Tree species fine classification method based on nested graph convolutional network

Abstract

The application discloses a tree species fine classification method based on a nested graph convolution network, and the method comprises the following steps: acquiring a hyperspectral image and point cloud data of a target tree area, respectively performing preprocessing to obtain a target hyperspectral image superpixel set and a target point cloud data set; obtaining a hyperspectral local graph according to the target hyperspectral image superpixel set, obtaining a point cloud local graph according to the target point cloud data set, and respectively performing feature extraction to obtain hyperspectral local features and point cloud local features; performing feature fusion on the hyperspectral local features and the point cloud local features to obtain global graph node features, and obtaining a global graph according to the global graph node features; performing adaptive node feature aggregation to obtain target node features, and performing classification processing on the target node features to obtain a tree species classification result. The application realizes the interaction of global context information and local detail information while improving data utilization, and realizes effective aggregation between global graph node features.

Description

Technical Field

[0001] This invention relates to the field of tree species classification technology, and in particular to a method, system, terminal, and storage medium for fine tree species classification based on nested graph convolutional networks. Background Technology

[0002] Precise tree species classification is fundamental to forestry resource research and utilization. Spatial distribution information of forest tree species provides basic information for research on biodiversity conservation and invasive alien species, and is of great significance for macro-monitoring of forest ecosystems and biodiversity. Tree species information is an important component of forestry resource surveys. Precise tree species classification can help formulate more accurate logging quotas and reasonable afforestation plans, thereby maintaining the sustainable supply of forest resources and ecological balance, and improving the quality and stability of ecosystems.

[0003] Currently, existing methods for fine-grained tree species classification, such as those based on hyperspectral images, LiDAR data, and cross-modal observation data, still have some problems. The feature coupling between hyperspectral images and LiDAR data is weak, and the interaction between global and local information is poor, making it impossible to effectively interact and fuse global and local features, thus affecting the classification results. There is also insufficient correlation between global features, which limits the understanding of the overall scene and makes it difficult to capture subtle differences between tree species.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a fine-grained tree species classification method based on nested graph convolutional networks. This method aims to address the problems in existing technologies, such as weak feature coupling between hyperspectral images and LiDAR data, poor interaction between global and local information, which prevents effective interaction and fusion between global and local features, and insufficient correlation between global features, resulting in poor tree species classification performance.

[0006] To achieve the above objectives, this invention provides a method for fine-grained tree species classification based on nested graph convolutional networks. The method includes the following steps:

[0007] Acquire hyperspectral images and point cloud data of the target tree region, and preprocess the hyperspectral images and point cloud data respectively to obtain the target hyperspectral image superpixel set and the target point cloud dataset;

[0008] A hyperspectral local map is obtained from the superpixel set of the target hyperspectral image, and a point cloud local map is obtained from the target point cloud dataset. Feature extraction is performed on the hyperspectral local map and the point cloud local map respectively to obtain hyperspectral local features and point cloud local features.

[0009] The hyperspectral local features and the point cloud local features are fused to obtain global map node features, and a global map of the target tree region is obtained based on the global map node features.

[0010] Adaptive node feature aggregation is performed on the global graph to obtain target node features, and the target node features are then classified to obtain tree species classification results.

[0011] Optionally, the tree species fine classification method based on nested graph convolutional networks, wherein acquiring the hyperspectral image and point cloud data of the target tree region, and preprocessing the hyperspectral image and the point cloud data respectively to obtain the target hyperspectral image superpixel set and the target point cloud dataset, specifically includes:

[0012] Acquire hyperspectral images and point cloud data of the target tree region, perform superpixel segmentation on the hyperspectral images to obtain multiple hyperspectral image superpixel blocks, and obtain a hyperspectral image superpixel set based on all the hyperspectral image superpixel blocks;

[0013] The location information of each hyperspectral image superpixel in the hyperspectral image superpixel set is obtained. The point cloud data is divided according to the location information to obtain multiple point cloud data subsets. A point cloud dataset is obtained based on all the point cloud data subsets.

[0014] Masking and elevation constraint processing are performed on the hyperspectral image superpixel set and the point cloud dataset, respectively, to obtain the target hyperspectral image superpixel set and the target point cloud dataset.

[0015] Optionally, the step of obtaining a hyperspectral local map based on the superpixel set of the target hyperspectral image, obtaining a point cloud local map based on the target point cloud dataset, and extracting features from the hyperspectral local map and the point cloud local map respectively to obtain hyperspectral local features and point cloud local features specifically includes:

[0016] Based on the hyperspectral superpixel set of the target hyperspectral image, a first vertex set and a first node feature matrix of the hyperspectral local map are constructed. Then, based on the first node feature matrix, spectral angle mapping is calculated between each node of the first vertex set to obtain the spectral similarity result.

[0017] Based on the spectral similarity results, a spectral angle similarity matrix is ​​obtained. The spectral angle similarity matrix is ​​then processed using the K-nearest neighbor algorithm to obtain the first edge set and the first adjacency matrix of the hyperspectral local graph. Finally, the hyperspectral local graph is obtained based on the first vertex set and the first edge set.

[0018] Based on the point cloud subset of the target point cloud dataset, a second vertex set and a second node feature matrix of the point cloud local graph are constructed, and spatial Euclidean distance between each node of the second vertex set is calculated based on the second node feature matrix to obtain the distance calculation result.

[0019] The spatial Euclidean distance matrix is ​​obtained based on the distance calculation results, and the spatial Euclidean distance matrix is ​​processed according to the K-nearest neighbor algorithm to obtain the second edge set and the second adjacency matrix of the point cloud local graph. The point cloud local graph is obtained based on the second vertex set and the second edge set.

[0020] Feature extraction is performed on the hyperspectral local image using simple graph convolution to obtain hyperspectral local features, and feature extraction is also performed on the point cloud local image using simple graph convolution to obtain point cloud local features.

[0021] Optionally, the tree species fine classification method based on nested graph convolutional networks, wherein obtaining a hyperspectral local map based on the target hyperspectral image superpixel set and obtaining a point cloud local map based on the target point cloud dataset, further includes:

[0022] Perform a union operation on the first adjacency matrix and the second adjacency matrix to obtain the target adjacency matrix of the global graph.

[0023] Optionally, the tree species fine classification method based on nested graph convolutional networks, wherein fusing the hyperspectral local features and the point cloud local features to obtain global graph node features, and obtaining a global map of the target tree region based on the global graph node features, specifically includes:

[0024] Global average pooling is performed on the hyperspectral local features and the point cloud local features respectively to obtain the hyperspectral local feature matrix and the point cloud local feature matrix. Max pooling and average pooling are then performed on the hyperspectral local feature matrix and the point cloud local feature matrix respectively to obtain the target hyperspectral local feature matrix and the target point cloud local feature matrix.

[0025] The attention weight distribution is calculated on the target hyperspectral local feature matrix and the target point cloud local feature matrix to obtain multiple attention weights. All the attention weights are weighted and summed to obtain a summation result. The summation result is then normalized to obtain a first attention coefficient and a second attention coefficient.

[0026] The first attention coefficient and the hyperspectral local features are combined to obtain the target hyperspectral local features. The second attention coefficient and the point cloud local features are combined to obtain the target point cloud local features. The target hyperspectral local features and the target point cloud local features are combined to obtain the global graph node features.

[0027] The global graph node features and the target adjacency matrix are transformed to obtain the third vertex set and the third edge set, and the global graph of the target tree region is obtained based on the third vertex set and the third edge set.

[0028] Optionally, the tree species fine classification method based on nested graph convolutional networks, wherein the adaptive node feature aggregation of the global graph to obtain target node features, and the classification processing of the target node features to obtain tree species classification results, specifically includes:

[0029] The global graph feature matrix is ​​obtained based on the global graph node features. Attention coefficients are then calculated on the global graph feature matrix according to the first formula and the second formula, respectively, to obtain the Pearson correlation attention coefficient and the graph attention coefficient.

[0030] The Pearson correlation attention coefficients and the global graph feature matrix are combined to obtain the Pearson attention coefficient matrix, and the graph attention coefficients and the global graph feature matrix are combined to obtain the graph attention coefficient matrix.

[0031] The Pearson attention coefficient matrix and the graph attention coefficient matrix are combined to obtain the target attention matrix. The attention coefficients of the target attention matrix are then calculated according to the third formula to obtain the target attention coefficients.

[0032] The target attention coefficient is combined with the global graph node features to obtain the target node features. The target node features are then classified using a graph convolutional neural network classifier to obtain the tree species classification result.

[0033] Optionally, in the tree species fine classification method based on nested graph convolutional networks, the first formula is:

[0034]

[0035] The second formula is:

[0036]

[0037] The third formula is:

[0038]

[0039] in, Let Pearson correlation attention coefficient be the feature of the i-th global graph node. Let σ be the graph attention coefficient of the i-th global graph node feature, and σ be the activation function. For the i-th global graph node feature, Let $\frac{\pi}{\pi}$ be the feature of the $j$-th global graph node, $\pi$ be the learnable parameter, $\pi$ be the Pearson correlation coefficient between global node features, $N(i)$ be the set of neighboring nodes of the $i$-th node, and $FC$ be the fully connected layer. Let be the target attention coefficient, exp be the natural exponential function, and LeakyReLU be the activation function. Let be the target attention matrix for the features of the i-th and j-th global graph nodes. Let i be the target attention matrix for the i-th and k-th global graph node features, where i, j, and k are the number of global node features.

[0040] Optionally, the tree species fine classification method based on nested graph convolutional networks, wherein the tree species fine classification system based on nested graph convolutional networks includes:

[0041] The data acquisition module is used to acquire hyperspectral images and point cloud data of the target tree area, and preprocess the hyperspectral images and point cloud data respectively to obtain the target hyperspectral image superpixel set and the target point cloud dataset.

[0042] The feature extraction module is used to obtain a hyperspectral local map based on the superpixel set of the target hyperspectral image, and a point cloud local map based on the target point cloud dataset, and to extract features from the hyperspectral local map and the point cloud local map respectively to obtain hyperspectral local features and point cloud local features;

[0043] The feature fusion module is used to fuse the hyperspectral local features and the point cloud local features to obtain global map node features, and to obtain a global map of the target tree region based on the global map node features;

[0044] The tree species classification module is used to perform adaptive node feature aggregation on the global graph to obtain target node features, and to classify the target node features to obtain tree species classification results.

[0045] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps of the tree species fine classification method based on nested graph convolutional networks as described above.

[0046] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a tree species fine classification program based on nested graph convolutional networks, and the tree species fine classification program based on nested graph convolutional networks, when executed by a processor, implements the steps of the tree species fine classification method based on nested graph convolutional networks as described above.

[0047] In this invention, hyperspectral images and point cloud data of a target tree region are acquired. The hyperspectral images and point cloud data are preprocessed to obtain a target hyperspectral image superpixel set and a target point cloud dataset. A hyperspectral local map is obtained from the target hyperspectral image superpixel set, and a point cloud local map is obtained from the target point cloud dataset. Feature extraction is performed on both the hyperspectral local map and the point cloud local map to obtain hyperspectral local features and point cloud local features. The hyperspectral local features and the point cloud local features are fused to obtain global map node features, and a global map of the target tree region is obtained based on the global map node features. Adaptive node feature aggregation is performed on the global map to obtain target node features, and the target node features are classified to obtain a tree species classification result. This invention enables the three-dimensional representation of graph-structured data from hyperspectral images and LiDAR point cloud data. It constructs superpixel segmentation local maps based on pixels and point clouds, respectively, and embeds these local maps into the global map. This improves data utilization while enabling the interaction between global contextual information and local detail information. Furthermore, it employs a channel attention fusion method for cross-modal data local map features, highlighting and amplifying useful information in each channel while reducing interference from irrelevant information. This improves the accuracy of cross-modal feature extraction and fusion. Additionally, it constructs a global graph attention mechanism that incorporates Pearson correlation coefficients to better capture key feature information of nodes in the global graph structure, achieving more effective aggregation between global graph node features. Attached Figure Description

[0048] Figure 1 This is a flowchart of a preferred embodiment of the tree species fine classification method based on nested graph convolutional networks in this invention;

[0049] Figure 2 This is an overall schematic diagram of the cross-modal embedding graph convolutional network of a preferred embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram illustrating the principle of the cross-modal feature attention fusion module in a preferred embodiment of the present invention;

[0051] Figure 4 This is a schematic diagram illustrating the principle of the global graph correlation aggregation module based on Pearson correlation coefficient in a preferred embodiment of the present invention.

[0052] Figure 5This is a schematic diagram illustrating the principle of a preferred embodiment of the tree species fine classification system based on nested graph convolutional networks in this invention.

[0053] Figure 6 This is a schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0055] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0056] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0057] The preferred embodiment of the present invention describes a tree species fine classification method based on nested graph convolutional networks, such as... Figure 1 As shown, the tree species fine classification method based on nested graph convolutional networks includes the following steps:

[0058] Step S10: Obtain hyperspectral images and point cloud data of the target tree region, and preprocess the hyperspectral images and point cloud data respectively to obtain the target hyperspectral image superpixel set and the target point cloud dataset.

[0059] Step S10 includes:

[0060] Step S11: Obtain hyperspectral images and point cloud data of the target tree region, perform superpixel segmentation on the hyperspectral images to obtain multiple hyperspectral image superpixel blocks, and obtain a hyperspectral image superpixel set based on all the hyperspectral image superpixel blocks;

[0061] Step S12: Obtain the position information of each hyperspectral image superpixel in the hyperspectral image superpixel set, divide the point cloud data according to the position information to obtain multiple point cloud data subsets, and obtain a point cloud dataset based on all the point cloud data subsets.

[0062] Step S13: Perform masking and elevation constraint processing on the hyperspectral image superpixel set and the point cloud dataset respectively to obtain the target hyperspectral image superpixel set and the target point cloud dataset.

[0063] Specifically, traditional machine learning methods cannot fully extract the information contained in HSI (Hyperspectral Image) and the structural information of each tree species in LiDAR (Light Detection and Ranging) data, thus failing to effectively interact and fuse global and local features, resulting in limited improvement in classification accuracy. Furthermore, they are primarily used for feature extraction from hyperspectral images and point cloud data, lacking understanding of the inherent coupling between features across modal data. Insufficient correlation between global features also limits the model's understanding of the overall scene and makes it difficult to capture subtle differences between tree species. This invention implements a Cross-Modal Nested Graph Convolutional Network (CmNGCN) using hyperspectral images and point cloud data, with the specific structure as follows: Figure 2 As shown, this method aims to better accomplish the task of fine classification of tree species. The specific steps are: acquiring hyperspectral images and point cloud data of the target tree region, and defining the hyperspectral images and point cloud data as follows: and Among them, X H For hyperspectral data, X L For point cloud data, Let H, W, and B be the height, width, and number of bands of the hyperspectral image, respectively, and N and C be the number of points and features of the point cloud data, respectively. During image construction, the SNIC (Simple Non-Iterative Clustering) superpixel segmentation method is used to convert the pixel-level hyperspectral image into a superpixel-level hyperspectral image. Then, the superpixel-level hyperspectral image is segmented to obtain multiple hyperspectral image superpixel blocks. Based on all these hyperspectral image superpixel blocks, a hyperspectral image superpixel set is obtained. Let H be the set of superpixels of the hyperspectral image, and h be the value of the superpixel set. u For the i-th hyperspectral image superpixel block, m i Let be the number of pixels contained in the i-th hyperspectral image superpixel, and s be the total number of superpixel segments. Next, matching points are selected from the point cloud data based on the geographic location information of each hyperspectral image superpixel block in the hyperspectral image superpixel set to form a point cloud dataset. That is, the point cloud data is divided into an equal number of point cloud datasets based on geographic location information. In this representation, L is the point cloud dataset, and l i Represents the i-th subset of point cloud data, n i This represents the number of points contained in the i-th point cloud dataset.

[0064] In this embodiment of the invention, considering that the target of the research task is trees, NDVI (Normalized Difference Vegetation Index) is also used to mask non-vegetation areas, making the cross-modal embedding graph convolutional network more focused on learning the features of various tree species. Furthermore, since the target tree area also contains many grasslands, elevation constraints are added to exclude hyperspectral image superpixel blocks and point cloud data subsets that meet the NDVI conditions but have a height of less than 1 meter. Finally, the target hyperspectral image superpixel set can be obtained, using H... M This represents the target point cloud dataset, represented by L. M express.

[0065] Step S20: Obtain a hyperspectral local map based on the superpixel set of the target hyperspectral image, obtain a point cloud local map based on the target point cloud dataset, and extract features from the hyperspectral local map and the point cloud local map respectively to obtain hyperspectral local features and point cloud local features.

[0066] Step S20 includes:

[0067] Step S21: Construct a first vertex set and a first node feature matrix of a hyperspectral local map based on the hyperspectral superpixel set of the target hyperspectral image, and perform spectral angle mapping calculation between each node of the first vertex set based on the first node feature matrix to obtain the spectral similarity result;

[0068] Step S22: Obtain the spectral angle similarity matrix based on the spectral similarity results, process the spectral angle similarity matrix according to the K-nearest neighbor algorithm to obtain the first edge set and the first adjacency matrix of the hyperspectral local graph, and obtain the hyperspectral local graph based on the first vertex set and the first edge set;

[0069] Step S23: Construct a second vertex set and a second node feature matrix of the local point cloud map based on the point cloud subset of the target point cloud dataset, and calculate the spatial Euclidean distance between each node in the second vertex set based on the second node feature matrix to obtain the distance calculation result;

[0070] Step S24: Obtain the spatial Euclidean distance matrix based on the distance calculation result, process the spatial Euclidean distance matrix according to the K nearest neighbor algorithm to obtain the second edge set and the second adjacency matrix of the point cloud local graph, and obtain the point cloud local graph based on the second vertex set and the second edge set;

[0071] Step S25: Extract features from the hyperspectral local image using simple graph convolution to obtain hyperspectral local features, and extract features from the point cloud local image using simple graph convolution to obtain point cloud local features.

[0072] Specifically, in this embodiment of the invention, the nested graph structure in the cross-modal embedded graph convolutional network is mainly implemented by nesting local graphs within a global graph, wherein the local graphs are respectively composed of H M and L M Composition, and use of H M To construct local atlases of hyperspectral images in, For the i-th hyperspectral image, s ′ The number of local maps. and These are the vertex set and edge set of the local hyperspectral image graph, respectively. Specifically, the hyperspectral superpixels of the target hyperspectral image superpixel set are used as nodes of the local hyperspectral image graph, forming the vertex set of the local hyperspectral image graph, i.e., the first vertex set. The representation, and the node feature matrix, that is, the first node feature matrix, are expressed as follows: It means that, among them, Let m be the j-th vertex of the local map of the hyperspectral image. i The first vertex set contains the number of vertices. Then, using spectral angle mapping as a metric between hyperspectral nodes, spectral angle mapping is calculated for each node in the first vertex set based on the first node feature matrix to obtain the spectral similarity result.

[0073]

[0074] in, The smaller the value, the more similar the spectra of the two nodes are, and the greater the likelihood that they belong to the same tree species. Let be the transpose matrix of the node features of the j-th vertex. Given the node feature matrix of the k-th vertex, the spectral angle similarity matrix is ​​obtained based on the spectral similarity results. The KNN algorithm (K-Nearest Neighbors) is used to determine the neighboring nodes of each node, thus obtaining the edge set of the hyperspectral local graph, i.e., the first edge set. Let and the corresponding adjacency matrix, i.e., the first adjacency matrix, be represented by . This indicates that the hyperspectral local map is obtained based on the first vertex set and the first edge set. Similarly, using L... M To construct a local atlas of point cloud data in, For the i-th point cloud local image, and These are the vertex set and edge set of the local point cloud graph, respectively. Specifically, a subset of the target point cloud dataset is used as the nodes of the local point cloud graph to obtain the vertex set of the local point cloud graph, i.e., the second vertex set. This represents the corresponding node feature matrix, i.e., the second node feature matrix, expressed as... It means that, among them, Let n be the j-th vertex of the local point cloud graph. i The number of vertices is used; then, spatial Euclidean distance is used to measure the correlation between each node in the local point cloud image to obtain the distance calculation results.

[0075]

[0076] Among them, (x j ,y j ,z j ) and (x k ,y k ,z k ) respectively represent and The spatial location is used to obtain the spatial Euclidean distance matrix based on the distance calculation results. Similarly, by using the KNN algorithm, the edge set of the local point cloud graph, i.e., the second edge set, can be constructed from the spatial Euclidean distance matrix. This represents the adjacency matrix, i.e., the second adjacency matrix, denoted by . This indicates that the local map of the point cloud is obtained based on the second vertex set and the second edge set.

[0077] After obtaining the hyperspectral local map and the point cloud local map, SGC (Simple Graph Convolution) is used as a feature extractor to extract local features from different data sources. Taking the hyperspectral local map as an example, the SGC operation for extracting the corresponding local map features can be represented as:

[0078]

[0079] in, For hyperspectral local features, To add a self-looping adjacency matrix, Let k be the number of neighboring nodes that a node can aggregate, and σ and Θ be the activation function and learnable parameters, respectively. Similarly, the local features corresponding to the local map of the point cloud, i.e., the local features of the point cloud, can also be obtained. express.

[0080] Furthermore, in this embodiment of the invention, a bridge is established between global and local information by nesting a global graph with a local graph. The adjacency matrices of the hyperspectral local graph and the point cloud local graph, i.e., the union of the first and second adjacency matrices, are taken as the adjacency matrix of the global graph, i.e., the target adjacency matrix. A is used as the target adjacency matrix. C This can be expressed by the formula:

[0081] A C =KNN(D H )∪KNN(D L ); where D H and D L The purpose of obtaining the target adjacency matrix is ​​to use the spectral angular similarity matrix and the spatial Euclidean distance matrix, respectively, to obtain the global graph by combining it with the node features of the fused global graph.

[0082] Step S30: Perform feature fusion on the hyperspectral local features and the point cloud local features to obtain global map node features, and obtain a global map of the target tree region based on the global map node features.

[0083] Step S30 includes:

[0084] Step S31: Perform global average pooling on the hyperspectral local features and the point cloud local features respectively to obtain the hyperspectral local feature matrix and the point cloud local feature matrix. Then, perform max pooling and average pooling on the hyperspectral local feature matrix and the point cloud local feature matrix respectively to obtain the target hyperspectral local feature matrix and the target point cloud local feature matrix.

[0085] Step S32: Calculate the attention weight distribution of the target hyperspectral local feature matrix and the target point cloud local feature matrix to obtain multiple attention weights. Sum all the attention weights by weighted summation to obtain a summation result. Normalize the summation result to obtain the first attention coefficient and the second attention coefficient.

[0086] Step S33: Combine the first attention coefficient and the hyperspectral local features to obtain the target hyperspectral local features; combine the second attention coefficient and the point cloud local features to obtain the target point cloud local features; and combine the target hyperspectral local features and the target point cloud local features to obtain the global graph node features.

[0087] Step S34: Transform the global graph node features and the target adjacency matrix to obtain the third vertex set and the third edge set, and obtain the global graph of the target tree region based on the third vertex set and the third edge set.

[0088] Specifically, after obtaining the hyperspectral local features and the point cloud local features, global average pooling is performed on the hyperspectral local features and the point cloud local features respectively to obtain the hyperspectral local feature matrix, which is then used with F... h The representation, and the local feature matrix of the point cloud, are expressed using F. l In this embodiment of the invention, inspired by the channel attention mechanism, a cross-modal feature attention fusion module (CmFM) is proposed. This module adaptively assigns weights to local features from different data sources through attention learning to achieve cross-modal data fusion. The specific process is as follows: Figure 3 As shown, firstly, the hyperspectral local feature matrix and the point cloud local feature matrix are subjected to max pooling and average pooling to obtain the target hyperspectral local feature matrix and the target point cloud local feature matrix. Then, the pooled target hyperspectral local feature matrix and the target point cloud local feature matrix are passed through two fully connected layers with nonlinear activation functions to calculate the attention weight distribution of the features. That is, the attention weight distribution of the target hyperspectral local feature matrix and the target point cloud local feature matrix is ​​calculated to obtain multiple attention weights. All the attention weights are weighted and summed to obtain the summation result. Then, the summation result is normalized by the softmax function to obtain the first attention coefficient and the second attention coefficient. The corresponding calculation formula is as follows:

[0089]

[0090] Where, α h α is the first attention coefficient. l The second attention coefficient is denoted by , and softmax is the normalized exponential function. This is the hyperspectral local feature matrix after average pooling. This is the local feature matrix of the point cloud after average pooling. This is the hyperspectral local feature matrix after max pooling. The result is the local feature matrix of the point cloud after max pooling. Finally, the attention coefficients of different modalities are combined with their corresponding features to obtain cross-modal local features, which are then embedded into the global graph as global graph node features. Fi is used to calculate these features. CThis means that the first attention coefficient and the hyperspectral local features are combined to obtain the target hyperspectral local features; the second attention coefficient and the point cloud local features are combined to obtain the target point cloud local features; and the target hyperspectral local features and the target point cloud local features are combined to obtain the global graph node features. This can be expressed by the formula: F C =α h ·F h +α l ·F l The target adjacency matrix A obtained from the above calculations. C This allows the construction of a global map of the study area, using G. C ={V C E C} indicates that, where V C For the corresponding vertex set, namely the third vertex set, The corresponding edge set, namely the third edge set, This can be obtained by converting between the target adjacency matrix and the target adjacency matrix, where, Let s be the j-th vertex of the global graph. i The number of vertices.

[0091] Step S40: Perform adaptive node feature aggregation on the global graph to obtain target node features, and classify the target node features to obtain tree species classification results.

[0092] Step S40 includes:

[0093] Step S41: Obtain the global graph feature matrix based on the global graph node features of the global graph, and calculate the attention coefficients of the global graph feature matrix according to the first formula and the second formula respectively to obtain the Pearson correlation attention coefficient and the graph attention coefficient;

[0094] Step S42: Combine the Pearson correlation attention coefficients and the global graph feature matrix to obtain the Pearson attention coefficient matrix, and combine the graph attention coefficients and the global graph feature matrix to obtain the graph attention coefficient matrix;

[0095] Step S43: Combine the Pearson attention coefficient matrix and the graph attention coefficient matrix to obtain the target attention matrix. Calculate the attention coefficients of the target attention matrix according to the third formula to obtain the target attention coefficients.

[0096] Step S44: Combine the target attention coefficient with the global graph node features to obtain target node features, and classify the target node features according to the graph convolutional neural network classifier to obtain tree species classification results.

[0097] Specifically, in this embodiment of the invention, the cross-modal embedding graph convolutional network, based on the traditional GAT (Graph Attention Network), incorporates the Pearson correlation coefficient between features to implement the Pearson Graph Attention mechanism PRGAT (Pearson Relation Graph Attention, a global graph correlation aggregation module based on the Pearson correlation coefficient) for aggregating global node features. Its structure is as follows: Figure 4 As shown, specifically, unlike traditional GAT, after inputting the node and its neighboring nodes into a fully connected layer with shared weights, PRGAT does not directly map the feature vectors to real numbers as attention weights between the node and its neighboring nodes. Instead, it integrates the Pearson correlation coefficient between the features of the node and its neighboring nodes. Therefore, based on the global graph node features of the global graph, a global graph feature matrix is ​​obtained. Attention coefficients are then calculated on the global graph feature matrix according to the first and second formulas to obtain the Pearson correlation attention coefficient and the graph attention coefficient. The first formula is:

[0098]

[0099] The second formula is:

[0100]

[0101] in, Let Pearson correlation attention coefficient be the feature of the i-th global graph node. Let σ be the graph attention coefficient of the i-th global graph node feature, and σ be the activation function. For the i-th global graph node feature, Let $\frac{j}{j}$ be the global graph node feature, $\frac{Θ}{p}$ be the learnable parameter, $\frac{ρ}{p}$ be the Pearson correlation coefficient between global node features, $N(i)$ be the set of neighboring nodes of the $i$-th node, and $FC$ be the fully connected layer. Then, the Pearson correlation attention coefficient and the global graph feature matrix are combined to obtain the Pearson attention coefficient matrix. The graph attention coefficient matrix is ​​then combined with the global graph feature matrix to obtain the graph attention coefficient matrix. Finally, the target attention matrix is ​​obtained by combining the Pearson attention coefficient matrix and the graph attention coefficient matrix. The attention coefficients of the target attention matrix are calculated according to the third formula to obtain the target attention coefficients. The third formula is:

[0102] in, Let be the target attention coefficient, exp be the natural exponential function, and LeakyReLU be the activation function. Let be the target attention matrix for the features of the i-th and j-th global graph nodes. Let the target attention matrix be the feature matrix of the i-th and k-th global graph nodes; then, combine the target attention coefficients with the global graph node features to obtain the target node features, and use... This can be expressed as a formula:

[0103] In the final stage of classifier selection, the cross-modal embedding graph convolutional network does not use the MLP (Multilayer Perceptron) as the classifier, which is used in other classification networks. Instead, it uses GCN (Graph Convolutional Network) as the classification module. The target node features are classified using the graph convolutional neural network classifier to obtain the tree species classification result. This is because MLP only considers the attributes of the nodes themselves, ignoring the connections between nodes, while GCN is better able to capture the relationships between nodes and can effectively utilize the structural information of the global graph. Furthermore, GCN avoids computationally complex fully connected layers, reducing the network's computational cost and improving classification efficiency while maintaining accuracy.

[0104] This invention enables the three-dimensional representation of graph-structured data from hyperspectral images and LiDAR point cloud data. It constructs superpixel segmentation local maps based on pixels and point clouds, respectively, and embeds these local maps into the global map. This improves data utilization while enabling the interaction between global contextual information and local detail information. Furthermore, it employs a channel attention fusion method for cross-modal data local map features, highlighting and amplifying useful information in each channel while reducing interference from irrelevant information. This improves the accuracy of cross-modal feature extraction and fusion. Additionally, it constructs a global graph attention mechanism that incorporates Pearson correlation coefficients to better capture key feature information of nodes in the global graph structure, achieving more effective aggregation between global graph node features.

[0105] Furthermore, such as Figure 5 As shown, based on the above-mentioned tree species fine classification method based on nested graph convolutional networks, the present invention also provides a tree species fine classification system based on nested graph convolutional networks, the tree species fine classification system based on nested graph convolutional networks comprising:

[0106] Data acquisition module 51 is used to acquire hyperspectral images and point cloud data of the target tree area, and preprocess the hyperspectral images and point cloud data respectively to obtain target hyperspectral image superpixel set and target point cloud dataset;

[0107] The feature extraction module 52 is used to obtain a hyperspectral local map based on the superpixel set of the target hyperspectral image, and to obtain a point cloud local map based on the target point cloud dataset, and to extract features from the hyperspectral local map and the point cloud local map respectively to obtain hyperspectral local features and point cloud local features;

[0108] The feature fusion module 53 is used to fuse the hyperspectral local features and the point cloud local features to obtain global map node features, and to obtain a global map of the target tree region based on the global map node features.

[0109] The tree species classification module 54 is used to perform adaptive node feature aggregation on the global graph to obtain target node features, and to classify the target node features to obtain tree species classification results.

[0110] Furthermore, such as Figure 6 As shown, based on the above-mentioned tree species fine classification method based on nested graph convolutional networks, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 6 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0111] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a tree species fine classification program 40 based on nested graph convolutional networks, which can be executed by the processor 10 to implement the tree species fine classification method based on nested graph convolutional networks in this application.

[0112] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the tree species fine classification method based on nested graph convolutional networks.

[0113] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.

[0114] In one embodiment, when processor 10 executes program 40 in memory 20 for fine-grained tree classification based on nested graph convolutional networks, the following steps are performed:

[0115] Acquire typhoon trajectory data, identify and process the typhoon trajectory data to obtain target time information and target typhoon intensity information;

[0116] Acquire hyperspectral images and point cloud data of the target tree region, and preprocess the hyperspectral images and point cloud data respectively to obtain the target hyperspectral image superpixel set and the target point cloud dataset;

[0117] A hyperspectral local map is obtained from the superpixel set of the target hyperspectral image, and a point cloud local map is obtained from the target point cloud dataset. Feature extraction is performed on the hyperspectral local map and the point cloud local map respectively to obtain hyperspectral local features and point cloud local features.

[0118] The hyperspectral local features and the point cloud local features are fused to obtain global map node features, and a global map of the target tree region is obtained based on the global map node features.

[0119] Adaptive node feature aggregation is performed on the global graph to obtain target node features, and the target node features are then classified to obtain tree species classification results.

[0120] The step of acquiring hyperspectral images and point cloud data of the target tree region, and preprocessing the hyperspectral images and point cloud data to obtain a target hyperspectral image superpixel set and a target point cloud dataset, specifically includes:

[0121] Acquire hyperspectral images and point cloud data of the target tree region, perform superpixel segmentation on the hyperspectral images to obtain multiple hyperspectral image superpixel blocks, and obtain a hyperspectral image superpixel set based on all the hyperspectral image superpixel blocks;

[0122] The location information of each hyperspectral image superpixel in the hyperspectral image superpixel set is obtained. The point cloud data is divided according to the location information to obtain multiple point cloud data subsets. A point cloud dataset is obtained based on all the point cloud data subsets.

[0123] Masking and elevation constraint processing are performed on the hyperspectral image superpixel set and the point cloud dataset, respectively, to obtain the target hyperspectral image superpixel set and the target point cloud dataset.

[0124] Specifically, the process of obtaining a hyperspectral local map based on the superpixel set of the target hyperspectral image, obtaining a point cloud local map based on the target point cloud dataset, and extracting features from both the hyperspectral local map and the point cloud local map to obtain hyperspectral local features and point cloud local features includes:

[0125] Based on the hyperspectral superpixel set of the target hyperspectral image, a first vertex set and a first node feature matrix of the hyperspectral local map are constructed. Then, based on the first node feature matrix, spectral angle mapping is calculated between each node of the first vertex set to obtain the spectral similarity result.

[0126] Based on the spectral similarity results, a spectral angle similarity matrix is ​​obtained. The spectral angle similarity matrix is ​​then processed using the K-nearest neighbor algorithm to obtain the first edge set and the first adjacency matrix of the hyperspectral local graph. Finally, the hyperspectral local graph is obtained based on the first vertex set and the first edge set.

[0127] Based on the point cloud subset of the target point cloud dataset, a second vertex set and a second node feature matrix of the point cloud local graph are constructed, and spatial Euclidean distance between each node of the second vertex set is calculated based on the second node feature matrix to obtain the distance calculation result.

[0128] The spatial Euclidean distance matrix is ​​obtained based on the distance calculation results, and the spatial Euclidean distance matrix is ​​processed according to the K-nearest neighbor algorithm to obtain the second edge set and the second adjacency matrix of the point cloud local graph. The point cloud local graph is obtained based on the second vertex set and the second edge set.

[0129] Feature extraction is performed on the hyperspectral local image using simple graph convolution to obtain hyperspectral local features, and feature extraction is also performed on the point cloud local image using simple graph convolution to obtain point cloud local features.

[0130] The process of obtaining a hyperspectral local map based on the superpixel set of the target hyperspectral image and a point cloud local map based on the target point cloud dataset further includes:

[0131] Perform a union operation on the first adjacency matrix and the second adjacency matrix to obtain the target adjacency matrix of the global graph.

[0132] The step of fusing the hyperspectral local features and the point cloud local features to obtain global map node features, and obtaining a global map of the target tree region based on the global map node features, specifically includes:

[0133] Global average pooling is performed on the hyperspectral local features and the point cloud local features respectively to obtain the hyperspectral local feature matrix and the point cloud local feature matrix. Max pooling and average pooling are then performed on the hyperspectral local feature matrix and the point cloud local feature matrix respectively to obtain the target hyperspectral local feature matrix and the target point cloud local feature matrix.

[0134] The attention weight distribution is calculated on the target hyperspectral local feature matrix and the target point cloud local feature matrix to obtain multiple attention weights. All the attention weights are weighted and summed to obtain a summation result. The summation result is then normalized to obtain a first attention coefficient and a second attention coefficient.

[0135] The first attention coefficient and the hyperspectral local features are combined to obtain the target hyperspectral local features. The second attention coefficient and the point cloud local features are combined to obtain the target point cloud local features. The target hyperspectral local features and the target point cloud local features are combined to obtain the global graph node features.

[0136] The global graph node features and the target adjacency matrix are transformed to obtain the third vertex set and the third edge set, and the global graph of the target tree region is obtained based on the third vertex set and the third edge set.

[0137] Specifically, the process of adaptively aggregating node features in the global graph to obtain target node features, and then classifying the target node features to obtain tree species classification results, includes:

[0138] The global graph feature matrix is ​​obtained based on the global graph node features. Attention coefficients are then calculated on the global graph feature matrix according to the first formula and the second formula, respectively, to obtain the Pearson correlation attention coefficient and the graph attention coefficient.

[0139] The Pearson correlation attention coefficients and the global graph feature matrix are combined to obtain the Pearson attention coefficient matrix, and the graph attention coefficients and the global graph feature matrix are combined to obtain the graph attention coefficient matrix.

[0140] The Pearson attention coefficient matrix and the graph attention coefficient matrix are combined to obtain the target attention matrix. The attention coefficients of the target attention matrix are then calculated according to the third formula to obtain the target attention coefficients.

[0141] The target attention coefficient is combined with the global graph node features to obtain the target node features. The target node features are then classified using a graph convolutional neural network classifier to obtain the tree species classification result.

[0142] The first formula is:

[0143]

[0144] The second formula is:

[0145]

[0146] The third formula is:

[0147]

[0148] in, Let Pearson correlation attention coefficient be the feature of the i-th global graph node. Let σ be the graph attention coefficient of the i-th global graph node feature, and σ be the activation function. For the i-th global graph node feature, Let $\frac{\pi}{\pi}$ be the feature of the $j$-th global graph node, $\pi$ be the learnable parameter, $\pi$ be the Pearson correlation coefficient between global node features, $N(i)$ be the set of neighboring nodes of the $i$-th node, and $FC$ be the fully connected layer. Let be the target attention coefficient, exp be the natural exponential function, and LeakyReLU be the activation function. Let be the target attention matrix for the features of the i-th and j-th global graph nodes. Let i be the target attention matrix for the i-th and k-th global graph node features, where i, j, and k are the number of global node features.

[0149] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a tree species fine classification program based on nested graph convolutional networks, and the tree species fine classification program based on nested graph convolutional networks, when executed by a processor, implements the steps of the tree species fine classification method based on nested graph convolutional networks as described above.

[0150] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0151] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0152] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for fine-grained tree classification based on nested graph convolutional networks, characterized in that, The tree species fine classification method based on nested graph convolutional networks includes: Acquire hyperspectral images and point cloud data of the target tree region, and preprocess the hyperspectral images and point cloud data respectively to obtain the target hyperspectral image superpixel set and the target point cloud dataset; A hyperspectral local map is obtained from the superpixel set of the target hyperspectral image, and a point cloud local map is obtained from the target point cloud dataset. Feature extraction is performed on the hyperspectral local map and the point cloud local map respectively to obtain hyperspectral local features and point cloud local features. A hyperspectral local map is obtained from the superpixel set of the target hyperspectral image, and a point cloud local map is obtained from the target point cloud dataset. Feature extraction is then performed on both the hyperspectral local map and the point cloud local map to obtain hyperspectral local features and point cloud local features, specifically including: The nested graph structure in the cross-modal embedding graph convolutional network is implemented by nesting a global graph with a local graph. The first vertex set and the first node feature matrix of the hyperspectral local graph are constructed based on the hyperspectral superpixel set of the target hyperspectral image. The spectral angle mapping between each node of the first vertex set is calculated based on the first node feature matrix to obtain the spectral similarity result. Based on the spectral similarity results, a spectral angle similarity matrix is ​​obtained. The spectral angle similarity matrix is ​​then processed using the K-nearest neighbor algorithm to obtain the first edge set and the first adjacency matrix of the hyperspectral local graph. Finally, the hyperspectral local graph is obtained based on the first vertex set and the first edge set. Based on the point cloud subset of the target point cloud dataset, a second vertex set and a second node feature matrix of the point cloud local graph are constructed, and spatial Euclidean distance between each node of the second vertex set is calculated based on the second node feature matrix to obtain the distance calculation result. The spatial Euclidean distance matrix is ​​obtained based on the distance calculation results, and the spatial Euclidean distance matrix is ​​processed according to the K-nearest neighbor algorithm to obtain the second edge set and the second adjacency matrix of the point cloud local graph. The point cloud local graph is obtained based on the second vertex set and the second edge set. Feature extraction is performed on the hyperspectral local map using simple graph convolution to obtain hyperspectral local features, and feature extraction is performed on the point cloud local map using simple graph convolution to obtain point cloud local features; The hyperspectral local features and the point cloud local features are fused to obtain global map node features, and a global map of the target tree region is obtained based on the global map node features. Adaptive node feature aggregation is performed on the global graph to obtain target node features, and the target node features are then classified to obtain tree species classification results.

2. The tree species fine classification method based on nested graph convolutional networks according to claim 1, characterized in that, The process of acquiring hyperspectral images and point cloud data of the target tree region, and preprocessing the hyperspectral images and point cloud data to obtain a target hyperspectral image superpixel set and a target point cloud dataset, specifically includes: Acquire hyperspectral images and point cloud data of the target tree region, perform superpixel segmentation on the hyperspectral images to obtain multiple hyperspectral image superpixel blocks, and obtain a hyperspectral image superpixel set based on all the hyperspectral image superpixel blocks; The location information of each hyperspectral image superpixel in the hyperspectral image superpixel set is obtained. The point cloud data is divided according to the location information to obtain multiple point cloud data subsets. A point cloud dataset is obtained based on all the point cloud data subsets. Masking and elevation constraint processing are performed on the hyperspectral image superpixel set and the point cloud dataset, respectively, to obtain the target hyperspectral image superpixel set and the target point cloud dataset.

3. The tree species fine classification method based on nested graph convolutional networks according to claim 1, characterized in that, The process of obtaining a hyperspectral local map based on the superpixel set of the target hyperspectral image and a point cloud local map based on the target point cloud dataset further includes: Perform a union operation on the first adjacency matrix and the second adjacency matrix to obtain the target adjacency matrix of the global graph.

4. The tree species fine classification method based on nested graph convolutional networks according to claim 3, characterized in that, The step of fusing the hyperspectral local features and the point cloud local features to obtain global map node features, and obtaining a global map of the target tree region based on the global map node features, specifically includes: Global average pooling is performed on the hyperspectral local features and the point cloud local features respectively to obtain the hyperspectral local feature matrix and the point cloud local feature matrix. Max pooling and average pooling are then performed on the hyperspectral local feature matrix and the point cloud local feature matrix respectively to obtain the target hyperspectral local feature matrix and the target point cloud local feature matrix. The attention weight distribution is calculated on the target hyperspectral local feature matrix and the target point cloud local feature matrix to obtain multiple attention weights. All the attention weights are weighted and summed to obtain a summation result. The summation result is then normalized to obtain a first attention coefficient and a second attention coefficient. The first attention coefficient and the hyperspectral local features are combined to obtain the target hyperspectral local features. The second attention coefficient and the point cloud local features are combined to obtain the target point cloud local features. The target hyperspectral local features and the target point cloud local features are combined to obtain the global graph node features. The global graph node features and the target adjacency matrix are transformed to obtain the third vertex set and the third edge set, and the global graph of the target tree region is obtained based on the third vertex set and the third edge set.

5. The tree species fine classification method based on nested graph convolutional networks according to claim 1, characterized in that, The adaptive node feature aggregation of the global graph to obtain target node features, and the classification processing of the target node features to obtain tree species classification results, specifically include: The global graph feature matrix is ​​obtained based on the global graph node features. Attention coefficients are then calculated on the global graph feature matrix according to the first formula and the second formula to obtain the Pearson attention coefficient and the graph attention coefficient. The Pearson attention coefficients and the global graph feature matrix are combined to obtain the Pearson attention coefficient matrix, and the graph attention coefficients and the global graph feature matrix are combined to obtain the graph attention coefficient matrix. The Pearson attention coefficient matrix and the graph attention coefficient matrix are combined to obtain the target attention matrix. The attention coefficients of the target attention matrix are then calculated according to the third formula to obtain the target attention coefficients. The target attention coefficient is combined with the global graph node features to obtain the target node features. The target node features are then classified using a graph convolutional neural network classifier to obtain the tree species classification result. The first formula is: ； The second formula is: ； The third formula is: ； in, For the first Pearson attention coefficients for features of global graph nodes. For the first Graph attention coefficients of global graph node features For activation function, For the first Features of global graph nodes For the first Features of global graph nodes For learnable parameters, The Pearson coefficient is the ratio between global node features. For the first The set of neighboring nodes of a node. It is a fully connected layer. For the target attention coefficient, It is a natural exponential function. For activation function, For the first The and the first The target attention matrix for global graph node features. For the first The and the first The target attention matrix for global graph node features. , and All of these represent the number of global node features.

6. A tree species fine classification system based on nested graph convolutional networks, characterized in that, The tree species fine classification system based on nested graph convolutional networks is used to implement the tree species fine classification method based on nested graph convolutional networks as described in any one of claims 1-5. The tree species fine classification system based on nested graph convolutional networks includes: The data acquisition module is used to acquire hyperspectral images and point cloud data of the target tree area, and preprocess the hyperspectral images and point cloud data respectively to obtain the target hyperspectral image superpixel set and the target point cloud dataset. The feature extraction module is used to obtain a hyperspectral local map based on the superpixel set of the target hyperspectral image, and a point cloud local map based on the target point cloud dataset, and to extract features from the hyperspectral local map and the point cloud local map respectively to obtain hyperspectral local features and point cloud local features; The feature fusion module is used to fuse the hyperspectral local features and the point cloud local features to obtain global map node features, and to obtain a global map of the target tree region based on the global map node features; The tree species classification module is used to perform adaptive node feature aggregation on the global graph to obtain target node features, and to classify the target node features to obtain tree species classification results.

7. A terminal, characterized in that, The terminal includes a memory, a processor, and a program stored in the memory and executable on the processor. When executed by the processor, the program implements the steps of the tree species fine classification method based on nested graph convolutional networks as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which stores a tree species fine classification program based on nested graph convolutional networks. When the tree species fine classification program based on nested graph convolutional networks is executed by a processor, it implements the steps of the tree species fine classification method based on nested graph convolutional networks as described in any one of claims 1-5.

Images