Laser radar individual tree segmentation method based on multi-scale feature fusion

Abstract

The invention discloses a laser radar individual tree segmentation method based on multi-scale feature fusion, and relates to the technical field of laser radar remote sensing forestry resource investigation, and the method comprises the steps: point cloud data preprocessing and multi-scale adaptive sampling, combined denoising in combination with geometric distance and intensity information, and sampling parameter adjustment according to stand density; multi-scale feature extraction and cross-scale association are carried out, and local, morphological and global features and cross-scale association features are extracted; fusing and enhancing dynamic attention features, and constructing a feature pyramid; performing density peak clustering and initial segmentation, adaptively adjusting clustering parameters, screening core points and aggregating adjacent points; segmentation result optimization and morphological constraint are carried out, boundaries are optimized through curvature adjustment, ellipse fitting and closed operation, and abnormal units are corrected in combination with morphological verification. According to the method, the individual tree segmentation precision and form conformity are improved, and different stand densities and point cloud conditions are dynamically adapted; the method is suitable for diversified scenes from low-density pure forests to high-density mixed forests.

Description

Technical Field

[0001] This invention relates to the field of lidar remote sensing forestry resource survey technology, and in particular to a lidar single-tree segmentation method with multi-scale feature fusion. Background Technology

[0002] In forestry resource surveys, accurate acquisition of individual tree parameters is the core foundation for estimating forest stock volume, assessing biomass, and monitoring forest dynamics. Traditional individual tree surveys rely on manual field measurements, which are not only time-consuming, labor-intensive, and costly, but also limited by factors such as topography and stand density, making it difficult to achieve large-scale, rapid, and efficient surveys. With the development of remote sensing technology, lidar, with its advantages of high precision and strong penetration, has become the mainstream technology for individual tree segmentation. However, early lidar individual tree segmentation methods were mostly based on single-scale point cloud processing, focusing only on the geometric features of the canopy surface and ignoring detailed features such as the trunk and branches, as well as the global features of the spatial distribution of trees. This resulted in segmentation results being easily affected by stand structure, leading to fragmented segmentation in sparsely wooded areas and clustered individual trees in densely wooded areas, failing to meet the needs of high-precision forestry surveys.

[0003] Existing multi-scale feature fusion methods have significant drawbacks. Feature extraction lacks cross-scale correlation design; small-scale local features, medium-scale morphological features, and large-scale global features are often simply concatenated without considering the intrinsic relationships between features at different scales, resulting in insufficient feature discriminative power. Clustering algorithm parameters are mostly fixed values, failing to adaptively adjust according to point cloud density and forest stand structure. In complex terrain and mixed forest stands, the mismatch between cluster radius and density threshold exacerbates segmentation errors. Furthermore, noise removal relies solely on geometric distance judgments without incorporating lidar point cloud intensity information, leading to some tree edge points being misclassified as noise, or interference points such as ground and low vegetation not being effectively removed, further impacting segmentation accuracy.

[0004] The segmentation optimization process lacks effective morphological constraints and closed-loop adjustment mechanisms. Existing methods often employ simple boundary smoothing or hole filling without verifying against the natural morphological patterns of tree growth, resulting in the inability to identify and correct some abnormal segmentation units. Furthermore, for complex forest stand scenarios such as high-density mixed forests and dense understory vegetation, there is a lack of targeted interference suppression strategies. Low-lying understory vegetation is confused with the point clouds of target trees, and occlusion between trees leads to incomplete feature extraction, significantly reducing segmentation integrity and accuracy. This makes it difficult to adapt to diverse forestry survey scenarios and limits the practical application scope of lidar single-tree segmentation technology. Summary of the Invention

[0005] This invention proposes a multi-scale feature fusion method for single-tree segmentation in lidar systems to address the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a multi-scale feature fusion method for single-tree segmentation in lidar, comprising the following steps:

[0007] In the point cloud data preprocessing and multi-scale adaptive sampling stage, the original point cloud is jointly denoised by combining geometric distance and intensity information. Statistical filtering is used to calculate the average Euclidean distance between each point and 50 neighboring points, and noise points with a mean distance exceeding 3 times the standard deviation and an intensity value outside the range are removed. The multi-scale sampling parameters are adjusted based on the stand density, and the key structures are preserved by uniform sampling at each scale, and the scale labels and original coordinates are recorded.

[0008] In the multi-scale feature extraction and cross-scale association stage, local, intensity gradient and texture features are extracted at the small scale; morphological features are extracted at the medium scale; global and associated features are extracted at the large scale; and cross-scale association features are added by calculating the correlation and fit of features at different scales.

[0009] In the dynamic attention feature fusion and enhancement stage, a multi-scale feature pyramid is constructed, and a dynamic attention mechanism is introduced to dynamically allocate weights according to regional features. The corresponding scale feature weights are increased for densely branched areas, clear canopy areas, and sparsely distributed areas, respectively.

[0010] In the density peak clustering and initial segmentation stage, an improved density peak clustering algorithm is used to screen core points with local density ≥60 and relative distance ≥0.5m. These core points are then clustered according to feature similarity ≥0.85 and spatial distance ≤0.8m. After removing noisy clusters, effective clusters with height ≥2m and crown width ≥0.5m are retained, and their affiliation labels and cluster center coordinates are recorded.

[0011] In the segmentation result optimization and morphological constraint stage, the boundaries of individual trees are optimized by Bayesian smoothing and ellipse fitting, holes are eliminated by morphological closing operation, and overlapping areas of adjacent clusters are processed; the tree height is corrected by digital elevation model, and the complete point cloud, three-dimensional coordinates, morphological parameters and high confidence segmentation results of individual trees are output.

[0012] Furthermore, it also includes a multi-scale feature weight dynamic adjustment step, calculating the weight coefficients based on the information entropy of features at each scale. The information entropy calculation formula is as follows: (f,s)df, where H(s) is the feature information entropy at scale s, and p(f,s) is the probability density function of the feature value f at scale s; the lower the information entropy value, the higher the feature discrimination, and the larger the corresponding weight coefficient.

[0013] Furthermore, it also includes an adaptive adjustment step for the clustering radius, which dynamically corrects the clustering radius based on the local point cloud density. The correction formula is as follows: Where r is the corrected cluster radius, r0 is the initial cluster radius of 0.8m, and ρ(ξ) is the point cloud density at a distance ξ from the center point. The average point cloud density of the region; the higher the point cloud density, the smaller the cluster radius, and the lower the density, the larger the radius.

[0014] Furthermore, in the point cloud data preprocessing and multi-scale adaptive sampling steps, intensity and geometry are used to jointly denoise and optimize the noise removal effect. Points with intensity values ​​exceeding the range and abnormal geometric distances are judged as noise. The hierarchical sampling strategy is further refined. When the height of the trunk and branches is less than 1 / 3 of the total height of the tree, the small-scale sampling density is increased to 120 points per square meter. When the height of the upper and middle parts of the canopy is between 1 / 3 and 2 / 3 of the total height, the medium-scale sampling density is 60 points per square meter. When the height of the top part of the canopy is greater than 2 / 3 of the total height, the large-scale sampling density is 30 points per square meter.

[0015] Furthermore, in the multi-scale feature extraction and cross-scale association steps, local topological feature extraction is optimized. When constructing a k-neighborhood graph with k=30, an intensity similarity weight is introduced. The connection strength between two points is the weighted sum of the cosine of the angle between the normal vectors (weight 0.6) and the intensity similarity weight (weight 0.4). The intensity similarity is the reciprocal of the difference in intensity values ​​between two points. The tree canopy volume feature is calculated using an improved AlphaShapes algorithm, with the Alpha parameter adaptively adjusted according to the tree canopy size. In the terrain adaptation feature extraction, the digital elevation model is generated by inverse distance weight interpolation with a search radius of 5m. The relative height of the trees is the difference between the elevation of the lidar point and the ground elevation at the corresponding location.

[0016] Furthermore, in the dynamic attention feature fusion and enhancement step, the feature pyramid network is optimized into a bidirectional fusion structure. The bottom-level small-scale features are fused with the middle-level features after upsampling bilinear interpolation, and the middle-level features are fused with the top-level features after downsampling max pooling. During the bidirectional fusion process, skip connections are used to retain detailed information at each scale. Principal component analysis is used to optimize feature selection, and the correlation between each principal component and the individual tree morphological parameters height and crown width is calculated. The fused feature vector is mapped to a high-dimensional space through a radial basis function kernel.

[0017] Furthermore, in the density peak clustering and initial segmentation steps, the time decay factor is optimized to a dynamic decay mode, and the decay coefficient is dynamically adjusted according to the variance of intra-class features; the intra-class distance constraint is added to the clustering termination condition; a new cluster center optimization step is added, and the initial cluster centers are fine-tuned by k-means iteration 10 times to minimize the intra-class sum of squares error.

[0018] Furthermore, in the segmentation result optimization and morphological constraint steps, the single-tree morphology verification step is refined. The height-to-diameter ratio threshold is dynamically adjusted according to the tree species type. For coniferous trees such as pine and spruce, and broad-leaved trees such as poplar and oak, the crown index is introduced to assist in verification. When the maximum or minimum crown width of the crown index is greater than 2.0, it is judged as an anomaly. The secondary segmentation of the anomaly unit adopts a height-layer strategy, dividing it into layers every 1m in height, and calculating the point cloud density and feature variance of each layer. The boundary smoothing uses Gaussian filtering with a standard deviation of 0.15m to process the boundary point coordinates, while constraining the filtered boundary points to be within the range of the crown ellipse fitting.

[0019] Furthermore, it also includes segmentation accuracy evaluation and closed-loop optimization steps, with boundary matching degree added to the evaluation indicators in addition to completeness and F1 score; the F1 score calculation formula is as follows: The completeness represents the integrity value; four indicators are calculated by comparing with 100 sample trees with the manually labeled results. When the F1 score is lower than 0.82 or the boundary matching degree is lower than 0.85, the dynamic attention feature fusion step is automatically returned, and the feature weight coefficients of each scale are adjusted according to the accuracy gap. The larger the gap, the larger the weight adjustment range, with a maximum adjustment of 0.2.

[0020] Furthermore, for complex forest stand scenarios such as high-density mixed forests and areas with dense understory vegetation, the understory vegetation suppression process is strengthened in the multi-scale feature extraction and cross-scale association steps. The height quantile threshold is optimized to the 25th percentile, and points below this threshold are marked as non-target points. A dual-weight mechanism is adopted during feature fusion, while reducing the features of non-target points. During initial segmentation, the clustering priority of non-target points is set lower than that of target points. A new forest stand density adaptive adjustment module is added, which reduces the mesoscale sampling radius to 0.35m when the high-density forest stand is ≥8 trees / 100㎡.

[0021] Compared with existing technologies, the beneficial effects of this invention are:

[0022] This invention significantly improves the accuracy and adaptability of LiDAR tree segmentation through multi-dimensional technological innovation, effectively addressing the core pain points of existing methods. Multi-scale adaptive sampling dynamically adjusts sampling parameters based on stand density, combined with intensity-geometric joint denoising, preserving detailed tree features while eliminating invalid interference points, laying a high-quality data foundation for subsequent feature extraction. Multi-scale feature extraction introduces cross-scale correlation design, establishing an intrinsic connection between local details, morphological features, and global distribution, enriching feature dimensions, improving feature discriminative power, and enabling trees of different species and growth stages to form unique feature expressions.

[0023] The dynamic attention feature fusion mechanism dynamically allocates feature weights based on the scene, strengthening the feature contribution of key regions and avoiding feature redundancy caused by simple splicing, making the fused features more targeted. The adaptive clustering algorithm adjusts clustering parameters based on local point cloud density, combined with cluster center optimization and dynamic decay factors, to improve the compactness and accuracy of cluster units and reduce segmentation fragmentation and single-tree adhesion. The segmentation result optimization process incorporates tree morphological constraints, verifying the rationality of segmentation through indicators such as height-to-diameter ratio and crown index. Combined with secondary segmentation and boundary smoothing, it ensures that the segmentation boundary conforms to the natural shape of the tree and corrects abnormal segmentation units.

[0024] This invention designs a specific optimization strategy for complex forest stand scenarios, strengthening understory vegetation suppression and adaptive adjustment of stand density. This effectively reduces interference from low-lying vegetation and tree shading, improving the segmentation effect of mixed and high-density forest stands. A segmentation accuracy assessment and closed-loop optimization mechanism achieve iterative improvement in accuracy, ensuring stable performance across different scenarios. Overall, this invention achieves breakthroughs in segmentation accuracy, scenario adaptability, and computational efficiency, enabling rapid and accurate acquisition of individual tree parameters. It provides reliable technical support for forestry resource surveys, forest dynamic monitoring, and biomass estimation, demonstrating significant practicality and promotional value. Attached Figure Description

[0025] Figure 1 This is a schematic block diagram of a multi-scale feature fusion lidar single-tree segmentation method proposed in this invention;

[0026] Figure 2 A bar chart comparing the entropy of multi-scale feature information;

[0027] Figure 3 Line graph comparing F1 scores under different stand densities;

[0028] Figure 4 A line graph comparing computation time and point cloud density;

[0029] Figure 5 This is a bar chart showing the impact of forest understory vegetation coverage on segmentation accuracy. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0032] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0033] Reference Figures 1 to 5 A multi-scale feature fusion method for single-tree segmentation using lidar includes the following steps:

[0034] Point cloud data preprocessing and multi-scale adaptive sampling are performed. The original point cloud from the LiDAR is jointly denoised. Combining geometric distance and intensity information, a statistical filtering algorithm is used to calculate the average Euclidean distance between each point and its 50 neighboring points. At the same time, the point cloud intensity value is extracted and the intensity distribution range is statistically analyzed. Noise points with a mean distance exceeding 3 times the standard deviation and an intensity value exceeding the typical intensity range of trees by 50-200 are removed. Based on the stand density, the multi-scale sampling parameters are adaptively adjusted. When the number of trees per unit area is ≥5 in densely wooded areas, the sampling radius is set to 0.15m for small scale, 0.4m for medium scale, and 0.8m for large scale. When the number of trees per unit area is <2 in sparsely wooded areas, the sampling radius is adjusted to 0.25m for small scale, 0.6m for medium scale, and 1.2m for large scale. At each scale, key geometric structures are preserved through uniform sampling. The small scale focuses on detailed features such as tree trunks and branches, the medium scale captures the crown morphology and canopy structure, and the large scale reflects the spatial distribution and neighborhood relationship of trees. The scale label and original coordinates of each point are recorded simultaneously during the sampling process.

[0035] Multi-scale feature extraction and cross-scale correlation are employed. For small-scale point clouds, multi-dimensional local features are extracted, including normal vectors calculated through principal component analysis with precision retained to three decimal places, curvature represented by eigenvalue ratios, density features of the number of neighboring points per unit volume, and point cloud intensity gradients and texture features. Texture features are calculated using parameters such as energy and entropy through the gray-level co-occurrence matrix. For mesoscale point clouds, core morphological features are extracted. The convex hull algorithm is used to delineate the canopy boundary and calculate the projected area. The canopy index is determined by the maximum and minimum values ​​of the canopy diameter, and the height quantile is selected. The 50th and 90th percentiles are used to construct height distribution features, supplementing the canopy volume and surface area calculated by the AlphaShapes algorithm; global and related features are extracted for large-scale point clouds, and the number of trees per unit area is counted to represent the distribution density by dividing the grid into 5m×5m grids, calculating the Euclidean distance and elevation difference between the cluster centers of adjacent trees, and adding cross-scale related features, namely the correlation between small-scale branch density and medium-scale canopy diameter, and the fit between medium-scale canopy height and large-scale regional elevation. The correlation is calculated using the Pearson coefficient and the threshold is set to 0.7.

[0036] Dynamic attention feature fusion and enhancement are employed to construct a multi-scale feature pyramid. The bottom layer consists of small-scale local and texture features, the middle layer consists of medium-scale morphological and volumetric features, and the top layer consists of large-scale global and relational features. A dynamic attention mechanism is introduced, dynamically assigning weights based on the local feature response values ​​of each point. The response values ​​are normalized using the Sigmoid function. When the local density in dense branch areas is ≥80 points / ㎡, the weight of small-scale features is increased to 0.4; when the crown shape index in areas with clear tree outlines is ≤1.5, the weight of medium-scale features is increased to 0.4; and when the distribution density in sparsely distributed areas is ≤3 trees / 100㎡, the weight of large-scale features is increased to 0.3. Through feature concatenation and 1×1 convolution dimensionality reduction, multi-dimensional features are fused into a 32-dimensional unified feature vector. After fusion, BatchNorm normalization is applied to stabilize the feature values ​​in the [0,1] range. At the same time, a Dropout layer with a dropout probability of 0.2 is used to reduce the risk of overfitting.

[0037] Density peak clustering and initial segmentation: An improved density peak clustering algorithm is adopted based on fusion features. The core points of the clusters are automatically identified by calculating the local density and relative distance of the point cloud. The local density is the number of points in a neighborhood with a radius of 0.8m, and the relative distance is the minimum Euclidean distance between the current point and the point with higher density. A density threshold of local density ≥60 and a distance threshold of relative distance ≥0.5m are set to filter core points. With the core point as the cluster center, neighboring points are aggregated according to feature similarity cosine similarity ≥0.85 and spatial distance ≤0.8m. The minimum number of cluster points is set to 50. Noisy clusters with fewer than 30 points are removed, and cluster units that meet the single tree size characteristics of height ≥2m and crown width ≥0.5m are retained. During the clustering process, the belonging label and cluster center coordinates of each point are recorded in real time.

[0038] The segmentation results were optimized and morphological constraints were applied. The curvature change rate of the boundary points of each cluster unit was calculated. When the curvature change rate of the abrupt curvature change region was ≥0.3, the boundary was adjusted using a Bayesian smoothing algorithm. The ratio of the major axis to the minor axis of the tree crown was ≤2.5 by fitting the crown outline with an ellipse, so that the boundary of the individual tree fit the actual growth form. Morphological closing operation was used to eliminate the void area in the segmentation results ≤0.1㎡ with a structural element radius of 0.2m. For the overlapping area of ​​the overlapping region of adjacent cluster units ≥0.2㎡, the feature similarity was calculated. When the similarity was higher than 0.85, they were merged into the same individual tree. When the similarity was lower than 0.7, the boundary points were assigned according to the nearest distance principle. The tree height was corrected by digital elevation model, and the influence of ground elevation was deducted. The complete point cloud set, 3D coordinate cluster center x, y, z, morphological parameters height, crown width, volume, and segmentation confidence score of each individual tree were output. The confidence score was calculated based on the intra-class feature variance. A score ≥0.8 was considered high confidence.

[0039] This invention also includes a multi-scale feature weight dynamic adjustment step, which calculates weight coefficients based on the information entropy of features at each scale. The information entropy calculation formula is as follows: (f,s)df, where H(s) is the feature information entropy at scale s, p(f,s) is the probability density function of feature value f at scale s, and the integration range is the feature value normalization interval [0,1]. The lower the information entropy value, the higher the feature discrimination and the larger the corresponding weight coefficient. This calculation realizes the adaptive allocation of feature weights under different scenarios, improves the recognition of tree features in complex terrain, and the weight coefficient is applied to the feature fusion process after being normalized by the Softmax function, so that the feature contribution of each scale is accurately adapted to the scene.

[0040] This invention also includes a clustering radius adaptive adjustment step, which dynamically corrects the clustering radius based on the local point cloud density, using the following formula: Where r is the corrected cluster radius, r0 is the initial cluster radius of 0.8m, and ρ(ξ) is the point cloud density at a distance ξ from the center point. The average point cloud density of the region is used, and the integration range is 1 / 2 to 3 / 2 times the initial radius. The higher the point cloud density, the smaller the cluster radius, which reduces the excessive merging of dense regions. The lower the density, the larger the radius, which reduces the probability of fragmentation of sparse regions. The corrected cluster radius range is limited to 0.5-1.2m to ensure the stability and rationality of clustering.

[0041] In this invention, during the point cloud data preprocessing and multi-scale adaptive sampling steps, intensity-geometric joint denoising is used to optimize noise removal. The typical intensity range of trees is set to 50-200 based on the 1550nm LiDAR band. Points with intensity values ​​exceeding this range and abnormal geometric distances are identified as noise. The hierarchical sampling strategy is further refined: when the height of the trunk and branches is less than 1 / 3 of the total tree height, the small-scale sampling density is increased to 120 points per square meter; when the height of the upper and middle parts of the canopy is between 1 / 3 and 2 / 3 of the total height, the medium-scale sampling density is 60 points per square meter; and when the height of the top part of the canopy is greater than 2 / 3 of the total height, the large-scale sampling density is 30 points per square meter. The intensity information retained during sampling is normalized to the [0,1] interval by Min-Max and concatenated with geometric features to form a joint feature vector. This vector is used in subsequent feature extraction to enhance the ability to distinguish between coniferous trees with intensity values ​​that are 0.6-1.0 higher and broad-leaved trees with intensity values ​​that are 0.2-0.6 lower, thereby improving the segmentation accuracy of mixed forest stands.

[0042] In this invention, during the multi-scale feature extraction and cross-scale association steps, local topological feature extraction is optimized. When constructing a k-neighborhood graph with k=30, an intensity similarity weight is introduced. The connection strength between two points is the weighted sum of the cosine of the angle between the normal vectors (weight 0.6) and the intensity similarity weight (weight 0.4). The intensity similarity is the reciprocal of the difference between the intensity values ​​of the two points. Topological relationships with a connection strength greater than 0.65 are retained, enhancing the feature expression of the branch structure. The crown volume feature calculation adopts an improved AlphaShapes algorithm. The Alpha parameter is adaptively adjusted according to the crown size. When the crown width is ≤3m, Alpha = 0.3, and when the crown width is >3m, Alpha = 0.5, improving the accuracy of volume calculation. In the terrain adaptation feature extraction, the digital elevation model is generated by inverse distance weight interpolation with a search radius of 5m. The relative height of the tree is the difference between the elevation of the lidar point and the ground elevation at the corresponding location, eliminating the interference of complex terrain with a slope >25° on the calculation of tree height. The relative height error is controlled within 0.3m.

[0043] In this invention, during the dynamic attention feature fusion and enhancement step, the feature pyramid network is optimized into a bidirectional fusion structure. The bottom-level small-scale features are fused with the middle-level features after upsampling bilinear interpolation, and the middle-level features are fused with the top-level features after downsampling max pooling. During the bidirectional fusion process, skip connections are used to retain detailed information at each scale and reduce feature loss. Principal component analysis optimizes feature selection, selecting principal components with a cumulative contribution rate exceeding 96%, typically the first 8-12 dimensions. At the same time, the correlation between each principal component and the individual tree morphological parameters height and crown width is calculated, retaining principal components with an absolute correlation coefficient ≥ 0.6 to further simplify the feature dimensions. The fused feature vector is mapped to a high-dimensional space through a radial basis function kernel with a kernel parameter set to 0.8, improving the nonlinear expressive power of the features and providing more discriminative feature support for subsequent clustering. The computational efficiency of the clustering algorithm is improved by more than 30%.

[0044] In this invention, during the density peak clustering and initial segmentation steps, the time decay factor is optimized to a dynamic decay mode. The decay coefficient is dynamically adjusted based on the intra-class feature variance. When the variance is ≥0.12, the decay coefficient is set to 0.92, and when the variance is <0.12, it is set to 0.98, avoiding excessive decay that could lead to insufficient weight for core points. The clustering termination condition is supplemented with an intra-class distance constraint. When the average distance from all points within a class to the cluster center exceeds 0.6m, the current clustering is forcibly stopped and a new cluster center search is initiated. The intra-class feature variance threshold is maintained at 0.15 to ensure the compactness of the clustering units. A new cluster center optimization step is added, which performs k-means fine-tuning iterations on the initial cluster centers 10 times to minimize the intra-class sum of squares error, making the cluster centers closer to the geometric center of individual trees, reducing the probability of misclassification of edge points, and improving the purity of individual tree segmentation to over 90%.

[0045] In this invention, the segmentation result optimization and morphological constraint steps refine the individual tree morphology verification process. The height-to-diameter ratio threshold is dynamically adjusted according to the tree species. For coniferous trees such as pine and spruce, the height-to-diameter ratio threshold is set to 3.0-5.5, while for broad-leaved trees such as poplar and oak, it is set to 1.8-3.2. The crown index is introduced to assist in verification. When the maximum / minimum crown width of the crown index is greater than 2.0, it is judged as an anomaly. The secondary segmentation of the anomaly unit adopts a height-layer strategy, dividing the unit into layers every 1m in height. The point cloud density and feature variance of each layer are calculated. New boundaries are defined at the density mutation layer with a mutation rate ≥50% or at the variance peak layer. After secondary segmentation, the individual tree morphology conformity rate is improved to 95%. The boundary smoothing uses Gaussian filtering with a standard deviation of 0.15m to process the boundary point coordinates. At the same time, the boundary points after filtering are constrained to be within the range of the crown ellipse fitting. The major and minor axes of the ellipse are set based on the crown width parameters, making the segmentation boundary smooth and conforming to the natural shape of the tree.

[0046] This invention also includes segmentation accuracy evaluation and closed-loop optimization steps. The evaluation metrics, in addition to completeness, accuracy, and F1 score, include boundary matching degree. Boundary matching degree is the ratio of the overlap length between the segmentation boundary and the manually labeled boundary to the total length of the labeled boundary. The F1 score is calculated using the following formula: Here, "complete" represents the integrity value, and "precise" represents the accuracy value. Four indicators are calculated by comparing with 100 sample trees with manually labeled results. When the F1 score is lower than 0.82 or the boundary matching degree is lower than 0.85, the system automatically returns to the dynamic attention feature fusion step. The feature weight coefficients at each scale are adjusted according to the accuracy gap. The larger the gap, the larger the weight adjustment. The maximum adjustment is 0.2. The fusion, segmentation and optimization process is re-executed to achieve closed-loop iterative optimization of segmentation accuracy. Finally, the F1 score is stabilized above 0.85.

[0047] In this invention, for complex forest stand scenarios such as high-density mixed forests and areas with dense understory vegetation, the understory vegetation suppression processing is strengthened in the multi-scale feature extraction and cross-scale association steps. The height quantile threshold is optimized to the 25th percentile, and points below this threshold are marked as non-target points. Non-target points include low vegetation such as herbs and shrubs. A dual-weight mechanism is adopted during feature fusion, with the target point weight coefficient set to 1.0 and the non-target point weight coefficient set to 0.25. At the same time, the feature response value of non-target points is reduced by multiplying by an inhibition factor of 0.3. Non-target points are set during initial segmentation. Clustering priority is lower than target points. During the clustering process, point clouds with a height ≥2m and intensity values ​​within the typical range of trees are prioritized for aggregation. For non-target points, a stricter clustering condition of similarity ≥0.9 is applied. A new stand density adaptive adjustment module is added. When there are ≥8 trees / 100㎡ in high-density stands, the mesoscale sampling radius is reduced to 0.35m and the weight of cross-scale associated features is increased to 0.35 to reduce feature confusion caused by tree occlusion. The completeness and accuracy of dense forest area segmentation are both improved to over 88%, effectively solving the problems of single tree adhesion and segmentation fragmentation in complex scenarios.

[0048] The following two examples further illustrate the specific implementation of this system:

[0049] Example 1: Application of single-tree segmentation in medium- and low-density coniferous forests (pure pine forest scenario)

[0050] This embodiment focuses on single-tree segmentation in a medium-to-low density pine pure forest. The stand density is 3 trees / 100㎡, the terrain is gentle with a slope of <15°, the average tree height is 12m, the average crown width is 3.5m, the lidar point cloud density is 80 points / ㎡, and the wavelength is 1550nm. The main purpose is to verify the segmentation accuracy and efficiency of the method in a conventional forestry scenario.

[0051] In the point cloud data preprocessing and multi-scale adaptive sampling stage, the original LiDAR point cloud is first subjected to intensity-geometric joint denoising. A statistical filtering algorithm is used to search 50 neighboring points for each point and calculate the average Euclidean distance. Assuming the average distance of a point's neighboring points is 0.3m, the standard deviation is 0.1m, and three times the standard deviation is 0.3m, the average distance of this point is 0.7m, exceeding the threshold. Simultaneously, its intensity value of 30 is lower than the typical intensity range of 50-200 for trees, thus it is identified as a noise point and removed. A stand density of 3 trees / 100㎡ is considered a sparse area; therefore, the sampling radius is set to 0.25m for small-scale, 0.6m for medium-scale, and 1.2m for large-scale. During stratified sampling, the trunk and branch areas are less than 4m high, with a small-scale sampling density of 120 points per square meter; the upper and middle canopy areas are 4-8m high, with a medium-scale sampling density of 60 points per square meter; and the top of the canopy is higher than 8m high, with a large-scale sampling density of 30 points per square meter. During the sampling process, the scale label and original coordinates of each point are retained. The intensity information is normalized to the [0,1] interval by Min-Max and then concatenated with the geometric features to form a joint feature vector.

[0052] In the multi-scale feature extraction and cross-scale association stages, multi-dimensional local features are extracted from the small-scale point cloud. Normal vectors are calculated using principal component analysis. A covariance matrix is ​​constructed for each point's 30 neighboring points, eigenvalues ​​and eigenvectors are solved, and the eigenvector corresponding to the smallest eigenvalue is taken as the normal vector, with precision retained to three decimal places. Curvature is the ratio of the smallest eigenvalue to the sum of the three eigenvalues. For a point with eigenvalues ​​of 0.002, 0.01, and 0.08, the curvature is calculated as 0.002 / (0.002+0.01+0.08)=0.02. Texture features are calculated using a 3×3 window gray-level co-occurrence matrix with a stride of 1, generating parameters such as an energy value of 0.8 and an entropy value of 0.3. The convex hull algorithm is used to delineate the tree canopy boundary in the mesoscale point cloud, calculating a projected area of ​​8.2㎡, a maximum canopy width of 4.2m, a minimum of 3.1m, and a canopy index of 4.2 / 3.1=1.35. The height quantiles were selected at the 5th, 50th, and 90th percentiles, which are 1.5m, 11.8m, and 13.2m, respectively. The canopy volume was calculated using a modified AlphaShapes algorithm; a canopy width of 3.5m corresponds to an Alpha of 0.3, resulting in a volume of 12.5m³. 3 Large-scale point clouds were divided into 5m×5m grids, with 3 trees per unit area, resulting in a distribution density of 3 trees / 100㎡. The Pearson coefficient was calculated for cross-scale correlation features, relating small-scale branch density to medium-scale crown diameter. With a branch density of 90 points / ㎡ and a crown diameter of 3.65m, a covariance of 1.2, and a standard deviation product of 1.71, the Pearson coefficient was 1.2 / 1.71 = 0.7, meeting the threshold requirement.

[0053] In the dynamic attention feature fusion and enhancement stage, a multi-scale feature pyramid is constructed. The bottom layer contains small-scale local and texture features, the middle layer contains medium-scale morphological and volumetric features, and the top layer contains large-scale global and relational features. Weights are assigned based on local feature response values: in densely branched areas with a local density of 90 points / m², the small-scale feature weight is set to 0.4; in areas with clear canopy outlines and a crown shape index of 1.35 ≤ 1.5, the medium-scale feature weight is set to 0.4; and in sparsely distributed areas with a distribution density of 3 trees / 100m², the large-scale feature weight is set to 0.3. Through feature concatenation and 1×1 convolution dimensionality reduction, a 32-dimensional feature vector is fused. After BatchNorm normalization, the feature values ​​are stabilized in the [0,1] interval, and the Dropout layer has a dropout probability of 0.2. The multi-scale feature weights are dynamically adjusted using the information entropy formula. The probability density function p(f,1) of the small-scale feature follows a normal distribution N(0.5,0.2), and the integral yields H(1) = 0.65; the medium-scale feature p(f,2) has N(0.4,0.15), and H(2) = 0.52; the large-scale feature p(f,3) has N(0.6,0.25), and H(3) = 0.71. The medium-scale feature has the lowest information entropy and the largest weight coefficient. After Softmax normalization, the weights are 0.3, 0.45, and 0.25, respectively, and are applied to feature fusion.

[0054] In the density peak clustering and initial segmentation stage, local density is defined as the number of points within a neighborhood with a radius of 0.8m, with a density threshold of 60 and a relative distance threshold of 0.5m. A point is identified as a core point if its neighborhood contains 75 points or more (≥60) and its minimum Euclidean distance to points with higher density is 0.6m or more (≥0.5m). The clustering radius is adaptively adjusted using a formula. r0 = 0.8m, the average point cloud density of the region ρ = 80 points / m², and the density ρ(ξ) at a distance ξ from the center point averages 90 points / m² in the range of 0.4-1.2m. Integral calculation yields... r = 0.8 × exp(0.9) = 1.94m, which exceeds the range of 0.5-1.2m, so r = 1.2m is finally chosen. Using the core point as the center, neighboring points are clustered based on cosine similarity (0.88) and spatial distance ≤ 1.2m, with a minimum cluster size of 50. Noisy clusters with 25 points are removed, retaining clusters with a height ≥ 2m and a crown width ≥ 0.5m. A dynamic time decay factor is used, with intra-cluster feature variance of 0.11 < 0.12, and a decay coefficient set to 0.98. The cluster centers are fine-tuned using k-means iteration 10 times to minimize the intra-cluster sum of squares error.

[0055] In the segmentation result optimization and morphological constraint stage, the curvature change rate of the boundary points of the cluster units was calculated. The curvatures of adjacent points of a certain boundary point were 0.03, 0.08, and 0.04, respectively, with a change rate of (0.08-0.03) / 0.03 = 1.67 ≥ 0.3. A Bayesian smoothing algorithm was used to adjust the boundary. The tree crown outline was fitted by ellipse, with a major axis of 4.2m, a minor axis of 3.1m, and a major-to-minor axis ratio of 1.35 ≤ 2.5. The morphological closing operation structuring element radius was 0.2m to eliminate holes with an area of ​​0.08㎡. The overlapping area of ​​adjacent cluster units was 0.25㎡, and the feature similarity was 0.87 > 0.85, so they were merged into the same tree. In the individual tree morphological verification, the height-to-diameter ratio of the pine tree was 13m / 3.65m = 3.56, which is in the range of 3.0-5.5, and the crown shape index was 1.35 ≤ 2.0, which was judged as normal. The digital elevation model is generated by inverse distance weighted interpolation with a search radius of 5m. After deducting the ground elevation, the relative height error of the trees is 0.2m.

[0056] In the segmentation accuracy evaluation and closed-loop optimization stage, 100 sample trees were selected for comparison with manually labeled trees. The completeness was 92 / 100 = 0.92, and the accuracy was 92 / 95 = 0.97. The F1 score was calculated using the formula... The calculated value is 2×(0.92×0.97) / (0.92+0.97)=0.94, and the boundary matching degree is 0.89. All indicators meet the requirements, and no iterative optimization is needed.

[0057] Table 1 Comparison of single-tree splitting performance in low-density pine pure stands

[0058] index Existing methods Method of the present invention Integrity 0.83 0.92 Accuracy 0.85 0.97 F1 score 0.84 0.94 Boundary matching degree 0.76 0.89 Calculation time (minutes / hectare) 18 10

[0059] Table 1 shows that the present invention has significant advantages in low- to medium-density pure forest scenarios. Existing methods have both completeness and accuracy below 0.85, due to the lack of consideration for cross-scale feature association and the resulting fragmentation and consolidation caused by fixed clustering parameters. The present invention retains key features through multi-scale adaptive sampling, improves discriminative power through cross-scale feature association, and enhances the contribution of effective features through dynamic attention fusion, resulting in a 9 percentage point improvement in completeness and a 12 percentage point improvement in accuracy. Boundary matching accuracy is improved by 13 percentage points, thanks to the morphological constraints of Bayesian smoothing and ellipse fitting, making the boundaries more closely resemble the natural form of trees. The computation time is reduced by 44%, attributed to feature dimensionality reduction and clustering algorithm optimization, improving efficiency while maintaining accuracy. Overall, the present invention achieves a dual improvement in accuracy and efficiency in conventional forestry scenarios, providing reliable support for the rapid acquisition of individual tree parameters.

[0060] Example 2: Application of single-tree segmentation in high-density mixed forests (pine-oak mixed forest scenario)

[0061] This embodiment focuses on individual tree segmentation in a high-density pine-oak mixed forest with a stand density of 9 trees / 100㎡, a complex terrain with a slope of 30°, and includes two tree species: pine and oak. The understory vegetation is mainly shrubs. The lidar point cloud density is 120 points / ㎡, with a wavelength of 1550nm. This verifies the adaptability and anti-interference ability of the method in complex scenarios.

[0062] In the point cloud data preprocessing and multi-scale adaptive sampling stages, intensity-geometric joint denoising was employed. Each point was searched for 50 neighboring points, with an average Euclidean distance of 0.2m, a standard deviation of 0.08m, and a three-fold standard deviation of 0.24m. Points with a mean distance of 0.5m and an intensity value exceeding 220 were considered noise and removed. A stand density of 9 trees / 100㎡ was classified as high-density, with sampling radii of 0.15m at small scale, 0.4m at medium scale, and 0.8m at large scale. During stratified sampling, the trunk and branch areas were sampled at 120 points per square meter at small scale (height below 5m); the upper and middle canopy areas were sampled at 60 points per square meter at medium scale (height 5-9m); and the top of the canopy was sampled at 30 points per square meter at large scale (height above 9m). After intensity information was normalized, it was concatenated with geometric features, with intensity values ​​of 0.7-1.0 for pine trees and 0.3-0.6 for oak trees, enhancing species differentiation.

[0063] In the multi-scale feature extraction and cross-scale association stages, a neighborhood graph with k=30 is constructed at the small scale. The cosine of the angle between the normal vectors of two points is 0.8, the strength difference is 0.1, the strength similarity is 1 / 0.1=10, and the connection strength is 0.8×0.6+10×0.4=4.48>0.65, preserving topological relationships. The mesoscale convex hull algorithm calculates the canopy projection area: 7.8㎡ for pine and 9.5㎡ for oak. The crown shape index is 4.0 / 3.2=1.25 for pine and 5.0 / 3.8=1.32 for oak. The height quantiles are 1.2m, 11.5m, and 14.0m for pine and 1.5m, 13.0m, and 15.5m for oak. In the canopy volume calculation, a pine canopy width of 3.6m corresponds to Alpha=0.3 and a volume of 11.2m³. 3 An oak tree with a crown width of 4.4m corresponds to an Alpha of 0.5 and a volume of 15.8m³. 3 A large-scale 5m×5m grid was used to count 9 trees per unit area, with an Euclidean distance of 8.5m between adjacent cluster centers and an elevation difference of 1.2m. Cross-scale association feature calculations showed that the Pearson coefficient for pine trees (small-scale branch density 100 points / m²) and meso-scale crown diameter 3.6m was 0.72, and for oak trees, it was 0.75, both meeting the threshold requirements. In the understory vegetation suppression treatment, the 25th percentile of height was 2.0m; points below this threshold were marked as non-target points.

[0064] In the dynamic attention feature fusion and enhancement stage, the feature pyramid adopts a two-way fusion structure. The bottom-level small-scale features are upsampled using bilinear interpolation and then fused with the middle-level features. The middle-level features are downsampled using max pooling and then fused with the top-level features. Principal component analysis selects the top 10 principal components with a cumulative contribution rate of 96%, whose absolute values ​​of correlation coefficients with individual tree height and crown width are both ≥0.62. The fused 32-dimensional feature vector is mapped using a radial basis function kernel with a kernel parameter of 0.8 to enhance nonlinear expressive power. In feature weight allocation, the local density in densely branched areas is 110 points / m², with a small-scale weight of 0.4; the crown shape index in areas with clear crown outlines is 1.25≤1.5, with a medium-scale weight of 0.4; and the distribution density in densely distributed areas is 9 trees / 100m², with a large-scale weight of 0.3. In the dynamic adjustment of multi-scale feature weights, the weights are H(1) = 0.68 for small scale, H(2) = 0.51 for medium scale, and H(3) = 0.65 for large scale. After Softmax normalization, the weights are 0.31, 0.46, and 0.23, respectively. The weight coefficient for non-target points is set to 0.25, and the feature response value is multiplied by the suppression factor of 0.3.

[0065] In the density peak clustering and initial segmentation stage, local density was statistically analyzed with a radius of 0.8m. The number of neighboring points for the pine core point was 78≥60, and for the oak it was 72≥60. The relative distances between the pine (0.55m) and oak (0.6m) were both ≥0.5m. During the adaptive adjustment of the clustering radius, r0 = 0.8m, and the average point cloud density of the region was... The average value of ρ(ξ) is 135 points / m², and the integral... r = 0.8 × exp(0.9) = 1.94m, taking the upper limit of 1.2m. Clustering priority is set higher for target points than non-target points, with a cluster similarity threshold of 0.9 for non-target points. In the time decay factor, the intra-cluster feature variance is 0.13 ≥ 0.12 for pine trees, with a decay coefficient of 0.92; and 0.11 < 0.12 for oak trees, with a decay coefficient of 0.98. The average distance from all points within a cluster to the cluster center is ≤ 0.6m for both pine (0.55m) and oak (0.58m), indicating good cluster compactness. After 10 iterations of k-means fine-tuning of the cluster centers, the misclassification probability of edge points is reduced.

[0066] In the segmentation result optimization and morphological constraint stage, the curvature change rate of the boundary points was 0.4 ≥ 0.3 for a certain point on the oak tree. Gaussian filtering with a standard deviation of 0.15m was used to process the boundary point coordinates. Ellipse fitting was performed on the pine tree with a major axis of 4.0m and a minor axis of 3.2m, and on the oak tree with a major axis of 5.0m and a minor axis of 3.8m, both meeting the major-to-minor axis ratio requirements. Morphological closing operations eliminated holes with an area of ​​0.09㎡. The overlap area of ​​adjacent cluster units was 0.3㎡, and the feature similarity was 0.83 > 0.8, so they were merged. In the individual tree morphology verification, the height-to-diameter ratio of the pine tree was 14.0m / 3.6m = 3.89, and that of the oak tree was 15.5m / 4.4m = 3.52, both falling within their respective threshold ranges. Abnormal units were further segmented into layers every 1m in height. A certain unit had a density mutation rate of 55%, and a new boundary was defined in this layer, increasing the morphological conformity rate to 95%.

[0067] In the segmentation accuracy evaluation and closed-loop optimization phase, 100 manually labeled sample trees were compared. The completeness was 88 / 100 = 0.88, the accuracy was 88 / 92 = 0.96, the F1 score was 2×(0.88×0.96) / (0.88+0.96) = 0.92, and the boundary matching degree was 0.86. All indicators met the standards, and no iteration was required.

[0068] Table 2 Comparison of individual tree segmentation performance in high-density pine-oak mixed forests

[0069] index Existing methods Method of the present invention Integrity 0.72 0.88 Accuracy 0.75 0.96 F1 score 0.73 0.92 Boundary matching degree 0.68 0.86 Calculation time (minutes / hectare) 25 14

[0070] Table 2 shows that the data demonstrates the superior performance of this invention in complex scenarios of high-density mixed forests. Existing methods have completeness and accuracy below 0.75, due to incomplete feature extraction caused by understory vegetation interference and tree occlusion, and fixed clustering parameters that cannot adapt to high-density scenarios. This invention reduces interference from non-target points through understory vegetation suppression, and adaptively adjusts sampling and feature weights based on stand density, improving completeness by 16 percentage points and accuracy by 21 percentage points. Boundary matching accuracy is improved by 18 percentage points, thanks to morphological constraints and secondary segmentation, correcting abnormal boundaries. Computation time is reduced by 44%, with feature dimensionality reduction and clustering optimization improving computational efficiency. This invention effectively solves the segmentation problem of high-density mixed forests, exhibits strong anti-interference capabilities and scenario adaptability, and can provide accurate technical support for single-tree surveys in complex forestry scenarios, demonstrating significant promotional value.

[0071] Reference Figure 2This figure visually demonstrates the high discriminative advantage of the multi-scale features in this invention. Lower information entropy equates to higher feature discriminative power, which is more beneficial for subsequent clustering and segmentation. Existing methods generally have high information entropy for features at all scales, especially since cross-scale association features are not effectively constructed, with an information entropy reaching 0.91, resulting in weak feature discriminative ability. This invention, through cross-scale association design and dynamic attention fusion, achieves information entropies as low as 0.52 and 0.58 for mesoscale morphological features and cross-scale association features, respectively, significantly outperforming existing methods. This result stems from multi-scale feature extraction, cross-scale association, and dynamic feature weight adjustment techniques. By exploring the intrinsic connections between features at different scales, it strengthens the expression of effective features, providing a solid foundation for accurate clustering and directly improving the core performance of single-tree segmentation.

[0072] Reference Figure 3 This figure highlights the strong scene adaptability of this invention. Increased forest stand density leads to increased tree adhesion and shading, making segmentation more difficult. Existing methods show a significant and continuous decrease in F1 score with increasing density, reaching only 0.62 at high densities, which fails to meet accuracy requirements. In contrast, the F1 score of this invention remains consistently above 0.88, with a decrease of less than 8%. This is due to multi-scale adaptive sampling and cluster radius adaptive adjustment techniques. As forest stand density changes, the sampling radius and clustering parameters are dynamically adapted synchronously, avoiding segmentation errors caused by fixed parameters. Cross-scale correlation features effectively capture the spatial distribution relationships of trees, reducing adhesion problems at high densities. This allows the method to stably adapt to diverse forest stand scenarios from low to high densities, broadening its application scope.

[0073] Reference Figure 4 This figure demonstrates the superior computational efficiency of this invention. Increased point cloud density leads to a surge in data volume and prolonged computation time. Existing methods exhibit exponentially increasing computation time with density, requiring 58 minutes per hectare at high densities, resulting in low efficiency. This invention, however, shows a more gradual increase in computation time, requiring only 22 minutes per hectare at high densities, demonstrating a significant efficiency improvement. The core reason lies in the optimized feature dimensionality reduction and clustering algorithms: principal component analysis is used to screen key features, eliminating redundant dimensions and reducing computational load; density peak clustering combined with k-means fine-tuning optimizes the clustering process and reduces the number of iterations; Dropout layers and BatchNorm normalization improve feature processing speed. This allows the method to maintain high accuracy while achieving significantly better computational efficiency than existing methods, meeting the rapid needs of large-scale forestry surveys.

[0074] Reference Figure 5This figure demonstrates the powerful anti-interference capability of this invention. Understory vegetation can be confused with the target tree point cloud, leading to decreased segmentation accuracy; the higher the coverage, the more severe the interference. Existing methods achieve an F1 score of only 0.51 at 50% coverage, failing to effectively distinguish between the target and interference. This invention maintains an F1 score above 0.85 even at 50% coverage, demonstrating significant anti-interference performance. This is attributed to the understory vegetation suppression processing technology: filtering non-target points by height quantiles, reducing their weight and feature response values, and prioritizing the aggregation of target point clouds during clustering; combining cross-scale correlation features to distinguish target trees from low-lying vegetation, reducing the impact of interference points on the segmentation results. This allows the method to maintain high accuracy even in complex scenes with dense understory vegetation, further expanding its applicability.

[0075] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-scale feature fusion method for single-tree segmentation using lidar, characterized in that, Includes the following steps: In the point cloud data preprocessing and multi-scale adaptive sampling stage, the original point cloud is jointly denoised by combining geometric distance and intensity information. Statistical filtering is used to calculate the average Euclidean distance between each point and 50 neighboring points, and noise points with a mean distance exceeding 3 times the standard deviation and an intensity value outside the range are removed. The multi-scale sampling parameters are adjusted based on the stand density, and the key structures are preserved by uniform sampling at each scale, and the scale labels and original coordinates are recorded. In the multi-scale feature extraction and cross-scale correlation stage, local, intensity gradient and texture features are extracted at the small scale. Mesoscale extraction of morphological features; Large-scale extraction of global and related features; addition of cross-scale related features, by calculating the correlation and fit of features at different scales; In the dynamic attention feature fusion and enhancement stage, a multi-scale feature pyramid is constructed, and a dynamic attention mechanism is introduced to dynamically allocate weights according to regional features. The corresponding scale feature weights are increased for densely branched areas, clear canopy areas, and sparsely distributed areas, respectively. In the density peak clustering and initial segmentation stage, an improved density peak clustering algorithm is used to screen core points with local density ≥60 and relative distance ≥0.5m. These core points are then clustered according to feature similarity ≥0.85 and spatial distance ≤0.8m. After removing noisy clusters, effective clusters with height ≥2m and crown width ≥0.5m are retained, and their affiliation labels and cluster center coordinates are recorded. In the segmentation result optimization and morphological constraint stage, the boundaries of individual trees are optimized by Bayesian smoothing and ellipse fitting, holes are eliminated by morphological closing operation, and overlapping areas of adjacent clusters are processed; the tree height is corrected by digital elevation model, and the complete point cloud, three-dimensional coordinates, morphological parameters and high confidence segmentation results of individual trees are output.

2. The multi-scale feature fusion method for single-tree segmentation in lidar according to claim 1, characterized in that, It also includes a multi-scale feature weight dynamic adjustment step, which calculates the weight coefficients based on the information entropy of features at each scale. The formula for calculating information entropy is: Where H(s) is the feature information entropy at scale s, and p(f,s) is the probability density function of feature value f at scale s; the lower the information entropy value, the higher the feature discrimination, and the larger the corresponding weight coefficient.

3. The multi-scale feature fusion method for single-tree segmentation in lidar according to claim 1, characterized in that, It also includes an adaptive adjustment step for the cluster radius, which dynamically corrects the cluster radius based on the local point cloud density. The correction formula is as follows: Where r is the corrected cluster radius, r0 is the initial cluster radius of 0.8m, and ρ(ξ) is the point cloud density at a distance ξ from the center point. The average point cloud density of the region; the higher the point cloud density, the smaller the cluster radius, and the lower the density, the larger the radius.

4. The multi-scale feature fusion method for single-tree segmentation in lidar according to claim 1, characterized in that, In the point cloud data preprocessing and multi-scale adaptive sampling steps, intensity and geometry are used to optimize the noise removal effect. Points with intensity values ​​exceeding the range and abnormal geometric distances are identified as noise. The hierarchical sampling strategy is further refined. When the height of the trunk and branches is less than 1 / 3 of the total height of the tree, the small-scale sampling density is increased to 120 points per square meter. When the height of the upper and middle parts of the canopy is between 1 / 3 and 2 / 3 of the total height, the medium-scale sampling density is 60 points per square meter. When the height of the top part of the canopy is greater than 2 / 3 of the total height, the large-scale sampling density is 30 points per square meter.

5. The multi-scale feature fusion method for single-tree segmentation in lidar according to claim 1, characterized in that, In the multi-scale feature extraction and cross-scale association steps, local topological feature extraction is optimized. When constructing a k-neighborhood graph with k=30, an intensity similarity weight is introduced. The connection strength between two points is the weighted sum of the cosine of the angle between the normal vectors (weight 0.6) and the intensity similarity weight (weight 0.4). The intensity similarity is the reciprocal of the difference between the intensity values ​​of the two points. The tree canopy volume feature is calculated using an improved AlphaShapes algorithm, with the Alpha parameter adaptively adjusted according to the tree canopy size. In the terrain-adaptive feature extraction, the digital elevation model is generated by inverse distance weight interpolation with a search radius of 5m. The relative height of the trees is the difference between the elevation of the lidar point and the ground elevation at the corresponding location.

6. The multi-scale feature fusion lidar single-tree segmentation method according to claim 1, characterized in that, In the dynamic attention feature fusion and enhancement step, the feature pyramid network is optimized into a bidirectional fusion structure. The bottom small-scale features are fused with the middle-layer features after upsampling bilinear interpolation, and the middle-layer features are fused with the top-layer features after downsampling max pooling. During the bidirectional fusion process, skip connections are used to retain detailed information at each scale. Principal component analysis is used to optimize feature selection, and the correlation between each principal component and the individual tree morphological parameters height and crown width is calculated. The fused feature vector is mapped to a high-dimensional space through a radial basis function kernel.

7. The multi-scale feature fusion method for single-tree segmentation in lidar according to claim 1, characterized in that, In the density peak clustering and initial segmentation steps, the time decay factor is optimized to a dynamic decay mode, and the decay coefficient is dynamically adjusted according to the variance of intra-class features. The clustering termination condition is supplemented with intra-cluster distance constraints; a new cluster center optimization step is added, which performs k-means fine-tuning iterations on the initial cluster centers 10 times to minimize the intra-cluster sum of squares error.

8. The multi-scale feature fusion method for single-tree segmentation in lidar according to claim 1, characterized in that, In the segmentation result optimization and morphological constraint steps, the single-tree morphology verification step is refined. The height-to-diameter ratio threshold is dynamically adjusted according to the tree species type. For coniferous trees such as pine and spruce, and broad-leaved trees such as poplar and oak, the crown index is introduced to assist in verification. When the maximum or minimum crown width of the crown index is greater than 2.0, it is judged as an anomaly. The secondary segmentation of the anomaly unit adopts a height-layer strategy, dividing it into layers every 1m in height, and calculating the point cloud density and feature variance of each layer. The boundary smoothing uses Gaussian filtering with a standard deviation of 0.15m to process the boundary point coordinates, while constraining the filtered boundary points to be within the range of the crown ellipse fitting.

9. The multi-scale feature fusion method for single-tree segmentation in lidar according to claim 1, characterized in that, It also includes segmentation accuracy evaluation and closed-loop optimization steps. The evaluation metrics, in addition to completeness and F1 score, include boundary matching degree. The F1 score calculation formula is... The completeness represents the integrity value; four indicators are calculated by comparing with 100 sample trees with the manually labeled results. When the F1 score is lower than 0.82 or the boundary matching degree is lower than 0.85, the dynamic attention feature fusion step is automatically returned, and the feature weight coefficients of each scale are adjusted according to the accuracy gap. The larger the gap, the larger the weight adjustment range, with a maximum adjustment of 0.

2.

10. The multi-scale feature fusion method for single-tree segmentation in lidar according to claim 1, characterized in that, For complex forest stand scenarios such as high-density mixed forests and areas with dense understory vegetation, the understory vegetation suppression process is strengthened in the multi-scale feature extraction and cross-scale association steps. The height quantile threshold is optimized to the 25th percentile, and points below this threshold are marked as non-target points. A dual-weight mechanism is adopted during feature fusion, while reducing the features of non-target points. During initial segmentation, the clustering priority of non-target points is set to be lower than that of target points. A new forest stand density adaptive adjustment module is added, which reduces the mesoscale sampling radius to 0.35m when the high-density forest stand is ≥8 trees / 100㎡.

Images