High-precision individual tree segmentation method based on multi-source point cloud data

Abstract

The invention relates to a high-precision single tree segmentation method based on multi-source point cloud data, and the method comprises the steps: building a high-density point cloud data set through the cooperative collection of an unmanned plane and a knapsack laser radar, carrying out the data preprocessing through the multi-site ICP registration and CSF filtering, carrying out the semantic segmentation based on an Enhanced Tree Instance Net network, and carrying out the segmentation of a single tree. Through feature contribution analysis and dynamic scene adaptive optimization segmentation data, semantic and geometric features are fused to realize individual tree instance segmentation, a physical constraint neural network is introduced to realize multi-parameter collaborative inversion of tree height, diameter at breast height, crown breadth, tree age and the like, it is ensured that the parameters conform to tree growth rules, precision evaluation and calibration are performed through a multi-level index system, and tree growth accuracy is improved. And finally realizing visual expression of a segmentation result and forest stand parameters. According to the method, the individual tree segmentation precision and the growth rule rationality are remarkably improved, and the method is suitable for urban forest carbon sink accurate measurement and intelligent forestry management.

Description

Technical Field

[0001] This invention relates to the field of forestry resource monitoring and computer vision interdisciplinary technology, specifically to a high-precision single-tree segmentation method based on multi-source point cloud data. Background Technology

[0002] With the increasing demands for urban ecological construction and the growing sophistication of forest resource management, the importance of accurate tree species identification in modern forestry monitoring systems is becoming increasingly prominent. The accuracy and efficiency of this technology not only affect the accuracy of ecosystem carbon sequestration but also directly impact the effectiveness of forest pest and disease control and biodiversity conservation. The rapid development of LiDAR technology has brought new opportunities to this field. By quickly acquiring three-dimensional point cloud data of trees, it can accurately capture key features such as canopy structure and trunk morphology, effectively overcoming the limitations of traditional optical imaging in acquiring three-dimensional information. This makes point cloud data-based tree species identification technology a major direction for current research and application.

[0003] However, the unique characteristics of complex forest environments present significant challenges to the implementation of these technologies. From the perspective of tree species characteristics, the high degree of morphological similarity among closely related species, coupled with individual variations within the same species due to differences in their growing environments, makes it difficult for traditional identification methods to establish stable classification models. Analysis of the data acquisition process reveals that factors such as changes in light intensity and vegetation obstruction in the field, combined with differences in scanning accuracy and data resolution among different LiDAR equipment models, result in significant fluctuations in the quality of the acquired point cloud data. This further increases the difficulty of feature extraction and severely restricts the improvement of final identification accuracy.

[0004] From a technical perspective, existing mainstream methods have significant limitations, as follows: At the data acquisition level, existing technologies largely rely on lidar data from a single platform, making it difficult to simultaneously acquire complete canopy structure and accurate tree trunk information. Airborne LiDAR has advantages in acquiring information about the top of the canopy, but its ability to detect understory structures is limited; while ground-based laser scanning can record tree trunks and understory vegetation in detail, it is difficult to fully describe the canopy morphology. This limitation of data sources leads to insufficient completeness of point cloud information, directly affecting the accuracy of subsequent segmentation and parameter inversion.

[0005] In the field of deep point cloud classification, while classic networks such as PointNet and PointNet++ achieve end-to-end processing of point cloud data, their processing efficiency drops significantly when dealing with large-scale forest point clouds. Furthermore, they struggle to effectively capture the multi-layered structural features of trees, from the canopy to the trunk. Simultaneously, the feature aggregation modules of these models, such as the Local Feature Aggregation module, are prone to dimensionality conflicts and lack filtering mechanisms for invalid indexes, leading to severe noise interference during local feature aggregation and ultimately affecting the reliability of the recognition results.

[0006] Existing methods for single-tree segmentation face numerous challenges. While traditional segmentation processes include complete steps such as point cloud denoising, ground filtering, and elevation normalization, and can effectively suppress false treetop interference through gradient descent, their complexity and poor adaptability to complex forest environments make them unsuitable. Deep learning-based solutions, such as open-source tools like Treelearn, while achieving instance segmentation through point coordinate offset projection, still face a challenge in balancing accuracy and efficiency in real-world complex forest applications. Unsupervised learning methods, although improving clustering results through contrastive learning, are prone to category confusion due to similar tree features; while multi-view clustering methods, while enhancing spatial information representation, often incorrectly segment a single tree into multiple cluster units vertically, severely impacting subsequent identification accuracy.

[0007] At the level of parameter inversion technology, existing methods have significant shortcomings. Traditional machine learning methods, such as random forest models based on tree factors, while having the advantages of being simple to operate and easy to implement, do not fully utilize the three-dimensional structural information of trees, and their recognition accuracy in complex forest scenarios is insufficient to meet the needs of practical applications. Of particular concern is that existing technologies rely excessively on empirical formulas and statistical models. For example, the widely used formula for estimating diameter at breast height (DBH) is DBH = 0.1 × tree height + 0.05 × crown width. This type of linear formula only considers a limited number of morphological parameters, which cannot accurately reflect the complex influence of tree species specificity, age stage, and environmental factors on tree growth, and lacks necessary consideration of biological mechanisms, resulting in a large deviation between the inverted parameters and the actual growth patterns of trees.

[0008] From a system integration perspective, existing single-tree segmentation systems typically treat segmentation, identification, and parameter estimation as independent processing steps, lacking an end-to-end collaborative optimization mechanism. This segmented processing mode not only easily leads to error accumulation effects but also results in low utilization of computing resources and limited processing efficiency. In engineering practice, existing technologies have poor compatibility with hardware environments, particularly under Windows operating systems, frequently encountering issues such as memory overflow and multi-process conflicts, making it difficult to meet the actual needs of rapid monitoring of modern urban forests in terms of processing efficiency for large-scale point cloud data.

[0009] In summary, existing forest tree species identification technologies have significant shortcomings in terms of environmental adaptability, feature fusion capability, anti-interference performance, and biological rationality. There is an urgent need to develop an innovative technical solution that can integrate multi-source data, integrate prior biological knowledge, and achieve end-to-end collaborative optimization in order to break through the current technical bottlenecks and meet the practical application needs of modern forest resource monitoring and management. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a high-precision single-tree segmentation method based on multi-source point cloud data. This method achieves fully automated and high-precision extraction of single-tree parameters conforming to biological laws from raw point clouds through multi-platform data collaboration, deep learning model optimization, and physical mechanism constraints.

[0011] To achieve the above objectives, the present invention employs a technical solution comprising the following steps: (1) Point cloud data acquisition: Use UAV airborne lidar and backpack lidar to collect point cloud data of complex forest land, perform point cloud data fusion, and construct a high-density point cloud dataset; (2) Data preprocessing: The collected raw point cloud data is subjected to denoising, registration, ground filtering, elevation normalization, outlier removal and voxel downsampling to obtain standardized point cloud data; (3) Semantic segmentation: Based on the Enhanced Tree Instance Net model, semantic segmentation is performed on the preprocessed point cloud to extract tree structure and category information; (4) Semantic segmentation data optimization: Optimize feature combination and segmentation parameters through feature contribution analysis and dynamic scene adaptation strategy; (5) Single tree segmentation: The semantic segmentation results are fused with geometric features, and single tree instance segmentation is achieved through instance embedding and classification branching; (6) Physical constraint neural network parameter inversion: The physical model of tree growth is introduced into the neural network loss function as a regularization constraint to realize the biological rationality inversion of tree height, diameter at breast height, crown width and tree age parameters; (7) Evaluation and calibration of single-tree segmentation accuracy: The segmentation results are quantitatively verified based on a multi-level index system and intelligent matching algorithm; (8) Visualization: Visualize the segmentation results and inversion parameters and calculate the stand parameters.

[0012] Furthermore, in step (1), the UAV airborne lidar adopts a multi-echo mode, with ≥4 echoes, a laser pulse frequency of 200-1000kHz, a scanning frequency of ≥400kHz, a flight altitude of 50-120 meters, a longitudinal overlap rate of ≥70%, a lateral overlap rate of ≥40%, and a point cloud density of 200-800 points / square meter. The backpack lidar has a scanning frequency of 200-600kHz, a point density of ≥1000 points / square meter, a positioning error of less than 3 cm, and adopts a combination of SLAM and IMU positioning.

[0013] Furthermore, the data preprocessing described in step (2) includes: Spatial alignment is performed using a multi-site ICP algorithm, with registration error controlled within 2 cm. Then, coordinate normalization and ground filtering are performed, and the CSF algorithm is used to separate ground and non-ground points. A digital elevation model (DEM) is generated based on the filtering results, and elevation normalization is performed to unify the point cloud height into a relative height system. Next, the canopy surface projection is generated, and false canopy regions are removed based on the gradient descent algorithm to eliminate false high-value points caused by noise or misregistration; Then, the normalized point cloud is sliced ​​at height to extract information about the trunk and middle layer structure. The SOR algorithm is applied for outlier removal. Finally, voxel downsampling is performed on the processed point cloud; The manually interpreted data is divided into a training set and a test set.

[0014] Furthermore, the semantic segmentation in step (3) adopts the Enhanced Tree Instance Net model. The standardized point cloud data obtained through the above preprocessing process is used as the input of the Enhanced Tree Instance Net model. It is loaded and converted into tensor format by the data_loaders function. After the data loader automatically completes the batch division and device transmission, it is sent to the network for training or inference. The semantic segmentation training process is as follows: First, enable CUDA synchronous debugging and device error detection by setting the environment variables CUDA_LAUNCH_BLOCKING and TORCH_USE_CUDA_DSA. The preprocessed standardized point cloud data is then loaded using the data_loaders function, with conservative settings of batch_size=1 and num_workers=0. The standardized point cloud data includes XYZ coordinates, RGB features, and manually interpreted labels. In the core training phase, the model receives [B, N, 6] dimensional input data and extracts features through a four-layer local feature aggregation module. At the same time, FocalLoss, which is designed specifically for class imbalance, is used to dynamically adjust the training weights of tree points and background points. Combined with the Adam optimizer and cosine annealing learning rate scheduler, stable convergence is achieved. After each training round, the total accuracy (OA) and mean intersection-to-union ratio (mIoU) are calculated synchronously using the calculate_metrics function, and the loss curve and accuracy metrics are visualized in real time using Tensor Board. Finally, save complete checkpoints containing model parameters, optimizer status, and training progress every 20-30 rounds.

[0015] Furthermore, the semantic segmentation data optimization in step (4) includes: First, the screening order is quantified by feature contribution metric, and the core features are identified by SHAP value analysis, including relative elevation, vertical density gradient and semantic segmentation confidence. Then, dynamic scene adaptation is implemented, and clustering parameters are dynamically adjusted according to point cloud density and tree canopy overlap rate, adopting an efficiency-first strategy or enabling a precision-first mode.

[0016] Furthermore, the single-wood segmentation described in step (5) includes: The data fusion pipeline optimized by semantic segmentation in step (4) is then imported into the Enhanced Tree InstanceNet single-tree segmentation model. The specific process is as follows: First, in the data preprocessing stage, the point-by-point category labels output by semantic segmentation, with trunk=0 and crown=1, are used as the fourth dimension feature and concatenated with the standardized three-dimensional coordinates to form a 4-channel input vector. Subsequently, during the forward propagation of the model, the hybrid features are processed by the encoder-decoder architecture of Enhanced Tree Instance Net, and point cloud features are extracted through multilayer perceptron (MLP), enabling semantic information and geometric features to interact deeply in the embedding space, and learning the spatial semantic pattern of vertical clustering of tree trunks and horizontal diffusion of tree crowns. Finally, multi-scale feature fusion is achieved through instance embedding branch and classification branch. Semantic information serves as a strong guiding signal in the instance embedding branch to distinguish tree boundaries, and provides contextual constraints in the classification branch, ultimately outputting the single tree segmentation result.

[0017] Furthermore, the physical constraint neural network parameter inversion in step (6) includes: By incorporating a physical model of tree growth as a regularization constraint into the neural network loss function, the allometric growth constraint represents the power-law relationship between tree height, diameter at breast height (DBH), and crown width. H = α × DBH β ×CD γ ; Where H is tree height, DBH is diameter at breast height, CD is crown width, and α, β, and γ are species-specific parameters; Growth rate constraints are based on the logistic growth model: dDBH / dt=r×DBH×(1–DBH / K); Where r is the inherent growth rate of diameter at breast height (DBH), K is the theoretical upper limit of DBH, which is determined by the tree species, and t is time; The morphological proportion constraint limits the crown width to tree height ratio to be between 0.2 and 0.8; The physical constraint loss function design incorporates physical constraints as soft regularization terms into the total loss function. L_total=L_data+λ_physics×L_physics; Where L_data is the data-driven error term; L_physics measures the difference between the network predictions and the physics model predictions. λ_physics represents the constraint weight coefficient.

[0018] Furthermore, in step (6), the physical constraint neural network is implemented in the following way: A multi-level constraint verification framework is constructed to monitor the biological rationality of prediction parameters in real time during the inference stage. When the deviation of the allometric growth equation exceeds 15-25%, the parameter correction mechanism is automatically triggered. We design a progressive feature decoupling architecture that separates and learns geometric features, semantic features, and physical features in a 256-512 dimension embedding space. Establish an adaptive constraint strength adjuster to dynamically adjust the strictness of physical constraints based on point cloud density and forest stand complexity. Use lenient constraints in simple isolated tree scenarios and enable strict constraints in complex dense forest scenarios. A cross-scale feature fusion mechanism is implemented, extracting local geometric features at three scales: 0.1m, 0.5m, and 1.0m, and then weighting and fusing them using attention weights. Deploy a real-time feedback loop system to automatically add samples with prediction errors greater than a preset value to the difficult example library for focused learning in subsequent training.

[0019] Furthermore, the accuracy assessment and calibration described in step (7) includes: The intelligent matching algorithm includes using KMeans clustering and adjusted RAND index (ARI) to match the predicted tree with the real tree; The aforementioned multi-level indicator system specifically involves constructing a four-dimensional evaluation framework, including indicators for point-level classification, clustering quality, instance detection, and segmentation quality.

[0020] Furthermore, the visualization in step (8) includes: The convex hull algorithm is used to draw the crown boundary of a single tree, and different colors are used to distinguish them. Automatically calculates forest stand parameters, and detects the number of trees, average tree height, crown diameter, diameter at breast height, and leaf area per tree; The leaf area index (LAI) is calculated based on the formula: LAI = total leaf area / plot area. The generated parameter distribution map shows the correlation between tree height-crown width and diameter at breast height-leaf area.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Innovative breakthroughs in multi-source data fusion and high-precision acquisition: Through the three-dimensional collaborative acquisition of UAV LiDAR and backpack LiDAR, a comprehensive coverage system from the canopy to the trunk was constructed, with a point cloud density of 200~800 points / square meter, which improved the information completeness by 40~50% compared with a single data source; the SLAM+IMU combined positioning technology still maintains a positioning accuracy of 2~3 cm in GNSS occlusion environment, solving the problem of signal loss in complex forest land; the multi-echo technology supports canopy penetration acquisition, increasing the amount of point cloud acquisition under the forest by 60~70%, providing an unprecedented data foundation for refined single tree segmentation.

[0022] (2) Revolutionary Architecture of Physically Constrained Neural Networks: Compared with traditional empirical formula estimation methods, the physically constrained neural network inversion estimation adopted in this invention has achieved a fundamental breakthrough in technology. The empirical formula DBH=0.1×tree height+0.05×crown width can only establish a simple linear relationship and cannot adapt to the differences in different tree species, terrain and forest stand conditions. Its estimation error is usually as high as ±5~8 cm. However, the physically constrained neural network of this invention integrates multi-source point cloud features through a deep learning architecture and innovatively integrates the three biological mechanisms of allometric growth equation, logistic growth model and morphological proportion constraint into the loss function, which fundamentally ensures that the parameter inversion conforms to the tree growth law. By balancing data-driven and physical consistency through an adaptive constraint weight adjustment mechanism, practical verification shows that this method improves the accuracy of diameter at breast height (DBH) estimation to ±2~3 cm, which is 60~65% higher than that of empirical formulas. At the same time, the biological rationality is improved from 65~70% to 90~95%. It can also simultaneously achieve the collaborative inversion of multiple parameters such as tree height (accuracy ±1~1.5 m), crown width (accuracy ±1~2 m), and tree age (accuracy ±8~15 years), providing reliable technical support for accurate measurement of forest carbon sequestration and smart forestry management.

[0023] (3) Technological leap in end-to-end intelligent segmentation: The input feature dimension is optimized from 7 dimensions to 4 dimensions (3 coordinates + 1 semantic), which improves the computational efficiency by 3 to 4 times while ensuring the integrity of information; the embedding space is enhanced from 16 dimensions to 24 to 32 dimensions, providing richer feature representation capabilities, and the single-tree segmentation ARI index reaches 0.85 to 0.92, which is 15 to 25% higher than the traditional method; a dual optimizer architecture is adopted to realize the collaborative training of segmentation and parameter inversion, and the training speed is accelerated by 2.5 to 3.0 times and the memory usage is reduced by 35 to 45% with the help of mixed precision technology.

[0024] (4) Breakthrough in precision of tree species specific processing: Establish a growth parameter database containing 3 to 5 major tree species, and achieve adaptive feature processing through 16 to 24-dimensional species embedding vectors; configure specific growth parameters for different tree species (coniferous forest α=1.8 to 2.0, broadleaf forest α=2.3 to 2.5), and the tree species identification accuracy reaches 90 to 95%, which is 25 to 30% higher than the general model; support online parameter optimization based on new data, and the model performance continues to improve by 3 to 5% annually, with good adaptive evolution capabilities.

[0025] (5) Practical innovation of engineering deployment: Deeply optimized for Windows system, setting num_workers=0 avoids multi-process conflicts, and the system operation stability reaches 98~99%; accelerates data transmission by pin_memory=True, and improves processing efficiency by 1.5~2.0 times; supports large-scale processing of 4096~8192 points / batch under 48GB video memory configuration, which increases throughput by 4~6 times compared with traditional methods; real-time constraint monitoring system tracks the satisfaction of indicators such as allometric growth and morphological ratio, and provides a visual basis for quality control.

[0026] (6) Synergistic improvement of accuracy and efficiency: Performance breakthroughs have been achieved in multiple dimensions, with RMSE of diameter at breast height inversion improved by 55-65%, tree height inversion accuracy improved by 50-60%, and crown width estimation error reduced by 55-60%; processing efficiency reached 3-4 times that of traditional methods, biological rationality increased from 68-72% to 92-96%; single tree detection recall rate increased from 80-85% to 94-97%, providing reliable technical support for precise forest monitoring.

[0027] (7) Significantly enhanced application value: It realizes the full-process automated processing from raw point cloud to single tree parameters, reduces manual intervention by 90-95%, and increases work efficiency by 8-10 times; improves the accuracy of leaf area index estimation by 30-40%, providing a scientific basis for carbon sink trading; supports generalized application in different regions and tree species, shortening the model migration adaptation time from 2-3 months to 2-3 weeks; and provides biological interpretability through physical constraints, enhancing the credibility and academic value of the results. Attached Figure Description

[0028] Figure 1 : A schematic diagram of the overall process of the method of the present invention.

[0029] Figure 2 Schematic diagram of the structure and loss function design of a physically constrained neural network.

[0030] Figure 3 Example of visualization of single-tree segmentation results (plan view and 3D view).

[0031] Figure 4 Comparison of segmentation effects of different methods in complex forest stands in the embodiments. Detailed Implementation

[0032] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments; it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the invention.

[0033] Example 1 like Figure 1 As shown, the technical solution adopted by the present invention includes the following steps: (1) Point cloud data acquisition: Use UAV airborne lidar and backpack lidar to collect point cloud data of complex forest land, perform point cloud data fusion, and construct a high-density point cloud dataset; (2) Data preprocessing: The collected raw point cloud data is subjected to denoising, registration, ground filtering, elevation normalization, outlier removal and voxel downsampling to obtain standardized point cloud data; (3) Semantic segmentation: Based on the Enhanced Tree Instance Net model, semantic segmentation is performed on the preprocessed standardized point cloud data to extract tree structure and category information; (4) Semantic segmentation data optimization: Through feature contribution analysis and dynamic scene adaptation strategy, optimize feature combination and segmentation parameters to improve segmentation accuracy and efficiency; (5) Single tree segmentation: The semantic segmentation results are fused with geometric features, and single tree instance segmentation is achieved through instance embedding and classification branching; (6) Physical constraint neural network parameter inversion: The physical model of tree growth is introduced into the neural network loss function as a regularization constraint to realize the biological rationality inversion of tree height, diameter at breast height, crown width and tree age parameters; (7) Evaluation and calibration of single-tree segmentation accuracy: The segmentation results are quantitatively verified based on a multi-level index system and intelligent matching algorithm; (8) Visualization: Visualize the segmentation results and inversion parameters and calculate the stand parameters.

[0034] In step (1), the UAV-borne lidar uses multi-echo technology with ≥4 echoes. The laser pulse frequency is 200-1000kHz, and the scanning frequency is ≥400kHz. The flight altitude is 50-120m, the flight path overlap rate is ≥70% longitudinally and ≥40% laterally, and the point cloud density should reach 200~800pts / m² to obtain complete canopy structure information. The backpack lidar requires a scanning frequency of 200~600kHz, a point density of ≥1000pts / m², a positioning error of <3cm, and uses SLAM+IMU combined positioning to adapt to GNSS obstruction environments.

[0035] In step (2), the point cloud is first denoised and registered. Spatial alignment is achieved using the multi-site ICP (Iterative ClosestPoint) algorithm, controlling the registration error to within 2cm. Then, coordinate normalization and ground filtering are performed. The CSF (Cloth Simulation Filter) algorithm is used to separate ground and non-ground points, with the cloth resolution (Cloth Resolution) set to 0.3~0.5m, the gravity direction being the positive Z-axis, and the number of iterations being 500, to obtain a high-precision terrain model. A digital elevation model (DEM) is generated based on the filtering results, and elevation normalization is performed using this model to unify the point cloud heights to a relative height system. Next, a canopy surface projection is generated, and a gradient descent algorithm is used to remove false canopy areas to eliminate pseudo-high-value points caused by noise or misregistration. Next, the normalized point cloud was height-sliced ​​from 0.6 to 2.1 m to extract trunk and mid-level structural information. Then, the Statistical Outlier Removal (SOR) algorithm was applied for statistical outlier denoising, with the number of neighborhood points (k) set to 20–30 and the standard deviation threshold (StdDevMulThresh) controlled between 1.0 and 1.5 to remove isolated points and anomalous noise. Finally, the processed point cloud was voxel-grid sampled, with voxel sizes set to 0.05–0.1 m to ensure a balance between data accuracy and computational efficiency. Simultaneously, the manually interpreted tree species and structure label data were divided into training and testing datasets in an 8:2 ratio to ensure the generalization ability and validation accuracy of subsequent tree segmentation and classification model training.

[0036] In step (3), the standardized point cloud data obtained through the preprocessing process, including XYZ coordinates, RGB features and manually interpreted labels, will be directly used as input to the Enhanced Tree Instance Net model. It will be loaded and converted into tensor format by the data_loaders function, and then automatically divided into batches and transmitted to the network for training or inference by the data loader.

[0037] Semantic segmentation training first enables CUDA synchronous debugging and device error detection through the environment variables CUDA_LAUNCH_BLOCKING and TORCH_USE_CUDA_DSA to ensure stability during large-scale point cloud computing. Then, it uses the data_loaders function to load preprocessed, standardized point cloud data, employing conservative settings of batch_size=1 and num_workers=0 to avoid memory overflow and multi-threading conflicts. In the core training phase, the model receives [B, N, 6]-dimensional input data and processes it through a four-layer local feature aggregation module (Local Feature... The system extracts features using aggregation and dynamically adjusts the training weights of tree points and background points using FocalLoss, which is designed specifically for class imbalance. This is combined with the Adam optimizer and cosine annealing learning rate scheduler to achieve stable convergence. After each training round, the total accuracy (OA) and mean intersection-over-union ratio (mIoU) are calculated synchronously using the calculate_metrics function, and the loss curve and accuracy metrics are visualized in real time using TensorBoard. Finally, a complete checkpoint containing model parameters, optimizer status, and training progress is saved every 20-30 rounds, forming an end-to-end training system that takes into account the special characteristics of forestry data, training stability, and model reproducibility.

[0038] In step (4), the screening order is first quantified by feature contribution. SHAP value analysis is used to identify three core indicators from the original 14-dimensional features: relative elevation, vertical density gradient, and semantic segmentation confidence, with a contribution rate of 78%. Then, dynamic scene adaptation is implemented. For simple scenes, such as isolated trees and sparse forests, the features are point cloud density <100 points / square meter and canopy overlap rate <10%. The system adopts an efficiency-first strategy, setting the neighborhood radius eps of the clustering algorithm to 0.7~1.0 meters, reducing the minimum number of points min_samples to 5~10, and only making The system uses core features such as elevation and semantic probability for rapid processing. For complex scenarios, such as dense forests with overlapping canopies (point cloud density > 500 points / m² or overlap > 30%), a precision-first mode is activated, strictly limiting EPS to 0.3-0.5 meters and increasing min_samples to 15-25. Simultaneously, the system incorporates the full feature set, including curvature and density gradient, for fine segmentation. In special scenarios, such as steep terrain with a slope > 30 degrees or mixed forests with multiple tree species, the system dynamically activates a terrain slope compensation algorithm and optimizes grouping parameters based on the morphological characteristics of different tree species. For example, for coniferous forests with narrow canopies, EPS is set to 0.4-0.6 meters; for broadleaf forests with larger canopies, EPS is set to 0.6-0.8 meters.

[0039] The algorithm automatically switches feature combinations based on point cloud density and forest stand complexity—for simple forest areas, only core indicators are retained to achieve 3x faster processing, while for complex overlapping areas, α-shape contour coefficients and local curvature entropy are introduced for enhanced analysis. Finally, through multi-level verification feedback, the segmentation results are cross-validated with manual annotations, and the feature weight thresholds are dynamically adjusted to form a closed-loop optimization system of "basic indicators to ensure efficiency - enhanced features to ensure accuracy - feedback mechanism to promote iteration". Ultimately, while ensuring segmentation accuracy of no less than 98.5%, the algorithm efficiency is increased to 3.2 times that of traditional methods.

[0040] In step (5), the semantic segmentation optimized data is imported into the Enhanced TreeInstance Net single-tree segmentation model through an end-to-end feature fusion pipeline: First, in the data preprocessing stage, the point-by-point category labels (trunk=0, crown=1) output by semantic segmentation are concatenated with the standardized three-dimensional coordinates as the fourth dimension feature to form a 4-channel input vector; then, during the forward propagation of the model, the hybrid feature is processed by the encoder-decoder architecture of Enhanced Tree Instance Net, and point cloud features are extracted through a multilayer perceptron (MLP) to enable semantic information and geometric features to interact deeply in the embedding space and learn the spatial semantic pattern of "vertical clustering of trunks and horizontal diffusion of crowns"; finally, multi-scale feature fusion is achieved through instance embedding branch and classification branch. Semantic information serves as a strong guiding signal to distinguish tree boundaries in the instance embedding branch and provides contextual constraints in the classification branch, ultimately outputting a single-tree segmentation result that maintains geometric accuracy and conforms to biological structure.

[0041] In step (6), the physical model of tree growth is incorporated as a regularization constraint into the neural network loss function. The biological rationality of the network output parameters is restricted by the growth equation, specifically including: Allometric growth constraint: Establishing a power-law relationship between tree height, diameter at breast height (DBH), and crown width: H = α × DBH β ×CD γ Where H is tree height, DBH is diameter at breast height (DBH), and CD is crown width; α, β, and γ are species-specific parameters obtained through field measurements or literature reviews, where α = 1.8~2.5, β = 0.60~0.68, and γ = 0.12~0.20.

[0042] Growth rate constraint: The annual growth conforms to the logistic growth model dDBH / dt = r × DBH × (1 - DBH / K), limiting the growth rate predicted by the network to a reasonable range. Here, r is the inherent growth rate of diameter at breast height (DBH), K is the theoretical upper limit of DBH determined by the tree species, and t is time.

[0043] Morphological proportion constraint: The crown width-to-tree height ratio is maintained within an ecologically reasonable range: CD / H∈[0.2, 0.8], and predictions that exceed the range are penalized.

[0044] Physical constraint loss function design: Physical constraints are added as soft regularization terms to the total loss function: L_total = L_data + λ_physics × L_physics, where L_data is the data-driven error term, L_physics measures the difference between the network prediction and the physical model prediction, and λ_physics is an adjustable constraint weight coefficient.

[0045] This framework effectively ensures that the tree parameters predicted by the neural network conform to biological laws, thereby improving the model's generalization ability and interpretability.

[0046] Furthermore, such as Figure 2 As shown, in step (6), the physical constraint neural network is implemented in the following way: a multi-level constraint verification framework is constructed to monitor the biological rationality of the prediction parameters in real time during the inference stage. When the deviation of the allometric growth equation exceeds 15~25%, the parameter correction mechanism is automatically triggered. A progressive feature decoupling architecture is designed to input geometric features, including coordinates (XYZ) and curvature, semantic features, including trunk and crown classification, and physical features, i.e. growth parameters, into the encoder-decoder network. Separation learning is performed in the 256~512-dimensional embedding space to avoid feature confusion, and output parameters H, DBH, CD, and Age. On the one hand, the physical constraint loss function is calculated, including allometric growth constraint, growth rate constraint, and morphological proportion constraint. Then, gradient backpropagation is performed, and the parameters are re-entered into the encoder-decoder network for learning. On the other hand, biological rationality verification is performed. If the parameters are reasonable, they are output. If they are unreasonable, the parameter correction mechanism is entered to adaptively adjust the constraint weights. The adjusted parameters are re-entered into the encoder-decoder network for learning.

[0047] An adaptive constraint strength adjuster was established to dynamically adjust the strictness of physical constraints based on point cloud density and stand complexity. The point cloud density ranged from 200 to 800 points / m², and the canopy overlap rate of the stand complexity ranged from 10% to 30%.

[0048] In simple scenarios, such as isolated trees, a relaxed constraint is adopted, with λ=0.03~0.05. In complex scenarios, such as dense forests, a strict constraint is adopted, with λ=0.15~0.20. A cross-scale feature fusion mechanism is implemented, extracting local geometric features at three scales: 0.1m, 0.5m, and 1.0m. Weighted fusion is performed using attention weights of 0.2~0.8, which improves the detection accuracy of small targets by 32~45%. A real-time feedback loop system is deployed to automatically add samples with prediction errors greater than 20% to the hard example library. These samples are then given 2~3 times the weight for focused learning in subsequent training, gradually eliminating system bias.

[0049] In step (7), the core of the single-tree segmentation accuracy evaluation system is to quantitatively verify the segmentation results through a multi-level indicator system and intelligent matching algorithm. Key points include using KMeans clustering and adjusted RAND index (ARI) to accurately match the predicted tree with the real tree, which solves the limitations of point-to-point matching in traditional evaluation; constructing a four-dimensional evaluation framework covering core indicators such as point-level classification, clustering quality, instance detection, and segmentation quality; and introducing spatial distribution analysis and oversegmentation / undersegmentation index to evaluate the rationality of the segmentation boundary.

[0050] The point-level classification accuracy was 0.85–0.95, and the Kappa coefficient was 0.80–0.90. The clustering quality ARI index is 0.75~0.90; The instance detection precision is 0.80~0.95, and the recall rate is 0.75~0.90; The segmentation quality IoU is 0.70~0.85, and the Dice coefficient is 0.75-0.88; The distance correlation in the spatial distribution analysis was 0.65~0.80; The ideal range for the oversegmentation / undersegmentation index is 0.05 to 0.15.

[0051] This system integrates various metrics into a single evaluation standard using a weighted comprehensive score (0.0~1.0), with point-level accuracy weighted at 15%, detection F1 score at 15%, and average IoU at 20%. Key adjustable parameters include a spatial tolerance radius of 0.3~0.8 meters, a minimum effective point threshold of 5~20 points, and a similarity matching threshold of 0.6~0.8. The system also provides a visual dashboard, including radar charts, confusion matrices, IoU distribution histograms, and scoring pointers, which can intuitively display the algorithm's comprehensive performance in tree number detection, boundary segmentation accuracy, and spatial distribution preservation, providing detailed quantitative data for algorithm optimization.

[0052] In step (8), the single-tree segmentation visualization and parameter calculation system displays the single-tree segmentation results through a planar diagram. The crown boundary of each tree is drawn using the convex hull algorithm, with 20 color distinctions, and the tree ID is marked at the center of the crown. The system automatically calculates key stand parameters: detects the number of trees, estimates the average tree height, crown diameter, and diameter at breast height (DBH) through physical constraint neural network inversion, and calculates the single-tree leaf area using the crown geometry method and allometric growth equation weighted calculation. The number of trees is 8-15 / 400㎡, the average tree height is 10-25 meters with an accuracy of ±0.5 meters, the crown diameter is 3-12 meters, the DBH is 15-45 centimeters, and the single-tree leaf area is 30-120㎡. Figure 3The diagram shows a schematic of the convex hull boundary obtained by using the single-tree segmentation algorithm of this invention under the convex hull algorithm.

[0053] Finally, the leaf area index (LAI) of the sample plot was calculated using the formula LAI = total leaf area / plot area, with a typical value of 1.5-3.5. The total leaf area measurement area is typically 20 × 20 meters. Simultaneously, parameter distribution maps were generated to display key indicators such as the tree height-crown width relationship (R² = 0.6-0.8) and the correlation between diameter at breast height (DBH) and leaf area. Figure 3 As shown, this provides quantitative data support for forest carbon sink assessment, with tree density ranging from 120 to 300 trees per hectare and total leaf area ranging from 600 to 1800 square meters.

[0054] Furthermore, in step (3), RandLA-Net is a model specifically optimized for complex forest single-tree segmentation tasks. Its semantic segmentation system has a code structure consisting of four core parts: data preprocessing, model architecture, training logic, and evaluation module. In the data preprocessing stage, the system loads point cloud data containing 6-dimensional features, including XYZ coordinates and RGB information, through the data_loaders function. The conservative settings of batch_size=1 and num_workers=0 are used to avoid memory overflow problems in large-scale forestry point clouds. At the same time, the data utilization rate is optimized through the active_learning sampling strategy. The model architecture consists of a Local Feature Aggregation module. This module first performs strict index safety processing: it uses torch.clamp(neighbors, 0, N-1) to restrict the neighbor index to a valid range, and then uses neighbors.view(B, 1, NK) to reshape the index dimension to ensure that the subsequent torch.gather operation can stably obtain neighbor features. In the feature aggregation stage, a dual-path MLP structure is used, with mlp1 processing neighbor features and mlp2 refining the output. Combined with an attention mechanism, nn.Conv1d(d_out2, 1, 1) is used to calculate weights and dynamically fuse local features and its own features. The training logic first uses a custom FocalLoss class to dynamically adjust the weights of easy and difficult samples using the (1-pt)gamma*ce formula, addressing the extreme imbalance between tree and background points in forestry scenarios. Secondly, a Cosine Annealing LR learning rate scheduler is configured to smooth the training process. Finally, the `calculate_metrics` function synchronously calculates the overall accuracy (OA) and mean intersection-over-union (mIoU), with the confusion matrix constructed using a class-by-class traversal approach to ensure evaluation accuracy. These key settings synergistically deliver significant benefits: the index safety mechanism prevents index out-of-bounds crashes in complex terrains, FocalLoss mitigates the impact of class bias on model performance, and the multi-metric evaluation system provides more comprehensive accuracy verification for forestry applications, ultimately enabling the system to achieve both robustness and accuracy in single-tree segmentation tasks.

[0055] Furthermore, in step (5), the Enhanced Tree Instance Net is a model specifically optimized for complex forest single-tree segmentation tasks. The single-tree segmentation system simplifies the feature dimension from the original 7 dimensions to 4 dimensions (coordinates + semantics), and enhances the embedding dimension to 24 dimensions; the batch size is set to 32 to adapt to large GPU memory configurations, and the number of points per batch is increased to 4096; the learning rate is adjusted to 0.002 and is coordinated with cosine annealing warm-up restarts (T_0=5, T_mult=2, η_min=1e-6); the loss function weights are set to a ratio of embedding loss of 0.7 and classification loss of 0.3, with a margin of 1.0. In terms of training strategy, the number of training rounds is controlled within 30 rounds, and mixed precision training (torch.cuda.amp.autocast and Grad Scaler) and KMeans clustering evaluation (sampling ratio 0.2~0.3) are adopted. At the same time, for Windows systems, num_workers is forced to be set to 0 to avoid multi-process conflicts.

[0056] Furthermore, in step (7), the training process of the physical constraint neural network includes: adopting a phased progressive training strategy; in the first stage, rounds 1-10 focus on instance segmentation tasks, with physical constraint weights set to 0.01-0.03; in the second stage, rounds 11-20 gradually introduce parameter regression, increasing the constraint weights to 0.08-0.12; in the third stage, rounds 21-30 strengthen physical consistency, with constraint weights reaching 0.15-0.20; implementing a gradient balancing mechanism, avoiding gradient conflicts in multi-task training through gradient normalization and task weight adaptation, improving training stability by 40-60%, wherein the gradient normalization norm is limited to 1.0-2.0, and the task weights are adaptively adjusted every 5 rounds.

[0057] A physical plausibility verification loop is introduced, and the validation set is screened for biological plausibility after each training round. Batches with a plausibility of less than 80% are retrained in a targeted manner, with the retraining ratio controlled at 10-20%. A mixed-precision training process is adopted, combining FP16 forward computation with FP32 gradient accumulation, which improves the training speed by 2.2-2.8 times and reduces memory usage by 35-45% while maintaining numerical accuracy. An early stopping mechanism and model ensemble are deployed, which automatically stops training when the physical consistency index of the validation set does not improve for 5-8 consecutive rounds, and integrates 3-5 best checkpoints for inference, improving the robustness of the model by 25-35%.

[0058] The method for constructing the tree species-specific growth parameter database includes: implementing a multi-seasonal data collection plan, conducting 2-3 repeated measurements during the growing season (April to September) and the dormant season (October to March of the following year) to capture seasonal growth dynamics, with the seasonal variation coefficient controlled at 8-15%; establishing a quality control system, setting an abnormal threshold of 0.01-0.05 for the diameter at breast height (DBH) to tree height ratio and a reasonable range of 0.2-0.8 for the crown height ratio, automatically removing abnormal measurement data that exceed the range, and improving the data qualification rate to 92-97%; and constructing a parameter uncertainty quantification model to provide a standard deviation estimate for each growth equation coefficient. The standard deviation of α is 0.15~0.25, and the standard deviation of β is 0.08~0.12, supporting probabilistic prediction and risk assessment; a parameter spatial interpolation system is developed, based on geographic information system (GIS) and environmental factors, including annual mean temperature and precipitation, to establish a regional parameter prediction model, achieving accurate parameter estimation in areas without measurement data, with an interpolation accuracy of 85-90%; a parameter version management and traceability mechanism is established to record the timestamp, data source and modification content of each update, supporting long-term tracking and scientific analysis of parameter evolution, and meeting the requirements of forestry scientific research for data repeatability.

[0059] Example 2 Forest plot single tree identification and comparison experiment Figure 4 The images show a comparison of the segmentation effects of different methods in complex forest stands, viewed from the front, top, and bottom views. Figure 3 The visual effects of different segmentation methods are fully demonstrated in this paper. (a) Method A: DBSCAN clustering algorithm; (b) Method B: seed point region growing algorithm; and (c) the method of this invention. The results show that this method outperforms traditional DBSCAN clustering and seed point region growing algorithms in terms of segmentation accuracy, boundary clarity, category integrity, multi-view adaptability, and algorithm stability. It is particularly suitable for 3D point cloud or image classification tasks with complex structures, rich details, and requiring multi-view analysis. The specific method is as follows: 1. Experimental plot setup An experimental plot measuring 100m × 100m was selected within the Haiwan Forest Park in Shanghai. This plot exhibits complex forest characteristics, comprising 45% Pinus tabuliformis, 30% Quercus variabilis, 15% Robinia pseudoacacia, and 10% other tree species in a mixed forest. The average tree height is 12.5 meters, ranging from 6 to 22 meters, with a stand density of 850 trees per hectare and a canopy closure of 0.7. This plot effectively represents the typical characteristics of complex urban forest land.

[0060] 2. Selection of comparison methods To systematically evaluate the overall performance of this invention, two representative traditional single-tree identification methods were selected as comparative benchmarks. Method A uses an improved DBSCAN clustering algorithm to segment single trees based on elevation and density features, with parameters set to eps=1.2m and min_samples=8. Method B uses a seed point region growth algorithm, utilizing canopy height model features, with a growth threshold of 0.8m and a minimum canopy width of 2m. These two methods represent the mainstream technical routes of traditional point cloud clustering and canopy model-based methods, respectively, and neither introduces a biological physical constraint mechanism. As a comparison, the method of this invention, namely Method C, adopts an Enhanced Tree Instance Net network architecture, combined with a physical constraint mechanism. The feature dimension is optimized to 4-dimensional input, the embedding space is enhanced to 24-dimensional, the physical constraint weight λ_physics=0.1, and the batch size is set to 32, forming a complete physical constraint single-tree segmentation technology system.

[0061] 3. Data Acquisition and Processing The experiment employed a multi-source data acquisition scheme. A DJI Zenmuse L1 drone with LiDAR was used to acquire point cloud data at a flight altitude of 80 meters, achieving a point cloud density of 280 points / m². Simultaneously, a Hummingbird Genesis backpack LiDAR was used for supplementary ground data acquisition, achieving a point cloud density of 1200 points / m². To obtain accurate verification data, a total station was used to precisely measure 125 standard trees within the sample plot. In the data preprocessing stage, the ICP algorithm was used for point cloud registration, with the registration error controlled within 2 cm. The CSF algorithm was used for ground filtering, with the data set to a resolution of 0.4 meters. Finally, elevation normalization was performed based on the DEM.

[0062] 4. Experimental environment configuration The system utilizes a hardware platform with an Intel i9-12900K processor, RTX 4090 graphics card, and 64GB of RAM, and is tested under Windows 11 using Python 3.9 and PyTorch 1.13 frameworks. Data acquisition employed a DJI Zenmuse L1 drone LiDAR and a Hummingbird Genesis backpack LiDAR to obtain multi-source datasets with point cloud densities ranging from 280 to 1200 points / square meter. Validation data was obtained by precisely measuring 125 standard trees using a total station. Performance improvement was calculated using a standardized formula: Performance Improvement (%) = (Result of this Invention - Result of Best Traditional Method) / Result of Best Traditional Method × 100, ensuring the objectivity and comparability of the evaluation results.

[0063] 5. Experimental Results and Analysis Table 1: Comprehensive Performance Comparison Analysis ; (1) Analysis of the effect of single wood detection Experimental results show that traditional methods have significant limitations in single-tree detection. Method A, based on point cloud clustering, achieved a detection rate of 78.3%, with oversegmentation and undersegmentation rates of 15.2% and 21.8%, respectively, and an F1 score of only 0.763. Method B, based on region growing, showed improved performance, achieving a detection rate of 82.5%, with oversegmentation and undersegmentation rates of 12.8% and 18.5%, respectively, and an F1 score of 0.795.

[0064] In contrast, Method C of this invention, through the introduction of physical constraint mechanisms and an optimized feature extraction architecture, achieves a detection rate of 87.8%, reduces oversegmentation and undersegmentation rates to 9.5% and 14.2%, respectively, and reaches an F1 score of 0.845. It is noteworthy that this invention primarily achieves significant improvements in reducing missegmentation (oversegmentation rate reduced by 25.8%, undersegmentation rate reduced by 23.2%), while the improvement in detection rate is relatively modest (6.4%), which aligns with the actual performance characteristics of deep learning methods in complex scenarios.

[0065] (2) Accuracy analysis of geometric parameter estimation Regarding the accuracy of geometric parameter estimation, the method of this invention shows a steady improvement. Method A has a positional error of 1.25 meters, a tree height estimation error of 1.82 meters, and a crown width estimation error of 2.35 meters. Method B shows improved estimation accuracy, with a positional error of 0.98 meters, a tree height error of 1.45 meters, and a crown width error of 1.86 meters.

[0066] Method C of this invention improves upon these indicators: positional error is reduced to 0.75 meters (an improvement of 23.5%), tree height error to 1.12 meters (an improvement of 22.8%), and crown width error to 1.48 meters (an improvement of 20.4%). This improvement ranges from 20% to 25%, which is in line with the expected results based on multi-source data fusion and deep feature learning. It also reflects a certain bottleneck in improving absolute accuracy in complex forest environments.

[0067] (3) Accuracy analysis of diameter at breast height inversion The comparison of the accuracy of the diaphragm diameter (DBD) inversion demonstrates the practical value of the physical constraint mechanism. Method A achieved a DBD inversion RMSE of 8.2 cm, a MAE of 6.5 cm, and a coefficient of determination (R²) of 0.68. Method B achieved slightly higher accuracy, with an RMSE of 6.8 cm, a MAE of 5.2 cm, and an R² of 0.73.

[0068] Method C of this invention, by introducing an allometric growth equation, a logistic growth model, and morphological proportion constraints, reduces the RMSE of diameter at breast height (DBH) inversion to 5.2 cm (an improvement of 23.5%), the MAE to 4.1 cm (an improvement of 21.2%), and the R² to 0.81 (an improvement of 11.0%). It should be noted that the error in DBH inversion mainly stems from objective factors such as uneven point cloud density and trunk occlusion. Even with the introduction of physical constraints, a systematic error of approximately ±5 cm still exists, a result consistent with the current capabilities of LiDAR technology.

[0069] (4) Analysis of biological rationality In the biological rationality assessment, this invention demonstrated stable improvement effects. Method A achieved a 65.3% compliance rate for allometric growth, a 68.7% compliance rate for morphological proportions, and a comprehensive biological score of 66.8. Method B achieved 71.2%, 73.5%, and 72.1 points for the relevant indicators, respectively.

[0070] Method C of this invention achieved an improvement of approximately 12% in these indicators, with an allometric growth accuracy of 79.8%, a morphological proportion reasonableness of 82.3%, and a comprehensive biological score of 80.9. This improvement is reasonable because the physical constraint mechanism mainly operates in the post-processing stage of parameters, which can correct predictions that clearly violate biological laws, but cannot completely eliminate the systematic bias of the model itself. In the experiment, approximately 20% of the samples still showed slight biological inconsistencies, mainly due to growth environment and individual differences.

[0071] (5) Typical scenario adaptability analysis In dense forest scenarios (canopy overlap > 30%), traditional methods face significant challenges. Method A's undersegmentation rate rises to 28.5%, and Method B's oversegmentation rate reaches 18.3%, while the overall error rate of Method C in this invention is controlled at 23.6%, a reduction of approximately 15-20% compared to traditional methods. In mixed forest scenarios, Method A achieves a detection rate of 82.1% for coniferous trees and 74.5% for broadleaf trees, a difference of 7.6%; the corresponding data for Method B are 84.3% and 79.2%, a difference of 5.1%; the detection rates of Method C in this invention are 88.2% and 85.6%, respectively, with the difference narrowing to 2.6%, indicating a more balanced adaptability to different tree species.

[0072] In complex terrain scenarios (slope > 25°), the detection rate of method A drops to 71.8%, method B drops to 76.4%, and the detection rate of method C of this invention remains at 82.3%. The performance drop compared to flat terrain scenarios (87.8%) is only 6.3%, which is significantly better than the 12-15% performance drop of traditional methods, but there is still some terrain sensitivity.

[0073] (6) Computational efficiency analysis From a computational efficiency perspective, method A takes 45 minutes to process and uses 8.2GB of memory, with limited parallelization support; method B takes 38 minutes to process and uses 6.5GB of memory, and has some parallelization capabilities.

[0074] The processing time of Method C in this invention is 42 minutes, slightly longer than Method B (an increase of 10.5%), and the memory usage increases to 11.8 GB (an increase of 81.5%). This result reflects the cost of the improved accuracy: the physically constrained neural network requires additional forward computation to evaluate biological consistency, and multi-source data fusion also increases memory overhead. Nevertheless, through optimized parallelization strategies and mixed-precision training, this invention can still meet the efficiency requirements of large-scale forest monitoring in practical applications. It is worth noting that in scenarios with limited hardware resources, a flexible balance between accuracy and efficiency can be achieved by adjusting the constraint weights and feature dimensions.

[0075] (7) Comprehensive performance evaluation This experiment, through systematic comparative analysis, verified the improved effect of the method of this invention in the identification of individual trees in forest plots. Compared with traditional methods, this invention achieves a stable improvement of 20-25% in core indicators such as accuracy of individual tree detection, geometric parameter estimation, and diameter at breast height (DBH) inversion, and an improvement of approximately 12% in biological rationality.

[0076] The improvements in this invention achieve a good balance between accuracy, efficiency, and interpretability: In terms of accuracy, a moderate but stable improvement was achieved through physical constraints; In terms of efficiency, although the computational overhead has increased slightly, it has been kept within an acceptable range through engineering optimization; In terms of interpretability, biological constraint mechanisms enhance the transparency and credibility of the model.

[0077] Experimental results show that this invention provides a practical and reliable technical solution for achieving precise forestry monitoring, and is particularly suitable for scientific research and operational application scenarios with high requirements for biological rationality.

[0078] The above description is merely a specific embodiment of the present invention. Any feature disclosed in this specification may be replaced by other equivalent or similar features unless otherwise specified. All disclosed features, or steps in all methods or processes, may be combined in any way except for mutually exclusive features and / or steps.

Claims

1. A high-precision single-tree segmentation method based on multi-source point cloud data, characterized in that, Includes the following steps: (1) Point cloud data acquisition: Use UAV airborne lidar and backpack lidar to collect point cloud data of complex forest land, perform point cloud data fusion, and construct a high-density point cloud dataset; (2) Data preprocessing: The collected raw point cloud data is subjected to denoising, registration, ground filtering, elevation normalization, outlier removal and voxel downsampling to obtain standardized point cloud data; (3) Semantic segmentation: Based on the Enhanced Tree Instance Net model, semantic segmentation is performed on the preprocessed standardized point cloud data to extract tree structure and category information; (4) Semantic segmentation data optimization: Optimize feature combination and segmentation parameters through feature contribution analysis and dynamic scene adaptation strategy; (5) Single tree segmentation: The semantic segmentation results are fused with geometric features, and single tree instance segmentation is achieved through instance embedding and classification branching; (6) Physical constraint neural network parameter inversion: The physical model of tree growth is introduced into the neural network loss function as a regularization constraint to realize the biological rationality inversion of tree height, diameter at breast height, crown width and tree age parameters; (7) Evaluation and calibration of single-tree segmentation accuracy: The segmentation results are quantitatively verified based on a multi-level index system and intelligent matching algorithm; (8) Visualization: Visualize the segmentation results and inversion parameters and calculate the stand parameters.

2. The method according to claim 1, characterized in that, In step (1), the UAV-borne lidar adopts a multi-echo mode with ≥4 echoes, a laser pulse frequency of 200-1000kHz, a scanning frequency of ≥400kHz, a flight altitude of 50-120 meters, a longitudinal overlap rate of ≥70%, a lateral overlap rate of ≥40%, and a point cloud density of 200-800 points / square meter; the backpack lidar has a scanning frequency of 200-600kHz, a point density of ≥1000 points / square meter, a positioning error of less than 3 cm, and adopts a combination of SLAM and IMU positioning.

3. The method according to claim 1, characterized in that, The data preprocessing described in step (2) includes: A multi-site ICP algorithm was used for spatial alignment, with registration errors controlled within 2 cm. Coordinate normalization and ground filtering were then performed, and the CSF algorithm was used to separate ground and non-ground points. A digital elevation model was generated based on the filtering results, and elevation normalization was performed to unify the point cloud heights to a relative height system. Next, a canopy surface projection was generated, and a gradient descent algorithm was used to remove false canopy regions and pseudo-high-value points caused by noise or misregistration. Then, the normalized point cloud was height-sliced ​​to extract trunk and mid-layer structure information. The SOR algorithm was applied to remove outliers. Finally, voxel downsampling was performed on the processed point cloud. The manually interpreted data was divided into training and test sets.

4. The method according to claim 1, characterized in that, The semantic segmentation in step (3) uses the Enhanced Tree Instance Net model. The standardized point cloud data obtained through the above preprocessing process is used as the input of the Enhanced Tree Instance Net model. It is loaded and converted into tensor format by the data_loaders function. The data loader automatically completes the batch division and device transmission and then sends it to the network for training or inference. The semantic segmentation training process is as follows: First, enable CUDA synchronous debugging and device error detection by setting the environment variables CUDA_LAUNCH_BLOCKING and TORCH_USE_CUDA_DSA. The preprocessed standardized point cloud data is then loaded using the data_loaders function, with conservative settings of batch_size=1 and num_workers=0. The standardized point cloud data includes XYZ coordinates, RGB features, and manually interpreted labels. In the core training phase, the model receives [B, N, 6] dimensional input data and extracts features through a four-layer local feature aggregation module. At the same time, FocalLoss, which is designed for class imbalance, is used to dynamically adjust the training weights of tree points and background points. Combined with the Adam optimizer and cosine annealing learning rate scheduler, stable convergence is achieved. After each training round, the total accuracy and average intersection-union ratio are calculated synchronously using the calculate_metrics function, and the loss curve and accuracy metrics are visualized in real time using Tensor Board; Finally, save complete checkpoints containing model parameters, optimizer status, and training progress every 20-30 rounds.

5. The method according to claim 1, characterized in that, The specific steps for semantic segmentation data optimization in step (4) are as follows: First, the screening order is quantified by feature contribution metric, and the core features are identified by SHAP value analysis, including relative elevation, vertical density gradient and semantic segmentation confidence. Then, dynamic scene adaptation is implemented, and the clustering parameters are dynamically adjusted according to the point cloud density and the tree canopy overlap rate. An efficient priority strategy or a precision priority mode is adopted.

6. The method according to claim 1, characterized in that, The single-wood splitting mentioned in step (5) includes: The data fusion pipeline optimized by semantic segmentation in step (4) is imported into the Enhanced Tree Instance Net single-tree segmentation model. The specific process is as follows: First, in the data preprocessing stage, the point-by-point category labels output by semantic segmentation, with trunk=0 and crown=1, are used as the fourth dimension feature and concatenated with the standardized three-dimensional coordinates to form a 4-channel input vector. Subsequently, during the forward propagation of the model, the hybrid features are processed by the encoder-decoder architecture of Enhanced Tree Instance Net, and point cloud features are extracted through a multilayer perceptron, enabling semantic information and geometric features to interact deeply in the embedding space, and learning the spatial semantic pattern of vertical clustering of the trunk and horizontal diffusion of the canopy. Finally, multi-scale feature fusion is achieved through instance embedding branch and classification branch. Semantic information serves as a strong guiding signal in the instance embedding branch to distinguish tree boundaries, and provides contextual constraints in the classification branch, ultimately outputting the single tree segmentation result.

7. The method according to claim 1, characterized in that, The physical constraint neural network parameter inversion in step (6) includes: By incorporating a physical model of tree growth as a regularization constraint into the neural network loss function, the allometric growth constraint represents the power-law relationship between tree height, diameter at breast height (DBH), and crown width. H=α×DBH β ×CD γ ; Where H is tree height, DBH is diameter at breast height, CD is crown width, and α, β, and γ are species-specific parameters; Growth rate constraints are based on the logistic growth model: dDBH / dt=r×DBH×(1–DBH / K); Where r is the inherent growth rate of diameter at breast height (DBH), K is the theoretical upper limit of DBH, which is determined by the tree species, and t is time; The morphological proportion constraint limits the crown width to tree height ratio to be between 0.2 and 0.8; The physical constraint loss function design incorporates physical constraints as soft regularization terms into the total loss function. L_total=L_data+λ_physics×L_physics; Where L_data is the data-driven error term; L_physics measures the difference between the network predictions and the physics model predictions. λ_physics represents the constraint weight coefficient.

8. The method according to claim 7, characterized in that, In step (6), the physical constraint neural network is implemented in the following way: A multi-level constraint verification framework is constructed to monitor the biological rationality of prediction parameters in real time during the inference stage. When the deviation of the allometric growth equation exceeds 15-25%, the parameter correction mechanism is automatically triggered. We design a progressive feature decoupling architecture that separates and learns geometric features, semantic features, and physical features in a 256-512 dimension embedding space. Establish an adaptive constraint strength adjuster to dynamically adjust the strictness of physical constraints based on point cloud density and forest stand complexity. Use lenient constraints in simple isolated tree scenarios and enable strict constraints in complex dense forest scenarios. A cross-scale feature fusion mechanism is implemented, extracting local geometric features at three scales: 0.1m, 0.5m, and 1.0m, and then weighting and fusing them using attention weights. Deploy a real-time feedback loop system to automatically add samples with prediction errors greater than a preset value to the difficult example library for focused learning in subsequent training.

9. The method according to claim 1, characterized in that, The accuracy assessment and calibration described in step (7) includes: The intelligent matching algorithm includes using KMeans clustering and adjusting the RAND index to match the predicted tree with the real tree; The aforementioned multi-level indicator system specifically involves constructing a four-dimensional evaluation framework, including indicators for point-level classification, clustering quality, instance detection, and segmentation quality.

10. The method according to claim 1, characterized in that, The visualization described in step (8) includes: The convex hull algorithm is used to draw the crown boundary of a single tree, and different colors are used to distinguish them. Automatically calculates forest stand parameters, and detects the number of trees, average tree height, crown diameter, diameter at breast height, and leaf area per tree; The leaf area index (LAI) is calculated based on the formula: LAI = total leaf area / plot area. The generated parameter distribution map shows the correlation between tree height-crown width and diameter at breast height-leaf area.

Images