Extreme offset deep learning-based individual tree segmentation feature extraction optimization method

Abstract

The invention discloses a single-tree segmentation feature extraction optimization method based on extreme value migration deep learning, relates to the related field of single-tree segmentation technologies, and provides a density adaptive down-sampling model combining adaptive point cloud density sampling, a dynamic multi-scale feature fusion module and a single-tree point cloud association enhancement module. The method comprises the following steps: firstly, introducing a self-adaptive point cloud density sampling method to generate point neighborhood local density aiming at the problem of omission of individual tree key feature points, and improving sparse point cloud sampling probability through a self-adaptive point cloud scoring strategy; secondly, aiming at insufficient multi-scale feature extraction, a dynamic multi-scale feature fusion module is introduced, different-scale features are coded, and dynamic fusion of the multi-scale features is realized by applying attention vectors; and finally, aiming at the problem of single tree edge information missing, introducing a sample association enhancement module, interacting with an external learnable memory unit, calculating an attention matrix, weighting and updating features, capturing global context information, and establishing correlation between single trees.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of single tree segmentation technology, and particularly relates to a single tree segmentation feature extraction optimization method based on extreme value offset deep learning. BACKGROUND

[0002] Accurate acquisition of single tree information plays a crucial role in promoting sustainable development of forestry. The emergence of airborne laser radar brings new opportunities and technical support for forest resource investigation and single tree segmentation. This method can quickly obtain three-dimensional spatial information of large-area forest, and can accurately identify the position, height, crown width and other key parameters of single tree by combining single tree segmentation algorithm, thereby solving the inherent limitations of field investigation.

[0003] Existing single tree segmentation algorithms based on deep learning have high abstraction ability and strong learning representation ability, and have stronger adaptability and generalization. The subjective limitations of manual feature extraction are overcome. The single tree segmentation method based on extreme value offset deep learning reduces the dependence of segmentation accuracy on the sparseness of trunk points and the non-homogeneity of point cloud density. However, the deep network used still has the following problems when extracting features from forest point cloud:

[0004] (1) Point cloud sampling strategy limitation: The density distribution of forest point cloud is uneven. The farthest point sampling (FPS) only considers the geometric distance between points, without fully considering the distribution characteristics of forest point cloud, which leads to over-sampling in high-density areas such as tree trunks and under-sampling in low-density areas such as branches, affecting feature extraction and segmentation performance.

[0005] (2) Insufficient feature fusion capability: Forest point cloud has significant multi-scale features. It reflects branch shape information at branch scale and provides single tree position and contour information at trunk scale. Different scale features are crucial for accurate identification of single trees. However, the original network fails to effectively fuse different scale feature information, affecting single tree segmentation accuracy.

[0006] (3) Single tree boundary information loss: There is a complex symbiotic relationship between adjacent single trees in the forest, which leads to blurred single tree boundary information. The original deep network ignores the spatial and semantic relationship between single trees, affecting the effective differentiation of single tree boundaries. SUMMARY

[0007] In view of the problems existing in the prior art, the present application provides a single tree segmentation feature extraction optimization method based on extreme value offset deep learning. In the process of using a deep network to extract features from forest point cloud, an adaptive point cloud density sampling method, a dynamic multi-scale feature fusion module and a single tree point cloud association enhancement module are introduced to establish an adaptive down-sampling model, enhance the network's understanding of single tree features and optimize the feature extraction process.

[0008] The application provides an optimization method for single tree segmentation feature extraction based on extreme value offset deep learning, comprising:

[0009] (1) An adaptive downsampling model is established in the extreme value offset stage of single tree segmentation based on extreme value offset deep learning, and an adaptive point cloud density sampling method is used to screen point clouds representative of the global structure from forest point clouds;

[0010] (2) For the sampling results of adaptive point cloud density sampling, a multi-scale dynamic feature fusion module is introduced to embed the adaptive downsampling model, point cloud features are extracted in the multi-scale neighborhood, and the multi-scale information of the point cloud is dynamically fused to enhance the feature representation capability of the model;

[0011] (3) After the multi-scale dynamic feature fusion module, a sample correlation enhancement module is applied to learn the typical features of single tree structure, model the global correlation between single tree samples, and capture the single tree boundary features.

[0012] Further, the specific process of single tree segmentation based on extreme value offset deep learning includes data preprocessing, extreme value offset, mean value offset, spatial mapping and point set aggregation.

[0013] Data preprocessing includes point cloud filtering, elevation normalization, segmentation subgraph, point cloud denoising, downsampling and coordinate normalization, which realizes data standardization and reduces computational complexity;

[0014] Extreme value offset uses a deep network to perform spatial transformation on point clouds to generate tree top offset point clouds, enhancing the discrimination between different single trees; mean value offset uses an adaptive mean value offset algorithm to cluster and label the offset point clouds, achieving preliminary single tree segmentation;

[0015] Spatial mapping includes mapping the clustered offset point clouds back to the original point cloud space to restore the preliminary segmentation results; point set aggregation aggregates adjacent crown points to reduce over-segmentation errors and obtain the final segmentation results.

[0016] Further, the extreme value offset realizes spatial transformation of forest point clouds through a deep network, offsets single tree point clouds to corresponding tree tops to enhance the discrimination between single trees, and the offset accuracy of the offset network depends on the feature extraction capability of the deep network. The extreme value offset stage uses a Point Transformer network for feature extraction, including a downsampling module, an upsampling module and a Transformer module:

[0017] First, the point cloud is preliminarily extracted by a multi-layer perceptron (MLP); then, a multi-level downsampling module is used to gradually reduce the number of points and improve the feature dimension, while a Transformer module is used to enhance feature representation; finally, a multi-level upsampling module is used to gradually restore the spatial resolution of the point cloud, completing the offset.

[0018] The role of the downsampling module is to extract high-dimensional features: first, sample the input point cloud to obtain a sampled point cloud; then, find the k nearest neighbors around each sampling point as the center point; finally, each center point and its k nearest neighbors form a whole, which is input into the deep network for high-dimensional feature extraction.

[0019] The role of the upsampling module is to assign high-dimensional features to non-sampling points, and its main steps are as follows: first, the point cloud composed of sampling points is subjected to feature extraction through MLP, batch normalization and ReLU activation function; then, the inverse distance weight interpolation algorithm is introduced to assign the features of each center point to its neighboring points; finally, the features corresponding to the downsampling block are aggregated to the corresponding upsampling block through the jump connection.

[0020] The downsampling stage directly determines the quality of point cloud feature extraction and feature representation ability, and has a significant impact on the final segmentation accuracy. The core task of the upsampling stage mainly depends on the high-level semantic features obtained in the downsampling stage. Therefore, by optimizing the downsampling stage, the overall network performance can be effectively improved.

[0021] The present application discloses the following technical effects:

[0022] The present application proposes an optimization method for single tree segmentation feature extraction based on extreme value offset deep learning, establishes a density adaptive downsampling model, optimizes the point cloud feature extraction process in the downsampling process of the extreme value offset network, and improves the point cloud segmentation accuracy. The specific optimization effects are as follows:

[0023] (1) In view of the limitation of point cloud sampling strategy, an adaptive point cloud density sampling method is designed to replace the farthest point sampling method used by the deep network. This method combines the global geometric distribution and density characteristics of the point cloud, improves the retention ability of key structure points in the sampling process, reduces the sampling probability of high-density area points, and increases the sampling probability of low-density area points.

[0024] (2) In view of the problem that the feature fusion ability is insufficient, resulting in low segmentation accuracy, a multi-scale dynamic feature fusion module is designed and embedded into the network downsampling stage. By dynamically adjusting the neighborhood scale information, the multi-scale feature fusion of points can be effectively realized.

[0025] (3) In view of the problem that there is a symbiotic relationship between adjacent single trees in the forest, resulting in the loss of boundary information, a sample correlation enhancement module is designed, which introduces an external attention mechanism in the downsampling stage to model the global correlation between single trees and enhance the understanding ability of the network to single tree features, so as to accurately distinguish the single tree boundary. BRIEF DESCRIPTION OF DRAWINGS

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0027] Figure 1 This is a flowchart illustrating the optimized method for single-tree segmentation feature extraction based on extreme value offset deep learning provided in this application embodiment.

[0028] Figure 2 This is a schematic diagram of the Point Transformer deep learning network used in the optimization method for single-tree segmentation feature extraction based on extreme value offset deep learning provided in the embodiments of this application.

[0029] Figure 3 This is a detailed structural diagram of the multi-scale dynamic feature fusion module in the optimized method for single-tree segmentation feature extraction based on extreme value offset deep learning provided in the embodiments of this application.

[0030] Figure labeling: SDG represents Scale Adaptive Generation, CMF represents capturing multi-scale feature information, DFF represents dynamic feature fusion; Conv2d represents a two-dimensional convolutional layer, BN2d represents a two-dimensional batch normalization layer, LeakyReLU is an activation function, MaxPool represents max pooling operation, Conv1d represents a one-dimensional convolutional layer, BN1d represents a one-dimensional batch normalization layer, and Sigmoid represents an activation function.

[0031] Figure 4 A detailed structural diagram of the external attention mechanism in the optimization method for single-tree segmentation feature extraction based on extreme value offset deep learning provided in the embodiments of this application.

[0032] Figure 5 A detailed structural diagram of the sample association enhancement module in the optimized method for single-tree segmentation feature extraction based on extreme value offset deep learning provided in the embodiments of this application.

[0033] Figure 6 The results of segmenting a subplot in plot G2 using an optimized method for single-tree segmentation feature extraction based on extreme value offset deep learning, as provided in the embodiments of this application, are visualized. Detailed Implementation

[0034] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application.

[0035] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] In the following description, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only.

[0037] Example 1: This application provides an optimized method for single-tree segmentation feature extraction based on extreme value shift deep learning, optimizing the feature extraction process of the extreme value shift deep learning network, such as... Figure 1 As shown, the method includes:

[0038] (1) An adaptive downsampling model is established in the extreme value shift stage of single tree segmentation based on extreme value shift deep learning. An adaptive point cloud density sampling method is adopted to select point clouds that are representative of the global structure from the forest point cloud.

[0039] In this embodiment, the extreme value offset deep learning network adopts the Point Transformer network structure, the detailed structure of which is as follows: Figure 2 As shown, a density adaptive downsampling model is introduced into the point cloud downsampling part of the network. The adaptive point cloud density sampling (APDS) method used in this model consists of four steps: global neighborhood construction, local density estimation, neighborhood distance adjustment, and global correlation scoring.

[0040] Step S1.1, Global Neighborhood Construction: Calculate the Euclidean distance between forest point clouds using the K nearest neighbor algorithm, select the K nearest neighbors for each point to form a neighborhood structure, and obtain multiple point cloud sets to capture the local geometric relationships between points. This process is represented by the following formula:

[0041] d i,j=||p i -p j ||

[0042]

[0043] N(p i )={p i |j∈argsort(d i,j )[:K]}

[0044] Where, d i,j Point p i to p j The Euclidean distance is given by , |·| represents the calculation process of the Euclidean distance, and max(·) and min(·) represent the functions for finding the maximum and minimum values, respectively. K min K represents the minimum number of nearest neighbors to prevent insufficient neighbors from yielding effective features. Experiments have verified that K is set to this minimum. min =7;K max This represents the maximum number of nearest neighbors to prevent feature redundancy caused by an excessive number of neighbors. Experiments have verified that setting K is appropriate. max =16; N represents the total number of point clouds; j represents the point cloud index, argsort(d i,j ) indicates obtaining point p i Distance sorting index to other points, [:K] indicates selecting the top K points in the distance sorting; N(p i ) represents point p i The set of its K nearest neighbors.

[0045] Step S1.2, Local density estimation: By calculating the local density of a point, the density of the space in which the point is located is quantified, providing a basis for the subsequent balance between dense and sparse regions.

[0046] First, a weighted summation is performed on the points in the neighborhood using a Gaussian kernel function to obtain point p. i Local density ρ i Then, normalization is performed, a process expressed by the following formula:

[0047]

[0048] Where σ is the Gaussian kernel scaling parameter, used to smooth point p. i and point p j To mitigate the impact of distance on Gaussian kernel weights and avoid abrupt changes in distance affecting subsequent correlation scoring, this embodiment sets σ = 1.2; ρ′ i Point p i Normalized local density;

[0049] Next, the density adjustment factor W is obtained by applying weights to the normalized density using an exponential function.i The calculation process of the adjustment factor is expressed by the following formula:

[0050] W i =exp(-ρ′) i ·λ)

[0051] The gain factor λ is used to control the importance of points in sparse regions, and λ = 1.5 was set after experimental verification.

[0052] Step S1.3, Neighborhood distance adjustment: Adjust point ρ using density weighting. i The distance to its K nearest neighbors is used to weaken the geometric influence of dense regions while increasing the structural contribution of coefficient regions, thus maintaining the global balance of the point cloud distribution. The density-weighted distance d′ is defined. i,j Distance adjustment is performed by combining local density information, expressed by the following formula:

[0053]

[0054] Step S1.4, Global Relevance Score: Quantify the importance of each point in the global geometric structure, and screen significant points through smooth weighted correlation calculation. Combined with density-weighted distance d′ i,j and density adjustment factor W i Calculate point p i Global relevance score R i This can be expressed as a formula:

[0055]

[0056] Finally, regarding point p i The global correlation scores of its K nearest neighbors are sorted, and the M points with the highest scores are selected as the final sampling results according to the set number of output points M.

[0057] Through the above steps, points that are representative of the global structure are effectively selected from all forest point clouds, and the sampled point set is passed as input to subsequent modules to complete the extraction of single tree segmentation features.

[0058] (2) For the sampling results of adaptive point cloud density sampling, a multi-scale dynamic feature fusion module is introduced to embed the adaptive downsampling model, extract point cloud features in the multi-scale neighborhood, and dynamically fuse the multi-scale information of the point cloud to enhance the feature representation capability of the model.

[0059] In this embodiment, the sampling results are mapped to the feature space via a multilayer perceptron, and a multi-scale dynamic feature fusion (MDFF) module is introduced to extract point cloud features. The detailed structure of the module is as follows: Figure 3As shown, it consists of three steps: Scale Adaptive Generation (SDG), Capture Multi-Scale Feature Information (CMF), and Dynamic Feature Fusion (DFF);

[0060] Step S2.1, Scale-adaptive generation: Given an input point cloud P = {P1, P2, ..., P...} N} Calculate the average distance d between the K nearest neighbors of each point. i,avg :

[0061]

[0062] Among them, K min =7 represents the number of least nearest neighbors, K max =16 represents the maximum nearest neighbor number; N is the total number of point clouds, ||P i -P j || represents point P i and P j The Euclidean distance, N i Point P i The set of its K nearest neighbors;

[0063] Next, the mean μ of the nearest average distance of the point cloud is calculated. d Then calculate the standard deviation σ d :

[0064]

[0065] Where, σ d A smaller σ value indicates a more uniform overall distribution of the point cloud and a smaller scale range. d The larger the value, the more significant the density difference in the point cloud, requiring a larger scale neighborhood to accommodate the sparse region.

[0066] Finally, calculate the normalized density of the points. Adaptively generated points corresponding to scales R1 and R2:

[0067]

[0068] Where D represents the set of average nearest neighbor distances for all points in the point cloud. The smaller P i The higher the local density, the greater the neighborhood radius r′. pi The radius r″ is smaller in areas of higher density. pi It is larger in areas of lower density; r″ pi Relative to r′ pi The expansion strategy maintains a larger expansion in high-density areas to enhance global feature modeling, while maintaining a smaller expansion in low-density areas to avoid excessive redundancy.

[0069] Step S2.2, Capture multi-scale feature information: For a given sampling point P i In radius r′ pi neighborhood N′ i Extract its features F′ on scale R1. pi :

[0070] F′ pi =MaxPool(MLP(F pi -F pi,j )),F pi,j ∈N′ i

[0071] Among them, F pi It is point P i In the feature representation of the feature space, F pi,j It is point P i In the characteristic space and at point P j The adjacency relationship is represented by MaxPool, which represents the max pooling operation, and MLP, which represents a multilayer perceptron, consisting of a convolutional layer, a batch normalization layer, and a LeakyReLU activation function.

[0072] Similarly, at radius r″ pi The neighborhood N″ i Extract its features F″ at scale R2. pi :

[0073] F″ pi =MaxPool(MLP(F pi -F pi,j )),F pi,j ∈N″ i .

[0074] Step S2.3, Dynamic Feature Fusion: Features F′ are fused by direct summation. pi and F″ pi The features are fused to form a new feature representation, which is then encoded into an attention vector W using a multilayer perceptron. pi :

[0075] W pi =MLP(F′) pi +F″ pi )

[0076] MLP stands for Multilayer Perceptron, which consists of two convolutional layers, each followed by batch normalization and a Sigmoid activation function.

[0077] Finally, at point P i When fusing features at scale R1 and scale R2, an attention vector is applied to achieve dynamic fusion of features at different scales:

[0078] M pi =F′ pi ⊙W pi +F″ pi ⊙(1-W pi )

[0079] Among them, M pi ⊙ represents the features after dynamic fusion, and ⊙ represents the Hadamard product.

[0080] (3) After the multi-scale dynamic feature fusion module, the sample association enhancement module is applied to learn the typical features of the single tree structure, model the global correlation between single tree samples, and capture the boundary features of the single tree.

[0081] In this embodiment, to address the problem of overlapping tree boundaries caused by symbiotic relationships between adjacent trees in a forest, which leads to blurred tree boundaries, an external attention mechanism is adopted to construct a sample association enhancement (SRE) module.

[0082] External attention mechanisms replace the intensive computation of all point pairs in traditional self-attention with learnable external memory units. This allows them to learn typical features of individual tree structures from training data, modeling global correlations between individual tree samples without significantly increasing computational cost. This enhances the network's ability to capture features of individual tree boundaries, thereby clarifying these boundaries and strengthening the feature representation of the point cloud. External attention is a mechanism based on two external learnable shared memory units (key memory and value memory). It generates an attention map by calculating the similarity between input features and the key memory, and updates features using this map and the value memory. It has linear complexity and can consider the correlations between data samples. Its structure is as follows: Figure 4 As shown:

[0083] The computational process of the external attention mechanism consists of three steps: initialization of external memory units, calculation of attention weights, and generation of output features;

[0084] Step S3.1, External Memory Unit Initialization: The external attention mechanism is implemented based on two independent external memory units: key memory and value memory; the key memory is denoted as... Used to compress input features and compute attention weights, the values ​​are denoted as follows: Used to reconstruct features based on attention weights, where Let S represent the matrix, where S is the number of elements in the memory cell and d is the feature dimension. The weights of the key memory and value memory are randomly initialized from a normal distribution with a mean of 0 and a variance of 0.001, and are used as learnable parameters that are adjusted during training based on backpropagation and the optimizer. M k and M v Created via linear layers:

[0085] Mk =Linear1(F In ),M v =Linear2(F In ),

[0086] Where Linear1 represents the linear layer that takes input feature F as input. In The dimension is transformed to S×d, and Linear2 indicates that the linear layer will convert the input features F. In The feature dimension is converted to d×S;

[0087] Step S3.2, Calculate attention weights: Given input features Where B represents the batch size and N represents the number of point clouds; the attention score is obtained by multiplying the input features by the key memory.

[0088]

[0089] in, Indicates matrix transpose;

[0090] Attention scores are column-normalized along the spatial dimension N to model the competition between different input locations for the same memory unit, thus identifying key regions:

[0091]

[0092] in, This represents the attention weight between the i-th element in the input feature map and the j-th element in the key memory. This corresponds to the attention weight between the k-th element in the input feature map and the t-th element in the key memory. The attention weights between the i-th element in the input feature map after column normalization and the j-th element in the key memory. This means that any value can be taken within the range;

[0093] L1 normalization is performed along the memory cell dimension S to balance the selection preference of each location for different memory cells and prevent excessive sparsity:

[0094]

[0095] in, The attention weight between the m-th element in the input feature map after column normalization and the l-th element in the key memory, A i,j This represents the final double-normalized attention weights.

[0096] Step S3.3, Generate output features: Apply attention weights to the value memory M v Weighted summation generates output feature F out:

[0097] F out =A·M v

[0098] Where A represents the final calculated attention weight vector; residual connections are used to fuse the original input features with the features processed by the external attention mechanism, avoiding the feature degradation problem caused by information updates introduced by the attention mechanism, and achieving complementary expression of features. This enhances the ability to distinguish individual tree boundaries while maintaining the network's perception accuracy of individual tree microstructures. Its structure is as follows: Figure 5 As shown.

[0099] Example 2: This example uses specific experimental data and evaluation metrics to verify the optimization method for single-tree segmentation feature extraction based on extreme value offset deep learning provided in Example 1.

[0100] The experimental environment for this embodiment is as follows: the system is Ubuntu 20.4; the deep learning framework is PyTorch; the GPU is Tesla V100; and the RAM is 32GB.

[0101] To test the applicability of the proposed method to different forest stand structures, airborne lidar (ALS) point cloud data provided by Weiser et al. in Germany and ALS point cloud data from the United States were used:

[0102] The ALS point cloud data was acquired by MILAN Geoservice GmbH using a RIEGL VQ-780i mounted on a Cessna C207 aircraft. The nominal specifications of this ALS instrument are as follows: accuracy of 20 mm within a 250 m scanning range, laser beam divergence angle less than 0.25 mrad, and point density of 115.2 points / m².

[0103] The acquisition of ALS point clouds in the United States was achieved using a small-spot, discrete-echo lidar system to map the study area. The Saab TopEye lidar system, mounted on a helicopter, collected data at the research site. The system operated at an altitude of 200m, a speed of 25m / s, a scan swath of 70m, a forward tilt of 8 degrees, a laser pulse rate of 7000 points / second, a maximum return count of 4 per pulse, a spot diameter of 40cm, and a point density of 4.86 points / square meter.

[0104] Since the Blue Ridge Forest in Washington, USA, is a low-density coniferous forest and the Bruten Forest in Germany is a high-density mixed forest, the two regions differ significantly in point cloud density and forest stand structure. Therefore, selecting forest point clouds from these two regions as datasets helps to verify the performance of the proposed method in forest environments of different densities and demonstrates the superiority of the method for forests of different densities.

[0105] 211 subplots, each 25m × 25m, were randomly divided into training and validation sets in an 8:2 ratio, resulting in 169 plots as the training set (containing 1558 individual trees) and 42 plots as the validation set (containing 382 trees). During training, the feature extractor deep network had 50 epochs and a training time of approximately 2.5 hours. The initial learning rate was set to 0.001, using the RMSprop optimizer with a smoothing constant alpha of 0.9. The StepLR method with dynamically adjusted learning rates was employed, where the learning rate adjustment factor gamma was set to 0.1.

[0106] The test set consisted of four distinct plots, labeled U1, U2, G1, and G2. U1 and U2 were located in the Blue Ridge area of ​​the Capitol National Forest in western Washington, while G1 and G2 were located in a mixed forest near Bruten, Germany.

[0107] In plots U1, U2, G1, and G2, the number of individual trees was 132, 186, 108, and 261, respectively. The reference canopy closure (RCD) was 61.58%, 69.09%, 93.25%, and 96.40%, respectively. Differences in vegetation distribution and data collection equipment between the two regions resulted in significant differences in point cloud density. Therefore, these two regions were selected for method testing to effectively verify the algorithm's applicability under different canopy closures. Detailed data descriptions are shown in Table 1.

[0108] Table 1. Detailed description of the test set.

[0109]

[0110] In Table 1, RCD is the reference canopy closure, representing the canopy closure conditions of different areas. The RCD is obtained by dividing the number of grid cells containing the point cloud by the total number of grid cells.

[0111]

[0112] Where n is the total number of grids, o i As a marker bit, if a point cloud exists in grid i, then o i =1, otherwise o i =0.

[0113] To test the single-tree segmentation accuracy of the optimized method for single-tree segmentation feature extraction based on extreme value shift deep learning, firstly, the distance difference D is calculated using a reference tree-estimater matching strategy, and the reference tree (ground value) is paired with the estimate tree (predicted value):

[0114]

[0115] Where (x) i ,yi ,z i ) and (x j ,y j ,z j ) represent the top coordinates of the reference tree and the estimated tree, respectively. The reference tree and the estimated tree are successfully matched when the two 3D spatial points representing the tree position information are closest; α is the weight of the height difference.

[0116] The segmentation accuracy of the proposed method is evaluated using three metrics: accuracy (p), recall (r), and F-score (F).

[0117]

[0118] Wherein, MT represents the matching tree, which is the number of reference trees and estimated trees that are in one-to-one correspondence; OE represents the missed tree, which is the number of reference trees that are not matched with estimated trees, that is, a tree is considered to be a branch of another tree, resulting in undersegmentation; CE represents the delegation error tree, which is the number of reference trees and estimated trees that have a one-to-many relationship, that is, a tree is split into multiple trees, resulting in oversegmentation.

[0119] Select one subplot from plot G2 for visualization of the segmentation results, such as... Figure 6 As shown: (a) represents the point cloud after data preprocessing, (b) represents the point cloud after extreme value offset, (c) represents the offset point cloud after clustering, and (d) represents the result of the offset point cloud after mapping and point set aggregation. The dark blue points are considered to be noise points.

[0120] The segmentation results of the proposed method in the four plots are shown in Table 2. A total of 687 individual trees were identified, of which 590 trees were correctly identified, 61 trees were under-segmented, and 36 trees were over-segmented. The F-scores of plots U1, U2, G1, and G2 were 0.93, 0.93, 0.92, and 0.91, respectively. The average p, r, and F of the four plots were 0.94, 0.90, and 0.92, respectively.

[0121] Table 2. Accuracy of the proposed method assessed in four sample plots.

[0122]

[0123] 1. Number of correctly divided individual trees; 2. Number of under-divided individual trees; 3. Number of over-divided individual trees.

[0124] To verify the contribution and effectiveness of the APDS method, MDFF module, and SRE module in the single-tree segmentation task, an ablation experiment was designed, as shown in Table 3. Each module was introduced step-by-step for comparative analysis. The experiment divided the model into four groups, each configured according to whether the APDS method, MDFF module, and SRE module were introduced. Specific configurations are shown in the table, where "×" indicates that the corresponding module is not introduced, and "√" indicates that the corresponding module is introduced.

[0125] Table 3. Ablation studies using the APDS method, MDFF module, and SRE module.

[0126]

[0127] As can be seen from Table 3, the segmentation accuracy of the model gradually improves with the gradual introduction of each module:

[0128] Using the original downsampling model of the Point Transformer deep learning network with extreme value offset as the baseline model, its F score is 0.90, indicating that the baseline model has a certain segmentation ability in the single tree segmentation task, but there is still a lot of room for improvement.

[0129] After introducing the APDS method, the recall rate improved to 0.897, indicating that the sampling method reduced missed detections. The adaptive point cloud density sampling method addresses the significant density differences in forest point clouds by combining global geometric structure and local density information to achieve point cloud sampling. This effectively solves the problems of oversampling in dense areas and undersampling in sparse areas in the traditional farthest point sampling method, making the network more likely to identify real trees and thus improving the recall rate.

[0130] Based on the introduction of the APDS method, the MDFF module was further introduced, and the F score of the model was improved to 0.91. This shows that the MDFF module, by dynamically fusing multi-scale features, ensures that the network can effectively handle the multi-scale feature differences in complex forest scenes, enhances the network's feature representation ability for single tree features, and effectively improves the segmentation accuracy, proving its effectiveness in processing multi-scale features of forest point clouds.

[0131] Based on the introduction of the APDS method and MDFF module, the SRE module was further introduced, which improved the accuracy of the model to 0.924 and the F score to 0.924. This shows that the SRE module, by introducing an external attention mechanism, models the global correlation between individual trees, further optimizes the global feature expression, improves the accurate differentiation of individual tree boundaries, and achieves feature complementarity, thereby enhancing the network's feature expression ability.

[0132] Ablation experiments show that the APDS method, MDFF module, and SRE module form a complete and collaborative feature extraction mechanism in the downsampling process.

[0133] The APDS method ensures the preservation of key structural points; the MDFF module enhances the fusion of local structure and global positional information; and the SRE module models the correlation between individual trees, further optimizing the global feature representation. The combined effect of these three methods significantly improves the accuracy of individual tree segmentation.

[0134] a. Point cloud sampling method

[0135] To verify the effectiveness of the proposed adaptive point cloud density sampling method, a comparative experiment was conducted with grid sampling and point sorting sampling (PRS). The experimental results are shown in Table 4.

[0136] Table 4. Comparison of single-tree segmentation accuracy using different sampling methods on four sample plots

[0137]

[0138] Grid sampling: This method involves dividing the point cloud into a fixed-size 3D grid, and selecting a representative point from each grid for sampling. This method is simple and efficient, and can uniformly cover the spatial structure of the point cloud. However, for forest point clouds with uneven density distribution, it easily overlooks key information in low-density areas, leading to the loss of points in sparse regions. Grid sampling performs reasonably well in low-density plots (U1, U2), but in high-density plots (G1, G2), information loss in complex structural areas such as tree branches is significant, resulting in decreased segmentation accuracy.

[0139] Point ranking sampling: This method calculates the geometric importance score (such as curvature, edge saliency, etc.) of each point using a deep learning model, prioritizing high-scoring points for sampling. Its advantage lies in preserving task-sensitive local structural features (such as canopy edges) and optimizing the sampling distribution through multi-scale features. However, in forest scenes, because the point cloud density in the trunk region is significantly higher than in the canopy region, the implicit scoring mechanism is prone to scoring confusion in dense trunk areas due to redundant point feature similarity, leading to oversampling of low-information points. In sparse canopy areas, the model struggles to accurately capture weak geometric features due to foliage occlusion and point cloud discreteness, resulting in missed key points. Point ranking sampling performs reasonably well in low-density plots (U1, U2), but its segmentation accuracy is low in high-density plots (G1, G2).

[0140] Adaptive Point Cloud Density Sampling (APDS): By combining global geometric structure and local density characteristics, APDS effectively solves the sampling problem caused by uneven density distribution in forest point clouds. In high-density regions, density weighting weakens the contribution of redundant points, while in low-density regions, density normalization enhances the sampling probability of key information points. Furthermore, APDS optimizes the balance of point cloud distribution globally, ensuring the preservation of key structural points. Experimental results show that the APDS method achieves the best segmentation results in all four sample plots.

[0141] b. Attention mechanism

[0142] To verify the effectiveness of the external attention mechanism used in the proposed sample association enhancement module, it was compared with sparse attention mechanism (SAM) and graph-based attention mechanism (GAM) through experiments. The experimental results are shown in Table 5.

[0143] Table 5. Comparison of single-tree segmentation accuracy under different attention mechanisms in four sample plots

[0144]

[0145]

[0146] Sparse attention mechanism: By filtering key points in the input point cloud, it focuses only on highly sparse feature regions, thereby reducing computational complexity. This mechanism assigns higher weights to sparse regions, improving the local feature representation capability of the point cloud. It performs well in low-density forest stands (U1, U2), but in high-density forest stands (G1, G2), its overall segmentation performance is slightly lower because it cannot effectively capture the global correlation between point cloud samples. By comparing the sparse attention mechanism, we can verify the advantages of external attention mechanisms in modeling the correlation between individual trees in high-density mixed forests.

[0147] Graph-based attention mechanism: This mechanism enhances the hierarchical modeling capability of point cloud geometric features by constructing a local point cloud graph structure and learning dynamic weight relationships between neighboring nodes. While this mechanism effectively captures the local geometric features of individual trees in regular forest stands, in high-density mixed forests (G1, G2), its over-reliance on static modeling of local neighborhood relationships makes it difficult to effectively establish long-distance spatial associations across the canopy, leading to decreased segmentation accuracy for severely occluded trees. Furthermore, the computational complexity of the dynamic graph structure increases exponentially with point cloud density, affecting the algorithm's real-time performance. By comparing with the point cloud graph attention mechanism, the advantage of the external attention mechanism—which utilizes learnable memory units—can be highlighted to further enhance the feature representation of point clouds without significantly increasing computational costs, thereby improving segmentation robustness in complex forest stand scenarios.

[0148] External attention mechanism: By using externally learnable key-value pair memory units, the correlation between point cloud samples is modeled globally, overcoming the limitations of traditional self-attention mechanisms. By avoiding complex pairwise relationship calculations through the external memory structure, it effectively captures the global correlation between individual trees. In high-density mixed forests (G1, G2), it can effectively model the global correlation between individual trees and accurately distinguish boundary points.

[0149] The following comparative experiment is introduced:

[0150] The chosen comparison methods include: Spatial Horizontal Distance Rule (SHDR), Improved DBSCAN and K-means method (Improved DK), and EOM method. The basic steps for processing point clouds using the above methods are as follows:

[0151] The Spatial Horizontal Distance Rule (SHDR) method segments the tree sequentially from the highest to the lowest point. First, the highest point is used as the seed point of the target tree. Then, the spatial horizontal distance rule is used to determine whether each point below the treetop belongs to the target tree's canopy layer. This process is iterated until all points are classified. This method offers high segmentation accuracy but has high time complexity and is highly sensitive to the selection of the horizontal distance threshold parameter, resulting in poor general applicability.

[0152] The improved DBSCAN and K-means method (Improved DK) is an improved strategy that integrates extreme value migration deep learning and the K-nearest neighbor algorithm. This method spatially transforms the original point cloud using extreme value migration to obtain a migrated point cloud. The seed points found in the migrated point cloud by the DBSCAN algorithm are then used as the initial parameters for the K-means method for single-tree segmentation. However, the uneven density distribution of the migrated point cloud affects the seed points found by the DBSCAN algorithm, thus impacting the segmentation accuracy.

[0153] The EOM method is a tree segmentation method based on extreme value offset deep learning. It increases the distinguishability between trees through a deep network, then introduces adaptive clustering and point set aggregation methods to complete the segmentation. While this method boasts high generality and segmentation accuracy, it fails to account for the randomness of deep network offsets, thus its generalization ability and segmentation accuracy need further improvement.

[0154] The above-described single-tree segmentation method was used to conduct tests in four test plots, and the results are shown in Table 6.

[0155] Table 6. Comparison of single-tree segmentation accuracy in four sample plots using the SHDR method, Improved DK method, EOM method, and the proposed method.

[0156]

[0157]

[0158] The SHDR method achieved average p, r, and F values ​​of 0.84, 0.85, and 0.84 in the four plots, respectively; the Improved DK method achieved average p, r, and F values ​​of 0.83, 0.86, and 0.85 in the four plots, respectively; the EOM method achieved average p, r, and F values ​​of 0.91, 0.88, and 0.90 in the four plots, respectively; and the proposed method achieved average p, r, and F values ​​of 0.94, 0.90, and 0.92 in the four plots, respectively.

[0159] To examine the accuracy and generalization of an optimized method for single-tree segmentation feature extraction based on extreme value shift deep learning for plots with different forest stand structures and point cloud densities, experiments were conducted on two low-density coniferous forest plots (U1, U2) in the Blue Ridge region of Washington, USA, and two high-density mixed forest plots (G1, G2) in Bruten, Germany. The experimental results showed that introducing the APDS method, MDFF module, and SRE module into the downsampling stage of the EOM method resulted in average precision, recall, and F-score of 0.94, 0.90, and 0.92, respectively. Compared with the EOM method (p = 0.91, r = 0.88, F = 0.90), p, r, and F were improved by 3%, 2%, and 2%, respectively.

[0160] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results.

Claims

1. An optimized method for single-tree segmentation feature extraction based on extreme value shift deep learning, characterized in that, The method includes the following steps: (1) An adaptive downsampling model is established in the extreme value shift stage of single tree segmentation based on extreme value shift deep learning. An adaptive point cloud density sampling method is adopted to select point clouds that are representative of the global structure from the forest point cloud. (2) For the sampling results of adaptive point cloud density sampling, a multi-scale dynamic feature fusion module is introduced to embed the adaptive downsampling model, extract point cloud features in the multi-scale neighborhood, and dynamically fuse the multi-scale information of the point cloud to enhance the feature representation capability of the model. (3) After the multi-scale dynamic feature fusion module, the sample association enhancement module is applied to learn the typical features of the single tree structure, model the global correlation between single tree samples, and capture the boundary features of the single tree.

2. The optimized method for single-tree segmentation feature extraction based on extreme value offset deep learning as described in claim 1, characterized in that, The specific process of single-tree segmentation based on extreme value offset deep learning includes data preprocessing, extreme value offset, mean offset, spatial mapping, and point set aggregation.

3. The optimized method for single-tree segmentation feature extraction based on extreme value shift deep learning as described in claim 2, characterized in that, The extreme value shift uses a deep network to spatially transform the forest point cloud, shifting the point cloud of a single tree to the corresponding treetop to enhance the distinguishability between individual trees. The extreme value shift stage uses a Point Transformer network for feature extraction, which includes a downsampling module, an upsampling module, and a Transformer module. The downsampling module is replaced with an adaptive downsampling model to optimize the feature extraction process for single tree segmentation.

4. The optimized method for single-tree segmentation feature extraction based on extreme value shift deep learning as described in claim 1, characterized in that, In step (1), the adaptive point cloud density sampling method is divided into four steps: global neighborhood construction, local density estimation, neighborhood distance adjustment, and global correlation scoring.

5. The optimized method for single-tree segmentation feature extraction based on extreme value offset deep learning as described in claim 4, characterized in that, The global correlation score quantifies the importance of each point cloud in the global geometric structure. Significant points are selected through correlation calculation using a Gaussian function with smooth weighting, and point p is calculated. i Global relevance score R i This can be expressed as a formula: Where, d′ i,j Point p i and p j The density-weighted Euclidean distance between them, W i σ represents the density adjustment factor, and σ represents the Gaussian kernel scale parameter for point p. i The global correlation scores of its K nearest neighbors are sorted, and the M points with the highest scores are selected as the sampling results according to the set number of output points M.

6. The optimized method for single-tree segmentation feature extraction based on extreme value offset deep learning as described in claim 1, characterized in that, In step (2), the multi-scale dynamic feature fusion module is used to extract point cloud features, which is divided into three steps: scale adaptive generation, capturing multi-scale feature information and dynamic feature fusion.

7. The optimized method for single-tree segmentation feature extraction based on extreme value offset deep learning as described in claim 6, characterized in that, The detailed steps of the scale-adaptive generation include: Given an input point cloud P = {P1, P2, ..., P...} N } Calculate the average distance d between the K nearest neighbors of each point. i,avg : Where max(·) represents the function for finding the maximum value, min(·) represents the function for finding the minimum value, and K min K represents the number of least nearest neighbors. max Represents the maximum nearest neighbor number; N is the total number of point clouds, ||P i -P j || represents point P i and P j The Euclidean distance, N i Point P i Given the set of its K nearest neighbors; calculate the mean μ of the nearest average distances to the point cloud. d Then calculate the standard deviation σ d : Where, σ d A smaller σ value indicates a more uniform overall distribution of the point cloud and a smaller scale range. d A larger value indicates significant density differences in the point cloud, requiring a larger scale neighborhood to accommodate sparse regions; calculate the normalized density of the points. Adaptively generated points corresponding to scales R1 and R2: Where D represents the set of average nearest neighbor distances for all points in the point cloud.

8. The optimized method for single-tree segmentation feature extraction based on extreme value shift deep learning as described in claim 6, characterized in that, The dynamic feature fusion includes: When capturing multi-scale feature information, features F′ are extracted at scales R1 and R2 respectively using a multilayer perceptron. pi and F″ pi The features at the two scales are directly summed to form a new feature representation, which is then encoded into an attention vector W using a multilayer perceptron. pi : W pi =MLP(F′ pi +F″ pi ) MLP stands for Multilayer Perceptron, which consists of two convolutional layers, each followed by batch normalization and a Sigmoid activation function. At point P i When fusing features at scale R1 and scale R2, an attention vector is applied to achieve dynamic fusion of features at different scales: M pi =F′ pi ⊙W pi +F″ pi ⊙(1-W pi ) Among them, M pi ⊙ represents the features after dynamic fusion, and ⊙ represents the Hadamard product.

9. The optimized method for single-tree segmentation feature extraction based on extreme value offset deep learning as described in claim 1, characterized in that, In step (3), an external attention mechanism is used to construct a sample association enhancement module. The calculation process of the external attention mechanism is divided into three steps: external memory unit initialization, calculation of attention weights and generation of output features.

10. The optimized method for single-tree segmentation feature extraction based on extreme value offset deep learning as described in claim 9, characterized in that, The calculation process for the attention weights is as follows: Given input features Where B represents the batch size and N represents the number of point clouds; the input features are compared with the key memory M. k Multiply to get attention score Where T represents the matrix transpose; Attention scores are column-normalized along the spatial dimension N to model the competition between different input locations for the same memory unit, thus identifying key regions: in, This represents the attention weight between the i-th element in the input feature map and the j-th element in the key memory. This corresponds to the attention weight between the k-th element in the input feature map and the t-th element in the key memory. The attention weights between the i-th element in the input feature map after column normalization and the j-th element in the key memory. This means that any value can be taken within the range; L1 normalization is performed along the memory cell dimension S to balance the selection preference of each location for different memory cells and prevent excessive sparsity: in, The attention weight between the m-th element in the input feature map after column normalization and the l-th element in the key memory, A i,j This represents the final double-normalized attention weights.

Images