Power transmission corridor individual tree identification method and system based on DGCNN fusion data

Abstract

The invention discloses a power transmission corridor single tree identification method and system based on DGCNN fusion data, and relates to the field of power transmission corridor tree identification and management, and the method comprises the steps: collecting point cloud data and hyperspectral image data, and obtaining first data after processing; based on the first data, constructing a graph structure through DGCNN and extracting spatial and spectral features; based on the spatial and spectral features, feature learning and classification of pine trees are carried out through a DGCNN-based deep neural network, and a preliminary classification result is obtained; based on the preliminary classification result, post-processing and precision optimization are carried out, and a final pine tree identification classification result is obtained; and identifying the single trees of the power transmission corridor based on the final pine tree identification and classification result. Through the deep learning technology and the fine point cloud data processing, the limitation of a traditional method is overcome, high-precision, automatic and high-robustness tree segmentation is realized, manual intervention is reduced, the segmentation quality is improved, and the method has a relatively strong application prospect.

Description

Technical Field

[0001] The present invention relates to the field of tree identification and management in power transmission corridors, and in particular to a method and system for identifying individual trees in power transmission corridors based on DGCNN fusion data. Background Art

[0002] With the development of intelligent power systems, transmission corridor safety monitoring and risk assessment have become key issues in power line inspections. Traditional transmission line inspections rely on manual labor or aerial photography, which is inefficient and difficult to accurately detect hidden dangers.

[0003] LiDAR and hyperspectral imaging technologies have been applied to transmission corridor monitoring, demonstrating progress in tree and obstacle identification, but their integration still faces challenges. Point cloud data acquired by LiDAR provides three-dimensional spatial information, but tree identification using single point cloud data is challenging. Traditional geometric feature extraction methods are prone to misidentification or omission in complex scenes. Hyperspectral imaging captures spectral information, providing rich features for tree classification. However, due to its lack of spatial structure, the images are two-dimensional, and their resolution is limited, making accurate tree feature extraction in complex environments challenging. Existing point cloud and hyperspectral image fusion techniques have shortcomings, often employing simple feature concatenation or decision fusion. These techniques lack the deep learning algorithms needed to efficiently model the complex relationships between the two. Consequently, recognition accuracy in complex scenes falls short of meeting practical requirements. In recent years, graph convolutional neural networks (GCNs) have achieved significant success in image recognition and point cloud processing. Dynamic graph convolutional neural networks (DGCNNs), in particular, can adaptively learn local and global features from point cloud data, overcoming the limitations of traditional point cloud processing methods, effectively handling noise and sparsity, and improving the robustness and accuracy of tree identification. To improve the accuracy of single tree recognition, researchers are attempting to deeply fuse DGCNN with hyperspectral imagery data to fully utilize the spatial information of point clouds and the spectral information of hyperspectral imagery. However, the current challenge is designing an efficient deep fusion framework that enables the two data types to collaboratively learn within the same network to improve recognition accuracy. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: how to overcome the problems of shallow integration, low recognition accuracy and high computational complexity of existing technologies, and provide an accurate, robust and efficient solution for single tree identification in transmission corridors.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, an embodiment of the present invention provides a method for identifying individual trees in a transmission corridor based on DGCNN fusion data, comprising:

[0008] Collecting point cloud data and hyperspectral image data, and obtaining first data after processing;

[0009] Based on the first data, a graph structure is constructed and spatial and spectral features are extracted through DGCNN;

[0010] Based on the spatial and spectral features, the feature learning and classification of pine trees were performed using a deep neural network based on DGCNN, and preliminary classification results were obtained;

[0011] Based on the preliminary classification results, post-processing and accuracy optimization are performed to obtain the final pine tree identification and classification results;

[0012] Based on the final pine tree identification and classification results, individual trees in the transmission corridor are identified.

[0013] As a preferred solution for the single tree identification method in the transmission corridor based on DGCNN fusion data, the following methods are proposed:

[0014] The constructing a graph structure and extracting spatial and spectral features based on the first data by using DGCNN includes:

[0015] The geometric features of point cloud data are fused with the spectral features of hyperspectral imagery to construct a multidimensional feature vector. A dynamic graph convolutional neural network is used to construct the graph structure, and graph convolution is used to extract spatial and spectral features.

[0016] As a preferred solution for the single tree identification method in the transmission corridor based on DGCNN fusion data, the following methods are proposed:

[0017] The constructing of a multidimensional feature vector comprises:

[0018] The geometric features of the point cloud extracted by dynamic graph convolution are spliced with the spectral features of the hyperspectral image to construct a multi-dimensional fusion feature vector.

[0019] As a preferred solution for the single tree identification method in the transmission corridor based on DGCNN fusion data, the following methods are proposed:

[0020] Based on the spatial and spectral features, the feature learning and classification of pine trees are performed using a deep neural network based on DGCNN, and the preliminary classification results include:

[0021] The fused features are input into the classification network to calculate the probability of each point belonging to the pine tree category, and the pine tree and non-pine tree classification are performed according to the preset threshold;

[0022] Cluster analysis is performed on the classified points, and combined with the geometric features of the point cloud, individual pine trees in the transmission corridor are identified, and the pine tree category labels and spatial locations are output.

[0023] As a preferred solution for the single tree identification method in the transmission corridor based on DGCNN fusion data, the following methods are proposed:

[0024] Based on the preliminary classification results, post-processing and accuracy optimization are performed to obtain the final pine tree identification and classification results including:

[0025] The initial classification results were smoothed and their boundaries corrected. Local noise and boundary errors were reduced using a sliding window and conditional random field algorithm. The pine tree recognition results were optimized locally and globally using geometric and spatial constraints.

[0026] As a preferred solution for the single tree identification method in the transmission corridor based on DGCNN fusion data, the following methods are proposed:

[0027] The category smoothing includes:

[0028] Assume that in a certain area, the classification results of point cloud data and hyperspectral data are that pine trees and non-pine trees appear alternately. Use a sliding window with a window size of w to calculate the average probability of pine trees in the area, and recalibrate the category of the area accordingly.

[0029] As a preferred solution for the single tree identification method in the transmission corridor based on DGCNN fusion data, the following methods are proposed:

[0030] The category smoothing also includes:

[0031] For each point in the region, define a neighborhood window N(i), including point p i and its k neighbor points;

[0032] The class probabilities of all points in the neighborhood window are averaged;

[0033] If the smoothed probability psmooth(i) is greater than the set threshold δ, the point p i Marked as pine tree category, otherwise marked as non-pine tree category.

[0034] In a second aspect, an embodiment of the present invention provides a system for identifying individual trees in a power transmission corridor based on DGCNN fusion data, comprising:

[0035] A preprocessing module, used for collecting point cloud data and hyperspectral image data, and obtaining first data after processing;

[0036] A feature extraction module, configured to construct a graph structure and extract spatial and spectral features based on the first data through a DGCNN;

[0037] A preliminary classification module is used to learn and classify the features of pine trees based on the spatial and spectral features through a deep neural network based on DGCNN to obtain preliminary classification results;

[0038] The final classification module is used to perform post-processing and accuracy optimization based on the preliminary classification results to obtain the final pine tree identification and classification results;

[0039] The recognition module is used to identify individual trees in the transmission corridor based on the final pine tree identification and classification results.

[0040] In a third aspect, an embodiment of the present invention provides a computing device, including:

[0041] memory and processor;

[0042] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the one or more programs are executed by the one or more processors, the one or more processors implement the transmission corridor single tree identification method based on DGCNN fusion data as described in any embodiment of the present invention.

[0043] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the method for identifying individual trees in a transmission corridor based on DGCNN fusion data.

[0044] Beneficial effects of the present invention: The present invention improves the accuracy of tree recognition by fusing the geometric features of point cloud data with the spectral features of hyperspectral images, making full use of the advantages of the two data sources. The DGCNN dynamic graph convolutional neural network can effectively capture the spatial features in point cloud data and the spectral features in hyperspectral data, avoiding the limitations of traditional single data source methods; through multimodal data fusion, especially the introduction of hyperspectral images, it effectively improves the recognition ability in complex environments. Through the graph structure modeling of the DGCNN network, the relationship and local features between trees can be better processed, and the recognition robustness and stability of different tree species are improved; post-processing and precision optimization steps are adopted, such as category smoothing, boundary correction, geometric constraints and spatial consistency optimization, which effectively reduce the cases of misidentification and missed identification. These optimization methods ensure recognition accuracy even in dense tree distributions and complex tree structures. They significantly improve recognition accuracy and boundary consistency, particularly when dealing with boundaries between trees and non-trees, and between tree species. By leveraging the multiple spectral features extracted from hyperspectral data and the geometric information provided by point cloud data, the method can distinguish the unique characteristics of different tree species (such as crown shape, trunk height, and spectral reflectance), adapting to the identification needs of various tree species. High recognition accuracy is achieved regardless of the tree species. The method boasts high computational efficiency and scalability, enabling it to handle large-scale tree recognition tasks. It operates stably and provides high-quality recognition results across a wide range of scenarios, from single transmission line corridors to large-scale scenarios. This makes the technology promising for broad application in various fields. The DGCNN-based method offers significant flexibility, allowing the network structure, feature selection, and optimization strategies to be tailored to suit different tree recognition tasks. This method can also be integrated with other remote sensing data (such as high-resolution satellite imagery and drone footage) to further improve recognition accuracy and expand its scope of application. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0046] Figure 1 This is the overall flow chart of the single tree identification method for transmission corridors based on DGCNN fusion data according to the present invention;

[0047] Figure 2 This is a DGCNN model architecture diagram of the transmission corridor single tree identification method based on DGCNN fusion data described in the present invention;

[0048] Figure 3 It is a fusion diagram of point cloud and hyperspectral image data of the transmission corridor single tree identification method based on DGCNN fusion data described in the present invention;

[0049] Figure 4 This is a voxelized three-dimensional point cloud scene classification method and flow chart of the transmission corridor single tree identification method based on DGCNN fusion data described in the present invention. DETAILED DESCRIPTION

[0050] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0051] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0052] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0053] Example 1

[0054] Reference Figure 1 , which is the first embodiment of the present invention, provides a method for identifying individual trees in a transmission corridor based on DGCNN fusion data, comprising:

[0055] S1: collect point cloud data and hyperspectral image data, and obtain first data after processing;

[0056] S2: Based on the first data, a graph structure is constructed and spatial and spectral features are extracted through DGCNN;

[0057] S3: Based on the spatial and spectral features, feature learning and classification of pine trees are performed using a deep neural network based on DGCNN to obtain preliminary classification results;

[0058] S4: Based on the preliminary classification results, post-processing and accuracy optimization are performed to obtain the final pine tree identification and classification results;

[0059] S5: Based on the final pine tree identification and classification results, identify individual trees in the transmission corridor.

[0060] It should be noted that, through steps S1-S5, this embodiment overcomes the limitations of traditional methods through deep learning technology and sophisticated point cloud data processing, achieves high-precision, automated, and highly robust tree segmentation, reduces manual intervention, and improves segmentation quality. It has strong application prospects and is particularly suitable for fields such as tree identification and management in transmission corridors.

[0061] Example 2

[0062] Reference Figures 1-4 , which is an embodiment of the present invention, provides a method for identifying individual trees in a transmission corridor based on DGCNN fusion data based on the previous embodiment, including:

[0063] In the embodiment of the present application, the point cloud data and hyperspectral image data are collected in the above step S1, and the first data obtained after processing includes:

[0064] Collect point cloud data and hyperspectral image data in the area, and use LiDAR and hyperspectral cameras to obtain ground feature information.

[0065] The point cloud data is processed by denoising, downsampling and normal estimation, while the hyperspectral image data is subjected to radiation correction, cloud removal, band selection and feature extraction to form fused data suitable for input into the neural network as the first data.

[0066] Specifically, the point cloud data and hyperspectral image data within the acquisition area include:

[0067] It is necessary to collect point cloud data and hyperspectral image data from the actual inspection environment of the transmission corridor. The data collection process usually involves the following two devices:

[0068] Laser Ranging and Detection Equipment (LiDAR): Used to obtain point cloud data, accurately scan and record the three-dimensional spatial structure of pine trees in the transmission corridor area.

[0069] Hyperspectral imaging equipment: used to obtain hyperspectral image data of the area where pine trees are located, usually including hundreds of continuous spectral bands, which can effectively distinguish the spectral characteristics of pine trees from other vegetation.

[0070] For example, the dataset, which covers the diverse morphological and distribution characteristics of pine trees across multiple transmission corridors, includes hyperspectral imagery (multispectral bands) and point cloud data. The data was collected using a LiDAR device with a high spatial resolution (approximately 10 cm). The point cloud includes the tree's geometric characteristics (such as position and normals). The data was collected using a drone-mounted hyperspectral camera, encompassing dozens of spectral bands and a resolution of 0.5 meters.

[0071] Specifically, the first data obtained after processing includes:

[0072] Point cloud data is obtained through LiDAR scanning, with each point having three-dimensional coordinates (x, y, z). Point cloud data often contains noise, outliers, and sparse distribution, so preprocessing is required to ensure its quality meets the requirements of subsequent algorithms.

[0073] LiDAR point cloud data is often affected by environmental noise, resulting in the inclusion of outliers. To remove this noise, a statistical outlier removal algorithm is used. This algorithm calculates the distance between each point and its neighbors and removes points that are far from the majority.

[0074] The algorithm operates as follows:

[0075] For each point, calculate the distance d(p i ,p j ), the distance is defined as the Euclidean distance:

[0076]

[0077] Among them, p i and p j are two points in the point cloud, and x, y, z are their spatial coordinates.

[0078] Calculate the mean distance and standard deviation of each point's neighborhood.

[0079] Set a threshold T. If the distance between a point and its neighboring points is greater than the threshold, the point is considered an outlier and will be removed.

[0080] After removing all outliers, the remaining data is the filtered point cloud data.

[0081] It should be noted that in this way, noise data irrelevant to the morphology of pine trees were removed, and the accuracy of subsequent processing was improved.

[0082] The geometric features of point cloud data (such as normals, curvature, etc.) are crucial for subsequent graph convolution operations. The normal reflects the local plane direction of the point, while the curvature indicates the degree of curvature of the local surface.

[0083] Use PCA (Principal Component Analysis) to estimate the normals in the point cloud. The specific process is as follows:

[0084] For each point p i , select its neighborhood point set N(p i ) (The number of neighborhood points can be set as needed, usually 20).

[0085] Calculate the covariance matrix C of the neighborhood points:

[0086]

[0087] in, For point p i The mean of the neighborhood points.

[0088] Perform eigenvalue decomposition on the covariance matrix C to obtain eigenvalues λ1, λ2, λ3 and eigenvectors v1, v2, v3, where λ1≤λ2≤λ3.

[0089] The normal direction is given by the eigenvector v3 (corresponding to the largest eigenvalue).

[0090] Curvature calculation: The local curvature κ can be calculated from the second smallest eigenvalue λ2 and the largest eigenvalue λ3:

[0091]

[0092] It should be noted that these geometric features (normal direction and curvature) provide rich local spatial information for the subsequent graph convolutional network, enabling the model to better understand the geometric structure of the point cloud and help distinguish pine trees from other objects.

[0093] like Figure 4 As shown in FIG, a voxelized 3D point cloud scene classification method and flowchart; point cloud data is usually sparse, and the voxelization method is used to convert it into a uniformly distributed voxel grid to facilitate subsequent calculations.

[0094] Choose a voxel size δ (e.g. 0.1m).

[0095] The point cloud data is divided into cubic voxels of size δ (a voxel is a small unit in a 3D grid).

[0096] Aggregate the point cloud data within each voxel into the center point of the voxel.

[0097] Generate a new point cloud dataset that contains representative points for each voxel, reducing the amount of data and improving processing speed.

[0098] Hyperspectral image data provides spectral information about pine trees and their surroundings, helping to distinguish different types of vegetation. Hyperspectral image preprocessing can remove environmental noise and extract the spectral characteristics of pine trees.

[0099] Perform spectral normalization on hyperspectral image data to eliminate the influence of different lighting conditions on spectral data. The commonly used normalization method is Z-score normalization:

[0100] The pixel values of each band bk are normalized. The normalized image data can eliminate the influence of most environmental factors and make the spectral characteristics of pine trees more stable.

[0101] Hyperspectral imagery contains multiple spectral bands, some of which are important for identifying pine trees. Therefore, it is necessary to extract discernible features. Common feature extraction methods include principal component analysis (PCA) and spectral indices (such as NDVI).

[0102] ① Principal Component Analysis (PCA): PCA is used to map high-dimensional spectral data into a low-dimensional space, retaining the most representative principal components. The goal of PCA is to find the direction with the largest variance in the data, namely:

[0103] Y=XW

[0104] Among them, X is the original spectral data matrix, W is the eigenvector matrix, and Y is the spectral feature matrix after dimension reduction.

[0105] Spectral index: By calculating the spectral difference between pine trees and their surroundings, some discriminative spectral indices can be obtained, such as the Normalized Difference Vegetation Index (NDVI):

[0106]

[0107] NIR is the spectral value of the near-infrared band, and RED is the spectral value of the red band. NDVI can effectively distinguish between vegetation and non-vegetated areas and help identify pine trees.

[0108] Since point cloud data and hyperspectral image data come from different sources, their spatial resolutions and coordinate systems may also be different, so spatial registration is required.

[0109] Use a feature-based registration method (such as the ICP algorithm) to preliminarily align the point cloud data with the image data to ensure that they are in the same coordinate system.

[0110] During the registration process, one usually aligns by selecting some known points (such as the base or top of a tree) and then calculating the transformation matrix based on these known points.

[0111] Methods such as optical matching or regional mutual information are used to correspond the spectral information in the hyperspectral image to the three-dimensional position in the point cloud data.

[0112] It should be noted that through the above registration steps, it can be ensured that each point cloud point corresponds to a pixel in the hyperspectral image, thereby achieving effective data fusion.

[0113] In the embodiment of the present application, in the above step S2, based on the first data, constructing a graph structure and extracting spatial and spectral features through DGCNN includes:

[0114] like Figure 3 Figure 1 shows the fusion of point cloud and hyperspectral image data. The geometric features of the point cloud data (such as position and normal) are fused with the spectral features of the hyperspectral image (such as NDVI and spectral index) to construct a multidimensional feature vector. A dynamic graph convolutional neural network (DGCNN) is used to construct a graph structure and extract spatial and spectral features through graph convolution. This allows for the fusion of point cloud and spectral data at a higher level for pine tree classification.

[0115] It should be noted that point cloud data naturally has the characteristics of a graph structure. Each point can be regarded as a node in the graph, and the connection relationship between nodes reflects the spatial relationship between different points in the point cloud. DGCNN extracts local and global features of the point cloud by performing convolution operations on the graph structure. Figure 2 The following figure shows the DGCNN model architecture diagram.

[0116] Specifically, during the point cloud data processing, each point is treated as a node in a graph, and the adjacency matrix of the graph is constructed based on the neighborhood relationship of each point. The graph structure can be constructed by calculating the distance or similarity between each point in the point cloud.

[0117] Calculate the distance between points: The coordinates of each point in the point cloud data of the pine tree area are p i =(x i ,y i , z i ), where i∈{1,2,...,N} is the index of the point and N is the number of points in the point cloud. i and its neighbor point p j The Euclidean distance between:

[0118]

[0119] For the point cloud of the area where the pine tree is located, the neighbor point p j is the distance in space from point p i Points that are close to each other. A distance threshold δ can be defined. If the distance between two points is less than the threshold, they are considered neighbors.

[0120] Based on the distance information, construct the adjacency matrix A, where A ij Represents point p i and point p j The Gaussian kernel function is usually used to measure the similarity between two points.

[0121] In addition to calculating the adjacency matrix based on distance, the k-nearest neighbor (k-NN) method can also be used to determine the neighborhood of each point in the graph. The specific steps are as follows:

[0122] For each point p i , find the k points p closest to it j (j ≠ i).

[0123] Calculate the distance between points and build the neighborhood relationship of each point.

[0124] Through the k-NN method, the edge weight of the graph no longer depends on the distance threshold, but on the number k of neighborhood points of each point.

[0125] The graph convolution operation of DGCNN is to gradually extract local features by transferring information on the graph structure, aggregate global features through a multi-layer network, and update node features through the connection relationship between nodes and their neighborhoods.

[0126] Specifically, node feature representation: Assume that each point p i The initial feature is its coordinate p i =(x i ,y i ,z i ), as well as geometric features such as normal direction and curvature. These features are used as the initial input features of the point and are recorded as:

[0127]

[0128] Among them, n x ,n y ,n z It's point p i The normal direction of the point, κ is the curvature of the point.

[0129] In the k-th layer of graph convolution, node features It is calculated by weighting the features of the node itself and its neighboring nodes. The update formula of graph convolution is:

[0130]

[0131] Among them, A ij is the adjacency matrix, which represents the connection strength (i.e., similarity) between node i and node j; N(i) is the set of neighbors of node i; is the feature of node j in the k-1th layer; W (k) is the weight matrix of the kth layer; σ is the activation function (usually ReLU is used).

[0132] Each graph convolution operation performs a weighted summation on the features of the current node and its neighbors to calculate the updated node features. After multiple graph convolutions, the features of each node will gradually aggregate neighborhood information from a larger range, thereby extracting the global and local structural features of the point cloud.

[0133] It should be noted that in traditional graph convolution, the graph structure (adjacency matrix) is fixed, while in DGCNN, the graph structure changes dynamically. After each convolution layer, the adjacency relationship is recalculated based on the changes in node features, making information transfer more flexible. This dynamic adjustment of the adjacency matrix enables DGCNN to better capture the complex geometric relationships in point cloud data.

[0134] Based on the node features after each convolution layer, the adjacency matrix is updated and the structure of the graph is adjusted:

[0135] Calculate the new adjacency matrix: After each layer of graph convolution, based on the current node features Recalculate the similarity between nodes and update the adjacency matrix A (k) .

[0136] Dynamic update: new adjacency matrix A (k) This can be obtained by calculating the similarity between node features (e.g., through Euclidean distance or cosine similarity), so that the adjacency matrix is adjusted after each convolution layer.

[0137] In DGCNN, graph convolution extracts geometric features based on point cloud data, while hyperspectral imagery provides spectral information for each point. When identifying pine trees, the spatial features of the point cloud data and the spectral features of the hyperspectral imagery can be effectively fused to form a joint feature input into the model.

[0138] Specifically, the feature fusion process is as follows:

[0139] The geometric features of each point extracted by DGCNN are combined with the spectral features in the hyperspectral image to obtain the fusion features:

[0140] f fusion =[f point ;f spectral ]

[0141] Among them, f point It is the geometric feature extracted based on graph convolution, f spectral is the spectral feature extracted from the hyperspectral image. The fusion feature f after splicing fusion It will be passed as input to the subsequent classification network for pine tree recognition.

[0142] After graph convolution and feature fusion, pine trees are classified using traditional classification methods (such as fully connected networks). Through the multi-layer fully connected network, the model can determine whether a point belongs to the pine tree category based on the fused features.

[0143] In the embodiment of the present application, in the above step S3, based on the spatial and spectral features, the feature learning and classification of pine trees are performed by a deep neural network based on DGCNN, and the preliminary classification results obtained include:

[0144] A deep neural network based on DGCNN is used to learn and classify pine tree features. The network outputs the probability of each point belonging to the pine tree category and classifies pine trees from non-pine trees based on a threshold. A fully connected neural network (FCN) is introduced to further optimize pine tree recognition results, and classification decisions are made using a multilayer perceptron.

[0145] Specifically, the fused features extracted by the DGCNN are input into a deep classification network (such as a fully connected layer or convolutional neural network) for pine tree identification. Assuming that pine trees have significantly different characteristics from other vegetation, the classification model can effectively distinguish pine trees from other tree species using these features. A multi-layer fully connected neural network (FCN) is used to classify the fused features, and a softmax function is used to calculate the classification probability of the pine tree.

[0146] In the previous DGCNN graph convolution step, a fused feature ffusion was obtained for each point. A fully connected neural network consists of multiple layers, each of which processes the input data through linear transformations and activation functions, ultimately outputting a pine tree classification result. Consider a two-layer fully connected network. The first layer has an output dimension of n1, the second layer has an output dimension of n2, and the final layer uses the Softmax function to output the probability that each point belongs to the pine tree class.

[0147] Input layer: The input is the feature vector f after the fusion of point cloud data and hyperspectral data fusion .

[0148] Hidden layer: The output of the first fully connected layer is h1=W1f fusion +b1, where W1 is the weight matrix and b1 is the bias term. Nonlinear transformation is performed through activation functions (such as ReLU):

[0149] h1=ReLU(W1f fusion +b1)

[0150] The output of the second fully connected layer is h2=W2 h2+b1, and the activation is also performed:

[0151] h2=ReLU(W2h1+b2)

[0152] Softmax output layer: The output of the fully connected network is passed through the Softmax function to calculate the probability that each point belongs to the pine tree category:

[0153] p class =Softmax(W3h2+b3)

[0154] Among them, W3 and b3 are the weight and bias of the last layer, p class is the probability that each point belongs to the pine tree class.

[0155] The calculation formula of the Softmax function is:

[0156]

[0157] Among them, i is the output of the last layer of the network, and C is the number of categories (considering two categories: pine trees and non-pine trees, C = 2).

[0158] During the training process, the network is trained using labeled data (point cloud and hyperspectral image data containing pine and non-pine tree labels). The loss function uses the cross entropy loss function to measure the gap between the predicted category and the true category:

[0159]

[0160] Among them, y i is the true label, p class,i is the probability of pine trees predicted by the model, and N is the number of samples.

[0161] During training, the gradient is calculated using the backpropagation algorithm, and the network parameters are updated through gradient descent (such as the Adam optimizer) until the loss function converges.

[0162] Once trained, the model can identify pine trees using new point cloud data and hyperspectral imagery. For each new input data point, the network outputs a probability of belonging to the pine tree class. Based on a set threshold (e.g., 0.5), if the probability of belonging to the pine tree class is greater than the threshold, the point is considered a pine tree; otherwise, it is considered a non-pine tree.

[0163] For example, given a test data point, the model might output the following classification probabilities:

[0164] Pine tree category probability p 松树 =0.85

[0165] Non-pine tree category probability p 非松树 =0.15

[0166] Because p 松树 It is greater than the threshold of 0.5, so the point is classified as a pine tree.

[0167] For the point cloud data of the entire transmission corridor area, each point needs to be classified and aggregated for overall identification. The specific method is as follows:

[0168] The features of each point (including point cloud and hyperspectral image features) are classified to obtain the category label of each point.

[0169] Perform cluster analysis on the classified points. Based on the geometric characteristics of the pine trees (such as height and crown width), the DBSCAN clustering algorithm or K-means clustering method can be used to identify individual pine trees in the area.

[0170] Combining the clustering results with the geometric features in the point cloud data (such as the height and shape of the points), individual pine trees in the transmission corridor can be accurately identified.

[0171] The model outputs a class label for each pine tree and its spatial location within the transmission corridor. This information allows the precise location of pine trees, providing power companies with detailed tree data for further maintenance and optimization of power lines.

[0172] In the embodiment of the present application, in step S4, post-processing and accuracy optimization are performed based on the preliminary classification results to obtain the final pine tree identification and classification results including:

[0173] The initial classification results were smoothed and their boundaries corrected, using a sliding window and conditional random field (CRF) algorithm to reduce local noise and boundary errors. The pine tree recognition results were then optimized locally and globally, combining geometric and spatial constraints to enhance the consistency of tree morphology and the rationality of spatial distribution.

[0174] It should be noted that after completing pine tree identification and classification, preliminary results were obtained. Due to environmental noise, data sparsity, and the complexity of tree morphology, the recognition results may contain misidentifications and omissions. To improve the accuracy and robustness of the recognition results, post-processing and accuracy optimization of the model output are generally required.

[0175] Specifically, due to the fusion characteristics of point cloud data and hyperspectral imagery, class labels may jump or become discontinuous between adjacent points. This phenomenon may stem from the discrete nature of point cloud data, local variations in spectral data, and other factors. Therefore, class smoothing and boundary correction methods are used to optimize recognition results.

[0176] The purpose of class smoothing is to reduce discontinuities in classification results caused by local noise or sensor errors. A sliding window method is used to smooth the recognition results. For example, if the point cloud data and hyperspectral data are classified as alternating pine trees and non-pine trees within a small area, a sliding window of size w can be used to calculate the average probability of pine trees within the area and recalibrate the category of the area accordingly.

[0177] The specific steps are as follows:

[0178] For each point, define a neighborhood window N(i) containing point p i and its k neighbor points.

[0179] Average the class probabilities for all points in the neighborhood window:

[0180]

[0181] If the smoothed probability psmooth(i) is greater than the set threshold δ, the point p i Marked as pine tree category, otherwise marked as non-pine tree category.

[0182] The goal of boundary correction is to optimize the boundaries between pine trees and non-pine trees, particularly the outlines of pine trees. Due to the discrete nature of point cloud data and local variations in spectral data, the boundaries of pine trees may be misclassified as non-pine trees. To address this issue, a conditional random field (CRF) can be used for boundary correction. CRF can correct classification boundaries based on contextual information.

[0183] The specific steps are as follows: Define an adjacency graph, where nodes represent each point in the point cloud data and edges connect adjacent points. Construct an energy function for the graph, which consists of two terms: a data term and a smoothing term.

[0184] Data item: represents the original classification probability of the point, and calculates the negative logarithmic probability that the point is a pine tree:

[0185] E data (i) = -log p class，i

[0186] Smoothness term: represents the similarity between neighboring points and encourages adjacent points to have the same category labels:

[0187] E smooth (i, j) = λ·A ij ·f i -f j

[0188] Among them, A ij is the similarity between point i and point j, f i and f jare the features of point i and point j, and λ is the smoothing parameter.

[0189] Minimize the energy function to obtain the corrected category labels. Through optimization algorithms (such as gradient descent or GraphCut), smoother classification boundaries can be obtained.

[0190] It should be noted that the goal of local and global optimization is to improve the accuracy of pine tree recognition, especially in complex scenarios (such as dense forest areas or overlapping areas between trees). The accuracy of the recognition results can be further improved through optimization methods.

[0191] In another possible implementation, in pine tree recognition, the morphological characteristics of different trees have certain regularities. For example, the trunk and crown of a pine tree usually have consistent geometric shapes, so geometric constraints can be introduced to perform local optimization:

[0192] Specifically, geometric constraints include:

[0193] Trunk height constraint: Assuming that pine tree trunk heights typically fall within a certain range, and the crown-to-trunk ratio is relatively constant, we can analyze the height characteristics in the point cloud data to filter out points that clearly do not conform to the pine tree morphology. If a point's height exceeds the typical trunk height of a pine tree, it is likely not part of the tree and can be removed.

[0194] Crown morphology constraints: Pine tree crowns typically exhibit a certain degree of symmetry and expansiveness. This can be combined with the spatial distribution of the point cloud to further optimize pine tree recognition results. By calculating the point density of the crown region within the point cloud, overly dispersed or low-density points can be removed.

[0195] In another possible implementation, spatial consistency constraints can also be used to optimize the recognition results. Global optimization takes into account the distribution patterns of pine trees in space. For example, in a power transmission corridor, the distribution of pine trees usually has a certain pattern, and multiple pine trees will not appear in very close areas.

[0196] Specifically, spatial consistency constraints include:

[0197] Spatial constraints: By performing spatial clustering analysis (such as DBSCAN or K-means) on the point cloud data, we can identify clusters of pine tree points within a local area. By setting a minimum distance threshold, we can prevent multiple pine trees that are too close from being mistakenly identified as the same tree.

[0198] Post-processing spatial adjustment: For multiple overlapping pine trees that may appear during the identification process, a spatial separation algorithm (based on the analysis of connected regions) is used to optimize the boundary of each tree to ensure that each identified tree point cloud conforms to a reasonable spatial distribution.

[0199] Conditional Random Fields (CRFs) can be used as a global optimization tool to further improve the global consistency of classification in pine tree identification across the entire region. By incorporating spatial context, CRFs can smooth classification results globally and correct local classification errors.

[0200] During the optimization of CRF, the geometric structure of the point cloud (such as the shape and distribution of trees) and the spectral characteristics of each point are taken into account, and the overall accuracy of pine tree recognition can be improved through a global optimization process.

[0201] In the embodiment of the present application, the identification of individual trees in the transmission corridor based on the final pine tree identification and classification results in step S5 includes:

[0202] Experiments were conducted across multiple transmission corridors, using point cloud data and hyperspectral imagery to identify pine trees. Evaluation metrics such as accuracy, precision, recall, and F1-score were used to verify the method's high accuracy, robustness, and potential for application.

[0203] For example, after post-processing and accuracy optimization, the following standard evaluation metrics were used to evaluate the accuracy of pine tree recognition:

[0204] Accuracy: The ratio of correctly identified pine tree points to the total number of points.

[0205]

[0206] Precision: The proportion of points identified as pine trees that are actually pine trees.

[0207]

[0208] Recall: The proportion of points that are correctly identified as pine trees among the points that are actually pine trees.

[0209]

[0210] F1-score: The harmonic mean of precision and recall.

[0211]

[0212] The experimental results are as follows:

[0213] Accuracy: The recognition accuracy rate was 92.5%. Through category smoothing, boundary correction, and geometric constraint optimization, the misidentification rate was significantly reduced, especially at the boundary between pine trees and non-pine trees.

[0214] Precision: Since pine trees are recognized with a high accuracy of 90.8%, experiments show that using DGCNN fusion features can effectively improve the accuracy of pine tree recognition and reduce the number of instances where trees are mistakenly identified as non-pine trees.

[0215] Recall: The recall rate is 93.2%, indicating that the method performs well in detecting the coverage of pine trees and can effectively identify most pine trees, especially in dense forest areas and complex environments.

[0216] F1-score: The F1-score is 92%, indicating that the method achieves a good balance between precision and recall.

[0217] Experimental results demonstrate that this method offers significant advantages in pine tree identification, effectively leveraging the fusion features of point cloud and hyperspectral imagery to achieve high-precision pine tree identification. Post-processing and precision optimization techniques further enhance the stability and robustness of the identification results. Compared to traditional methods, this method demonstrates superior performance in accuracy, precision, recall, and F1-score, making it suitable for pine tree identification and classification in complex environments.

[0218] Example 3

[0219] The above is a schematic diagram of the method for identifying individual trees in transmission corridors based on DGCNN fusion data in this embodiment. It should be noted that the technical solution of the system for identifying individual trees in transmission corridors based on DGCNN fusion data and the technical solution of the method for identifying individual trees in transmission corridors based on DGCNN fusion data are based on the same concept. For details not described in detail in the technical solution of the system for identifying individual trees in transmission corridors based on DGCNN fusion data in this embodiment, please refer to the description of the technical solution of the method for identifying individual trees in transmission corridors based on DGCNN fusion data.

[0220] This embodiment also provides a system for identifying individual trees in a transmission corridor based on DGCNN fusion data, including:

[0221] A preprocessing module, used for collecting point cloud data and hyperspectral image data, and obtaining first data after processing;

[0222] A feature extraction module, configured to construct a graph structure and extract spatial and spectral features based on the first data through a DGCNN;

[0223] A preliminary classification module is used to learn and classify the features of pine trees based on the spatial and spectral features through a deep neural network based on DGCNN to obtain preliminary classification results;

[0224] The final classification module is used to perform post-processing and accuracy optimization based on the preliminary classification results to obtain the final pine tree identification and classification results;

[0225] The recognition module is used to identify individual trees in the transmission corridor based on the final pine tree identification and classification results.

[0226] This embodiment further provides a computing device applicable to a method for identifying individual trees in a transmission corridor based on DGCNN fusion data, including:

[0227] Memory and processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the single-tree identification method for the transmission corridor based on DGCNN fusion data proposed in the above embodiment.

[0228] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the method for identifying individual trees in a transmission corridor based on DGCNN fusion data as proposed in the above embodiment.

[0229] The storage medium proposed in this embodiment and the single tree identification method for transmission corridors based on DGCNN fusion data proposed in the above embodiment belong to the same inventive concept. For technical details not described in detail in this embodiment, please refer to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0230] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for identifying individual trees in power transmission corridors based on DGCNN fusion data, characterized in that: include: Collecting point cloud data and hyperspectral image data, and obtaining first data after processing; Based on the first data, a graph structure is constructed and spatial and spectral features are extracted through DGCNN; Based on the spatial and spectral features, the feature learning and classification of pine trees were performed using a deep neural network based on DGCNN, and preliminary classification results were obtained; Based on the preliminary classification results, post-processing and accuracy optimization are performed to obtain the final pine tree identification and classification results; Based on the final pine tree identification and classification results, individual trees in the transmission corridor are identified.

2. The method for identifying individual trees in a power transmission corridor based on DGCNN fusion data according to claim 1, characterized in that: The constructing a graph structure and extracting spatial and spectral features based on the first data by using DGCNN includes: The geometric features of point cloud data are fused with the spectral features of hyperspectral images to construct a multidimensional feature vector; a dynamic graph convolutional neural network is used to construct a graph structure, and spatial and spectral features are extracted through graph convolution.

3. The method for identifying individual trees in a transmission corridor based on DGCNN fusion data according to claim 2, characterized in that: The constructing of a multidimensional feature vector comprises: The geometric features of the point cloud extracted by dynamic graph convolution are spliced with the spectral features of the hyperspectral image to construct a multi-dimensional fusion feature vector.

4. The method for identifying individual trees in a power transmission corridor based on DGCNN fusion data according to claim 3, characterized in that: Based on the spatial and spectral features, the feature learning and classification of pine trees are performed using a deep neural network based on DGCNN, and the preliminary classification results include: The fused features are input into the classification network to calculate the probability of each point belonging to the pine tree category, and the pine tree and non-pine tree classification are performed according to the preset threshold; Cluster analysis is performed on the classified points, and combined with the geometric features of the point cloud, individual pine trees in the transmission corridor are identified, and the pine tree category labels and spatial locations are output.

5. The method for identifying individual trees in a transmission corridor based on DGCNN fusion data according to claim 4, characterized in that: Based on the preliminary classification results, post-processing and accuracy optimization are performed to obtain the final pine tree identification and classification results including: The preliminary classification results are smoothed and boundary corrected, and the local noise and boundary errors are reduced through the sliding window and conditional random field algorithms. The recognition results of pine trees are locally and globally optimized by combining geometric and spatial constraints.

6. The method for identifying individual trees in a power transmission corridor based on DGCNN fusion data according to claim 5, characterized in that: The category smoothing includes: Assume that in a certain area, the classification results of point cloud data and hyperspectral data are that pine trees and non-pine trees appear alternately. Use a sliding window with a window size of w to calculate the average probability of pine trees in the area, and recalibrate the category of the area accordingly.

7. The method for identifying individual trees in a power transmission corridor based on DGCNN fusion data according to claim 6, characterized in that: The category smoothing also includes: For each point in the region, define a neighborhood window N(i), which contains point p i and its k neighbor points; The class probabilities of all points in the neighborhood window are averaged; If the smoothed probability psmooth(i) is greater than the set threshold δ, the point p i Marked as pine tree category, otherwise marked as non-pine tree category.

8. A system using the method for identifying individual trees in a power transmission corridor based on DGCNN fusion data as described in any one of claims 1 to 7, characterized in that: include: A preprocessing module, used for collecting point cloud data and hyperspectral image data, and obtaining first data after processing; A feature extraction module, configured to construct a graph structure and extract spatial and spectral features based on the first data through a DGCNN; A preliminary classification module is used to learn and classify the features of pine trees based on the spatial and spectral features through a deep neural network based on DGCNN to obtain preliminary classification results; The final classification module is used to perform post-processing and accuracy optimization based on the preliminary classification results to obtain the final pine tree identification and classification results; The recognition module is used to identify individual trees in the transmission corridor based on the final pine tree identification and classification results.

9. A computing device comprising: memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the single-tree identification method for transmission corridors based on DGCNN fusion data as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the method for identifying individual trees in a transmission corridor based on DGCNN fusion data as described in any one of claims 1 to 7.

Images