Method for extracting tree information in power transmission line channel based on multi-source remote sensing data

Abstract

The invention discloses a method for extracting tree information in a power transmission line channel based on multi-source remote sensing data, and relates to the technical field of remote sensing data processing, and the method comprises the following steps: screening a tree distribution space based on the spectral characteristics of a target region, and inputting image data with the tree distribution space as a constraint into an improved D-UNet model; a convolution attention branch in the improved D-UNet model is used to generate an attention feature map; extracting global and local context features by improving parallel multi-scale pooling branches in a D-UNet model; performing weighted fusion on the context features by using the attention feature map to generate context enhanced features, and outputting a tree species classification result after decoding; and generating comprehensive tree information in combination with the tree species classification result and the tree geometric height obtained based on the target area point cloud. The method is used for solving the problem that the refined hidden danger evaluation requirement is difficult to meet due to low tree identification precision in the prior art.

Description

Technical Field

[0001] This invention relates to the field of remote sensing data processing technology, and more specifically, to a method for extracting tree information in power transmission line corridors based on multi-source remote sensing data. Background Technology

[0002] With the advancement of smart grid construction, using drones equipped with multi-source remote sensing equipment to inspect trees in power transmission line corridors has become the mainstream approach.

[0003] However, in the face of long-distance channels and complex surface environments, existing tree information extraction technologies still have the following problems in practical applications: First, the growth status and distribution patterns of trees in transmission line channels vary significantly. The existing segmentation network has a fixed receptive field, making it difficult to take into account both large-scale trees and small-scale seedlings, resulting in the omission of small-scale trees or the fragmentation of large tree crown edges and internal cavities. Second, the environment of transmission channels is complex, with rich background noise from ground shrubs, weeds, shadows, and artificial facilities. Background interference can easily lead to rough tree edge segmentation or misclassification, making it difficult to meet the technical requirements for refined hidden danger assessment and differentiated management in power operation and maintenance.

[0004] Therefore, there is an urgent need for a method for extracting tree information within power transmission line corridors that can achieve refined extraction of tree features at multiple scales. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for extracting tree information in power transmission line corridors based on multi-source remote sensing data. By utilizing the convolutional attention and multi-scale pooling branches of the improved D-UNet model, the method solves the problem that the low tree recognition accuracy of the prior art makes it difficult to meet the requirements of refined hazard assessment.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for extracting tree information within power transmission line corridors based on multi-source remote sensing data includes the following steps: Tree distribution space is screened based on the spectral features of the target area, and image data constrained by tree distribution space is input into the improved D-UNet model; An attention feature map is generated by utilizing the convolutional attention branch in the improved D-UNet model. By improving the parallel multi-scale pooling branch in the D-UNet model, global and local contextual features are extracted. The attention feature map is used to perform weighted fusion of the context features to generate context-enhanced features, and the tree species classification result is output after decoding. By combining the tree species classification results with the tree geometric heights obtained from the point cloud of the target area, comprehensive tree information is generated.

[0007] In a preferred embodiment, the process of screening the tree distribution space includes: acquiring spectral image data of the target area and extracting spectral reflectance information; calculating the vegetation index of each pixel in the spectral image data and constructing a spectral feature set based on the spectral reflectance information and the vegetation index; using a pre-constructed land cover recognition model to perform binary classification of each pixel in the spectral image data into trees and non-trees based on the spectral feature set; and extracting the pixel regions whose classification results are classified as trees to generate the tree distribution space.

[0008] In a preferred embodiment, the convolutional attention branch includes: performing a convolution operation on the high-level semantic feature map output by the encoder in the improved D-UNet model to extract a local feature map containing local information; and learning weights for the channel dimension and spatial dimension of the local feature map based on the convolutional attention mechanism to generate an attention feature map.

[0009] In a preferred embodiment, the multi-scale pooling branch includes a parallel global pooling branch and an average pooling branch; the global pooling branch performs global average pooling and upsampling on the high-level semantic feature map to obtain global contextual features; the average pooling branch performs local average pooling and upsampling on the high-level semantic feature map to obtain local contextual features.

[0010] In a preferred embodiment, generating context-enhanced features includes: performing element-wise weighted fusion of the global context features and the local context features using the attention feature map; performing channel concatenation and convolutional compression on the fused features to obtain multi-scale context information; and performing residual fusion of the multi-scale context information with the high-level semantic feature map to generate context-enhanced features.

[0011] In a preferred embodiment, the tree species classification result includes: fusing the context enhancement features with the corresponding level of the encoded feature map in the encoder; classifying the channels of the fused feature map into tree species feature channels based on a preset number of tree species categories and generating a probability score for the corresponding tree species; normalizing the probability scores to obtain the tree species probability of each pixel in the fused feature map, and then generating the tree species classification result.

[0012] In a preferred embodiment, the tree geometric height includes: acquiring point cloud data of a target area, and using the tree distribution space for spatial filtering to generate a tree area point cloud; separating ground points and non-ground points in the tree area point cloud, and calculating the elevation difference between the non-ground points and the ground points as the tree geometric height.

[0013] The technical effects and advantages of the method for extracting tree information in power transmission line corridors based on multi-source remote sensing data of the present invention are as follows: 1. This invention utilizes convolutional attention branches to generate attention feature maps. By suppressing non-tree background noise responses and enhancing the feature representation of tree textures and boundaries, it reduces the false detection rate caused by the complex background of the power transmission channel environment. While improving the segmentation accuracy and edge integrity under complex lighting and surface environments, it further enhances the discrimination robustness and semantic purity of features under complex surface environments.

[0014] 2. This invention utilizes multi-scale pooling branches to extract global and local contextual features, ensuring the consistency of semantics within large-scale trees, avoiding fragmentation or voids in large tree canopy segmentation, and enhancing the ability to capture small-scale seedlings or scattered trees. This enables refined extraction of full-scale tree information within the power transmission channel, ensuring the precision and reliability of the final output comprehensive tree information. Attached Figure Description

[0015] Figure 1 This is a schematic flowchart of a method for extracting tree information in a power transmission line corridor based on multi-source remote sensing data, provided by an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of the overall architecture of the improved D-UNet model in an embodiment of the present invention.

[0017] Figure 3 This is a schematic diagram of the structure of the DAPP module in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1, Figure 1 This invention presents a method for extracting tree information within power transmission line corridors based on multi-source remote sensing data, comprising the following steps: S1. Based on the spectral characteristics of the target area, the tree distribution space is screened, and the image data constrained by the tree distribution space is input into the improved D-UNet model; S101. Acquire multispectral imagery, UAV imagery, and LiDAR point cloud data of the target area, as detailed below: The original multispectral images of the target area for power transmission line construction, including the red and near-infrared bands, were acquired by satellites such as GF-2, Landsat-8, and Sentinel-2. Radiometric correction, atmospheric correction, and geometric correction were performed on the original multispectral images to eliminate the effects of sensor noise and atmospheric scattering. The images were then cropped and registered to obtain the multispectral images. A drone equipped with a high-resolution camera is used to acquire low-altitude images of the target area, obtaining the original drone images of the target area; the original drone images are then subjected to orthorectification, image stitching and radiometric equalization to generate high-resolution orthorectified images, which are used as drone images. The raw 3D point cloud data of the target area is acquired by the lidar system carried by the UAV; noise points are removed and ground points are filtered on the raw 3D point cloud data to obtain preprocessed LiDAR point cloud data. Through coordinate transformation and projection matching, the multispectral imagery, UAV imagery, and LiDAR point cloud data are unified into the CGCS2000 geographic coordinate system.

[0020] S102. Acquire spectral image data of the target area, extract spectral reflectance information, calculate the vegetation index of each pixel in the spectral image data, and construct a spectral feature set based on the spectral reflectance information and the vegetation index, as follows: Select typical ground feature samples containing both trees and non-trees from the multispectral images, such as trees, grasslands, water bodies, bare land, roads, buildings, etc. From the multispectral image, the reflectance information of the typical ground cover samples in the red and near-infrared bands is extracted, and the normalized vegetation index is calculated. The calculation formula is as follows: , in, Normalized Difference Vegetation Index (NDVI) This represents the reflectance value in the near-infrared band. This represents the reflectivity value in the red light band. The reflectance information of ground cover samples is combined with the normalized vegetation index to form a multidimensional feature vector. The multidimensional feature vectors of all typical ground cover samples are combined to form a typical spectral feature set. The multidimensional feature vectors of all pixels in the multispectral image are combined to form a spectral feature set. S103. Based on the spectral feature set, a pre-constructed land cover recognition model is used to perform binary classification of trees and non-trees on each pixel of the spectral image data, extracting the pixel regions that are classified as trees, and generating the tree distribution space, as follows: The typical spectral feature set is randomly divided into a training set and a validation set according to a preset ratio, for example, a preset ratio of 7:3; The training set is input into a Support Vector Machine (SVM) model. A kernel function, specifically an RBF function, is used to map the low-dimensional spectral feature vectors to a high-dimensional feature space. The optimal hyperplane that maximizes the distinction between trees and non-trees is then searched in the high-dimensional feature space. During the search, a grid search combined with K-fold cross-validation is used to iteratively optimize the penalty coefficient and kernel function parameters of the SVM. Training stops when the loss function of the SVM converges, and the model with the current penalty coefficient and kernel function parameters is used as the land cover recognition model. The multidimensional feature vectors of all pixels in the multispectral image are combined into a spectral feature set. The land cover recognition model is then used to perform binary classification of trees and non-trees and generate land cover category labels. Extract the pixel region labeled "tree" and set the pixel value within the "tree" region to a specific identifier value, such as 1, while setting the rest of the region to a background value, such as 0, thereby generating a binary mask image that identifies the spatial location of trees, i.e., the tree distribution space.

[0021] S104. Improve the D-UNet model by taking the image input constrained by the tree distribution space, as follows: On the CGCS2000 geographic coordinates, the tree distribution space is used as the spatial constraint of the UAV image. The UAV image data of the tree area within the tree distribution space is retained, and the pixel values ​​of the background area outside the space are uniformly set to the background value. The UAV image after the tree distribution space constraint processing is used as the input image of the improved D-UNet model.

[0022] This step utilizes spectral features to generate a tree distribution space as a spatial constraint, reducing the subsequent computational load and suppressing the false alarm rate in complex backgrounds, thereby improving the overall detection efficiency and accuracy.

[0023] S2. Generate attention feature maps by utilizing the convolutional attention branch in the improved D-UNet model; S201. Construct and train the improved D-UNet model, as follows: like Figure 2 As shown, the improved D-UNet model includes an encoder, a multi-scale information extraction (DAPP) module, and a decoder. The encoder consists of four cascaded convolutional layers with identical internal structures. Each layer contains two sequentially stacked convolutional block units [(Conv3×3→BN→ReLU)×2]. The 3×3 convolution is used to extract spatial texture features within the local neighborhood; BN (Batch Normalization) is used to normalize the features; and the ReLU activation function is used to introduce non-linearity. The decoder has a symmetrical structure with the encoder. The drone imagery, which is located in the same area as the tree distribution space, is combined with visual features such as the color, texture, and crown shape of ground features in the drone imagery, as well as the actual distribution of local tree species. The different tree species are distinguished by manual visual interpretation and pixel-level semantic tags are created. The number of target classification categories to be identified is determined based on the semantic labels of different tree species. For example, the categories include: purple-leaf plum, fir, oak, paulownia, chestnut, mulberry, bamboo forest, and background, etc. Using a sliding window cropping technique, the UAV images and corresponding pixel-level semantic labels are cropped in pairs into fixed-size image blocks, such as 256×256 pixels. The cropped image blocks are then randomly divided into training and validation sets according to a preset ratio, which can be 8:2. Data augmentation operations are performed on the training set, such as horizontal flipping, vertical flipping, and random rotation, to expand sample diversity and improve the model's generalization ability. The improved D-UNet model is trained end-to-end using the training set to determine the network weight parameters, as follows: Set the training batch size, initial learning rate, and maximum number of iterations. For example, the training batch size can be 8, the initial learning rate can be 0.001, and the maximum number of iterations can be 100. Use a multi-class cross-entropy loss function to calculate the difference between the model's predictions and the true labels. Use the adaptive moment estimator optimizer (Adam) to backpropagate based on the calculated gradients and update the convolutional kernel parameters and bias terms. During training, use the validation set to periodically evaluate the model's accuracy. When the validation set loss function converges or the accuracy no longer improves, stop training and save the model with the current optimal weight parameters as the final improved D-UNet model.

[0024] S202. The attention feature map is generated in S2 as follows: The high-level semantic feature map output by the encoder is specifically achieved by: downsampling the feature data output by the previous layer through max pooling between two adjacent convolutional layers; with each pooling operation, the spatial resolution of the feature data output by the previous layer is halved, that is, the height and width become half of the original, while the channel dimension becomes twice the number of channels of the previous layer; the input image of the improved D-UNet model is processed sequentially through "convolutional layer 1 → pooling → convolutional layer 2 → pooling → ... → convolutional layer 4", retaining the feature data at the corresponding resolution at the end of each convolutional layer as the encoding feature map of the current layer, and using the encoding feature map of the last layer as the high-level semantic feature map; like Figure 3 As shown, the DAPP module includes four parallel convolutional attention branches, a global pooling branch, an average pooling branch, a 1×1 convolution branch, and a main input and residual branch. The high-level semantic feature map output by the encoder is used as the input to the DAPP module through the input and residual main branch, and distributed to the parallel branch for processing. At the same time, the high-level semantic feature map is passed to the end of the DAPP module through a direct connection path to preserve the original semantic information. Local contextual information of the high-level semantic feature map is extracted through a 3×3 convolutional layer in the convolutional attention branch to obtain a local feature map containing local information; the local feature map is then input into the convolutional attention (CBAM) module, and two-level weight learning of channel attention and spatial attention is performed sequentially to obtain the attention feature map; Simultaneously, a 1×1 convolution operation is performed on the high-level semantic feature map through a 1×1 convolution branch to complete linear transformation and channel rearrangement, resulting in a simplified feature representation.

[0025] This step introduces a convolutional attention mechanism to learn dual channel and spatial weights, which suppresses noise interference from complex backgrounds, enhances the feature representation of tree textures and boundaries, and improves the model's feature discrimination robustness and semantic purity in complex terrain environments.

[0026] S3. By improving the parallel multi-scale pooling branch in the D-UNet model, global and local contextual features are extracted, as follows: The global pooling branch specifically includes: performing a global average pooling operation on the high-level semantic feature map to obtain global statistical features describing the overall response of the channel dimension; and upsampling the global statistical features to restore them to the same spatial size as the high-level semantic feature map to generate global contextual features that are uniformly spread in space but have obvious differences between channels. The average pooling branch specifically includes: performing local average pooling on the high-level semantic feature map to obtain local statistical features, such as local texture intensity and regional average reflectance; upsampling the local statistical features to restore them to the same spatial size as the high-level semantic feature map to generate local contextual features.

[0027] This step extracts global and local contextual features through parallel multi-scale pooling branches, overcoming the limitations of a fixed receptive field, ensuring semantic consistency within large-scale trees to avoid segmentation gaps, enhancing the ability to capture small-scale seedlings, and achieving refined extraction of tree information across all scales.

[0028] S4. Use the attention feature map to perform weighted fusion of the context features to generate context-enhanced features, and output the tree species classification result after decoding; In this embodiment, generating context-enhanced features in S4 includes: The attention feature map is used as a weight, and then multiplied element-wise with the global context feature and the local context feature, respectively. Attention-weighted global and local context features, attention feature maps themselves, and simplified feature representations are concatenated and aggregated along the channel dimension. The concatenated aggregated features are input into a 1×1 convolutional layer for feature compression and fusion to obtain multi-scale contextual information containing multi-scale attention information. The multi-scale contextual information is added element-wise to the high-level semantic feature map directly transmitted from the input and the residual main branch to achieve residual fusion, thereby obtaining context-enhanced features.

[0029] In this embodiment, the tree species classification results output in S4 include: S401. The context enhancement features are fused with the corresponding layer's encoded feature map in the encoder, as follows: The context-enhanced features are used as input features to the decoder in the improved D-UNet model; At each decoding level, the input features are upsampled to double their spatial size to match the size of the encoded feature map at the corresponding level in the encoder. Simultaneously, the upsampled input features are used as the decoded feature map for the current level. By using a skip connection mechanism, the decoded feature map is concatenated with the encoded feature map of the corresponding level of the encoder through channel splicing; The concatenated feature map is input into the convolutional block unit (structure Conv3×3→BN→ReLU×2) inside the decoder for convolutional fusion and refinement; Repeat the process of "upsampling → channel stitching → convolutional fusion" until the spatial resolution of the feature map after convolutional fusion is the same as that of the UAV image before spatial constraints.

[0030] S402. Based on the preset number of tree species categories, classify the channels of the fused feature map into tree species feature channels and generate corresponding probability scores for each tree species; normalize the probability scores to obtain the tree species probability of each pixel in the fused feature map, and then generate the tree species classification result, as follows: The final feature map after convolutional fusion is input into a classification layer consisting of a 1×1 convolutional layer. The number of channels in the feature map is converted into tree species feature channels that are the same as the number of tree species in the target region, and a probability score for each tree species feature channel is generated. The probability score of each pixel is normalized using Softmax, and the probability of each pixel belonging to each tree type is output in the final feature map. The tree type with the highest probability is taken as the prediction result for that pixel, thus generating the final tree type classification result.

[0031] This step efficiently combines deep semantic features with shallow spatial details through attention-weighted fusion and skip connection mechanisms, restoring image resolution and tree spatial geometric details, ensuring the integrity and spatial fineness of segmentation edges, and achieving pixel-level accurate classification of tree species in complex scenes.

[0032] S5. Combine the tree species classification results with the tree geometric heights obtained based on the point cloud of the target area to generate comprehensive tree information; In this embodiment, the geometric height of the tree in S5 includes: S501. Obtain point cloud data of the target area, and perform spatial filtering using the tree distribution space to generate a tree area point cloud, as follows: Traverse the 3D LiDAR point cloud data with a unified coordinate system to obtain the planar coordinates of each laser point; Determine the pixel value corresponding to the planar coordinates in the tree distribution space: if the pixel value is a specific identifier value representing a tree (such as 1), then retain the laser point; if it is a background value, then discard it; use the retained point cloud data as the point cloud of the tree region.

[0033] S502. Separate the ground points and non-ground points in the point cloud of the tree region, and calculate the elevation difference between the non-ground points and the ground points as the geometric height of the trees, as follows: The point cloud of the tree area is flipped, and a virtual cloth mesh covering the entire area is initialized above the flipped point cloud; then a virtual cloth with a specific hardness that can fall under the action of gravity is set, and the specific hardness can be set to 0.5 to 1.0. Initiate an iterative calculation loop and perform the following operations in each iteration until the maximum number of iterations is reached, which can be set to 500: Calculate the displacement of the cloth mesh nodes under gravity, causing them to approach the downward point cloud; detect whether the cloth nodes penetrate the flipped LiDAR point cloud. If penetration occurs, fix the position of the cloth node at the corresponding LiDAR point position (i.e., consider it as being supported); constrain and adjust the relative positions between cloth nodes according to a preset specific hardness parameter to prevent the cloth from undergoing excessive deformation. Flip the iterated cloth mesh back to its original orientation and calculate the distance from each point in the original point cloud to the cloth surface; mark points with a distance less than a threshold as the ground point set, and mark the remaining points as the non-ground point set. The threshold can be 0.5m. Based on the set of ground points, an inverse distance weighted interpolation algorithm is used to construct a digital elevation model covering the tree-covered area. The construction process is as follows: The grid resolution of the digital elevation model to be generated is set, for example, 0.5m × 0.5m; for each grid node to be interpolated, ground points within a preset radius are searched, with the grid node as the center. The radius range is 2m to 5m; the weight of each searched ground point relative to the central grid node is calculated, and the weight is inversely proportional to the square of the Euclidean distance from each ground point to the central grid node; the original elevation coordinates of each searched ground point in the point cloud data are read, i.e., the Z-value in the coordinate system; the original elevation coordinates are multiplied by the weight of the corresponding ground point, summed, and then divided by the sum of all weights to obtain the terrain elevation value of the grid node to be interpolated, thereby generating a continuous digital elevation model describing the surface undulations. The elevation values ​​of non-ground points are read based on point cloud data, and the corresponding ground elevation values ​​are found in the digital elevation model according to their planar coordinates; the difference between the elevation values ​​of non-ground points and the ground elevation values ​​is calculated, which is the geometric height of the tree.

[0034] S503. Combining the tree species classification results with the tree geometric height, generate comprehensive tree information as follows: Based on the unified CGCS2000 spatial coordinate system, non-ground points containing tree geometric height information are projected onto a two-dimensional tree species classification result plane; the pixel category label of each non-ground point at the projection position is read, and the tree species category attribute at that position is obtained; the tree species category attribute is assigned to the corresponding non-ground point, thereby generating a tree point cloud data set containing three-dimensional height information and tree species category information. Based on the spatial distribution characteristics of the tree species classification results, spatially continuous point cloud sets belonging to the same tree species are divided into independent tree objects, and the tree species category corresponding to the point cloud set is identified as the tree species category of the tree object. For each tree object, the calculation includes: extracting the geometric height values ​​of all point cloud data within the range of the tree object, and selecting the maximum value as the current height of the tree object; calculating the arithmetic mean of the planar coordinates of all point cloud data within the range of the tree object, and using the calculated coordinates as the center spatial position of the tree object; Furthermore, a pre-set tree species growth characteristic database is introduced, which contains the annual growth rates of common tree species in power transmission channels (such as purple-leaf plum, fir, paulownia, bamboo, etc.). Based on the tree species category of the tree object, the corresponding annual growth rate is searched and matched in the growth characteristic database. For example, if the tree object is identified as "Paulownia", a higher growth rate, such as 2.0 m / year, is associated; if it is "Pine", a lower growth rate, such as 0.4 m / year, is associated. By combining the central spatial location, tree species, current height, and matched growth rate of the tree object, comprehensive tree information containing "location, tree species, current height, and growth rate" is generated as the basis data for subsequent power line tree obstruction hazard analysis.

[0035] This step integrates tree species, geometric height, and differentiated growth rates to construct comprehensive tree information that combines spatial geometry and biological attributes, providing a complete data foundation for the refined assessment and dynamic prediction of tree barrier risks.

[0036] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0037] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0038] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0039] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0040] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0041] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting tree information within a power transmission line corridor based on multi-source remote sensing data, characterized in that, Includes the following steps: Tree distribution space is screened based on the spectral features of the target area, and image data constrained by tree distribution space is input into the improved D-UNet model; An attention feature map is generated by utilizing the convolutional attention branch in the improved D-UNet model. By improving the parallel multi-scale pooling branch in the D-UNet model, global and local contextual features are extracted. The attention feature map is used to perform weighted fusion of the context features to generate context-enhanced features, and the tree species classification result is output after decoding. By combining the tree species classification results with the tree geometric heights obtained from the point cloud of the target area, comprehensive tree information is generated.

2. The method for extracting tree information in a power transmission line corridor based on multi-source remote sensing data according to claim 1, characterized in that, The selected tree distribution space includes: Extract reflectance information of the target region in multiple spectral bands; Calculate the vegetation index of each pixel in the multispectral image, and construct a spectral feature set based on the reflectance information and the vegetation index; Based on the spectral feature set, a pre-constructed land cover recognition model is used to perform binary classification of trees and non-trees in each pixel of the multispectral image; Extract the pixel regions that are classified as tree categories to generate the tree distribution space.

3. The method for extracting tree information in a power transmission line corridor based on multi-source remote sensing data according to claim 2, characterized in that, The convolutional attention branch includes: Convolutional operations are performed on the high-level semantic feature maps output by the encoder in the improved D-UNet model to extract local feature maps containing local information; Based on the convolutional attention mechanism, weights in the channel dimension and spatial dimension are learned on the local feature map to generate an attention feature map.

4. The method for extracting tree information in a power transmission line corridor based on multi-source remote sensing data according to claim 3, characterized in that, The multi-scale pooling branch includes a parallel global pooling branch and an average pooling branch; the global pooling branch performs global average pooling and upsampling on the high-level semantic feature map to obtain global contextual features. The average pooling branch performs local average pooling and upsampling on the high-level semantic feature map to obtain local contextual features.

5. The method for extracting tree information in a power transmission line corridor based on multi-source remote sensing data according to claim 4, characterized in that, The generated context-enhanced features include: The attention feature map is used to perform element-wise weighted fusion of the global context features and the local context features; The fused features are then subjected to channel concatenation and convolutional compression to obtain multi-scale contextual information. The multi-scale contextual information is residually fused with the high-level semantic feature map to generate context-enhanced features.

6. The method for extracting tree information in a power transmission line corridor based on multi-source remote sensing data according to claim 5, characterized in that, The tree species classification results include: The context enhancement features are fused with the corresponding layer's encoded feature map in the encoder; Based on the preset number of tree species categories, the channels of the fused feature map are classified into tree species feature channels and a probability score for the corresponding tree species is generated. After normalizing the probability scores to obtain the tree species probability of each pixel in the fused feature map, the tree species classification result is generated.

7. The method for extracting tree information in a power transmission line corridor based on multi-source remote sensing data according to claim 6, characterized in that, The geometric height of the tree includes: Obtain point cloud data of the target area, and perform spatial filtering using the tree distribution space to generate a tree area point cloud; Separate ground points and non-ground points in the point cloud of the tree region, and calculate the elevation difference between the non-ground points and the ground points as the geometric height of the trees.

Images