Point cloud tree species identification method based on DASA-Net deep learning network

Abstract

The invention provides a point cloud tree species identification method based on a DASA-Net deep learning network, and belongs to the technical field of forestry remote sensing monitoring. According to the method, a deep learning network DASA-Net based on differential attention sampling is proposed, and the limitation of the prior art is broken through through multi-level feature optimization: a dynamic feature retention mechanism driven by geometric difference is introduced in a sampling stage, so that the problem of key information loss caused by uneven density is solved; a density self-adaptive convolution module is designed in the feature extraction stage, and the modeling capacity for local difference features such as the branch curvature and the canopy concave-convex structure is enhanced; a hierarchical feature recovery strategy is developed in the feature aggregation stage, and the global expression ability of the model for the tree overall form and the multi-scale space relation is improved. According to the method, higher classification accuracy is realized while high recall rate is kept, and reliable technical support is provided for smart city ecological monitoring and greening resource digital management.

Description

Technical Field

[0001] The present invention belongs to the technical field of forestry remote sensing monitoring, and in particular relates to a point cloud tree species recognition method based on a DASA-Net deep learning network. Background Art

[0002] Urban greening trees play an important role in maintaining urban ecological stability, reducing carbon emissions and improving the living environment. Accurate identification of tree species is a core link in assessing the ecological value of urban greening resources, guiding carbon sink management and biodiversity conservation. Traditional tree species identification mainly relies on manual field surveys, and classification is carried out through visual judgment or measurement of tree morphological parameters. It has the disadvantages of low efficiency, strong subjectivity, and difficulty in large-scale application. With the development of remote sensing technology, laser radar (LiDAR) technology has become an important means of urban tree monitoring because it can obtain high-precision three-dimensional point cloud data. However, how to achieve automatic and high-precision tree species identification from complex and heterogeneous LiDAR point clouds remains a technical challenge that needs to be overcome in the field of forestry remote sensing.

[0003] Early inventions of tree species identification using LiDAR point clouds were mainly based on traditional machine learning methods, which artificially extracted structural features such as tree height, crown width, and branch curvature, and combined them with classification models such as support vector machines and random forests to achieve tree species differentiation. Although this type of method has improved recognition efficiency to a certain extent, it relies on artificial feature engineering and has difficulty capturing the deep correlation characteristics of the three-dimensional morphology of trees. It also has poor adaptability to complex scenes such as uneven point cloud density and morphological similarities between tree species. In addition, artificially designed features are often limited to specific tree species or fixed scenes, and their generalization capabilities are insufficient, making it difficult to meet the actual needs of multi-species, multi-scale, and dynamic changes in urban greening monitoring. In recent years, deep learning methods have made significant progress in the field of point cloud processing. Models based on voxelization, multi-view projection, and raw point cloud input have provided a new technical path for tree species identification. However, existing deep learning models still face multiple bottlenecks when applied to urban tree point clouds: First, point cloud sampling mechanisms are limited. Traditional random sampling or farthest point sampling methods are prone to missing key morphological features such as branch bifurcations and leaf distribution in areas with uneven tree point cloud density (such as crown edges and interlaced branches and leaves), resulting in a loss of information integrity in the input data. Second, standard convolution operations are insufficiently capable of modeling local geometric features such as changes in trunk curvature and differences in canopy three-dimensional structure, making it difficult to effectively distinguish morphologically similar tree species (such as camphor and camphor among broad-leaved trees). Third, existing pooling strategies overly rely on maximum pooling, ignoring less significant but discriminative fine-grained features, which limits the model's ability to globally perceive the overall crown morphology and branch spatial distribution patterns. Furthermore, the widespread scale variation, noise interference, and occlusion issues in point cloud data further exacerbate the difficulty of model feature extraction and classification.

[0004] Therefore, how to accurately identify tree species in complex urban scenes based on LiDAR point clouds and use deep learning methods has become a research focus. Summary of the Invention

[0005] In view of the problem of insufficient recognition accuracy in complex scenarios that exists in the background technology, especially the poor ability to distinguish morphologically similar tree species, the purpose of the present invention is to provide a point cloud tree species recognition method based on the DASA-Net deep learning network. The method proposes a deep learning network based on Differential Attention Sampling Network (DASA-Net), which breaks through the limitations of existing technologies through multi-level feature optimization: a dynamic feature retention mechanism driven by geometric differences is introduced in the sampling stage to solve the problem of key information loss caused by uneven density; a density-adaptive convolution module is designed in the feature extraction stage to enhance the modeling ability of local difference features such as branch curvature and canopy concave-convex structure; a hierarchical feature recovery strategy is developed in the feature aggregation stage to enhance the model's global expression ability of the overall morphology and multi-scale spatial relationship of trees. The method of the present invention achieves higher classification accuracy while maintaining a high recall rate, providing reliable technical support for smart city ecological monitoring and digital management of greening resources.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] A point cloud tree species recognition method based on the DASA-Net deep learning network includes the following steps:

[0008] Step 1. Use a vehicle-mounted LiDAR to perform multi-view scanning of urban street trees to obtain high-density 3D point cloud data containing reflection intensity information. Then, use a cloth simulation filtering algorithm to filter out the ground point cloud.

[0009] Step 2. A tree segmentation algorithm based on point cloud geometric feature constraints, combined with the density-based spatial clustering (DBSCAN) algorithm and principal direction indexing technology, is used to separate complete tree point clouds from the scene point cloud obtained in Step 1. Tree species labels are added to each individual tree point cloud file. The resulting labeled individual tree point clouds can be used to construct a single tree sample library containing several urban tree species.

[0010] Step 3. Standardize and preprocess the single tree sample library: use farthest point sampling to unify the sampling, and eliminate scale differences through geometric centroid alignment and unit sphere normalization; divide the preprocessed data into training and test sets according to a certain ratio;

[0011] Step 4. Build the DASA-Net deep learning network, use the data in the training set as training data, and the data in the test set as validation data, and train the DASA-Net deep learning network until the DASA-Net deep learning network converges;

[0012] Step 5. Collect, separate, and preprocess the data of the urban street trees to be identified according to steps 1-3. Then input the preprocessed single tree point cloud data into the DASA-Net deep learning network trained in step 4 to obtain the tree species identification results.

[0013] Furthermore, in step 4, the DASA-Net deep learning network includes a first multi-layer perceptron MLP, several layered sampling aggregation modules, a global pooling module and a connector, which are connected in sequence;

[0014] The stratified sampling aggregation module includes a feature aggregation unit and an inverse residual MLP unit connected in sequence. The feature aggregation unit in the first stratified sampling aggregation module includes an attention-based feature downsampling layer (AFDS), a difference graph convolution layer (DiffConv), a feature grouping layer (Grouping), a second multi-layer perceptron MLP, and a feature recycling pooling layer (FRP) connected in sequence. The feature aggregation units in the remaining stratified sampling aggregation modules only include an attention-based feature downsampling layer, a feature grouping layer (Grouping), a third multi-layer perceptron MLP, and a recycling pooling layer connected in sequence.

[0015] The first multi-layer perceptron MLP performs initial feature embedding, the feature retention sampling layer based on the attention mechanism dynamically retains key points through attention weights, the difference graph convolution layer extracts local difference features through density-adaptive graph convolution, the recovery pooling layer recovers secondary significant features through iterative pooling, and the feature grouping layer (Grouping) is used to group the input features. The second multi-layer perceptron MLP, the third multi-layer perceptron MLP and the first multi-layer perceptron MLP have the same structure and are used for feature embedding; the global pooling module and the connector are used for global feature aggregation.

[0016] Furthermore, the number of stratified sampling aggregation modules is preferably 5.

[0017] Furthermore, the specific process of the feature-preserving sampling layer of the attention mechanism is as follows:

[0018] The input point cloud is first mapped to a high-dimensional space through an embedding layer to improve the ability to express spatial structure and geometric details; then, the module constructs a local neighborhood set and calculates the difference between the neighborhood points and the center point features to capture subtle geometric changes in the local area of ​​the point cloud; based on the Transformer's self-attention mechanism, the center point feature is used as the query vector (Query), and the neighborhood feature differences are used to generate a key vector (Key) and a value vector (Value), respectively. Finally, the deep correlation between the center point and the neighborhood points is calculated through the attention weight, and the normalized attention weight is generated through the Softmax function to dynamically evaluate the importance of each neighborhood point to the center point; finally, the importance score of each point is calculated based on the combination of the obtained attention weight and the neighborhood features, and the most representative point cloud subset is accurately selected and retained through the Top-M operation to achieve effective point cloud feature retention and efficient downsampling, and finally the sampled point cloud data is obtained.

[0019] Furthermore, the specific process of calculating the deep correlation between the center point and the neighborhood points through attention weight is as follows:

[0020] For each point i in the point cloud, in its local neighborhood set S i The correlation measure within is defined as shown in formula (1):

[0021] h l (p i ,p ij )=Q(p i ) T K(p ij -p i ) (1)

[0022] Among them, Q and K represent linear mappings, p i is the feature of the center point, p ij is the feature of the points in the neighborhood, p ij -p i Indicates the feature difference between the neighborhood points and the center point;

[0023] The specific calculation formula of normalization is shown in formula (2):

[0024]

[0025] Among them, d is the feature dimension.

[0026] Furthermore, the specific working process of the difference graph convolution layer is as follows:

[0027] First, this layer takes point cloud data of dimension (B, C, N) as input, where B is the batch size, C is the feature dimension of the point, and N is the number of points. The module calculates the spatial Euclidean distance between all points in the input point cloud and obtains a distance matrix of dimension (B, N, N) to accurately describe the spatial relationship between points in the point cloud. Subsequently, the module adaptively determines the radius based on the density characteristics of the local area where each point is located and stores it in the (B, N, 1) matrix, thereby realizing dynamic neighborhood query and effectively dealing with the problem of uneven density of point clouds; after determining the dynamic neighborhood of each point, the spatial coordinate features of the point are further encoded and mapped to a high-dimensional space, thereby enhancing the representation ability of spatial structure information; at the same time, the adaptive radius is processed using a shielded attention mechanism, and the features of each center point and its neighborhood points are dynamically weighted to obtain a local smooth feature with a dimension of (B, C, N); then, the local difference feature is obtained by calculating the feature difference between the center point feature and the position encoding of its neighborhood; the local difference feature and the original feature are spliced ​​and fused to form a more discriminative combined feature (B, C_out, N); finally, the final feature expression is output through feature combination processing.

[0028] Furthermore, the specific process of adaptive radius adjustment based on the density characteristics of the local area is as follows:

[0029] DiffConv first uses Gaussian kernel density estimation (KDE) to calculate the kernel density of each point. The formula is shown in formula (3):

[0030]

[0031] Among them, i represents the target point whose density is currently being calculated, j is the summation variable, N is the total number of points in the point cloud, h is the bandwidth parameter used to control the degree of smoothness, ‖p i -p j ‖ represents the Euclidean distance between point pairs, and the output d i The larger the value, the denser the points in the area, so the radius is appropriately expanded to enhance the diversity of local features;

[0032] In order to make the density of different points have a uniform scale, d i Normalized to [0,1] to get the normalized density Embed the normalized density into the radius calculation to obtain the density-driven search radius r for each point i , and its formula is shown in formula (4):

[0033]

[0034] Among them, r is the initial fixed sphere query radius. The square root is used to slow down the radius growth rate and prevent excessive expansion of high-density areas.

[0035] The specific process of shielding the attention mechanism for processing is:

[0036] The learned irregular neighborhood selection is introduced; the attention weight of each point in the neighborhood is calculated by formula (5) and then normalized by formula (6):

[0037]

[0038]

[0039] Furthermore, the specific working process of the feature recovery pooling layer is as follows:

[0040] First, traditional maximum pooling is used to extract salient feature information from local areas to generate an initial permutation-invariant feature vector. Then, for the point feature matrix not selected by the first pooling, maximum pooling is iteratively repeated to continuously extract secondary salient features, forming a multi-level permutation-invariant feature set. By designing a hierarchical loss function, these recovered features are dynamically weighted to achieve detailed expression and efficient utilization of feature information, ultimately improving the richness and discriminative ability of the overall feature representation. Finally, the features obtained by maximum pooling and the features obtained by recycled pooling are combined to obtain the required aggregated features.

[0041] Furthermore, the calculation formula of the layered loss function is,

[0042]

[0043] Among them, λ is a hyperparameter used to balance the weight between the two losses; the classification loss L c It is composed of the prediction results of each permutation invariant feature set, and its cross entropy loss function is expressed as formula (8):

[0044]

[0045] Feature refinement loss L r The calculation is based on the predicted difference between the recovered features and the original maximum pooling features, as shown in formula (9):

[0046]

[0047] Among them, α i is an adjustable parameter, y m and They are the true label and the predicted probability based on the initial maximum pooling feature.

[0048] Furthermore, in step 4, the DASA-Net deep learning network is trained using the AdamW optimizer, with the initial learning rate set to 0.001. The cosine annealing strategy is adopted, and multi-objective optimization is performed in combination with label smoothing cross entropy loss and feature refinement loss. The model is converged through multiple rounds of iterative training.

[0049] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0050] The present invention first constructed a point cloud dataset for tree species identification tasks and performed standardized preprocessing to ensure data balance and network adaptability. Subsequently, comprehensive experiments were conducted on point cloud datasets of 24 tree species. The results showed that the DASA-Net deep learning network of the present invention achieved an OA of 96.62% and a mAcc of 96.58% in tree species identification tasks, both outperforming the comparison networks PointNeXt, ASSANet, and DeepGCN. It achieved the best classification results in the identification tasks of most tree species, fully demonstrating the accuracy of the DASA-Net deep learning network of the present invention in distinguishing tree species characteristics and providing a solid foundation for subsequent tree species identification work. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Schematic diagram of the structure of the DASA-Net network of the present invention.

[0052] Figure 2 It is a schematic diagram of the structure of the AFDS layer in the DASA-Net network of the present invention.

[0053] Figure 3 Schematic diagram of the structure of the Head DiffConv layer in the DASA-Net network of the present invention.

[0054] Figure 4 Schematic diagram of the structure of the FRP layer in the DASA-Net network of the present invention.

[0055] Figure 5 This is a schematic diagram simulating the acquisition effect of a vehicle-mounted lidar.

[0056] Figure 6 This is a diagram showing the changes in loss, OA, and mAcc during DASA-Net model training. DETAILED DESCRIPTION

[0057] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below in conjunction with the implementation methods and drawings.

[0058] A point cloud tree species recognition method based on the DASA-Net deep learning network includes the following steps:

[0059] Step 1. Use a vehicle-mounted LiDAR to perform multi-angle scanning of urban roadside trees, acquiring high-density 3D point cloud data containing reflection intensity information. (As the vehicle-mounted LiDAR moves along the road, it continuously rotates and scans to acquire 360° 3D point cloud data from the base of the trunk to the top of the canopy.) The ground point cloud is then filtered out using a cloth-simulation filtering algorithm.

[0060] Step 2. Based on the point cloud geometric feature constraints, a single tree segmentation algorithm is used, combined with DBSCAN clustering and principal direction indexing technology, to separate complete single tree point clouds from the scene point cloud and construct a single tree sample library containing several urban tree species;

[0061] Step 3. Standardize and preprocess the single tree sample library. The specific process is as follows:

[0062] Step 3.1. Perform point cloud downsampling:

[0063] In order to match the input point count requirement of the model and keep the spatial distribution of point cloud data as uniform as possible, this paper adopts Farthest Point Sampling (FPS) to downsample all point cloud data to 2048 points. The FPS sampling strategy can ensure that the sampling points evenly cover the entire geometric space, avoiding the problem of feature information loss that may be caused by traditional random sampling. In this process, if the number of points of some point cloud samples is less than 2048, they will be eliminated to ensure the consistency of data dimension, thereby avoiding instability of the network during training due to different input sizes;

[0064] Step 3.2. Dataset Balancing:

[0065] Because the number of point cloud samples of different tree species in the public dataset varies significantly, if they are used directly for training, it may lead to an imbalance in data distribution between categories, which in turn affects the overall classification performance. To ensure a balanced training sample for each tree species, for tree species point clouds with a small sample size, a secondary independent downsampling method is used. Two different point cloud samples are taken from the original point cloud to ensure that the spatial distribution of the sampling points has a certain degree of difference, thereby enhancing the diversity of the data and amplifying the sample. Finally, 24 tree species were obtained, with a total of 6,312 point cloud samples, which effectively alleviated the data category imbalance problem and improved the network's learning and generalization capabilities for different tree species.

[0066] Step 3.3. Point cloud normalization:

[0067] Since the point cloud data used comes from different sources, there may be significant differences in coordinate scale and spatial position. To eliminate these effects, the present invention normalizes all point cloud data. This process is divided into two parts: centering and scale normalization. Centering is mainly achieved by calculating the geometric centroid of the point cloud and subtracting the centroid coordinates from the coordinates of all points. The purpose is to align the point cloud to the origin, thereby eliminating displacement deviations caused by different data collection locations. Scale normalization is achieved by calculating the maximum Euclidean distance from all points to the geometric centroid and dividing the coordinates of all points by this maximum distance to scale the point cloud to the unit sphere. The normalization process not only eliminates the scale differences between different tree point clouds, but also enhances the consistency of the data, enabling the network to more stably learn the geometric characteristics of different tree species.

[0068] Step 3.4. Divide the data set:

[0069] To ensure scientific model training and testing, all point cloud data, after adding tree species category labels, was stored in the HDF5 file format for efficient storage and access. Subsequently, the point cloud data for each category was randomly divided into a 70% training set and a 30% test set to ensure an even distribution of training and test data and minimize category bias. This process ultimately resulted in a series of standardized data files with the suffix .h5, which served as input files for subsequent deep learning models.

[0070] Step 4. Build the DASA-Net deep learning network, use the data in the training set as training data, and the data in the test set as labels, and train the DASA-Net deep learning network until the DASA-Net deep learning network converges;

[0071] The structural diagram of DASA-Net deep learning network is as follows Figure 1 As shown, it includes a multi-layer perceptron MLP, five layered sampling aggregation modules, a global pooling module and a connection head connected in sequence;

[0072] Among them, the stratified sampling aggregation module includes feature aggregation units and inverse residual MLP units connected in sequence; the feature aggregation unit in the first stratified sampling aggregation module includes an attention-based feature downsampling layer (AFDS), a difference graph convolution layer (DiffConv), a feature grouping layer (Grouping), a second multi-layer perceptron MLP and a recycling pooling layer (FRP) connected in sequence; the feature aggregation units in the remaining stratified sampling aggregation modules only include an attention-based feature downsampling layer, a feature grouping layer (Grouping), a third multi-layer perceptron MLP and a recycling pooling layer connected in sequence.

[0073] In point cloud deep learning networks, layer-by-layer downsampling aims to gradually reduce the point cloud size while preserving key information. This allows for multi-scale feature extraction, optimizes computational efficiency and memory usage, and effectively captures both local and global geometric structural features. In tree species identification tasks, the trunk and crown outlines of different tree species often differ significantly, and this outline information constitutes a crucial feature of individual tree point clouds. However, due to variations in scanning angles and distances when acquiring individual tree point clouds using LiDAR, as well as inter-branch occlusion and varying canopy densities, point cloud data can exhibit significant density unevenness. This manifests as dense point clouds on trunks and crowns facing the scan direction and at lower heights, while sparse or even absent points are found on trunks facing away from the scan direction and at the edges of higher crowns. This density unevenness can lead to the loss of key local features during point cloud downsampling, compromising subsequent tree species identification accuracy. Therefore, the downsampling process should prioritize the preservation of geometric details across tree regions to avoid loss of key feature information.

[0074] Existing sampling methods, such as Random Sampling (RS) and Farthest Point Sampling (FPS), have certain limitations when processing point cloud data. In tree species identification tasks, especially the structural feature analysis of single tree point clouds, traditional sampling methods may not be able to fully retain key information, affecting subsequent recognition accuracy. Although random sampling is simple to implement, it relies entirely on randomly selected points, often losing a lot of detailed information in complex point clouds, especially in areas with complex or small structures such as tree branches and leaves. In addition, random sampling cannot guarantee the representativeness of sampling points for different regions, making it difficult for the network to capture fine-grained local features during subsequent feature extraction. In contrast, although the FPS method can ensure the uniform distribution of the point cloud by selecting points far away from the sampled points, thereby improving the representation of global structure, this method often ignores local details in the point cloud while focusing on the global distribution of the point cloud. For example, the structure of trees may show significant local features in different areas (such as the distribution of branches, the arrangement of leaves, etc.), and these details may be ignored in the farthest point sampling, making it difficult for the model to distinguish subtle differences between different tree species. Therefore, the present invention designs a feature-preserving sampling layer (AFDS) based on the attention mechanism. The feature-preserving sampling layer of the attention mechanism uses the Transformer self-attention mechanism to perform feature modeling on the local neighborhood of the point cloud, dynamically allocates attention weights by constructing a point cloud neighborhood correlation graph, and then calculates the importance score of each point based on the combination of the obtained attention weights and neighborhood features, selects the most representative point cloud subset, and realizes effective point cloud feature retention and efficient downsampling, and finally obtains the sampled point cloud data.

[0075] The structure of the AFDS layer is shown as follows Figure 2As shown in the figure, the first call to AFDS is used as an example. The input raw 3D point cloud data has the dimensions (B, 3, N), where B represents the batch size, 3 represents the 3D spatial coordinates of the point cloud, and N is the number of initial point clouds. The input point cloud is first mapped to a high-dimensional space through the embedding layer to improve the representation of spatial structure and geometric details. Subsequently, this layer constructs a local neighborhood set and calculates the difference between the features of the neighborhood points and the center point to capture subtle geometric changes in the local area of ​​the point cloud. A Transformer-based self-attention mechanism is then introduced, in which the center point feature is used as the query vector (Query), and the neighborhood feature differences are used to generate the key vector (Key) and value vector (Value) respectively. Finally, the deep correlation between the center point and the neighborhood points is calculated through the attention weight, and the normalized attention weight is generated through the Softmax function to dynamically evaluate the importance of each neighborhood point to the center point; finally, the importance score of each point is calculated based on the combination of the obtained attention weight and the neighborhood feature (that is, the local feature discrimination is evaluated based on the weight standard deviation), and the most representative point cloud subset (point cloud with high discrimination area) is accurately selected and retained through the Top-M operation to achieve effective point cloud feature preservation and efficient downsampling, and finally the sampled point cloud data (B, 3, M) is obtained.

[0076] This module reduces the number of points by 50% (2048 → 64) at each level of sampling. By dynamically sensing density and structural changes, it achieves refined preservation of key geometric features, overcoming the information loss issues associated with traditional random / furthest point sampling. In practice, the AFDS layer will be called multiple times in the Set Abstraction layer at different locations, and the values ​​of N and M will be adjusted to meet the sampling requirements of different levels.

[0077] In summary, the AFDS layer can achieve refined extraction and dynamic retention of key structural information of point clouds through the deep fusion of high-dimensional feature embedding and self-attention mechanism, overcome the shortcomings of traditional sampling methods that are difficult to adapt to local density differences and structural complexity, and enhance the network's feature expression ability and generalization performance for complex point cloud scenes.

[0078] The task of tree species identification requires detailed analysis of single tree point cloud data, and this type of data often has complex and highly diverse structural features, including multi-scale, irregular spatial forms such as trunks, branches and leaves. Traditional standard spatial convolution is suitable for regular grid structures such as images, and assumes that the relationship between adjacent pixels is fixed and uniform; however, this method is difficult to effectively adapt to the disorder of point cloud data and the non-uniform distribution characteristics of local density, resulting in its inability to accurately capture local geometric differences in point clouds. Specific to the task of tree species identification, due to limitations such as the laser radar scanning angle, distance differences, and occlusion of branches and leaves, there are usually significant density differences in various areas of the tree point cloud, which places higher demands on the model's ability to extract local features. Therefore, the design of the present invention is based on the difference graph convolution DiffConv layer, and its structure is shown in the figure below. Figure 3 The figure shows the complete convolutional feature extraction process of the Head DiffConv layer for point cloud data. First, this layer takes point cloud data of dimension (B, C, N) as input, where B is the batch size, C is the feature dimension of the point, and N is the number of points. By calculating the spatial Euclidean distance between all points in the input point cloud, a distance matrix of dimension (B, N, N) is obtained, which precisely describes the spatial relationships between points within the point cloud. This involves traversing each point sequentially and calculating the distance between it and all other points (including itself). The resulting output dimension is (B, N, N), effectively increasing the feature dimension from C to N. Subsequently, based on the density characteristics of the local area where each point resides, the radius is adaptively determined and stored in a (B, N, 1) matrix, enabling dynamic neighborhood queries and effectively addressing uneven point cloud density. After determining the dynamic neighborhood of each point, the spatial coordinate features of the point are further positionally encoded (the coordinate difference between the point and all points within the spherical query radius is calculated as the position encoding result) and mapped to a high-dimensional space, thereby enhancing the representation of spatial structural information. At the same time, a masked attention mechanism is introduced to dynamically weight the features of each center point and its neighboring points, generating a local smooth feature with the dimension (B, C, N). Local difference features are then obtained by calculating the feature difference between the center point feature and the encoding of its neighboring positions. These two types of features are concatenated and fused to form a more discriminative combined feature (B, C_out, N). Finally, the feature combination module outputs the final feature representation.

[0079] Overall, the Head DiffConv layer achieves efficient capture of multi-scale geometric features in point cloud data through dynamic density-adaptive neighborhood query, position encoding, and attention mechanisms, effectively improving feature extraction accuracy and model robustness, making it suitable for the complex point cloud analysis task of tree species identification.

[0080] DiffConv adopts a masked attention mechanism. In areas where local spatial density changes greatly, it dynamically selects and focuses on the most representative neighborhood key points, thereby flexibly capturing the local structural information of the point cloud, specifically strengthening the geometric features in the point cloud that are most important for classification, and improving adaptability to complex scenes. In the scenario of the present invention, DiffConv can adaptively adjust the selection weights of neighborhood points in the convolution process according to the local structural features of the single tree point cloud (such as the outline shape of the crown, the spatial distribution of branches, etc.). Unlike traditional convolution methods with fixed neighborhood size or radius, DiffConv does not rely on a fixed view, but uses a ball query method based on density expansion to dynamically adjust the neighborhood size of each point.

[0081] Most traditional point cloud deep learning networks choose to use maximum pooling to capture the most significant local features in the point cloud. However, this method only relies on the maximum response value in the local area, and is prone to ignoring less significant but still important local information for classification. In the task of tree species identification, the crowns, branches and other parts of different tree species often have similar overall shapes, but there are small but critical differences in fine-grained structural features (such as branch angles, leaf distribution, etc.). These subtle features may directly affect the accuracy of tree species identification. Therefore, a simple maximum pooling method may not be able to effectively retain these important local details, thereby reducing the recognition accuracy. Therefore, the present invention designs a FRP layer based on recycled pooling, and its structure is shown in the figure below. Figure 4 As shown in the figure, the feature recovery pooling layer adopts the iterative maximum pooling method. Through the three-layer feature recovery network, it gradually captures the less significant features discarded by the conventional pooling operation. Combined with the feature refinement loss function, it constructs a joint optimization goal of global features and local details, so that the model can accurately depict fine-grained features such as leaves while maintaining the overall morphological integrity of the tree crown.

[0082] The core idea of ​​the FRP layer is to dynamically recycle suboptimal but still informative features discarded by traditional max pooling. Specifically, first, traditional max pooling is used to extract salient feature information from the local area to generate an initial permutation-invariant feature vector; then, for the point feature matrix not selected by the first pooling, max pooling is repeated iteratively to continuously extract sub-salient features, forming multiple levels of permutation-invariant feature sets F1, F2, ..., F k ; By designing a hierarchical loss function, these recovered features are dynamically weighted to achieve detailed expression and efficient utilization of feature information, ultimately improving the richness and discriminative ability of the overall feature representation.

[0083] First, the input point cloud data undergoes multi-layer feature aggregation and convolution operations to extract point cloud structural information at different scales. The point features of each layer are aggregated through local neighborhoods and calculated through convolution to form a layer-by-layer refined point feature representation, providing rich feature input for subsequent pooling operations. Next, the feature recovery pooling module performs hierarchical pooling on the input features, first applying maximum pooling to extract the most significant features from the input feature matrix. It then performs Top-k selection on the remaining feature set, sequentially extracting less significant features to ensure that feature information at different levels is fully utilized. Finally, all pooled features are combined through the feature fusion module and optimized using the feature refinement loss.

[0084] By introducing the FRP module, the network is able to more comprehensively retain the subtle and key structural feature extraction results in tree species point cloud data, ultimately improving the classification accuracy and generalization performance of the tree species identification task.

[0085] Step 5. Perform data collection and preprocessing on the urban street trees to be identified according to steps 1-3. Then, input the preprocessed single tree point cloud data into the DASA-Net deep learning network trained in step 4 to obtain the quadruple results of the tree species.

[0086] Furthermore, because the point cloud data in the TreeNet dataset is constructed using virtual tree models rather than directly collected from the real environment, to ensure that its point cloud features more closely resemble those collected by mobile LiDAR, this paper employed HELIOS software for simulation. During this simulation, the scanning path and sensor parameters of the mobile LiDAR were configured, and a simulated scene of urban street trees was constructed to more realistically simulate the LiDAR scanning effect on trees during vehicle travel. Figure 5 The simulated point cloud data is displayed, with the red arrow indicating the vehicle's trajectory. This simulated point cloud data not only better matches the distribution characteristics of actual LiDAR point clouds, but also provides more possibilities for subsequent tree species identification tasks and biological parameter estimation.

[0087] Example 1

[0088] During training, the specific parameters of the DASA-Net network of the present invention are set as follows: the model uses ReLU as the activation function, and uses BatchNorm for batch normalization to improve stability. The sampling method of point cloud data adopts spherical query, the sampling radius is set to 0.15, the expansion multiple is 1.5, and 32 neighboring points are sampled each time to fully capture local geometric information. In terms of hyperparameters, the optimizer selects AdamW, the weight decay is set to 0.05, the initial learning rate is set to 0.001, and the cosine scheduling strategy is used for dynamic adjustment to improve the convergence efficiency and generalization ability of the model. The loss function adopts the cross entropy loss with label smoothing, and the smoothing coefficient is set to 0.2. At the same time, in order to ensure the stability of training, the gradient clipping threshold is set to 1. The batch size of the training data is 32, and a total of 600 epochs are trained. During the training process, the log is recorded every 10 rounds, and verification is performed once every round.

[0089] The present invention adopts overall classification accuracy (OA) and mean accuracy (mAc c) as evaluation indicators of recognition results.

[0090] Overall classification accuracy (OA) can directly evaluate the proportion of correct tree species identification, and its definition is shown in formula (9):

[0091]

[0092] Among them, TP is the number of samples predicted as positive, FP is the number of samples predicted as negative, and FN is the number of samples predicted as negative.

[0093] mAcc represents the mean of the recognition accuracy of all tree species, and its calculation formula is shown in formula (10):

[0094]

[0095] Among them, C is the total number of categories, TP i and FN i They represent the number of samples predicted correctly and the number of samples misclassified for category i, respectively.

[0096] After determining the evaluation indicators, training is carried out according to the set training parameters. The changes in loss, OA, and mAcc of the DASA-Net deep learning network during the 600 epoch training process are shown in the figure below. Figure 6 As shown, it can be seen that the model has reached convergence after 600 epochs.

[0097] In addition, in order to more comprehensively evaluate the performance of the DASA-Net deep learning network of the present invention, the classic model PointNet++ and the highly representative and high-performance point cloud classification networks DeepGCN and ASSANet networks were introduced for comparison. Training and testing were carried out on 6313 point cloud samples of 24 tree species. The final tree species recognition results are shown in Table 1, and the bold represents the best results.

[0098] Table 1 Tree species recognition results of DASA-Net model and comparison model

[0099]

[0100]

[0101] As shown in Table 1, the DASA-Net deep learning network demonstrated excellent performance in this tree species identification task, achieving an overall classification accuracy (OA) of 96.62% and an average precision (mAcc) of 96.58%, ranking first among all tested deep learning networks. It also outperformed the original PointNeXt network (95.79%, 95.88%), as well as ASSANet (95.25%, 95.21%) and DeepGCN (96.20%, 95.71%), two other leading point cloud classification networks. In particular, DASA-Net achieved improvements of 2-10% over ASSANet and DeepGCN for species such as camphor, quince, Douglas fir, European larch, and European sweet cherry.

[0102] It is worth noting that in the self-collected vehicle-mounted LiDAR dataset, the three tree species of Jacaranda, Metasequoia, and Ginkgo all achieved a classification accuracy of 100%, and the accuracy rate for Phoebe zhennan also reached 97.92%, which fully demonstrates its accuracy in distinguishing tree species characteristics and provides a solid foundation for subsequent biological parameter estimation work.

[0103] The DASA-Net deep learning network also demonstrated robust recognition capabilities for more difficult-to-classify tree species. For example, on the quince tree (96.00%), its accuracy was higher than ASSANet (95.00%) and DeepGCN (92.00%), while PointNet++ performed only 74.00%. On the beech (83.33%), its accuracy was close to PointNeXt (86.67%) and better than DeepGCN (84.44%). On the European spruce (90.20%), although lower than PointNeXt (94.12%), it was higher than ASSANet (86.27%) and PointNet++ (79.41%). These results demonstrate that DASA-Net can stably extract effective features and reduce misclassification even when tree species have complex morphologies and large variations in point cloud density.

[0104] The superior performance of the DASA-Net deep learning network of the present invention is mainly due to its optimized network architecture, including the attention mechanism for enhancing point cloud sampling, the difference graph convolution to optimize local feature extraction, and the feature recovery pooling module to improve the global feature expression capability. These improvements enhance the learning ability of DASA-Net on complex tree species point clouds, making it more discriminative in high-dimensional feature space. Therefore, in this tree species recognition task, DASA-Net not only achieved the highest overall accuracy, but also achieved optimal or equivalent performance in the recognition of multiple tree species, providing a more accurate and robust technical solution for large-scale urban tree point cloud recognition.

[0105] Compared to existing networks, the core of the proposed DASA-Net deep learning network lies in the design of the AFDS, DiffConv, and FRP layers. To verify the effectiveness of the proposed network design, ablation experiments were conducted on these three layers, with the results shown in Table 2.

[0106] Table 2

[0107]

[0108] As shown in Table 2, each improvement layer has different degrees of impact on the recognition accuracy of different tree species and improves the overall classification performance. First, AFDS, as a key module for optimizing point cloud sampling, improves the recognition accuracy of multiple tree species. For example, the recognition accuracy of small-leaved olive increased from 98.89% to 100.00%, that of European larch from 82.50% to 95.00%, and that of European sweet cherry from 85.71% to 92.86%. Although there was a certain decline in some categories, such as camphor (97.78% → 95.56%) and European spruce (94.12% → 86.27%), this may be related to the sampling method's incompatibility with the specific point cloud distribution. Overall, AFDS improved the overall classification accuracy (OA) from 95.79% to 96.15% and mAcc from 95.88% to 96.12%, demonstrating its advantage in retaining key features.

[0109] Head DiffConv further enhances the ability to extract local geometric features, improving the classification accuracy of the model on multiple tree species. Specifically, the recognition rate of Tibetan cherry increased from 98.89% to 100.00%, the Eurasian maple increased from 88.89% to 88.89%, and the European larch remained at 95.00%. The recognition rate of European Scots pine decreased slightly from 100.00% to 99.08%, and the Douglas fir decreased from 100.00% to 98.90%, possibly due to the module's different adaptability to point cloud features on some tree species. Overall, the introduction of Head DiffConv further improved OA to 96.47% and mAcc to 96.32%, indicating that it has played a positive role in improving local feature extraction capabilities.

[0110] As a key component of the optimized pooling strategy, FRP plays an important role in the final performance of the model. Compared with the Head DiffConv layer, FRP further improves the classification accuracy of multiple tree species, such as the Eurasian maple from 88.89% to 97.22%, the European spruce from 91.18% to 94.12%, the European sweet cherry from 92.86% to 89.29%, and the camphor tree from 97.92% unchanged. Although there is a slight decrease in some tree species, such as the rock oak (88.50% → 84.07%), FRP still maintains a high level of overall performance, maintaining OA at 96.36% and mAcc at 96.30%.

[0111] Overall, AFDS primarily improves the ability to retain key features in point cloud sampling, resulting in improved classification performance for most tree species. Head DiffConv further optimizes local feature extraction, enabling the network to more reliably identify different tree species. FRP further improves the classification accuracy of some complex tree species by optimizing the pooling strategy. Although the classification accuracy of individual tree species fluctuates, overall, DASA-Net's recognition performance across different tree species is more stable and accurate, enhancing the model's generalization capabilities and ensuring more reliable tree species recognition.

[0112] As can be seen from the above examples, the present invention first constructs a point cloud dataset for tree species identification tasks and performs standardized preprocessing to ensure data balance and network adaptability. Subsequently, comprehensive experiments were conducted on a point cloud dataset of 24 tree species. The results show that DASA-Net achieves 96.62% OA and 96.58% mAcc in tree species identification tasks, both outperforming the comparison networks PointNeXt, ASSANet, and DeepGCN. It achieves the best classification results in most tree species identification tasks. In particular, on a self-collected vehicle-mounted LiDAR dataset, it achieves 100% classification accuracy for Jacaranda, Metasequoia, and Ginkgo, and 97.92% accuracy for Phoebe zhennan. This fully demonstrates its accuracy in distinguishing tree species characteristics and provides a solid foundation for subsequent tree species identification work. In addition, module ablation experiments further verified the contribution of the three key layers of AFDS, Head DiffConv, and FRP to recognition performance, demonstrating that the proposed method effectively improves the DAS A-Net deep learning network's ability to distinguish different tree species morphologies.

[0113] The above description is only a specific embodiment of the present invention. Any feature disclosed in this specification, unless otherwise stated, can be replaced by other equivalent or alternative features with similar purposes; all disclosed features, or all steps in the methods or processes, except for mutually exclusive features and / or steps, can be combined in any way.

Claims

1. A point cloud tree species identification method based on DASA-Net deep learning network, characterized in that: The following steps are involved: Step 1. Use a vehicle-mounted LiDAR to perform multi-view scanning of urban street trees to obtain high-density 3D point cloud data containing reflection intensity information. Then, use a cloth simulation filtering algorithm to filter out the ground point cloud. Step 2. A tree segmentation algorithm based on point cloud geometric feature constraints, combined with a density-based spatial clustering algorithm and a principal direction indexing method, is used to separate complete tree point clouds from the scene point cloud obtained in Step 1. A tree species label is added to each individual tree point cloud file. The resulting labeled tree point clouds are then used to construct a single tree sample library containing several urban tree species. Step 3. Standardize and preprocess the single tree sample library: use farthest point sampling to unify the sampling, and eliminate scale differences through geometric centroid alignment and unit sphere normalization; divide the preprocessed data into training and test sets according to a certain ratio; Step 4. Build the DASA-Net deep learning network, use the data in the training set as training data, and the data in the test set as validation data, and train the DASA-Net deep learning network until the DASA-Net deep learning network converges; Step 5. Collect, separate, and preprocess the data of the urban street trees to be identified according to steps 1-3. Then input the preprocessed single tree point cloud data into the DASA-Net deep learning network trained in step 4 to obtain the tree species identification results.

2. The point cloud tree species identification method based on DASA-Net deep learning network according to claim 1, characterized in that: In step 4, the DASA-Net deep learning network includes a first multi-layer perceptron MLP, several layered sampling aggregation modules, a global pooling module and a connector connected in sequence; Among them, the stratified sampling aggregation module includes a feature aggregation unit and an inverse residual MLP unit connected in sequence; the feature aggregation unit in the first stratified sampling aggregation module includes a feature retention sampling layer based on the attention mechanism, a difference graph convolution layer, a feature grouping layer, a second multi-layer perceptron MLP and a recycling pooling layer connected in sequence; the feature aggregation units in the remaining stratified sampling aggregation modules only include a feature retention sampling layer based on the attention mechanism, a feature grouping layer, a third multi-layer perceptron MLP and a recycling pooling layer connected in sequence; The first multi-layer perceptron MLP performs initial feature embedding, the feature retention sampling layer based on the attention mechanism dynamically retains key points through attention weights, the difference graph convolution layer extracts local difference features through density-adaptive graph convolution, the recovery pooling layer recovers secondary significant features through iterative pooling, and the feature grouping layer is used to group the input features. The second multi-layer perceptron MLP, the third multi-layer perceptron MLP and the first multi-layer perceptron MLP have the same structure and are used for feature embedding; the global pooling module and the connector are used for global feature aggregation.

3. The point cloud tree species identification method based on DASA-Net deep learning network according to claim 2, characterized in that: The number of stratified sampling aggregation modules is 5.

4. The point cloud tree species identification method based on DASA-Net deep learning network according to claim 2, characterized in that: The specific process of the feature-preserving sampling layer of the attention mechanism is: The input point cloud is first mapped to a high-dimensional space through an embedding layer, and then a local neighborhood set is constructed to calculate the difference between the features of the neighborhood points and the center point. Based on the self-attention mechanism of Transformer, the center point feature is used as the query vector, and the neighborhood feature differences are used to generate key vectors and value vectors respectively. Finally, the deep correlation between the center point and the neighborhood points is calculated through the attention weight, and the normalized attention weight is generated through the Softmax function to dynamically evaluate the importance of each neighborhood point to the center point. Finally, the importance score of each point is calculated based on the combination of the obtained attention weight and the neighborhood features, and the most representative point cloud subset is selected and retained through the Top-M operation to finally obtain the sampled point cloud data.

5. The point cloud tree species identification method based on DASA-Net deep learning network according to claim 4, characterized in that: The specific process of calculating the deep correlation between the center point and the neighborhood points through attention weight is: For each point i in the point cloud, in its local neighborhood set S i The correlation measure within is defined as shown in formula (1): h l (p i ,p ij )=Q(p i ) T K(p ij -p i ) (1) Among them, Q and K represent linear mappings, p i is the feature of the center point, p ij is the feature of the points in the neighborhood, p ij -p i Indicates the feature difference between the neighborhood points and the center point; The specific calculation formula of normalization is shown in formula (2): Among them, d is the feature dimension.

6. The point cloud tree species identification method based on DASA-Net deep learning network according to claim 2, characterized in that: The specific working process of the difference graph convolution layer is: First, this layer takes point cloud data of dimension (B, C, N) as input, where B is the batch size, C is the feature dimension of the point, and N is the number of points. By calculating the spatial Euclidean distance between all points in the input point cloud, a distance matrix of dimension (B, N, N) is obtained. Then, according to the density characteristics of the local area where each point is located, the radius is adaptively determined and stored in a (B, N, 1) matrix. After determining the dynamic neighborhood of each point, the spatial coordinate features of the point are positionally encoded and mapped to a high-dimensional space. At the same time, the adaptive radius is processed using a masked attention mechanism, and the features of each center point and its neighborhood points are dynamically weighted to obtain a local smooth feature of dimension (B, C, N). Then, by calculating the feature difference between the center point feature and the position encoding of its neighborhood, the local difference feature is obtained. The local difference feature and the original feature are spliced ​​and fused to form a combined feature (B, C_out, N). Finally, the final feature expression is output through feature combination processing.

7. The point cloud tree species identification method based on DASA-Net deep learning network according to claim 6, characterized in that: The specific process of adaptive radius adjustment based on the density characteristics of the local area is as follows: DiffConv first uses Gaussian kernel density estimation to calculate the kernel density of each point. The formula is shown in formula (3): Among them, i represents the target point whose density is currently being calculated, j is the summation variable, N is the total number of points in the point cloud, h is the bandwidth parameter, ‖p i -p j ‖ represents the Euclidean distance between point pairs; In order to make the density of different points have a uniform scale, d i Normalized to [0,1] to get the normalized density Embed the normalized density into the radius calculation to obtain the density-driven search radius r for each point i , and its formula is shown in formula (4): Among them, r is the initial fixed sphere query radius; The specific process of shielding the attention mechanism for processing is: The learned irregular neighborhood selection is introduced; the attention weight of each point in the neighborhood is calculated by formula (5) and then normalized by formula (6):

8. The point cloud tree species identification method based on DASA-Net deep learning network according to claim 2, characterized in that: The specific working process of the feature recovery pooling layer is: First, traditional maximum pooling is used to extract salient feature information from the local area to generate an initial permutation-invariant feature vector. Then, for the point feature matrix not selected in the first pooling, maximum pooling is repeated iteratively to continuously extract secondary salient features, forming a multi-level permutation-invariant feature set. By designing a hierarchical loss function, these recycled features are dynamically weighted, ultimately improving the richness and discriminative ability of the overall feature representation; finally, the features obtained by maximum pooling and the features obtained by recycled pooling are combined to obtain the required aggregated features.

9. The point cloud tree species identification method based on DASA-Net deep learning network according to claim 8, characterized in that: Layer-wise loss function L total The calculation formula is as follows: Among them, λ is a hyperparameter, L c is the classification loss, which is composed of the prediction results of each permutation invariant feature set, and is expressed as formula (8): is the cross entropy loss function of layer I; L r is the feature refinement loss, which is calculated based on the predicted difference between the recovered features and the original maximum pooling features, and is expressed as follows: Among them, α i is an adjustable parameter, y m and They are the true label and the predicted probability based on the initial maximum pooling feature.

10. The point cloud tree species identification method based on DASA-Net deep learning network according to claim 1, characterized in that: In step 4, the DASA-Net deep learning network is trained using the AdamW optimizer. The initial learning rate is set to 0.001, and the cosine annealing strategy is adopted. The label smoothing cross entropy loss and feature refinement loss are combined to perform multi-objective optimization. The model is converged through multiple rounds of iterative training.

Images