filtering, and the final output is wood_only.las; at the same time, the complete classification point cloud classified_new_forest.las is also output, which supports visualization and subsequent processing.

[0034] Step 6: Take the result from Step 5 Spatial consistency filtering is performed on the point-level preliminary prediction results formed by P(y=1|x).

[0035] After obtaining the preliminary prediction results at the point level, spatial consistency filtering is performed on each point p in the point cloud to suppress speckle noise and enhance the spatial connectivity and stability of the wooden structure points: set the filtering radius r_filter (unit m) and the voting threshold tau (range 0 to 1), and in this embodiment, tau=0.7 is taken; retrieve the neighborhood point set N_filter(p) with the point p as the center and the Euclidean distance does not exceed r_filter, and count the proportion of points in the neighborhood point set whose preliminary prediction label is wooden structure (1), denoted as s(p).

[0036] When s(p)≥tau, the predicted label of point p is corrected to wood structure (1); when s(p)≤1-tau, the predicted label of point p is corrected to leaf structure (0); otherwise, the initial predicted label of point p remains unchanged, so as to obtain the final label of each point and realize the spatial consistency correction of isolated misclassified points.

[0037] Experimental data and effect analysis:

[0038] On labeled data, samples were stratified into training and test sets according to labels. The method of this invention (including spatial consistency filtering) was used on the test set to evaluate the following results: accuracy 0.8787, recall for wood structures 0.7198, and precision for wood structures 0.9246. Unsupervised geometric consistency metrics were: label consistency 0.7220, and boundary point ratio 41.7%.

[0039] On the real unlabeled point cloud new_forest.las, the trained random forest model was directly used for inference. After performing spatial consistency filtering and post-processing, the number of points in the wood structure was adjusted from 213,053 to 73,791, and the number of isolated points decreased from 3,980 to 0. Since real unlabeled point clouds cannot be used to calculate supervised metrics such as accuracy, this invention uses point count changes and spatial consistency diagnosis as engineering evaluation criteria. The extraction results are shown in [link to documentation]. Figure 5 .

[0040] As can be seen from the above embodiments, this invention addresses the challenges of mixed distribution of woody structures such as tree trunks and branches with leaf structures in forest point clouds, low point cloud density, and insufficient local neighborhood points, which lead to instability of traditional fixed-scale features and difficulty in reliably separating wood and leaves. It proposes a complete process of "density-adaptive multi-scale geometric feature extraction - random forest supervised classification - post-processing optimization such as spatial consistency" to achieve automatic identification and extraction of woody structure points and output confidence and spatial consistency related results that can be used for engineering diagnosis.

Claims

1. A method for extracting the wood structure from point clouds in low-density forests, characterized in that, Includes the following steps: Step 1: Inputting training data and constructing labels; Read the point cloud of the wood structure and the point cloud of the leaf in the training data, extract the three-dimensional coordinates (x,y,z) of all points to form the training sample point set P; Assign label 1 to the wood structure points and label 0 to the leaf points, and stitch them together to form labeled training samples. Step 2: Point cloud density statistics and adaptive scale determination; construct a KD tree for the training sample point set P from Step 1, and randomly sample several points to calculate the nearest neighbor distance, obtaining the mean nearest neighbor distance mu in m; generate a multi-scale radius sequence r_i = mu * m_i based on mu, where m_i is the scaling factor, and set a minimum radius lower limit r_min in m. Step 3: Multi-scale geometric feature extraction; For each point in the training sample point set P, perform a spherical neighborhood query at the corresponding radius r_i to obtain the neighborhood point set N_i; Calculate the neighborhood covariance matrix and perform eigenvalue decomposition on the neighborhood point set N_i that meets the neighborhood point number requirement to obtain the eigenvalues ​​lambda1>=lambda2>=lambda3 and the corresponding eigenvectors v1, v2, v3; Then, sort the eigenvalues ​​and eigenvectors in descending order to ensure that v1 corresponds to the largest eigenvalue lambda1 and v3 corresponds to the smallest eigenvalue lambda3. Based on eigenvalues, linearity, flatness, sphericity, curvature, anisotropy, and feature entropy features are constructed, and local point density is calculated; then, perpendicularity feature is further introduced: The perpendicularity of the normal direction is V3 = |v3_z|, and the perpendicularity of the main direction is V1 = |v1_z|. Then, the simplified PCA features and the reciprocal of the average nearest neighbor distance are extracted at the K-nearest neighbor KNN scale, and global position features are added to form a multi-scale feature vector for each point. Finally, the multi-scale feature vectors of all points in the training sample point set P are combined and concatenated row by row to construct a complete feature matrix. Step 4: Random Forest Classifier Training; Standardize the feature matrix obtained in Step 3, divide the training set and test set into layers according to the labels, train the random forest classifier, and save the trained random forest classifier and standardized parameters. The output of a trained random forest classifier is: predicted label. and the probability of belonging to the category of wooden structure Predicted Labels It belongs to {0,1}, where 1 represents woody structure and 0 represents leaf; Step 5: Classification and output of the forest point cloud to be classified; Apply the random forest classifier trained in Step 4 to each point in the forest point cloud to be classified, and output the predicted label. and the probability of belonging to the category of wooden structure The predicted label and class probability together constitute the classification result for each point; based on the predicted label... The set of points representing the wood structure is obtained by filtering, and the output is wood_only.las; at the same time, the complete classification point cloud classified_new_forest.las is also output, which supports visualization and subsequent processing.

2. The method for extracting wood structure from point clouds in low-density forests as described in claim 1, characterized in that: In step 2, when r_i is less than r_min, r_min is taken.

3. The method for extracting wood structure from point clouds in low-density forests as described in claim 1, characterized in that: In step 3, when the number of neighboring points is less than the minimum neighboring point threshold min_neighbors, the scale feature is set to 0.

4. The method for extracting wood structure from point clouds in low-density forests as described in claim 1, characterized in that: The complete classification point cloud classified_new_forest.las output in step 5 is used to search for neighboring points with a radius of r_filter within a spatial range centered on each point. The Euclidean distance between each point and the neighboring point is no more than r_filter. The proportion of points in the neighborhood that are predicted to be wood structures is counted and the proportion is used as the neighborhood consistency index s. When s≥tau, the label of this point is corrected to wood structure 1; when s≤1-tau, the label of this point is corrected to leaf 0; otherwise, the initial prediction label of this point remains unchanged.

Images