Vegetation image recognition method, system and device based on OfficientNet and medium

Abstract

The invention discloses a vegetation image recognition method, system, equipment and medium based on OfficientNet, and belongs to the technical field of computer vision and image processing, and the method comprises the steps: collecting data through image collection and synchronous utilization of a laser radar, and carrying out the preprocessing of the collected data; an improved OfficientNet network model is constructed, and training and optimization are carried out by using the improved OfficientNet network model; and carrying out vegetation classification and hidden danger detection on the input image, inputting the collected real-time data into a comprehensive risk scoring model for hidden danger detection for calculation, and carrying out risk early warning positioning. The hidden danger trees are accurately positioned based on the GIS technology, the omission ratio is reduced through a dynamic early warning system, and the manual inspection risk is reduced through automatic reporting; different climate monitoring requirements are adapted based on the expansibility of the model, the tree barrier hidden danger can be blocked and the maintenance efficiency can be improved in practical application, and an integrated solution integrating accurate identification and real-time positioning is formed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and image processing, and particularly relates to a vegetation image recognition method, system, device and medium based on EfficientNet. BACKGROUND

[0002] The power transmission corridor is a crucial component of the power system, and ensuring its safe operation is of great significance to the stability of the power grid. However, the excessive growth or dumping of vegetation in the power transmission corridor can pose a threat to the normal operation of the transmission line, causing major safety incidents such as line short circuits and tripping. Therefore, timely and accurate detection and identification of potential vegetation risk hazards in the power transmission corridor is a core task in power operation and maintenance. With the rapid development of unmanned aerial vehicle inspection, remote sensing image acquisition, and deep learning technology, automated vegetation risk hazard detection technology has gradually become a research hotspot.

[0003] Most current methods for detecting vegetation risk hazards in power transmission corridors rely on traditional image processing and deep learning techniques. Traditional image processing methods typically involve edge detection, texture analysis, or spectral feature extraction to identify vegetation and transmission lines. However, these methods require complex parameter tuning and often lack accuracy and robustness when dealing with high-resolution remote sensing images or complex scenes. Moreover, traditional machine learning methods rely heavily on feature engineering, requiring manual feature design and being difficult to adapt to the diversity of vegetation types and environments in different scenarios.

[0004] EfficientNet is a highly efficient convolutional neural network that optimizes the depth, width, and resolution of the network through a compound scaling strategy, significantly reducing computational costs while maintaining high accuracy. It has demonstrated excellent performance in multiple image classification and object detection tasks, particularly in resource-constrained edge computing and real-time processing scenarios. However, the application of EfficientNet in vegetation risk hazard detection in power transmission corridors has not been thoroughly researched. By combining the efficient forward propagation algorithm of EfficientNet with optimization strategies specific to the power transmission corridor scenario, detection efficiency and accuracy can be significantly improved.

[0005] With the development of unmanned aerial vehicle inspection and remote sensing image technology, the methods of acquiring power transmission corridor monitoring data have become increasingly diverse. However, the diversity and complexity of these data pose challenges to existing detection methods. Vegetation morphology may vary significantly in different regions and seasons, and traditional methods are less adaptable to these changes. Meanwhile, during high-resolution image processing, models are often limited by computational resources, making it difficult to meet real-time requirements. Therefore, developing an efficient, robust, and adaptable vegetation risk hazard detection method for power transmission corridors has become a pressing problem. SUMMARY

[0006] In view of the problems in the prior art described above, the present application is proposed.

[0007] Therefore, the technical problem to be solved by the present application is that, in view of the problems of large consumption of computing resources, poor real-time performance, insufficient accuracy and lack of multi-dimensional risk assessment in traditional methods for vegetation risk detection in power transmission corridors, a solution based on an EfficientNet high-efficiency forward propagation algorithm is proposed, the model structure is optimized through a composite scaling technique, multi-modal data and risk assessment mechanisms are fused, edge computing deployment and adaptive training strategies are combined, efficient and accurate detection, real-time positioning and early warning of vegetation hazards are realized, and the computing load and artificial maintenance cost are reduced, thereby improving the safety operation efficiency of power transmission lines.

[0008] To solve the above technical problems, the present application provides the following technical solutions, through image acquisition and simultaneous laser radar data acquisition, and preprocessing of the collected data; An improved EfficientNet network model is constructed and used for training and optimization; The improved EfficientNet network model integrates a cross-modal attention fusion mechanism and a multi-scale convolutional layer, and performs end-to-end feature fusion and feature extraction on the preprocessed multi-modal data; The improved EfficientNet network model after training and optimization is used for vegetation classification and hazard detection on input images; The hazard detection is based on a comprehensive risk score model constructed based on tree height, distance between tree and power transmission line, tree growth health state score and tree inclination angle; The collected real-time data is input into the comprehensive risk score model for hazard detection for calculation, and risk warning positioning is performed based on the calculation results.

[0009] As a preferred solution of the EfficientNet-based vegetation image recognition method of the present application, wherein the image acquisition and simultaneous laser radar data acquisition, and preprocessing of the collected data include, Vegetation data in the power transmission corridor is collected through remote sensing image and laser radar point cloud technology, including remote sensing image data and laser radar point cloud data; The remote sensing image data includes true color remote sensing image and multispectral image; The laser radar point cloud data includes radar point cloud data.

[0010] As a preferred scheme of the vegetation image recognition method based on EfficientNet, the preprocessing comprises denoising processing, image cropping and standardization, image enhancement and color space conversion on the remote sensing image data, to generate standardized image data suitable for feature extraction by the alternating multi-scale convolution layer. The noise removal, ground point segmentation and point cloud downsampling are performed on the laser radar point cloud data, to provide point cloud data suitable for three-dimensional feature extraction by the point cloud adapter, and the registration of the remote sensing image data and the laser radar point cloud data is completed through the coordinate conversion and registration method, to provide spatially aligned multi-modal data input for the cross-modal attention fusion unit.

[0011] As a preferred scheme of the vegetation image recognition method based on EfficientNet, the construction and training and optimization using the improved EfficientNet network model comprise, The cross-modal attention fusion mechanism and the multi-scale convolution layer are integrated in the basic EfficientNet network model to form the improved EfficientNet network model, and the multi-modal data after preprocessing is subjected to end-to-end feature fusion and feature extraction; The cross-modal attention fusion mechanism is realized by placing the point cloud adapter and the cross-modal attention unit; The point cloud adapter is used to convert the laser radar point cloud data into two-dimensional geometric features; The cross-modal attention unit is used to fuse the image features and the two-dimensional geometric features based on the attention weight; The multi-scale convolution layer is realized by the alternating multi-scale convolution strategy, and is used for the vegetation risk identification task; Different sizes of convolution kernels are configured in the shallow, medium and deep layers of the network; The shallow layer uses small-scale convolution kernels to focus on capturing the texture details of the leaf level of vegetation; The medium layer uses large-scale convolution kernels to expand the receptive field, and is used for modeling the geometric shape of the tree crown level and the spatial relationship of the corresponding power transmission line; The top layer continues to use small-scale convolution kernels to stabilize the semantic level feature expression.

[0012] As a preferred scheme of the vegetation image recognition method based on EfficientNet, the construction and training and optimization using the improved EfficientNet network model further comprise, The training and optimization of the improved EfficientNet network model optimize the network architecture through the composite scaling technology, balance the depth, width and resolution, and process the true color remote sensing image and the laser radar point cloud data of multi-modal fusion; The preprocessed data is divided into a training set and a test set, a multi-task learning is used to jointly optimize the model, and a self-adaptive estimation optimizer, a learning rate decay and a regularization strategy are combined to complete the training; The improved EfficientNet network model is pruned to remove redundant parameters and quantized to reduce the calculation precision and improve the inference efficiency. The accuracy, recall, F1-score and precision of the test set are evaluated to verify the performance of the improved EfficientNet network model in vegetation classification and risk assessment tasks.

[0013] As a preferred scheme of the EfficientNet-based vegetation image recognition method, the improved EfficientNet network model is used to classify and detect vegetation in the input image, and the improved EfficientNet network model is output simultaneously to obtain the comprehensive risk score and spatial positioning information. Vegetation classification is to identify the type of vegetation in the power transmission corridor through remote sensing image data and laser radar point cloud data. Based on the comprehensive risk score, the risk level of vegetation is evaluated. When the comprehensive risk score is greater than or equal to the corresponding preset risk threshold, a risk hazard warning signal is automatically triggered. According to the distance between the tree and the power transmission line, the tree height, the tree inclination angle and the tree growth state, the risk level is evaluated. The distance between the tree and the power transmission line is represented as: , wherein, represents the horizontal distance between the tree and the power transmission line, is the horizontal coordinate of the tree spatial coordinate, is the vertical coordinate of the tree spatial coordinate, is the horizontal coordinate of the power transmission line coordinate, is the vertical coordinate of the power transmission line coordinate. If the distance of the tree is less than the preset height threshold, there is a risk of hitting the power transmission line. If the tree height exceeds the preset height threshold, it will touch the power transmission line. The preset height threshold of the tree is set according to different tree species, climate conditions and power industry standards. The tree inclination angle is represented as: , wherein, is the tree inclination angle, is the tree top height, is the tree root height, The horizontal distance from the base to the top of the tree. It is the inverse function of the tangent function; Tree growth status is assessed by evaluating the degree of tree mortality and pest and disease conditions to obtain a corresponding health score, thus quantifying the impact of vegetation physiological activity on risk. The comprehensive risk score is expressed as: , in, For comprehensive risk scoring, Tree height, Score the health status of tree growth. These are the weighting coefficients for the horizontal distance between the tree and the power transmission line, the tree height, the tree tilt angle, and the tree's growth health status score.

[0014] As a preferred embodiment of the vegetation image recognition method based on EfficientNet described in this invention, the risk warning and positioning includes risk hazard warning and risk hazard positioning. The risk and hazard early warning system sets three preset risk thresholds, including low risk, medium risk, and high risk thresholds. It makes judgments based on a comprehensive risk score. When the comprehensive risk score is greater than or equal to the corresponding preset risk threshold, an early warning is automatically triggered and the inspection personnel are notified through GIS map marking, SMS, and email. Risk and hazard location is based on the spatial location information synchronously output by the optimized and improved EfficientNet network model. The geographical coordinates of the trees with potential hazards are marked by spatial location technology and then visualized on a GIS map.

[0015] Another objective of this invention is to provide an EfficientNet-based vegetation image recognition system that integrates multimodal data acquisition, cross-modal attention fusion, efficient model training, and dynamic early warning and positioning functions. Through modular design, it enables automated identification, multi-dimensional risk assessment, and real-time hazard location of vegetation in power transmission corridors, improving detection efficiency and accuracy while reducing computational resource consumption, thus meeting the intelligent requirements for safe operation and maintenance of power transmission lines.

[0016] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an EfficientNet-based vegetation image recognition system, comprising: an acquisition and preprocessing module, a training and optimization module, a classification and detection module, and an early warning and positioning module; The acquisition and preprocessing module acquires data through image acquisition and synchronous LiDAR acquisition, and preprocesses the acquired data. The training optimization module is used to build and train and optimize the improved EfficientNet network model. The classification and detection module is an improved EfficientNet network model after training and optimization, which is used for vegetation classification and hidden danger detection on the input image. The early warning positioning module inputs the collected real-time data into a comprehensive risk scoring model for hidden danger detection for calculation, and performs risk early warning positioning based on the calculation result.

[0017] A computer device comprises a memory and a processor, the memory stores a computer program, characterized in that the processor executes the computer program to realize the steps of the EfficientNet-based vegetation image recognition method as described above.

[0018] A computer readable storage medium stores a computer program, characterized in that the computer program is executed by a processor to realize the steps of the EfficientNet-based vegetation image recognition method as described above.

[0019] The beneficial effects of the present application are: by introducing the EfficientNet deep neural network and data enhancement technology, the vegetation classification accuracy and environmental adaptability are significantly improved, the high parameter efficiency reduces the consumption of computing resources, and the robust recognition of sunny, cloudy, seasonal changes and other scenes is realized. The risk assessment model based on multi-source data fusion breaks through the limitation of traditional single visual detection, generates a comprehensive risk score by quantitatively analyzing vegetation species, growth state, spatial distance and other multi-dimensional parameters, accurately marks the hidden danger coordinates combined with GIS geographic positioning technology, and effectively reduces the missed detection rate. The system integrates automatic evaluation report generation and remote early warning mechanism, greatly reduces the frequency of manual inspection of high-voltage transmission lines and safety risks, and relies on the model expansion to adapt to the monitoring needs of different climates and geographical characteristics. Practical application shows that this technology can block tree barrier hidden dangers in advance, reduce power interruption accidents, and effectively improve maintenance efficiency through intelligent positioning, forming a comprehensive solution integrating accurate identification, dynamic early warning and intelligent decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0021] Figure 1 The flowchart of the EfficientNet-based vegetation image recognition method provided by an embodiment of the present application.

[0022] Figure 2This is a schematic diagram of a composite scaling strategy for a vegetation image recognition method based on EfficientNet, provided in an embodiment of the present invention. Detailed Implementation

[0023] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0024] Example 1, referring to Figures 1-2 This is one embodiment of the present invention, which provides a vegetation image recognition method based on EfficientNet, including: S1: Data is acquired through image acquisition and synchronously using LiDAR, and the acquired data is preprocessed.

[0025] It should be noted that, such as Figure 1 As shown in step S1, the data acquisition process involves collecting vegetation data within the power transmission corridor using remote sensing images and lidar point cloud technology, including remote sensing image data and lidar point cloud data; wherein, the remote sensing image data includes true-color remote sensing images and multispectral images; and the lidar point cloud data includes radar point cloud data.

[0026] Further preprocessing includes denoising, image cropping and normalization, image enhancement, and color space conversion for remote sensing image data; noise removal, ground point subdivision, and point cloud downsampling for lidar point cloud data; and registration between remote sensing image data and lidar point cloud data through coordinate transformation and registration methods. The denoising process for remote sensing image data is represented as follows: ; in, Indicates standard deviation, , These are respectively represented as their positions in the two-dimensional coordinate system of the image. Used for normalization to ensure that the sum of the filter weights is 1, without changing the overall brightness of the image; It is the value of the Gaussian function on a two-dimensional plane (image), used to construct Gaussian filters, which are used in image processing (especially denoising) to smooth images and reduce the impact of noise; The remote sensing image data is cropped and normalized as follows: ; in, Represents the original image. This represents the normalized image. Represents the image mean. Indicates the standard deviation of the image; Furthermore, to achieve the fusion of remote sensing image data and lidar point cloud data, a registration method is used to transform both into the same coordinate system, which can be represented as: Select starting point cloud: Set the source point cloud and target point cloud ,in, It is the total number of points in the source point cloud. It is the total number of points in the target point cloud. , These are the last points in the point cloud and the target point cloud, respectively.

[0027] Closest point matching: for source point cloud and target point cloud For each source point It is necessary to find the target point cloud. One of the closest points , is represented as: ; in, and It is the source cloud And the nearest point in the target point cloud , Represents the Euclidean distance between two points; by finding each source point A nearest point in the target point cloud Establish a set of corresponding points ; Transformation calculation: In each iteration, the optimal rigid transformation between the source point cloud and the target point cloud is calculated, expressed as: ; in, It is the objective function. It is a rotation matrix. It is a translation vector. It is the total number of points in the source point cloud; It is a parameter index used to identify the source point currently being processed during the summation process. Values ​​range from 1 to This is used to access each source point and calculate the error with the corresponding target point. The value ranges from 1 to m.

[0028] The rotation matrix and translation vector are calculated as follows: ; in, It is the centroid of the source point cloud, representing the average position of all points in the source point cloud, used to describe the overall position of the point cloud; This is the total number of points in the source point cloud. When calculating the centroid, the coordinates of all points are added together and then divided by this number. ; It is the index of the current point in the source point cloud. It is the centroid of the target point cloud; It is the total number of points in the target point cloud, calculated. At that time, the coordinates of all points will be added together and then divided by . ; It is the index of the current point in the target point cloud; Centralized point cloud: ; in, It is the first in the source cloud. The original coordinates of a point are usually represented in three-dimensional space, such as... ; This means subtracting the centroid of the source point cloud from the original coordinates of each point in the source point cloud to center the point cloud; These are the original coordinates of the j-th point in the target point cloud, typically represented as 3D coordinates such as... ; This means subtracting the centroid of the target point cloud from the original coordinates of each point in the target point cloud to center the corresponding point cloud; Calculate the rotation matrix : ; ; in, It is the first in the source point set The coordinate vector centered by the mean of each point. It is the mean-centered coordinate vector of the j-th point in the target point set. yes The transpose of is used to calculate the outer product; H is the total number of points in the source point cloud, and H is the value of the covariance matrix, obtained by summing the weighted outer product of the source point set and the target point set; SVD decomposes matrix H into matrix H. The product of the two points is used to calculate the optimal rotation matrix R between the two point sets. Calculate the translation vector : ; Update source point cloud: ; in, The matrix formed by the outer product of the centered point pairs of the source point cloud and the target point cloud. The matrix is ​​decomposed into a rotation matrix and a scaling matrix, using a new source point cloud. Proceed to the next iteration until the objective function is achieved. Convergence indicates that the source point cloud and the target point cloud have been registered. If the value of the objective function is lower than the set point cloud threshold after each iteration, the algorithm is considered to have converged. To prevent the algorithm from getting stuck in an infinite loop, a maximum number of iterations is set, and the algorithm is forcibly terminated when the number of iterations reaches the upper limit.

[0029] It should be further noted that the point cloud threshold is derived from experience and has a range of [range missing]. Magnitude, when satisfied When the point cloud registration error no longer decreases significantly, the algorithm is considered to have reached convergence. The threshold value is the point cloud threshold.

[0030] S2: Build and train and optimize the improved EfficientNet network model.

[0031] It should be noted that, such as Figure 1 As shown in step S2, the core of this step is to construct an improved EfficientNet network model specifically for identifying vegetation risks in power transmission corridors, and to train and optimize the model using preprocessed data.

[0032] This invention restructures and optimizes the training process based on the standard EfficientNet framework, making it more suitable for cross-modal fusion scenarios of remote sensing images and lidar point clouds, as well as the engineering needs of identifying and assessing vegetation hazards in power transmission corridors.

[0033] The improved EfficientNet network model achieves end-to-end feature fusion and extraction from preprocessed multimodal data by integrating a cross-modal attention fusion mechanism and multi-scale convolutional layers. While maintaining lightweight design and high inference efficiency, the improved network significantly enhances its sensitivity to minute spatial distances in vegetation and conductors, fine-grained textures, and height features. It also supports deep semantic alignment with 3D point cloud features, achieving unified output across multiple tasks and edge deployment capabilities.

[0034] Unlike the standard EfficientNet, which relies solely on the mobile-side inverse bottleneck convolutional block MBConv and composite scaling, this invention improves the structure from three aspects: convolutional structure, multimodal fusion, and output design.

[0035] Basic feature extraction layer: The input is a preprocessed (registered, cropped, normalized) true-color or multispectral image, with resolution determined by a composite scaling strategy. The initial layer uses 1×1 convolution + BN + SiSwish activation function to complete channel mapping and low-level semantic enhancement, providing a unified feature basis for multi-scale convolution.

[0036] Alternating multi-scale convolutional block groups: Enhance feature extraction capabilities through an alternating multi-scale convolution strategy.

[0037] The moving convolutional block consists of the following sequence: depthwise separable convolution (3×3 or 5×5, with large kernels that can include dilation), batch normalization (BN), activation function SiSwish, pointwise convolution (1×1), and batch normalization (BN).

[0038] It should be further explained that BN is an automatic regulating valve. It adjusts all the semi-finished products from the previous depthwise convolution process to uniform and stable specifications.

[0039] Unlike the fixed 3×3 depthwise convolution of standard EfficientNet, this invention adopts an alternating strategy of 3×3 and 5×5, and introduces large kernel dilated convolution to capture the unique canopy span and spatial extension direction characteristics of power transmission channels.

[0040] Based on the texture, shape, and geometric distribution of vegetation in power transmission corridors, this invention inserts convolutional kernels of different scales at different depths in the network backbone.

[0041] Shallow network layer (after Stem): (3×3)×2 stacked, used to enhance the capture of high-frequency information such as leaf texture and bark details.

[0042] The middle layer of the network (main feature extraction area): (5×5)×2 or ×3 stacked, using large kernels to model the crown geometry and power line direction structure; improve the ability to capture the "tree obstacle approach trend".

[0043] Top layer of the network (close to the output): Restores 3×3 to enhance semantic consistency and improve the stability of classification and risk regression heads.

[0044] By inserting convolutional kernels of different scales at different depths in the network backbone, an adaptive feature extraction path with texture enhancement, geometric expansion, and semantic convergence is achieved, making it more suitable for tree obstacle risk scenarios compared to the standard EfficientNet.

[0045] The improved EfficientNet network model achieves end-to-end feature fusion and feature extraction of preprocessed multimodal data by integrating a cross-modal attention fusion mechanism and multi-scale convolutional layers.

[0046] The cross-modal attention fusion mechanism is embedded in the convolutional backbone layer of the model and consists of four collaborative parts: feature extraction, cross-modal alignment, attention mechanism, and feature fusion. Specifically, it includes: Feature extraction involves using convolutional neural networks to extract multi-level features from remote sensing image data, while using the PointNet deep learning neural network architecture to extract three-dimensional spatial features from lidar point cloud data.

[0047] Cross-modal alignment uses coordinate transformation to map lidar point cloud data to a spatial scale consistent with remote sensing image data.

[0048] Attention mechanisms include self-attention and cross-modal attention.

[0049] Self-attention captures the relationships between local features through a self-attention mechanism, applying this mechanism to the input feature vector. Calculate self-attention weights , is represented as: ; in, It is the input feature vector matrix with dimension . , It is the number of features. It is the dimension of the feature; The transpose of the matrix has the following dimensions: ; It is used to calculate the similarity between features (dot product operation); Regarding feature dimensions Perform a square root operation, using it as a scaling factor to normalize the dot product results; apply it to each row. The function ensures that the sum of the weights in each row is 1, ultimately yielding the attention weight matrix. .

[0050] Cross-modal attention captures the relationship between remote sensing image data and lidar point cloud data through a cross-modal attention mechanism. It calculates the correlation between the features of the remote sensing image data and the features of the lidar point cloud data to obtain the cross-modal attention weights. , is represented as: ; in, Represents the feature matrix of a true-color remote sensing image. Represents the feature matrix of radar point cloud data. This indicates the similarity between features of true-color remote sensing images and features of radar point cloud data; Feature fusion uses cross-modal attention weights to weight and fuse features from remote sensing image data and lidar point cloud data, generating fused multi-source features, represented as: ; in, This indicates the features after fusion.

[0051] This invention restructures and optimizes the training process based on the standard EfficientNet framework, making it more suitable for cross-modal fusion scenarios of remote sensing images and lidar point clouds, as well as the engineering needs of identifying and assessing vegetation hazards in power transmission corridors.

[0052] The cross-modal attention mechanism is responsible for aligning the scale features of the image and point cloud in the semantic layer through the Cross-Modal Attention Fusion Unit (CMA). Its position is not placed in the input layer or the output layer, but inserted in the middle layer of the convolutional backbone. The feature scale here not only preserves the local spatial structure, but also contains preliminary semantic information, which is most suitable for alignment with the geometric features of the point cloud.

[0053] CMA internally includes: 1. Image feature query. 2. Geometric feature key / value obtained from the point cloud adapter. 3. Interactive attention matrix A(i,j): representing the semantic coupling strength between image pixel position i and point cloud structuring element j. Weighted fusion is performed through A(i,j) to enable the improved EfficientNet network model to actively pay attention to key risk areas such as canopy regions close to the wire spacing and rapidly growing trees.

[0054] The point cloud adapter aims to transform the three-dimensional spatial structure features in LiDAR point cloud data into two-dimensional feature representations that can be directly aligned with the layer features in a convolutional neural network, in order to support deep structural coupling of cross-modal attention fusion units.

[0055] The adapter first performs spatial segmentation and single-tree level extraction on the collected point cloud. It then uses the PointNet network to extract geometric descriptors such as height, crown width, trunk direction vector, local density gradient, and spatial PCA features from the point cloud set of each tree, enabling the network to capture the three-dimensional semantic structure related to tree barrier risk.

[0056] Subsequently, based on the spatial registration relationship between point cloud and image, the extracted three-dimensional geometric features are mapped to the image grid according to the projection coordinates. Through depth channel encoding and position embedding, the height difference, distance information and local morphology of point cloud in space are injected into the two-dimensional image feature space in the form of continuous features.

[0057] To achieve channel consistency with the backbone features of the improved EfficientNet network model, this invention further uses 1×1 convolution for channel adjustment, so that the point cloud adaptation features and the mid-layer features of the image have consistent scale, dimension and spatial encoding rules.

[0058] Through the above steps, a trainable mapping from the three-dimensional structure of point clouds to two-dimensional feature maps is achieved, enabling point cloud geometric information to participate in mid-level semantic modeling rather than just serving as post-processing input, significantly enhancing the expressive power of cross-modal feature fusion.

[0059] The multi-task output mechanism enables the multi-task output head to simultaneously perform three tasks: vegetation category identification, risk level regression, and hazard location, in order to adapt to the integrated business needs of identification, risk assessment, and location in the power transmission corridor scenario.

[0060] The classification output head employs fully connected layers and softmax activation to perform multi-class identification of vegetation species appearing in the input scene. Its weights share the mid-to-high-level semantic features of the backbone, enabling the model to distinguish canopy texture, color, and spatial structure of different tree species. The risk regression output head constructs a comprehensive risk scoring model based on feature vectors, outputting a risk score of 0 to 1, which can be automatically mapped to a three-level risk level according to a preset risk threshold. This continuous regression mechanism, compared to traditional discrete classification, better characterizes the non-linear relationship between tree height growth, canopy expansion, and changes in guide distance, thus providing higher early warning sensitivity for trees in a critical state.

[0061] The positioning output head uses heatmap regression or coordinate regression to predict the spatial location of high-risk trees in the image. By linking with the GIS coordinate transformation module, the recognition results are converted into actual latitude and longitude coordinates in the transmission line coordinate system, facilitating rapid location of potential hazards by on-site inspection personnel. This multi-task design not only integrates classification, risk assessment, and spatial positioning into an end-to-end training framework, but also enhances the information complementarity between tasks through shared convolutional backbones, achieving a comprehensive understanding and accurate representation of tree-related hazards along the transmission line.

[0062] The current improved structure maintains the efficiency of EfficientNet while being deeply customized for power transmission scenarios, laying a unified, stable, and scalable feature representation foundation for subsequent model training optimization and risk detection processes.

[0063] The training and optimization of the improved EfficientNet network model includes: The network architecture is optimized through composite scaling techniques to balance depth, width, and resolution. Simultaneously, the composite scaling strategy reallocates weights according to the scenario, achieving a more balanced and optimal configuration for the specific task.

[0064] Specifically, this involves enhancing depth and focusing on improving the expressive power of mid-layer large-kernel convolutions; The width is moderate, controlling the overall model complexity; Resolution adaptive, better adapting to the distance to tree obstacles.

[0065] This differs from the general scaling strategy of the standard EfficientNet, reflecting domain-specific optimizations.

[0066] The preprocessed data was divided into training and test sets. A multi-task learning joint optimization model was adopted, and training was completed by combining an adaptive estimation optimizer, learning rate decay and regularization strategies. Redundant parameters were removed by model pruning and computational precision was reduced by quantization to improve inference efficiency. Accuracy, recall, F1-score and precision were evaluated based on the test set to verify the performance of the model in vegetation classification and risk assessment tasks.

[0067] Through systematic improvements to the EfficientNet backbone structure, multimodal feature fusion mechanism, and task output format, the network possesses stronger geometric representation capabilities, cross-modal fusion capabilities, and task collaboration capabilities when processing remote sensing imagery and LiDAR point clouds. This enables accurate identification and quantitative assessment of vegetation hazards in power transmission corridors while maintaining a lightweight design. This improved structure retains the high efficiency of EfficientNet while deeply customizing key modules for power transmission scenarios, laying a unified, stable, and scalable feature representation foundation for subsequent model training optimization and risk detection processes.

[0068] Based on the improved model architecture defined in S2, a composite scaling technique is used to optimize the network's depth, width, and input resolution, achieving the optimal configuration for the vegetation risk identification task. The multimodal data preprocessed in S1 is divided into training and testing sets for training and independent evaluation of the improved EfficientNet network model.

[0069] The training process employs a multi-task learning strategy, using a composite loss function to jointly optimize vegetation classification and risk regression tasks.

[0070] like Figure 2 As shown, it is used to process true-color remote sensing images and lidar point cloud data in remote sensing image data.

[0071] Data input and partitioning involves taking the preprocessed data as input and dividing it into training and test sets.

[0072] The EfficientNet model training process includes loss function selection, model optimization, and training.

[0073] loss function Represented as: ; in, The cross-entropy loss for a classification task is expressed as: ; in, The mean squared error of the regression task is expressed as: ; in, Indicates the sample index. Indicates the first The true label of each sample Indicates the first The predicted probability of a sample. Indicates the number of categories. The true risk level is determined based on on-site measurement rules. The values ​​are based on tree height, minimum safe distance between the tree canopy and the transmission line, tilt angle, and growth health status indicators. After normalization, the values ​​are mapped to continuous values ​​in the range of [0,1] to reflect the true degree of potential threat of vegetation to the transmission line. The model predicts the risk level. The continuous output automatically generated by the regression head based on the fusion features of the improved EficientNet network model is located in the [0,1] interval and is used to characterize the model's estimation of the risk level of the input vegetation sample. The difference between the two is optimized by mean square error, so that the improved EficientNet network model gradually approaches the real risk distribution during the training process, and improves the sensitivity and quantification accuracy of critical tree barrier hazards. Indicates the number of samples. This represents the trade-off coefficient between the classification task and the regression task. This means taking the logarithm of the predicted probability to make the loss function more consistent with the optimization objective of maximum likelihood estimation.

[0074] The improved EficientNet network model optimizations include employing an adaptive estimation optimizer, a learning rate decay strategy, setting the batch size, and regularization techniques. The optimization process includes inputting image data into the network and extracting features using the improved EficientNet network model; using the model's output to perform joint calculations for classification and regression tasks to obtain classification probabilities and risk level predictions; calculating the error based on the loss function and adjusting the network weights through backpropagation. The improved EficientNet network model optimization includes improved EficientNet network model pruning and improved EficientNet network model quantization. Improved EficientNet network model pruning involves pruning the optimized EficientNet network model by sorting the weights of neurons and removing neurons with small weights. Improved EficientNet network model quantization involves converting the floating-point parameters of the improved EficientNet network model to low-precision integers. During training, the learning rate is initially set to 0.001 and decays using a learning rate decay mechanism. The batch size is 64 images per batch. The training epochs consist of 20 training cycles, with the learning rate gradually decreasing.

[0075] The evaluation of the optimized and improved EfficientNet network model includes assessing the model's performance using a test set after training, including evaluating accuracy, recall, F1 score, and precision; the optimized and improved EfficientNet network model is derived from the evaluation results.

[0076] Furthermore, accuracy Represented as: ; Recall rate Represented as: ; F1 score is represented as: ; Accuracy Represented as: ; in, Indicates the true value. Indicates a true negative value. Indicates a false positive example. This indicates a false negative.

[0077] S3: Use the improved EfficientNet network model after training and optimization to perform vegetation classification and hazard detection on the input image.

[0078] It should be noted that the vegetation classification and hazard detection identifies the vegetation type of the power transmission corridor through remote sensing image data and lidar point cloud data, and assesses the risk level of the vegetation based on the comprehensive risk score; when the comprehensive risk score is greater than or equal to the corresponding preset risk threshold, the tree obstacle risk hazard warning signal is automatically triggered; the risk level assessment is completed based on the distance between the tree and the power transmission line, the tree height, the tree tilt angle, and the tree growth status.

[0079] Furthermore, the risk level assessment considers the distance between the trees and the power transmission lines, the height of the trees, the angle of inclination of the trees, and the growth status of the trees; The distance between trees and power transmission lines is expressed as: ; in, This indicates the horizontal distance between trees and power lines. The x-coordinate of the tree's spatial coordinates. The vertical coordinate of the tree's spatial coordinates. The x-coordinate of the transmission line coordinates These are the x and y coordinates of the transmission line.

[0080] If the distance between the trees is less than a preset height threshold, there is a risk of them colliding with the power transmission lines.

[0081] If a tree's height exceeds a set height threshold, it will touch the power transmission line.

[0082] Preset height thresholds for trees are set according to different tree species, climate conditions and power industry standards. For example, trees that exceed 10 meters need to be monitored.

[0083] The angle of tree tilt is expressed as: ; in, Indicates the angle of tree tilt. Indicates the height of the treetop. Indicates the height of the tree roots. Indicates the horizontal distance from the base to the top of the tree. ( ) is the tangent function The inverse function is used to calculate the angle corresponding to a certain tangent value. Here, the angle value is calculated from the known ratio of horizontal projection to height, thus obtaining the tree's tilt angle. Tree growth status is assessed by evaluating the degree of mortality and pest and disease conditions to obtain a corresponding health score, which is used to quantify the impact of vegetation physiological activity on risk. Among them, dead, weakened, or obviously diseased plants have significantly lower health scores, thus having an enhancing effect on the final risk value in the comprehensive risk scoring model.

[0084] Comprehensive Risk Score Represented as: ; in, Indicates tree height, This system represents the health status score of trees, calculated using multi-source information such as remote sensing imagery, hyperspectral data, and lidar point clouds, through feature extraction and evaluation methods. It utilizes leaf spectral reflectance characteristics, canopy density, foliage integrity, and pest and disease identification results, combined with predefined scoring rules, to map trees of different health states to quantitative values, typically set between 0 and 1, where 0 represents complete death or severe pest and disease damage, and 1 represents healthy growth. Through this method, each tree can obtain a corresponding health status score, which serves as one of the input factors for comprehensive risk calculation. , , The weighting coefficients are for the horizontal distance between the tree and the power transmission line, the tree height, the tree tilt angle, and the tree's health status score. The values ​​of each weighting coefficient can be set between 0 and 1, and the sum of all weights can be normalized to 1 to ensure the stability and comparability of the comprehensive score. The relationship between tree height and the probability of accident occurrence can be fitted by historical inspection data to obtain the weights of horizontal distance, tree height, and tree tilt angle. The weight of the tree's health status score is obtained by analyzing the correlation between the health status score and the potential incident. Thus, the final comprehensive risk score reflects both the physical characteristics of the tree and the impact of physiological activity on the risk.

[0085] It should be further explained that the tree growth health status score achieves its mapping from multi-source features to quantitative values ​​by combining predefined scoring rules.

[0086] The predefined scoring rules are implemented through a quantitative scoring system, which standardizes and scores four core features extracted from remote sensing data: NDVI (Normalized Differential Vegetation Index) assesses vegetation greenness, with the first assessment threshold set at [0.3, 0.8]; Red Edge Position Ratio (REIP) monitors chlorophyll content, with the second assessment threshold set at [725, 735] nanometers; Tree height extracted by lidar characterizes growth status, with the third assessment threshold set at [8, 15] meters; Canopy integrity is calculated through point cloud penetration rate, with the fourth assessment threshold set at [0.6, 0.9]. Each feature uses the same scoring logic: a measured value below the corresponding lower assessment threshold receives 0 points, and a value above the corresponding upper assessment threshold receives 1 point. Within the interval, a linear formula is used for calculation: Assessment score of the corresponding feature = (Measured value - Lower limit) / (Upper limit - Lower limit).

[0087] The comprehensive score of vegetation health status is calculated using a weighted summation formula, the specific expression of which is: , in, The score is a comprehensive score in the range of 0-1, where Sᵢ represents the evaluation score of the corresponding feature, and L... i The corresponding weights satisfy the constraint that the sum is 1.

[0088] The four core features include NDVI, which reflects vegetation greenness and biomass; Red Edge Position Ratio (REIP), which monitors chlorophyll content; Tree height extracted by lidar, which characterizes growth status; and Canopy integrity, which is measured by point cloud penetration rate.

[0089] S4: Input the collected real-time data into the comprehensive risk scoring model for hazard detection for calculation, and conduct risk warning and location based on the calculation results.

[0090] Furthermore, risk and hazard early warning includes preset risk threshold settings, early warning triggering mechanisms, and early warning responses.

[0091] The preset risk thresholds are set based on a comprehensive risk score; the higher the score, the greater the risk the vegetation poses to the power transmission lines. The preset risk thresholds include low-risk, medium-risk, and high-risk thresholds.

[0092] The low-risk threshold is a comprehensive risk score between 0 and 0.3, indicating that the distance between the trees and the power transmission lines is relatively far, the trees are in good health, and their height or tilt angle does not pose a threat; no immediate intervention is required, and regular inspections are sufficient.

[0093] The medium-risk threshold is a comprehensive risk score between 0.3 and 0.6, indicating that the trees may be close to power lines or are in a critical stage of growth and require monitoring.

[0094] The high-risk threshold is a comprehensive risk score higher than 0.6, indicating that the trees pose a significant threat to the power transmission lines and may touch or affect the operation of the power transmission lines, triggering an early warning from the system.

[0095] Early warning responses are communicated to inspection personnel via GIS (Geographic Information System) map display, SMS, email, and reports.

[0096] Risk and hazard location is determined through spatial positioning technology (combining image data, point cloud data, and vegetation risk analysis results, and using a geographic information system (GIS)) and precise location calibration and visualization (using precise spatial positioning technology, specific spatial coordinates can be provided for each high-risk tree and displayed on a GIS map).

[0097] Example 2 is an embodiment of the present invention, which provides a vegetation image recognition method based on EfficientNet. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0098] The optimized and improved EfficientNet network model was used to classify nine vegetation types, and the precision, recall, and F1 score for each vegetation type were calculated. Table 1 below shows the evaluation results of the classification model for each vegetation type. Table 1 Evaluation Results , The accuracy rate of the improved EfficientNet network model is above 96% for all tree species, demonstrating its stability and efficiency in vegetation classification. The highest accuracy rate is achieved for Platycladus orientalis, reaching 99%, indicating that the characteristics of this tree species are relatively easier for the model to recognize.

[0099] Recall rate reflects the model's ability to identify positive samples. The recall rate of all tree species exceeded 90%, with paper mulberry having the highest recall rate at 96.5%. This indicates that the model can identify all vegetation in most vegetation classifications, especially for species with relatively stable tree growth.

[0100] The F1 score takes into account both precision and recall, evaluating the balance between these two factors in the model. All tree species achieved F1 scores above 94%, demonstrating good balance, with the paper mulberry and Chinese arborvitae showing F1 scores exceeding 97%.

[0101] The risk factors were compared with preset thresholds to obtain a risk score for each tree. Table 2 below shows the risk assessment results for each type of vegetation: Table 2 Risk Assessment Results , The experiment revealed that tung trees and pine trees had high risk scores, both exceeding 0.5, indicating a high risk. Tung trees are relatively tall and have a significant tilt angle, while pine trees are also tall and located close to power transmission lines, thus posing a potential hazard of tree obstruction.

[0102] Medium-risk trees include eucalyptus, paper mulberry, cypress, and fir, with risk scores between 0.3 and 0.5, requiring regular monitoring. Paper mulberry and fir have high growth potential, and their future growth should be closely monitored.

[0103] Low-risk trees such as camphor trees, oak trees, and bamboo have low risk scores and are considered low-risk trees, posing no significant threat.

[0104] Based on the risk assessment results, spatial positioning technology was used to accurately mark the specific locations of high-risk trees on a GIS map to locate potential hazards. Table 3 below shows the location results of high-risk trees: Table 3 Location Results , Experiments showed that the optimized EfficientNet network model performed excellently in vegetation classification, with an accuracy rate generally exceeding 95%, ensuring the accuracy of subsequent risk assessments. By comprehensively considering factors such as the distance, height, and tilt angle between trees and power transmission lines, the risk level of trees can be effectively assessed, and high-risk vegetation can be identified. Using GIS maps and spatial positioning technology, the location of high-risk trees can be accurately pinpointed, providing inspection personnel with precise risk location information and reducing the blind spots of manual inspections.

[0105] Example 3 is the third embodiment of the present invention, which differs from the previous two embodiments in that: If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] The logic and / or steps otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0107] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0108] Example 4 is the fourth embodiment of the present invention. This embodiment provides a vegetation image recognition system based on EfficientNet, characterized in that it includes: an acquisition and preprocessing module, a training and optimization module, a classification and detection module, and an early warning and positioning module. The acquisition and preprocessing module acquires data through image acquisition and synchronous LiDAR acquisition, and preprocesses the acquired data. The training optimization module is used to build and train and optimize the improved EfficientNet network model. The classification and detection module uses the improved EfficientNet network model, after training and optimization, to classify vegetation and detect potential hazards in the input image. The early warning and positioning module inputs the collected real-time data into the comprehensive risk scoring model for hazard detection for calculation, and performs risk early warning and positioning based on the calculation results.

[0109] This embodiment also provides a computing device applicable to the EfficientNet vegetation image recognition method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the EfficientNet vegetation image recognition method proposed in the above embodiment.

[0110] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the vegetation image recognition method based on EfficientNet as proposed in the above embodiments.

[0111] The storage medium proposed in this embodiment belongs to the same inventive concept as the vegetation image recognition method based on EfficientNet proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0112] 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 it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A vegetation image recognition method based on EfficientNet, characterized in that: include: The data is acquired through image acquisition and simultaneous use of LiDAR, and the acquired data is preprocessed. We construct and train and optimize the improved EfficientNet network model. The improved EfficientNet network model integrates a cross-modal attention fusion mechanism and multi-scale convolutional layers to perform end-to-end feature fusion and feature extraction on preprocessed multimodal data. The improved EfficientNet network model, after training and optimization, is used to classify vegetation and detect potential hazards in the input image. The hazard detection is based on a comprehensive risk scoring model constructed from tree height, distance between the tree and the power transmission line, tree health status score, and tree tilt angle. The collected real-time data is input into the comprehensive risk scoring model for hazard detection for calculation, and risk warning and location are carried out based on the calculation results.

2. The vegetation image recognition method based on EfficientNet as described in claim 1, characterized in that: The process of acquiring data through image acquisition and simultaneously using LiDAR to acquire data, and preprocessing the acquired data, includes: Vegetation data within the power transmission corridor was collected using remote sensing images and lidar point cloud technology, including remote sensing image data and lidar point cloud data. Remote sensing image data includes true-color remote sensing images and multispectral images; LiDAR point cloud data includes radar point cloud data.

3. The vegetation image recognition method based on EfficientNet as described in claim 2, characterized in that: The preprocessing includes denoising the remote sensing image data, image cropping and normalization, image enhancement, and color space conversion to generate standardized image data suitable for feature extraction using alternating multi-scale convolutional layers. The system performs noise removal, ground point subdivision, and point cloud downsampling on lidar point cloud data to provide point cloud data suitable for 3D feature extraction by a point cloud adapter. It also completes the registration of remote sensing image data and lidar point cloud data through coordinate transformation and registration methods, providing spatially aligned multimodal data input for cross-modal attention fusion units.

4. The vegetation image recognition method based on EfficientNet as described in claim 3, characterized in that: The construction and training / optimization of the improved EfficientNet network model includes, An improved EfficientNet network model is formed by integrating a cross-modal attention fusion mechanism and multi-scale convolutional layers into the basic EfficientNet network model, which performs end-to-end feature fusion and feature extraction on preprocessed multimodal data. The cross-modal attention fusion mechanism is implemented by incorporating a point cloud adapter and a cross-modal attention unit; Point cloud adapters are used to convert LiDAR point cloud data into two-dimensional geometric features; Cross-modal attention units are used to fuse image features and two-dimensional geometric features based on attention weights; Multi-scale convolutional layers are implemented through an alternating multi-scale convolution strategy, targeting vegetation risk identification tasks. Different sizes of convolutional kernels are configured in the shallow, medium, and deep layers of the network; Shallow layers use small-scale convolution kernels to focus on capturing texture details at the leaf level of vegetation; The middle layer uses large-scale convolutional kernels to expand the receptive field for modeling tree canopy geometry and its spatial relationship with transmission lines; The top layer continues to use small-scale convolutional kernels to stabilize semantic-level feature representations.

5. The vegetation image recognition method based on EfficientNet as described in claim 4, characterized in that: The construction and training / optimization of the improved EfficientNet network model also includes, The improved EfficientNet network model was trained and optimized. The network architecture was optimized through compound scaling technology to balance depth, width and resolution in processing multimodal fusion of true-color remote sensing images and LiDAR point cloud data. The preprocessed data is divided into training and test sets. A multi-task learning joint optimization model is adopted, which combines an adaptive estimation optimizer, learning rate decay and regularization strategy to complete the training. The optimized and improved EfficientNet network model removes redundant parameters through pruning and reduces computational precision through quantization to improve inference efficiency. The accuracy, recall, F1-score, and precision metrics were evaluated using a test set to validate the model's performance in vegetation classification and risk assessment tasks.

6. The vegetation image recognition method based on EfficientNet as described in claim 5, characterized in that: The process of using the improved EfficientNet network model after training and optimization to classify vegetation and detect potential hazards in the input image involves obtaining the comprehensive risk score and spatial location information synchronously output by the improved EfficientNet network model after training and optimization. Vegetation classification identifies vegetation types in power transmission corridors using remote sensing image data and lidar point cloud data. The risk level of vegetation is assessed based on the comprehensive risk score. When the comprehensive risk score is greater than or equal to the corresponding preset risk threshold, a tree obstacle risk warning signal is automatically triggered. Risk level assessment is completed based on the distance between the trees and the power transmission lines, the height of the trees, the angle of inclination of the trees, and the growth status of the trees; The distance between trees and power transmission lines is expressed as: ， in, This indicates the horizontal distance between trees and power transmission lines. The x-coordinate of the tree's spatial coordinates. The vertical coordinate of the tree's spatial coordinates. The x-coordinate of the transmission line coordinates The x and y coordinates represent the coordinates of the transmission line. If the distance between the trees is less than the preset height threshold, there is a risk of them colliding with the power transmission line. If the height of a tree exceeds a preset height threshold, it will touch the power transmission line; Preset height thresholds for trees are set according to different tree species, climate conditions, and power industry standards; The angle of tree tilt is expressed as: ， in, The angle at which the trees are tilted. The height of the treetop. The height of the tree roots. The horizontal distance from the base to the top of the tree. It is the inverse function of the tangent function; Tree growth status is assessed by evaluating the degree of tree mortality and pest and disease conditions to obtain a corresponding health score, thus quantifying the impact of vegetation physiological activity on risk. The comprehensive risk score is expressed as: ， in, For comprehensive risk scoring, Tree height, Score the health status of tree growth. These are the weighting coefficients for the horizontal distance between the tree and the power transmission line, the tree height, the tree tilt angle, and the tree's growth health status score.

7. The vegetation image recognition method based on EfficientNet as described in claim 6, characterized in that: The risk warning and location system includes risk hazard warning and risk hazard location. The risk and hazard early warning system sets three preset risk thresholds, including low risk, medium risk, and high risk thresholds. It makes judgments based on a comprehensive risk score. When the comprehensive risk score is greater than or equal to the corresponding preset risk threshold, an early warning is automatically triggered and the inspection personnel are notified through GIS map marking, SMS, and email. Risk and hazard location is based on the spatial location information synchronously output by the optimized and improved EfficientNet network model. The geographical coordinates of the trees with potential hazards are marked by spatial location technology and then visualized on a GIS map.

8. An EfficientNet-based vegetation image recognition system, using the EfficientNet-based vegetation image recognition method as described in any one of claims 1 to 7, characterized in that it includes: an acquisition and preprocessing module, a training and optimization module, a classification and detection module, and an early warning and positioning module; The acquisition and preprocessing module acquires data through image acquisition and synchronous LiDAR acquisition, and preprocesses the acquired data. The training optimization module is used to build and train and optimize the improved EfficientNet network model. The classification and detection module uses the improved EfficientNet network model, after training and optimization, to classify vegetation and detect potential hazards in the input image. The early warning and positioning module inputs the collected real-time data into the comprehensive risk scoring model for hazard detection for calculation, and performs risk early warning and positioning based on the calculation results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the vegetation image recognition method based on any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the vegetation image recognition method based on any one of claims 1 to 7.

Images