Forest laser radar point cloud tree species classification method and system based on lightweight network

Abstract

The invention belongs to the field of tree species classification, and discloses a forest laser radar point cloud tree species classification method and system based on a lightweight network, and the method comprises the steps: carrying out the farthest point sampling of an input single tree point cloud; the local feature extraction module captures local geometric information and detail features in the point cloud of the single tree through a layered multi-scale feature extraction structure; the global feature extraction module processes a long-distance dependency relationship between features through an attention mechanism, and obtains global features of the point cloud of the single tree; combining the extracted local features and global features in a channel connection mode, and performing feature aggregation by using maximum pooling and average pooling; and outputting a tree species category by using a multi-layer sensor and a linear mapping layer. According to the method, lightweight design is adopted, and the tree species classification precision of the forest point cloud can be improved while relatively low parameter quantity and calculation complexity are kept.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of tree species classification technology, and particularly relates to a forest lidar point cloud tree species classification method and system based on a lightweight network. Background Technology

[0002] Accurate tree species information is crucial for assessing forest condition. Whether used as a standalone parameter or combined with other tree characteristics, it effectively reflects indicators such as biodiversity, disturbance levels, and reserves. Precise tree species information also provides a theoretical and practical basis for forest management planning and policy development. However, obtaining detailed tree species information remains a significant challenge. Forests are often widely distributed and located in remote or inaccessible areas, making field surveys time-consuming and labor-intensive. Furthermore, accurate tree species identification requires specialized knowledge. Many tree species share similar characteristics, making differentiation particularly difficult in dense or highly diverse forests. Seasonal variations, environmental variability, and high species diversity also increase the difficulty of assessment.

[0003] LiDAR (Light Detection and Ranging) can penetrate vegetation canopy to acquire high-precision three-dimensional structural information (such as tree height, canopy morphology, and branch distribution), which is crucial for distinguishing different tree species. Furthermore, LiDAR is unaffected by lighting conditions and can operate in all weather, overcoming the limitations of optical imagery in cloudy or low-light conditions. Compared to passive optical remote sensing (such as high-resolution imagery and hyper / multispectral imagery), LiDAR point clouds can achieve higher-precision segmentation at the single-tree scale, providing a more refined data foundation for tree species classification.

[0004] Currently, the most commonly used methods for tree species classification using LiDAR point clouds in forests can be summarized into two categories: those based on traditional machine learning models and those based on deep learning models. Traditional machine learning methods are typically represented by models such as K-Nearest Neighbor (KNN), Support Vector Machine (SVM), and Random Forest (RF). These methods use extracted individual tree parameters (such as structure, texture, and intensity) for tree species classification. These methods rely on domain knowledge and the stability of individual tree parameter extraction, and are usually conducted under simple experimental conditions (such as fewer than 5 tree species). The robustness of these methods still needs to be verified.

[0005] Existing deep learning methods can be divided into two categories: (1) indirect methods—converting single tree point clouds into image processing; and (2) direct methods—directly using point clouds as input. Indirect methods classify tree species by projecting point clouds, which is computationally efficient, but the two-dimensional conversion will lose three-dimensional structural information. Due to the unstructured nature of point clouds, the processes of neighborhood search, feature extraction, and spatial relationship modeling are computationally complex. Although direct methods avoid the loss of structural information, they have a high computational cost. Currently, most existing direct methods for tree species classification directly transfer computer vision models. Due to the high cost of forest point cloud processing, the sample size for tree species classification is often limited, which leads to overfitting problems when directly applying existing computer vision models. In addition, high classification accuracy is only applied to simple scenarios (such as the classification of coniferous and broad-leaved tree species), and the robustness and computational efficiency of the methods still need to be further improved.

[0006] Based on the above analysis, the urgent technical problems that need to be solved in the existing technology are:

[0007] Existing direct-method forest LiDAR point cloud tree species classification models are complex and computationally expensive. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention provides a method and system for classifying tree species in forest lidar point clouds based on lightweight networks.

[0009] This invention is implemented as follows: a forest lidar point cloud tree species classification method based on a lightweight network, characterized in that the method specifically includes:

[0010] S1: Perform farthest point sampling on the input single tree point cloud;

[0011] S2: The local feature extraction module captures local geometric information and detailed features in the point cloud of a single tree through a hierarchical multi-scale feature extraction structure;

[0012] S3: The global feature extraction module uses an attention mechanism to handle long-distance dependencies between features and obtains global features of a single tree point cloud;

[0013] S4: The extracted local and global features are combined using channel concatenation through the feature fusion module, and feature aggregation is performed using max pooling and average pooling.

[0014] S5: Use a multilayer perceptron and a linear mapping layer to output the tree species category.

[0015] Furthermore, in S1, given a point cloud of a single tree consisting of N points, its three-dimensional coordinates can be represented as P = [p1,...,p...]. i ,...,p N ]∈RN×3 Each point is represented as p i ∈R N×3 A subset is selected using the farthest point sampling method to represent the local center point.

[0016]

[0017] in, The selected point is represented by N, which represents the number of individual tree points, and M is the number of local center points.

[0018] Furthermore, the S2 local feature extraction module mainly includes KNN, MLP, and max-pooling:

[0019] (1) Select a subset using the farthest point sampling method to represent the local center points. The local neighborhood relations of each center point f are constructed using KNN.

[0020]

[0021] It is the relative position offset of the selected center point for constructing local neighborhood relations. The neighborhood relations are constructed using three scales: 5, 10, and 15.

[0022] (2) Processed by multiple stacked MLP layers An MLP layer consists of a convolutional (Conv) layer, a normalization (BN) layer, and a ReLU activation function.

[0023]

[0024] δ d″′ This represents the output of the MLP, where k is set to 5, 10, and 15, and the corresponding feature dimensions d are set to 16, 32, and 64, respectively. Each MLP is stacked three times.

[0025] (3) Apply the Max-Pooling(MaxP) function for aggregation.

[0026] v d =MaxP(δ d″′ (4)

[0027] Then, local features at different scales are combined by connecting channels.

[0028] E(d)=Concat{v d},d∈{16,32,64} (5)

[0029] High-dimensional features are combined with a subset of local centroids through residual connections.

[0030]

[0031] Output Feature E c After two MLP processing steps

[0032] T d1′ =ReLU(BN(Conv(Ε) c ) d1′ (7)

[0033] T d2′ =ReLU(BN(Conv(T) d1′ ) d2′ ))

[0034] The feature dimensions of d1 and d2 are 256 and 512, respectively.

[0035] Furthermore, in step S3, the global feature extraction module includes an MLP layer and an attention module. Input objects used for extracting global features:

[0036] (1) MLP for local center point Perform two processes.

[0037]

[0038] ρ d3″ =ReLU(BN(Conv(ρ) d3′ ) d3″ ))

[0039] The feature dimension d3 is set to 64;

[0040] (2) Use linear transformation to process ρ d3″ We obtain the query (Q), key (K), and value (V) matrix.

[0041] (Q,K,V)=Linear((ρ d3″ )·(W q W k W v )) d3 / 4 (9)

[0042] W q W k W v It is the weight matrix. Linear is a linear transformation operation, and the output dimension is set to d3 / 4(16).

[0043] (3) The attention weights are obtained by performing a dot product operation on the query matrix (Q) and the key matrix (K), and then normalized using the softmax function.

[0044] A = softmax(Q·K) T (10)

[0045] The output feature U is a weighted sum of the value vectors of the attention weights.

[0046] U = A·V (11)

[0047] The offset-attention layer is calculated as follows:

[0048]

[0049] X d3′ It is processed by stacking multiple attention modules.

[0050] {X d3″ ,X d3″′ ,X d3″″}={Attention(X d3′ ),Attention(X d3″ ),Attention(X d3″′ (13)

[0051] Global features are combined together through channel connections.

[0052] E X =Concat{Xd 3′ ,X d3″ ,X d3″′ ,X d3″″} (14)

[0053] Then, the output feature E is processed by MLP. X ,

[0054] ρ d4′ =ReLU(BN(Conv(EX)) d4′ (15)

[0055] The mapping dimension of d4 is set to 512.

[0056] Furthermore, in S4, the feature fusion module includes max pooling, average pooling, MLP, and a linear mapping layer, and local features T d2′ and global features ρ d4′ Global features F are obtained through channel connection. m :

[0057] F m =Concat{T d2′ ,ρ d4′} (16)

[0058] Feature aggregation employs max pooling and average pooling; the fusion process can be described as follows:

[0059] F c =MaxP(F m )+AveP(F m (17)

[0060] Furthermore, in step S5, the fused features are processed using an MLP and linear mapping layer to complete tree species classification.

[0061] L = Linear(ReLU(BN(Linear(F)) c ) d5 ))) c (18)

[0062] The dimension of d5 is 256, and c is the number of tree species categories.

[0063] Another objective of this invention is to provide a forest lidar point cloud tree species classification system based on a lightweight network, the system specifically comprising:

[0064] The sampling module is used to sample the farthest point of the input single tree point cloud;

[0065] The local feature extraction module is used to capture local geometric information and detailed features in the point cloud of a single tree;

[0066] The global feature extraction module is used to obtain the global features of the point cloud of a single tree.

[0067] The feature fusion module is used to fuse extracted local and global features;

[0068] The output module is used to output the tree species category.

[0069] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0070] First, to improve the ability to identify subtle differences between tree species, this invention utilizes a local-global collaborative feature extraction strategy. It employs a hierarchical multi-scale feature extraction architecture to progressively acquire local geometric details and semantic information, and uses a lightweight attention mechanism to perceive overall structural features. This invention adopts a lightweight design, maintaining relatively low parameter count and computational complexity while improving the tree species classification accuracy of forest point clouds, exhibiting the following significant advantages:

[0071] (1) Lightweight network: The network model has only 600,000 parameters and a computational cost as low as 0.3 GFLOPs (billion floating-point operations per second), which significantly improves computational efficiency compared to existing similar forest point cloud tree species classification models.

[0072] (2) Simplified process: The end-to-end network model architecture can directly process point clouds, effectively avoiding cumbersome preprocessing steps such as extraction of complex forest / tree parameters and data type conversion;

[0073] (3) Strong robustness: The network model was validated on a complex forest dataset containing 30 types of trees and compared with 7 cutting-edge methods, proving that the method proposed in this invention is more applicable to complex scenarios.

[0074] Secondly, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:

[0075] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0076] This invention proposes a method and system for classifying tree species using forest lidar point clouds based on lightweight networks, enabling automatic, efficient, and accurate tree species classification. The technical solution of this invention not only significantly improves the accuracy and computational efficiency of tree species identification but also provides reliable technical support for dynamic monitoring of forest resources, biodiversity assessment, and carbon sequestration. This invention has significant application value in fields such as ecological environmental protection, smart forestry management, and the achievement of carbon neutrality goals. Its industrial application will effectively promote the digital transformation of forestry and the development of a green economy, with broad market prospects.

[0077] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0078] In the field of lidar remote sensing and forestry application research both domestically and internationally, there are currently no publicly available research reports on automatic tree species classification based on lightweight deep learning networks for forest lidar point clouds. The innovation of this invention lies in its first proposal and implementation of a lightweight network architecture for forest lidar point cloud tree species classification. Through optimized network structure and innovative feature learning methods, it effectively solves the bottleneck problems of computational efficiency and classification accuracy inherent in traditional methods. This technological breakthrough not only fills a technological gap in this field but also provides an innovative solution for forest resource remote sensing monitoring that combines high efficiency, accuracy, and practicality.

[0079] (3) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully:

[0080] This invention solves a long-standing technical challenge in the field of forestry remote sensing—how to achieve efficient and accurate automatic tree species classification from forest LiDAR point clouds. Traditional methods generally face the dilemma of high computational complexity and low classification accuracy when processing massive point cloud data. The breakthrough innovation of this technical solution lies in: the pioneering construction of a lightweight deep network architecture tailored to the characteristics of forest point clouds, which significantly improves computational efficiency while maintaining high accuracy through computational simplification; and the innovative design of a lightweight local-global feature collaborative fusion mechanism, effectively extracting multi-level geometric features of different tree species in point clouds. This series of technological innovations not only fills a technological gap in this field but also provides reliable technical support for the development of smart forestry.

[0081] (4) The technical solution of the present invention overcomes technical bias:

[0082] This invention breaks through the long-standing technical paradigm limitations in the field of LiDAR point cloud tree species classification. A common technical bias in traditional research methods is the simple application or modification of existing models from computer vision, relying excessively on complex deep network architectures. These methods often fail to meet the practical needs of forestry applications due to their high computational complexity and large number of parameters. Addressing the characteristics of large forest point cloud data volumes and complex scenarios, this invention creatively designs a novel lightweight network architecture. This technical solution fundamentally overcomes the inherent shortcomings of traditional methods, such as high computational resource consumption and high deployment difficulty, providing a more efficient and practical solution for forestry remote sensing applications. Attached Figure Description

[0083] Figure 1 This is a flowchart of a forest lidar point cloud tree species classification method based on a lightweight network provided in an embodiment of the present invention;

[0084] Figure 2 This is the overall design diagram of the forest lidar point cloud tree species classification network architecture based on a lightweight network provided in the embodiments of the present invention;

[0085] Figure 3 This is a module diagram of a forest lidar point cloud tree species classification system based on a lightweight network, provided in an embodiment of the present invention.

[0086] Figure 4 This is a tree species classification confusion matrix diagram provided in an embodiment of the present invention.

[0087] Figure 5 This is a graph showing the precision and recall rates for various tree species provided in this embodiment of the invention.

[0088] Figure 6 This is a diagram illustrating the tree species classification effect of a forest plot provided in an embodiment of the present invention. Detailed Implementation

[0089] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0090] like Figure 1 As shown, this embodiment of the invention provides a method for classifying tree species in forest lidar point clouds based on lightweight networks. The method specifically includes:

[0091] S1: Perform farthest point sampling on the input single tree point cloud;

[0092] S2: The local feature extraction module captures local geometric information and detailed features in the point cloud of a single tree through a hierarchical multi-scale feature extraction structure;

[0093] S3: The global feature extraction module uses an attention mechanism to handle long-distance dependencies between features and obtains global features of a single tree point cloud;

[0094] S4: The extracted local and global features are combined using channel concatenation through the feature fusion module, and feature aggregation is performed using max pooling and average pooling.

[0095] S5: Use a multilayer perceptron (MLP) and a linear mapping layer to output tree species categories.

[0096] S1 involves sampling the farthest point of the input single-tree point cloud for two purposes: first, to set a center point in a local region to further construct local features; and second, to reduce the computational complexity and number of parameters required to extract global features. Given a single-tree point cloud consisting of N points, its three-dimensional coordinates can be represented as P = [p1,...,p...]. i ,...,p N ]∈R N×3 Each point is represented as p i ∈R N×3 A subset is selected using the farthest point sampling method to represent the local center point.

[0097]

[0098] in, The selected point is represented by N, which represents the number of individual tree points, and M is the number of local center points.

[0099] In step S2, the local feature extraction module effectively captures local geometric information and detailed features in the point cloud of a single tree through a hierarchical multi-scale feature extraction structure, adapting to the complexity and differences of different tree species. The local feature extraction module mainly includes KNN, MLP, and Max-Pooling. KNN enhances the network's ability to perceive local details by establishing the relationship between each local center point and its neighboring points. MLP extracts abstract and rich features to capture complex nonlinear relationships, enhancing the network's ability to understand complex geometric shapes. Max-Pooling extracts the extreme values ​​of local features to reduce redundant information.

[0100] The local neighborhood relationships of each center point f are constructed using KNN.

[0101]

[0102] This is the relative position offset of the selected center point when constructing local neighborhood relationships. Three scales are used to construct the neighborhood relationships: 5, 10, and 15.

[0103] Then, it is processed by multiple stacked MLP layers. An MLP layer consists of a convolutional (Conv) layer, a normalization (BN) layer, and a ReLU activation function.

[0104]

[0105] δ d″′ This represents the output of the MLP, where k is set to 5, 10, and 15, and the corresponding feature dimensions d are set to 16, 32, and 64, respectively. For example... Figure 2 As shown, each MLP is stacked three times. Aggregation is performed using the Max-Pooling (MaxP) function.

[0106] v d =MaxP(δ d″′ (4)

[0107] Then, local features at different scales are combined by connecting channels.

[0108] E(d)=Concat{υ d},d∈{16,32,64} (5)

[0109] By combining high-dimensional features with a subset of local centroids through residual connections, we can better capture the spatial distribution patterns of point clouds in individual trees and prevent the loss of geometric structure in deep features.

[0110]

[0111] Output Feature E cAfter two MLP processing steps

[0112] T d1′ =ReLU(BN(Conv(Ε) c ) d1′ (7)

[0113] T d2′ =ReLU(BN(Conv(T) d1′ ) d2′ ))

[0114] The feature dimensions of d1 and d2 are 256 and 512, respectively.

[0115] In step S3, the global feature extraction module uses an attention mechanism to handle long-distance dependencies between features, enhancing its ability to perceive global geometric structures and thus effectively acquiring global features of a single tree point cloud. The global feature extraction module includes an MLP layer and an attention module. Used as input for extracting global features. First, the MLP processes local center points. Perform two processes.

[0116]

[0117] ρ d3″ =ReLU(BN(Conv(ρ) d3′ )d 3″ ))

[0118] The feature dimension d3 is set to 64. Then, a linear transformation is used to process ρ. d3″ We obtain the query (Q), key (K), and value (V) matrix.

[0119]

[0120] W q W k W v It is the weight matrix, and Linear is a linear transformation operation. To improve computational efficiency, the output dimension is set to d3 / 4 (16).

[0121] Attention weights are obtained by performing a dot product operation on the query matrix (Q) and the key matrix (K), and then normalized using the softmax function.

[0122] A = softmax(Q·K) T (10)

[0123] The output feature U is a weighted sum of the value vectors of the attention weights.

[0124] U = A·V (11)

[0125] The offset-attention layer is calculated as follows:

[0126]

[0127] like Figure 2 As shown, X d3′ It is processed by stacking multiple attention modules.

[0128] {X d3″ ,X d3″′ ,X d3″″}={Attention(X d3′ ),Attention(X d3″ ),Attention(X d3″′ (13)

[0129] Global features are combined together through channel connections.

[0130] E X =Concat{X d3′ ,X d3″ ,X d3″′ ,X d3″″} (14)

[0131] Then, the output feature E is processed by MLP. X ,

[0132] ρ d4′ =ReLU(BN(Conv(E) X ) d4′ (15)

[0133] The mapping dimension of d4 is set to 512.

[0134] The S4 feature fusion module includes max pooling, average pooling, MLP, and a linear mapping layer. Local features T d2′ and global features ρ d4′ Global features F are obtained through channel connection. m :

[0135] F m =Concat{T d2′ ,ρ d4′} (16)

[0136] Feature aggregation employs max pooling and average pooling; the fusion process can be described as follows:

[0137] F c =MaxP(F m )+AveP(F m (17)

[0138] In step S5, the fused features are processed by an MLP and a linear mapping layer to complete the tree species classification. The process is summarized as follows:

[0139] L = Linear(ReLU(BN(Linear(F)) c ) d5 ))) c (18)

[0140] The dimension of d5 is 256, and c is the number of tree species categories.

[0141] like Figure 3 As shown in the figure, an embodiment of the present invention provides a forest lidar point cloud tree species classification system based on a lightweight network, specifically including:

[0142] The sampling module is used to sample the farthest point of the input single tree point cloud;

[0143] The local feature extraction module is used to capture local geometric information and detailed features in the point cloud of a single tree;

[0144] The global feature extraction module is used to obtain the global features of the point cloud of a single tree.

[0145] The feature fusion module is used to fuse extracted local and global features;

[0146] The output module is used to output the tree species category.

[0147] The lightweight network model proposed in this invention addresses the tree species classification needs of forest LiDAR point clouds and can be widely applied in fields such as smart forestry management and ecological research. In smart forestry, this model can be integrated into drones or ground-based LiDAR devices to achieve real-time tree species identification, assisting in precise silvicultural planning and biomass estimation, such as in portable forestry inspection terminals or cloud-based forest resource assessment platforms. In ecological research, the model's high-precision tree species identification can improve the accuracy of forest carbon sink assessment. Leveraging its lightweight advantage, the model can be efficiently deployed on edge devices (such as drone-borne / ground mobile platform computers), promoting the application of low-cost, high-frequency dynamic forest monitoring.

[0148] II. Evidence related to the technical effects obtained by the embodiments of the present invention.

[0149] The forest data used for testing and verification in this invention includes 30 tree species and 12,682 trees, most of which come from temperate forests. Table 1 summarizes the information for each tree species. All data were collected using terrestrial LiDAR, and the point clouds of individual trees were downsampled to 1024, 2048, and 4096 points, respectively.

[0150] Table 1 Tree Species Information

[0151]

[0152]

[0153] The lightweight network model proposed in this invention demonstrates excellent performance in forest plot tree species classification, with classification results as follows: Figure 6 As shown in the figure, different colors clearly indicate the distribution of different tree species, intuitively demonstrating the model's excellent classification ability. It is worth noting that this model can not only accurately identify the distribution characteristics of dominant trees (i.e., large trees in the forest), but also effectively capture the species information of understory saplings and sub-canopy trees, achieving a detailed analysis of the complex vertical structure of forests. This high-precision identification capability of understory vegetation types provides more comprehensive and accurate support for forest resource surveys.

[0154] When the input single tree point cloud contains 1024, 2048, and 4096 points, the overall accuracy (OA) of this invention for classifying 30 types of trees is 71.8%, 73.7%, and 74.0%, respectively, and the average accuracy (AA) is 53.0%, 55.9%, and 56.2%, respectively. The tree classification confusion matrix for the 30 types of trees, and the precision and recall for each type of tree are shown below. Figure 4 and Figure 5 As shown.

[0155] Table 2 shows a comparison between this invention and seven cutting-edge methods. This invention exhibits the highest average accuracy across all input numbers (1024, 2048, 4096), indicating stable and balanced classification performance for all tree species (including those difficult to classify). Compared to other cutting-edge methods, this invention demonstrates the highest overall accuracy in most cases, with the sole exception of 4096 input points, where PointNet++ slightly outperforms this invention, with an overall accuracy difference of only 1.2%.

[0156] Table 2 compares the tree species classification performance under different numbers of input points. The highest values ​​of the evaluation indicators are shown in bold. OA and AA represent the overall accuracy and average accuracy, respectively.

[0157]

[0158]

[0159] Table 3 compares the computational cost of this invention with other cutting-edge methods when the number of input points is 1024. The results show that this invention has the lowest computational cost, with 600,000 parameters (Million, M) and 0.3 giga (G) floating-point operations per second (FLOPs), even lower than the existing lightweight network Point-PN, making it suitable for resource-constrained applications. PointMLP has the highest number of parameters and significantly higher FLOPs than other methods, at 13.2M and 31.3G respectively; the most popular PointNet++ also has higher parameter counts and FLOPs, at 1.7M and 4.1G FLOPs. Overall, this invention maintains high classification performance while reducing computational cost, making it suitable for resource-constrained tasks with high accuracy requirements.

[0160] Table 3 compares the computational costs using 1024 points as input; the lowest values ​​for the evaluation metrics are shown in bold.

[0161]

[0162] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0163] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A forest lidar point cloud tree species classification method based on lightweight networks, characterized in that, The method includes: S1, sample the farthest point of the input single tree point cloud and select a representative center point; S2, construct multi-scale features for the local neighborhood of the center point using the local feature extraction module, process them through a multilayer perceptron and perform max pooling to generate local features; S3 utilizes a global feature extraction module to introduce an attention mechanism, models long-distance dependencies between point clouds, and outputs global features. S4, integrates local and global features, uses channel connection to stitch features together, and generates fused features by aggregating them through max pooling and average pooling; S5 outputs tree species classification results after the fused features are processed by a multilayer perceptron and linear mapping.

2. The method as described in claim 1, characterized in that, The farthest point sampling operation is based on the input point cloud set. Execute and output M center points. The condition M is less than N, and the center point is used to construct a local neighborhood.

3. The method as described in claim 1, characterized in that, The local feature extraction module includes: The K-nearest neighbor algorithm is used to construct local neighborhoods of multiple scales at each center point; Features are extracted using a three-layer stacked multilayer perceptron at each scale, with the feature dimensions increasing sequentially. Features at different scales are combined through channel splicing and connected with residual information at the center point. The output is processed by another multilayer perceptron to form a local feature tensor.

4. The method as described in claim 1, characterized in that, The global feature extraction module includes: By inputting the center point set into a two-layer multilayer perceptron, the basic feature representation is obtained; Construct a query, key, and value matrix, and execute a dot product attention mechanism; Global feature enhancement is achieved by stacking multiple attention modules to form an offset-attention network. The output features of multiple attention modules are concatenated and mapped to the final global features through a multilayer perceptron.

5. The method as described in claim 1, characterized in that, The feature fusion module performs the following steps: Concatenate local and global features along the channel dimension; Perform max pooling and average pooling operations respectively; The sum of the two results is used as the fusion feature F_c.

6. The method as described in claim 1, characterized in that, The classification output module consists of a multilayer perceptron and a linear mapping layer. The former maps the fused features to an intermediate dimension, while the latter outputs them as tree species category labels.

7. A forest lidar point cloud tree species classification system based on a lightweight network, characterized in that, The system includes: The sampling module is used to receive and sample the farthest point of the input single tree point cloud; The local feature extraction module is used to extract local features through multi-scale K-nearest neighbors and multilayer perceptrons; The global feature extraction module is used to generate global features through an attention mechanism; The feature fusion module is used to concatenate and aggregate local features with global feature channels into a fused feature; The output module is used to output the corresponding tree species category based on the fused features.

8. The system as described in claim 7, characterized in that, The local feature extraction module is configured with three scales of K-nearest neighbor structures, namely 5, 10 and 15, to capture geometric details under different receptive fields.

9. The system as described in claim 7, characterized in that, The global feature extraction module is configured to process through a triple-stacked attention mechanism module. Each module uses a linear transformation to construct a query, key, and value matrix and performs a normalized dot product.

10. The system as described in claim 7, characterized in that, The output module includes a multilayer perceptron, a normalization module, and a linear classifier, and the output dimension is consistent with the set number of tree species.

Images