Method for identifying forest tree species by using laser point cloud data

Abstract

The invention provides a method for identifying forest tree species by using laser point cloud data, and the method comprises the following steps: collecting three-dimensional laser point cloud data of a forest region, setting an elevation threshold, filtering ground points, and extracting a point cloud sample object; respectively extracting VFH, CVFH and ESF feature descriptors from the sample point cloud, and constructing three types of geometric feature vectors; performing supervised classification on the features by adopting a random forest and a support vector machine learning classifier; the output of each classifier is fused through strategies such as weighted voting, an average method or a stacking method, and a final tree species identification result is obtained; according to the method, three types of global or semi-global feature descriptors of VFH, CVFH and ESF are extracted for a point cloud sample object of a single tree, feature modeling is carried out on tree species from three dimensions of spatial attitude, local scale structure and global shape distribution, the advantage of real restoration of a target structure by using point cloud data is utilized, and the feature modeling efficiency is improved. And the problem of projection distortion of image features under multiple view angles is avoided.

Description

Technical Field

[0001] This invention relates to the field of remote sensing forestry information extraction technology, and in particular to a method for identifying forest tree species using laser point cloud data. Background Art

[0002] Forests, as an important component of terrestrial ecosystems, are the largest carbon sink system in the global carbon cycle and play an irreplaceable role in maintaining ecological balance and responding to climate change. Forests not only have ecological functions such as carbon sequestration, water conservation, climate regulation, and air purification, but also provide key supporting data for ecosystem assessment, biodiversity monitoring, and carbon storage estimation. Tree species information, as a basic parameter for forest surveys, is of great significance for forest biomass estimation, ecosystem modeling, and forestry decision-making. Traditional tree species identification methods largely rely on manual plot surveys or interpretation of two-dimensional remote sensing images, which suffer from problems such as high workload, poor scalability, and lack of information on tree canopy spatial structure. With the development of LiDAR technology, three-dimensional laser point cloud data, with its advantages of high spatial resolution and high structural fidelity, provides a new technical approach for acquiring structural information at the individual tree level in forests. Point cloud data can accurately reflect the three-dimensional morphological characteristics of individual trees, thus providing a data foundation for tree species identification based on geometric structure. Existing point cloud classification methods mostly focus on geometric feature division based on rule thresholds, or rely solely on a single feature descriptor for classification. These methods are difficult to adapt to problems such as similar morphology among tree species, high data dispersion, and uneven point cloud distribution in natural forest areas. There is still considerable room for improvement in both recognition accuracy and generalization ability. Therefore, a method for identifying forest tree species using laser point cloud data is proposed. Summary of the Invention

[0003] In view of this, the present invention provides a method for identifying forest tree species using laser point cloud data, in order to solve or alleviate the technical problems existing in the prior art, and at least provide a beneficial alternative.

[0004] The technical solution of this invention is implemented as follows: A method for identifying forest tree species using laser point cloud data, comprising the following steps: S1. Collect three-dimensional laser point cloud data of forest area, set elevation threshold to filter the point cloud data for ground points, extract the point cloud of target objects on the ground, and extract point cloud sample objects of various tree species based on manual selection or point cloud segmentation. S2. For each point cloud sample object, extract point cloud feature descriptors that characterize its geometric structure features. The point cloud feature descriptors include: VFH feature descriptor constructs a global geometric histogram by statistically analyzing the angle information between the normals of points in the point cloud and the viewing angle. The CVFH feature descriptor is enhanced by constructing multiple unconnected smooth regions in the target point cloud and then synthesizing the local VFH histogram for each region. ESF feature descriptors construct shape function histograms using geometric functions such as area, angle, and point-to-point distance, and then concatenate them to form a global descriptive vector with fixed dimensions. S3. Using the extracted feature descriptors as input, supervise the classification of point cloud sample objects using two different types of machine learning classifiers, including random forest classifier and support vector machine classifier. S4. The prediction results of the two machine learning classifiers are fused together. A fusion strategy is used to integrate and judge the output category to obtain the final tree species identification result.

[0005] More preferably, in step S1, the extraction of the point cloud sample objects employs a strategy combining spatial voxel partitioning and manual correction, including: By setting an elevation threshold, the point cloud is filtered to remove ground points, and the retained ground target object point cloud data is extracted. The point cloud data area is divided into a three-dimensional spatial mesh to generate multiple voxel units. Each voxel unit contains a local point cloud cluster. Based on the density distribution and normal consistency of the point cloud clusters, connectivity analysis is used to identify potential single-tree structures. Combined with manual annotation tools and a visualization interface, the sample areas with unclear boundaries or misjudgments are corrected.

[0006] More preferably, in step S2, The extraction of the VFH feature descriptors includes: For each point in the point cloud sample object, calculate its normal vector, count the angle between the normals of each point and its nearest neighbors and form an angle distribution histogram, and construct a joint angle feature histogram through the angle information between the viewpoint direction and the point normal, which is used to express local geometric relationships and viewpoint structure information at the same time. The joint angular feature histogram is globally accumulated and normalized to output a 128-dimensional VFH global description vector.

[0007] The extraction of the CVFH feature descriptors includes: Based on the degree of normal variation of point cloud sample objects, they are divided into multiple non-connected smooth sub-regions. Local VFH feature histograms are calculated in each smooth sub-region to obtain the geometric information representation of the smooth sub-region. High curvature noise points are removed, and the local VFH feature histograms corresponding to each smooth sub-region are fused using an independent scale normalization method to generate a composite feature vector with both size and shape representation capabilities.

[0008] The extraction of the ESF feature descriptors includes: By randomly selecting point pairs from point cloud sample objects, three types of shape functions—Euclidean distance function, angle function, and area function—are calculated sequentially between each pair of points. The 10 sets of 64-dimensional sub-histograms generated by the three types of shape functions are spliced ​​and fused to form a 640-dimensional fixed-length shape function vector, which is used to characterize the global geometric structure features of point cloud sample objects. The boundary structure is extracted by traversing the point cloud voxel grid using the 3D Bresenham algorithm. In summary, using VFH, CVFH, and ESF feature descriptors can comprehensively characterize the spatial morphological features of point cloud sample objects from different geometric dimensions. Among them, the VFH feature descriptor effectively enhances the model's ability to express the target structure under different viewing postures; the CVFH feature descriptor, through region partitioning and high curvature removal strategies, divides the point cloud structure into multiple sub-structural units according to normal continuity, independently extracts local VFH within each sub-structure, and normalizes and fuses them, improving the model's robustness to scale changes and structural noise; the ESF feature descriptor constructs a high-dimensional histogram description vector based on shape functions such as the distance between point pairs, the included angle, and the area of ​​triangles. The three features complement each other and can effectively improve feature separation and discrimination accuracy in multi-category tree species recognition tasks.

[0009] More preferably, in step S3, the random forest classifier is constructed from M sub-decision trees. During the construction process, a sub-feature set of size A is randomly extracted from the input point cloud feature vector set for the construction of each tree splitting basis, where A is less than the original feature dimension d. During classification, each tree Ti(x) outputs a category result for the input sample x. The ensemble classification output F(x) of the random forest is F(x)=mate{T1(x),T2(x),...,Tm(x)}; Where mate represents taking the mode of the categories output by M trees as the overall recognition result.

[0010] The support vector machine classifier uses a radial basis function kernel, which is mathematically expressed as k(x,y)=exp(-γ||x–y||2); Where x and y are the feature vectors corresponding to the two sets of point cloud sample objects, ||x–y||2 represents the squared distance in Euclidean space, m represents the number of training samples, mate is the majority voting function, which means that the class with the highest frequency in all Ti(x) outputs is used as the final classification result F(x), and γ is the width parameter of the kernel function, which is used to control the similarity decay rate. The support vector machine classifier is trained on a set T = {(x1,y1),(x2,y2),...,(xm,ym)}. In this context, the category label Li∈{-1,+1} corresponding to each point cloud sample object is set, and the penalty parameter C is set to balance the objective function term of maximizing the classification margin and minimizing the classification error, so as to obtain the optimal segmentation hyperplane P in the high-dimensional mapping space, which is used to identify tree species samples of different categories. Each element (xi, yi) represents the i-th training sample; m represents the number of training samples, i.e., the total number of samples in the training set. More preferably, in step S4, the fusion strategy includes at least one of voting, averaging, or stacking. The output of multi-classifier fusion is filtered by setting a fusion confidence threshold. Only classification results with a fusion confidence greater than the preset threshold are accepted; otherwise, they are marked as samples to be reviewed for manual intervention or subsequent verification.

[0011] The embodiments of the present invention have the following advantages due to the adoption of the above technical solutions: I. This invention uses the geometric structural features contained in three-dimensional laser point cloud data as the classification basis. For point cloud samples of a single tree, three types of global or semi-global feature descriptors, namely VFH, CVFH, and ESF, are extracted. The tree species are modeled from three dimensions: spatial pose, local scale structure, and global shape distribution. This application utilizes the advantage of point cloud data to realistically restore the target structure, avoiding the problem of projection distortion of image features under multiple views. It can adapt to the differences in canopy structure in different sampling densities, different angles, and complex forest stands, and has good versatility and feature stability in actual forest resource surveys.

[0012] Second, this invention constructs two classifiers, random forest and support vector machine, based on supervised learning methods, and integrates and judges the model output results by fusing multiple classifiers. This can give full play to the discrimination advantages of each classifier for different feature structures. Compared with single classifier schemes, the method of this invention has stronger recognition ability in scenarios with complex forest structure and similar tree species.

[0013] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0014] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0015] Figure 1 This is a flowchart of the steps of the present invention; Figure 2 This is a sample image of the point cloud of Pinus tabuliformis according to the present invention; Figure 3 This is a sample image of the Liaodong oak point cloud from the present invention. Detailed Implementation

[0016] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0017] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] like Figure 1 As shown, this embodiment of the invention provides a method for identifying forest tree species using laser point cloud data, which mainly includes four steps: S1. Collect and extract tree species point cloud sample objects; S2. Extract structural geometric features from point cloud sample objects; S3. Construct a supervised classifier model to complete tree species identification; S4. Fusion classifier output to improve recognition accuracy; Through the above four-step process, it is possible to achieve high-confidence identification of typical tree species under the premise of complex forest environment and subtle differences in tree morphology, thereby providing stable and structured data support for accurate forest resource surveys and carbon sink monitoring.

[0019] Specifically, In step S1, point cloud data acquisition employs a fixed ground laser scanning platform or a portable multi-view scanning device to set up stations and perform high-density scanning around the target trees to obtain complete 3D point cloud information. The acquired raw point cloud data typically includes ground, shrub, herbaceous layer, and other interfering points. Ground points need to be filtered out by setting an elevation threshold, retaining only the target vegetation points on the ground. The point cloud region is divided into equally spaced 3D grid units using a spatial voxel partitioning algorithm. Each voxel unit serves as a local structure candidate. Based on this, preliminary candidate single tree structures are identified using structural features such as point density, normal distribution continuity, and centroid height, along with region growing and connected component analysis algorithms. Finally, samples with unclear boundaries are fine-tuned and confirmed manually using 3D point cloud visualization tools, thereby establishing a high-confidence "point cloud sample object set." Each object represents a structurally complete tree sample with clear structural hierarchy and distinguishable canopy morphology. This step effectively eliminates redundant background interference points and provides physical support for subsequent feature extraction and classification using sample-level structural representation. In step S2, for each point cloud sample object, three geometric feature descriptors are extracted sequentially: Viewpoint Feature Histogram (VFH), Clustered Viewpoint Feature Histogram (CVFH), and Shape Function Set Descriptor (ESF). These three descriptors comprehensively characterize the geometric structure of the point cloud from different dimensions. The VFH descriptor constructs a 128-dimensional joint angle histogram by estimating the normal vector for each point and combining the statistical information of the angle between the normals of neighboring points and the angle between the normal and the viewpoint direction. This represents the overall shape of the point cloud object and its structural features under the observation viewpoint. The CVFH descriptor introduces structural division of the point cloud region based on the VFH, first dividing the point cloud object according to the normal variation... The region is divided into multiple non-interconnected smooth sub-regions. VFH is calculated separately within each region and then independently scaled and normalized. The results are then concatenated to form CVFH features that enhance the expression structure, exhibiting strong scale sensitivity and morphological stability. The ESF feature descriptor uses random point-pair sampling to calculate the Euclidean distance D2, the area S3 of the triangle formed by the three points, and the angle distribution A3, forming three types of shape function histograms and concatenating them to generate a 640-dimensional global shape description vector. Through these three feature extraction methods, the geometric structure expression of point cloud sample objects under interference factors such as pose changes, scale differences, and uneven point density can be enhanced.

[0020] In step S3, the extracted VFH, CVFH, and ESF feature vectors are combined and then input into the random forest classifier and the support vector machine classifier for supervised learning training. The random forest uses Bagging to construct multiple sub-decision trees. During the training phase, each tree randomly selects A-dimensional sub-features from the original feature set for node splitting. The classification result is determined by voting from M trees, and the output category is based on the majority vote. It has the advantages of high stability and low overfitting. The Support Vector Machine (SVM) classifier constructs a high-dimensional mapping space using the radial basis function k(x,y)=exp(-γ||x–y||2), where γ is the kernel width parameter used to control the intensity of nonlinear mapping, and C is the error penalty term, which adjusts the optimal weight between the training error and the classification margin. The training set is T={(x1,y1),(x2,y2),...,(xm,ym)}. The optimal splitting hyperplane P is obtained by solving a quadratic programming problem. It has good linear expansion capability for tree species with blurred boundaries between samples. This step enables each tree sample object to obtain independent judgment outputs under multiple classifiers. At the same time, it establishes the response of different classifiers to the same structural input, which has the complementarity of classification results.

[0021] In step S4, the predicted outputs from the random forest and support vector machine are fused into a multi-classifier, employing three mainstream fusion strategies with switching capabilities: The first is the voting method, which is suitable for scenarios where the classifier outputs discrete labels. The final category is decided by majority voting or by weighting the verification accuracy, which can balance the impact of differences in the accuracy of different models. The second method is the averaging method, which is suitable for cases where the output is a probability vector. It averages the probability scores of each category or introduces a truncation strategy to filter out extreme prediction values ​​and improve the stability of the fusion. Thirdly, the stacking method uses the predicted outputs of multiple base classifiers as new input features to train a meta-classifier to complete the final discrimination. When the number of base classifiers exceeds 5, cross-validation is used to generate a first-level output to control model complexity and avoid overfitting. This multi-fusion strategy achieves integrated optimization of different classifiers under the same tree structure input, further improving the overall recognition accuracy while maintaining generalization ability.

[0022] In a preferred embodiment of the present invention, in addition to using random forest classifiers and support vector machine classifiers, other common supervised classification algorithms can also be selected for tree species identification of point cloud sample objects, including but not limited to neural network methods, maximum likelihood methods and K-nearest neighbor classification methods; the above methods can all be used as candidate classifiers to participate in ensemble decision-making, or can replace the original classifier for independent discrimination.

[0023] In addition, this embodiment uses a typical mixed coniferous and broad-leaved forest as the test area, such as Figure 2 and Figure 3 Selecting Pinus tabuliformis and Quercus liaotungensis as target tree species, a process of field sampling, point cloud processing, feature extraction, classification training, and evaluation was conducted. Specifically, I. Test Sites and Data Collection This test was conducted in a mixed forest area in the northern mountainous region, including natural Pinus tabuliformis forests, planted Pinus tabuliformis forests, and natural Quercus liaodongensis forests. The sample plots covered typical structures with significant differences in individual tree posture. A ground-based laser scanner was used to conduct panoramic 3D scanning from 5–8 locations on the ground. The equipment used a laser wavelength of 1550nm, with a scanning accuracy better than 5mm. The measurement radius was set to 80 meters, and the repeated scanning time interval was 10ms. Finally, a complete 3D point cloud file was generated. To enhance the applicability of the model under different density sampling conditions, samples were collected from areas with different crown densities, tree ages, and leaf volume levels in the sample plots. II. Sample Extraction and Preprocessing The original point cloud data after coordinate registration was merged by a laser scanning station. Ground point filtering was performed based on a height threshold set to 1.5m to remove ground and low shrub points. Then, the remaining point cloud was divided into spatial voxels with a voxel side length set to 0.2m. Based on the point density and structural connectivity clustering within the voxel blocks, it was initially divided into potential single-tree structural blocks. Combined with point cloud visualization and annotation tools, point cloud fragments with clear structures were manually confirmed and extracted as standard point cloud sample objects. No less than 100 single trees were extracted for each tree species, and the point cloud of each tree was controlled between 2048 and 5120 points to ensure data stability and geometric representation integrity.

[0024] III. Feature Descriptor Extraction For the standard single-tree point cloud sample objects that have been obtained, the three types of feature descriptors defined in this invention are extracted sequentially: VFH features: Calculate the normal vector for each point (nearest neighbor number k=20), construct a histogram of the angle between normals between points, and combine the angle information between the viewpoint direction (scanner position) and the normal to construct a 128-dimensional joint angle feature histogram; CVFH features: The point cloud is divided into multiple smooth regions, the normal change threshold is set to 10°, high curvature points at the edges are removed, the VFH features of each region are calculated separately, and finally the CVFH vector of the enhanced structure is formed by splicing them together. ESF features: 5000 pairs of points are randomly selected, and three types of shape function histograms are calculated for distance D2, angle A3, and triangle area S3. Each type has 64 dimensions, and the total features are spliced ​​together to form a 640-dimensional feature vector.

[0025] IV. Implementation of Classifier Training and Fusion Strategies Random forest classifier and support vector machine classifier are used as the basic supervised classifiers: RF configuration: number of trees M=100, maximum depth=20, feature subset A=√d for each tree, where d is the input feature dimension, and Gini impurity is used as the splitting criterion; SVM configuration: The radial basis function (RBF) is selected as the kernel function, and the parameters C and γ are jointly optimized by grid search (C∈{1,10,100}, γ∈{0.01,0.1,1}) and 5-fold cross-validation. The data was divided into a 70% training set and a 30% test set, and stratified sampling was used to maintain the balance of each type of sample. The independent outputs of the two classifiers are fused to make a decision. The main fusion method is weighted voting, with weights dynamically allocated based on the accuracy of the validation set. The weights are RF 0.55 and SVM 0.45, and a confidence threshold of ≥0.7 is set as a valid recognition output.

[0026] V. Performance Evaluation and Index Calculation The performance evaluation uses the following three core metrics: Overall recognition accuracy: Number of correctly recognized samples / Total number of samples; Macro average F1 score: Calculate the F1 score for each tree species separately and then take the average; Kappa coefficient: Used to evaluate the degree of consistency between the classifier output and the true label.

[0027] The results show that in the binary classification task of Pinus tabuliformis and Quercus liaodongensis, the accuracy of SVM or RF classifiers alone is 84.2% and 86.7%, respectively. After adopting the fusion strategy of this invention, the overall accuracy is improved to 91.3%, and the macro F1 value reaches 87.6%, which is more than 10% higher than traditional image classification or single classifier methods. Under different point cloud densities, different scanning angles and changes in tree canopy posture, the classification accuracy remains fluctuating by less than ±3%, verifying the robustness of this method under structural uncertainty conditions.

[0028] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying forest tree species using laser point cloud data, characterized in that: Includes the following steps: S1. Collect three-dimensional laser point cloud data of forest area, set elevation threshold to filter ground points of point cloud data, extract point cloud of target objects on the ground, and extract point cloud sample objects of various tree species based on manual selection or point cloud segmentation. Each point cloud sample object is controlled between 2048 and 5120 points. S2. For the point cloud sample objects, extract point cloud feature descriptors that can characterize geometric structure features, wherein the point cloud feature descriptors include: The VFH feature descriptor calculates the normal vector of each point in the point cloud sample object, counts the angle between the normals of each point and its nearest neighbors and forms an angle distribution histogram, and constructs a joint angle feature histogram through the angle information between the viewpoint direction and the point normal. The CVFH feature descriptor divides the point cloud sample object into multiple unconnected smooth sub-regions based on the degree of normal variation. The local VFH feature histogram is calculated in each smooth sub-region to obtain the geometric information representation of the smooth sub-region. ESF feature descriptor randomly selects point pairs from point cloud sample objects, and calculates three types of shape functions between each pair of points in turn: Euclidean distance function, angle function and area function. The 10 sets of 64-dimensional sub-histograms generated by the three types of shape functions are spliced ​​and fused to form a 640-dimensional fixed-length shape function vector. The three point cloud features are complementary, respectively expressing spatial pose, structural size and overall shape, forming a multi-dimensional structural fusion expression; S3. Use the extracted feature descriptors as input; S4. The prediction results of the two machine learning classifiers are fused.

2. The method for identifying forest tree species using laser point cloud data according to claim 1, characterized in that: In step S1, the extraction of the point cloud sample objects employs a strategy combining spatial voxel partitioning and manual correction, including: By setting an elevation threshold, the point cloud is filtered to remove ground points, and the retained ground target object point cloud data is extracted. The point cloud data area is divided into a three-dimensional spatial mesh to generate multiple voxel units. Each voxel unit contains a local point cloud cluster. Based on the density distribution and normal consistency of the point cloud clusters, connectivity analysis is used to identify potential single-tree structures. Combined with manual annotation tools and a visualization interface, the sample areas with unclear boundaries or misjudgments are corrected.

3. The method for identifying forest tree species using laser point cloud data according to claim 1, characterized in that: In step S3, the supervised classifier includes a random forest classifier and a support vector machine classifier. The random forest classifier constructs a nonlinear tree structure model based on voting, while the support vector machine classifier constructs an margin maximization model based on a kernel function. The two classifiers have different decision boundary structures. The random forest classifier is constructed from M sub-decision trees. During the construction process, a sub-feature set of size A is randomly extracted from the input point cloud feature vector set to be used as the basis for splitting each tree, where A is less than the original feature dimension d. During classification, each tree T i For input sample x, output a category result. The ensemble classification output F(x) of the random forest is given by F(x) = mate{T1(x),T2(x),...,T...} m (x)}; Where mate represents taking the mode of the categories output by M trees as the overall recognition result.

4. The method for identifying forest tree species using laser point cloud data according to claim 1, characterized in that: In step S3, the support vector machine classifier uses a radial basis function kernel, which is mathematically expressed as k(x,y)=exp(-γ||x–y|| 2 ); Where x and y are the feature vectors corresponding to the two sets of point cloud sample objects, ||x–y|| 2 This represents the squared distance in Euclidean space, where γ is the width parameter of the kernel function, used to control the magnitude of similarity decay. The support vector machine classifier is trained on a set T={(x1,y1),(x2,y2),...,(x...}. m ,y m )}; In this system, each point cloud sample object has a corresponding category label Li∈{-1,+1}, and a penalty parameter C is set to balance the objective function terms of maximizing the classification margin and minimizing the classification error, thereby obtaining the optimal segmentation hyperplane P in the high-dimensional mapping space, which is used to identify tree species samples of different categories.

5. The method for identifying forest tree species using laser point cloud data according to claim 1, characterized in that: In step S4, the fusion strategy includes at least one of the voting method, the averaging method, or the stacking method.

6. The method for identifying forest tree species using laser point cloud data according to claim 1, characterized in that: In step S4, the output of the fusion of the two machine learning classifiers is filtered by setting a fusion confidence threshold. Only classification results with a fusion confidence greater than the preset threshold are accepted; otherwise, they are marked as samples to be reviewed for manual intervention or subsequent verification.

Images