A tree species identification method, device, equipment and storage medium
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGZHOU URBAN PLANNING & DESIGN SURVEY RES INST
- Filing Date
- 2026-05-15
- Publication Date
- 2026-08-07
AI Technical Summary
[0004]本发明提供一种树种识别方法、装置、设备及存储介质，以解决现有的树种识别方法效率低下，难以实现大面积的树种快速识别的技术问题
[0015] Compared with the prior art, the beneficial effect of the embodiments of the present invention is that by performing preliminary identification of tree species in the area to be identified based on spectral features in spectral images, and combining the spatial location information of each tree provided by laser point cloud data, it is possible not only to accurately identify the tree species of each tree in the area to be identified, but also to accurately identify the spatial distribution of different tree species, thereby improving the efficiency of tree species identification and enabling rapid identification of tree species over a large area.
Smart Images
Figure CN122530671A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of tree species identification technology, and in particular to a tree species identification method, apparatus, electronic device, and computer-readable storage medium. Background Technology
[0002] In the collective forest areas of southern my country, mixed forests of Masson pine and broadleaf trees are widely distributed. Accurate identification and differentiation between Masson pine and broadleaf trees is of great significance. On the one hand, pine wilt disease is a devastating disease caused by pine wilt nematodes, which mainly harms pine species such as Masson pine. Identifying the location and distribution of Masson pine is the basis for precise prevention and control of pine wilt disease and can provide a basis for the harvesting of Masson pine to effectively prevent the spread of the disease. On the other hand, broadleaf trees play an important role in maintaining biodiversity, conserving water resources, and protecting soil and water. Identifying the spatial distribution of broadleaf trees helps to assess the health status of forest ecosystems.
[0003] In existing technologies, sample plot surveys are commonly used to identify Masson pine and broad-leaved trees. However, this method is time-consuming, labor-intensive, inefficient, and makes it difficult to achieve rapid identification of tree species over large areas. Summary of the Invention
[0004] This invention provides a tree species identification method, apparatus, device, and storage medium to solve the technical problem that existing tree species identification methods are inefficient and difficult to achieve rapid identification of tree species over large areas.
[0005] To address the aforementioned technical problems, a first aspect of this invention provides a tree species identification method, comprising: Acquire spectral images and laser point cloud data of the region to be identified; A preset tree species identification model is used to perform semantic segmentation on the spectral image to obtain the initial tree species identification result of the region to be identified; wherein, the tree species identification model is obtained by training a deep learning model using a training set of spectral images with tree species labels; the tree species labels include Masson pine labels and broadleaf tree labels; Based on the laser point cloud data, obtain the spatial location information of each tree in the area to be identified; The initial tree species identification result is superimposed with the spatial location information to obtain the target tree species identification result for the area to be identified.
[0006] As a preferred embodiment, the step of using a preset tree species identification model to perform semantic segmentation on the spectral image to obtain the initial tree species identification result for the region to be identified specifically includes: The spectral image is cropped into several cropped images according to a preset image cropping size; Based on the tree species identification model, feature extraction is performed on the cropped image to obtain shallow spectral features and deep semantic features; The deep semantic features are processed by parallel convolution through the hollow pyramid pooling module to obtain multi-scale contextual features. The shallow spectral features are fused with the multi-scale context features, and the tree species probability distribution of each pixel in the fused feature map is calculated to obtain the tree species identification result of the cropped image. The tree species identification results of each cropped image are fused to obtain the initial tree species identification result of the region to be identified.
[0007] As a preferred embodiment, the dilated pyramid pooling module includes a convolution module, three dilated convolution modules, a global pooling module, and a feature fusion module; The process of performing parallel convolution on the deep semantic features using a hollow pyramid pooling module to obtain multi-scale contextual features specifically includes: The deep semantic features are processed by parallel convolution of the convolutional module and the three dilated convolutional modules respectively to obtain several branches for feature extraction. The deep semantic features are pooled using the global pooling module to obtain global features. The feature fusion module fuses the global features with the features extracted from several branches to obtain the multi-scale context features.
[0008] As a preferred embodiment, the step of fusing the shallow spectral features with the multi-scale contextual features and calculating the tree species probability distribution of each pixel in the fused feature map to obtain the tree species identification result of the cropped image specifically includes: The multi-scale context features are upsampled using bilinear interpolation. The shallow spectral features are channel compressed by a 1×1 convolutional layer so that the spatial resolution of the channel-compressed shallow spectral features is consistent with the spatial resolution of the upsampled multi-scale context features. The shallow spectral features after channel compression are fused with the upsampled multi-scale context features and enhanced to obtain a fused feature map. The fused feature map is upsampled using bilinear interpolation to restore its size to the original image size. Calculate the tree species probability distribution of each pixel in the fused feature map to obtain the tree species identification result of the cropped image.
[0009] As a preferred embodiment, fusing the tree species identification results of each of the cropped images to obtain the initial tree species identification result of the region to be identified specifically includes: Obtain the probability distribution of several tree species for each overlapping pixel within the overlapping region between the cropped images; A multiplication fusion strategy is used to fuse several tree species probability distributions for each overlapping pixel to obtain a fused tree species probability distribution for each overlapping pixel. Based on the tree species corresponding to the tree species with the highest probability in the fused tree species probability distribution, the tree species identification result of each overlapping pixel is determined; The initial tree species identification result of the region to be identified is obtained based on the tree species identification result of each overlapping pixel and the tree species identification result of each non-overlapping pixel in each cropped image.
[0010] As a preferred embodiment, the method further includes: Based on the laser point cloud data, each tree in the area to be identified is segmented into individual trees to obtain the structural features of each tree; Based on the structural features, calculate the tree parameters for each tree in the area to be identified; wherein the tree parameters include canopy closure, leaf area index, diameter at breast height (DBH), and tree height.
[0011] As a preferred embodiment, the method further includes: Based on the spectral image and the spatial location information, the red light band reflectance and near-infrared light band reflectance of each tree in the area to be identified are obtained; The vegetation index of each tree is calculated based on the red light band reflectance and the near-infrared light band reflectance. Based on the vegetation index and a preset vegetation index threshold, live and dead trees in the area to be identified are identified.
[0012] A second aspect of the present invention provides a tree species identification device, comprising: The data acquisition module is used to acquire spectral images and laser point cloud data of the area to be identified; The initial tree species identification module is used to perform semantic segmentation on the spectral image using a preset tree species identification model to obtain the initial tree species identification result of the region to be identified; wherein, the tree species identification model is obtained by training a deep learning model using a training set of spectral images with tree species labels; the tree species labels include Masson pine labels and broad-leaved tree labels; The tree location identification module is used to obtain the spatial location information of each tree in the area to be identified based on the laser point cloud data. The target tree species identification result acquisition module is used to overlay the initial tree species identification result with the spatial location information to obtain the target tree species identification result of the area to be identified.
[0013] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the tree species identification method described in any of the first aspects.
[0014] A fourth aspect of the present invention provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the tree species identification method described in any of the first aspects.
[0015] Compared with the prior art, the beneficial effect of the embodiments of the present invention is that by performing preliminary identification of tree species in the area to be identified based on spectral features in spectral images, and combining the spatial location information of each tree provided by laser point cloud data, it is possible not only to accurately identify the tree species of each tree in the area to be identified, but also to accurately identify the spatial distribution of different tree species, thereby improving the efficiency of tree species identification and enabling rapid identification of tree species over a large area. Attached Figure Description
[0016] Figure 1 This is a flowchart illustrating the tree species identification method in an embodiment of the present invention; Figure 2 This is an architecture diagram of the tree species identification model in an embodiment of the present invention; Figure 3 This is a schematic diagram of the tree species identification device in an embodiment of the present invention. Detailed Implementation
[0017] 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 skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] Please see Figure 1 The first aspect of this invention provides a tree species identification method, comprising the following steps S1 to S4: Step S1: Obtain the spectral image and laser point cloud data of the area to be identified; Step S2: The spectral image is semantically segmented using a preset tree species identification model to obtain the initial tree species identification result for the region to be identified; wherein, the tree species identification model is obtained by training a deep learning model using a training set of spectral images with tree species labels; the tree species labels include Masson pine labels and broadleaf tree labels; Step S3: Based on the laser point cloud data, obtain the spatial location information of each tree in the area to be identified; Step S4: Overlay the initial tree species identification result with the spatial location information to obtain the target tree species identification result for the area to be identified.
[0019] Specifically, this embodiment utilizes a UAV equipped with a spectral sensor and a lidar sensor to acquire spectral images and lidar point cloud data of the area to be identified. It is worth noting that the spectral sensor can capture the spectral reflectance characteristics of each tree within the area to be identified. By analyzing multiple spectral bands of data for each tree, its biophysical characteristics such as chlorophyll content and water status can be identified. The spectral image can be a hyperspectral image or a multispectral image; this embodiment does not specifically limit this. The lidar sensor, through precise lidar ranging technology, generates three-dimensional point cloud data, providing spatial location information of the trees and helping to determine the specific coordinates and height of each tree, thereby revealing vegetation structure and terrain features. This embodiment uses a spectral sensor and a lidar sensor to acquire high-resolution small-cell ground cover observation data to finely capture the spectral characteristics and spatial structure information of vegetation within each small-cell area. All data can be processed by UASMaster software, including spatial registration and image stitching, to generate spectral images and lidar point cloud data with a spatial resolution of 10 cm.
[0020] Furthermore, considering the differences in spectral characteristics among different tree species, this embodiment pre-constructs a tree species identification model for semantic segmentation of spectral images to obtain initial tree species identification results for the region to be identified. It is understood that the tree species identification model is obtained by training a deep learning model using a training set of spectral images labeled with tree species. Labeling different tree species can be done using ArcGIS software. For example, in the identification scenario of Masson pine and broadleaf trees, Masson pine can be represented by label "1", broadleaf trees by label "2", and other tree species by label "0". During model training, the Xavier strategy can be used to initialize weights, with the number of iterations set to 10,000. Stochastic gradient descent can be used for network optimization, with an initial learning rate set to 0.001. The learning rate descent method can employ a warm-up optimization strategy. Additionally, a random sampling strategy can be used to select a portion of the spectral image training set to construct a validation set for performance evaluation during model training.
[0021] Furthermore, in order to accurately locate each tree in the area to be identified and thus identify the species of each tree, this embodiment obtains the spatial location information of each tree based on laser point cloud data, and then superimposes it with the initial tree species identification result to obtain the tree species identification result of each tree in the area to be identified, and can clarify the spatial distribution of each tree species, determine the number, volume, and location of Masson pine and broad-leaved trees, and provide a basis for logging.
[0022] As a preferred embodiment, the step of using a preset tree species identification model to perform semantic segmentation on the spectral image to obtain the initial tree species identification result for the region to be identified specifically includes: The spectral image is cropped into several cropped images according to a preset image cropping size; Based on the tree species identification model, feature extraction is performed on the cropped image to obtain shallow spectral features and deep semantic features; The deep semantic features are processed by parallel convolution through the hollow pyramid pooling module to obtain multi-scale contextual features. The shallow spectral features are fused with the multi-scale context features, and the tree species probability distribution of each pixel in the fused feature map is calculated to obtain the tree species identification result of the cropped image. The tree species identification results of each cropped image are fused to obtain the initial tree species identification result of the region to be identified.
[0023] Specifically, considering that deep learning models have fixed requirements for the size of input data, this embodiment first crops the spectral image into several cropped images according to a preset image cropping size. For example, each cropped image is 800×800 pixels in size.
[0024] Furthermore, the tree species identification model in this embodiment is specifically a convolutional neural network semantic segmentation model based on an encoder-decoder architecture, to achieve accurate classification of tree species, such as... Figure 2 As shown in the diagram, in the encoder section, a convolutional neural network is used as the feature extraction backbone network, combined with a dilated pyramid pooling module to construct the core structure of the encoder. Based on the feature extraction backbone network in the tree species recognition model, features are extracted from the cropped image to obtain shallow spectral features and deep semantic features. It is worth noting that the feature extraction backbone network contains 13 convolutional layers, one average pooling layer, and a final fully connected layer, ultimately outputting a downsampled feature map with a spatial resolution of 1 / 16 of the original image. In this downsampled feature map, one part consists of shallow spectral features, which refer to the high-resolution feature maps output from the shallower layers of the feature extraction backbone network, containing rich spatial detail information, such as local features like edges and textures; the other part consists of deep semantic features, which contain deeper semantic information.
[0025] Furthermore, deep semantic features are processed by parallel convolution through a hollow pyramid pooling module to extract multi-scale contextual features.
[0026] In the decoder section, by fusing shallow spectral features with multi-scale contextual features, spatial details can be preserved while fully utilizing contextual semantic information, thereby improving the accuracy and robustness of tree species classification. It can be understood that after calculating the tree species probability distribution for each pixel, the tree species corresponding to the highest probability is the tree species identification result for that pixel, and integrating the tree species identification results for each pixel yields the tree species identification result for the cropped image.
[0027] Furthermore, since the cropped image is obtained by cropping from the original spectral image, it is necessary to fuse the tree species identification results of each cropped image to obtain the initial tree species identification result of the region to be identified.
[0028] As a preferred embodiment, the dilated pyramid pooling module includes a convolution module, three dilated convolution modules, a global pooling module, and a feature fusion module; The process of performing parallel convolution on the deep semantic features using a hollow pyramid pooling module to obtain multi-scale contextual features specifically includes: The deep semantic features are processed by parallel convolution of the convolutional module and the three dilated convolutional modules respectively to obtain several branches for feature extraction. The deep semantic features are pooled using the global pooling module to obtain global features. The feature fusion module fuses the global features with the features extracted from several branches to obtain the multi-scale context features.
[0029] Specifically, dilated convolution can arbitrarily expand the receptive field of filters at any depth of convolutional neural network layer. This method removes the pooling layers and corresponding upsampling layers in the last few layers of the network, and achieves a finer feature response by inserting 0 values between different filter weights without adding new learning parameters. Dilated convolution can also increase the receptive field of the filter, obtaining more content. The dilated pyramid pooling module uses four parallel convolutional layers to process deep semantic features, including three dilated convolutional layers with different sampling rates. The features extracted at each sampling rate are then processed by a separate branch and fused to generate the final result. Spatial pyramids can improve the accuracy of repeated sampling at different scales and are effective for segmentation at any scale. In addition, this embodiment also uses a global pooling module in the dilated pyramid pooling module to pool deep semantic features to extract image-level global features. Finally, a feature fusion module fuses the global features with the features extracted by several branches. The feature fusion module is a 1×1 convolutional layer to obtain multi-scale contextual features, which are then restored to the size of the input downsampled feature map through bilinear interpolation.
[0030] In one alternative embodiment, the three perforated convolutional layers are all 3×3 in size, with dilation rates of 6, 12, and 18, respectively.
[0031] As a preferred embodiment, the step of fusing the shallow spectral features with the multi-scale contextual features and calculating the tree species probability distribution of each pixel in the fused feature map to obtain the tree species identification result of the cropped image specifically includes: The multi-scale context features are upsampled using bilinear interpolation. The shallow spectral features are channel compressed by a 1×1 convolutional layer so that the spatial resolution of the channel-compressed shallow spectral features is consistent with the spatial resolution of the upsampled multi-scale context features. The shallow spectral features after channel compression are fused with the upsampled multi-scale context features and enhanced to obtain a fused feature map. The fused feature map is upsampled using bilinear interpolation to restore its size to the original image size. Calculate the tree species probability distribution of each pixel in the fused feature map to obtain the tree species identification result of the cropped image.
[0032] Specifically, for the multi-scale context features output by the encoder, bilinear interpolation is first used to upsample them, increasing their spatial resolution to 1 / 4 of the original image. Then, a 1×1 convolutional layer is used to compress the shallow spectral features, ensuring that the spatial resolution of the compressed shallow spectral features matches that of the upsampled multi-scale context features, while reducing computational complexity. Further, the channel-compressed shallow spectral features are fused with the upsampled multi-scale context features, and a 3×3 convolutional layer is used to enhance the fused feature map, further refining the feature representation. Finally, bilinear interpolation is used to upsample the fused feature map, restoring its size to the original image size.
[0033] As a preferred embodiment, fusing the tree species identification results of each of the cropped images to obtain the initial tree species identification result of the region to be identified specifically includes: Obtain the probability distribution of several tree species for each overlapping pixel within the overlapping region between the cropped images; A multiplication fusion strategy is used to fuse several tree species probability distributions for each overlapping pixel to obtain a fused tree species probability distribution for each overlapping pixel. Based on the tree species corresponding to the tree species with the highest probability in the fused tree species probability distribution, the tree species identification result of each overlapping pixel is determined; The initial tree species identification result of the region to be identified is obtained based on the tree species identification result of each overlapping pixel and the tree species identification result of each non-overlapping pixel in each cropped image.
[0034] Specifically, since there may be overlapping areas between the cropped images, in order to effectively fuse the tree species identification results of the various cropped images, this embodiment first obtains several tree species probability distributions for each overlapping pixel within the overlapping area between the cropped images, and then uses a multiplication fusion strategy to fuse the several tree species probability distributions for each overlapping pixel, as shown in the following expression: ; in, Let represent the probability that cropped image 1 and cropped image 2 belong to tree species c at pixel (i,j) and be merged into a single tree species c. Let represent the probability that the cropped image 1 at pixel (i,j) belongs to tree species c. Let represent the probability that the cropped image 2 at pixel (i,j) belongs to tree species c. Indicates the total number of tree species. Let represent the probability that the cropped image 1 at pixel (i,j) belongs to tree species k. This represents the probability that the cropped image 2 belongs to tree species k at pixel (i,j).
[0035] Furthermore, after obtaining the probability distribution of the fused tree species for each overlapping pixel, the probability of fused tree species for each overlapping pixel is determined. The tree species corresponding to the maximum tree species probability for each overlapping pixel is taken as the tree species identification result for that overlapping pixel. Finally, by combining the tree species identification results of each non-overlapping pixel in each cropped image, the initial tree species identification result for the region to be identified is obtained.
[0036] As a preferred embodiment, the method further includes: Based on the laser point cloud data, each tree in the area to be identified is segmented into individual trees to obtain the structural features of each tree; Based on the structural features, calculate the tree parameters for each tree in the area to be identified; wherein the tree parameters include canopy closure, leaf area index, diameter at breast height (DBH), and tree height.
[0037] Specifically, in this embodiment, to further obtain tree parameters for each tree within the area to be identified, each tree is first segmented based on laser point cloud data to obtain its structural characteristics, including the vertical distribution density of the canopy point cloud, the projected area of the canopy point cloud, and the vertical distribution characteristics of the trunk point cloud. Then, tree parameters including canopy closure, leaf area index (LAI), diameter at breast height (DBH), and tree height are calculated. For example, the LAI is estimated by analyzing the vertical distribution density of the canopy point cloud. Specifically, the point cloud is layered by height (usually with a layer interval of 0.5m), and based on the Beer-Lambert law, the density ratio of each layer is calculated to obtain the LAI. The canopy closure is calculated by vertically projecting the canopy point cloud onto a horizontal plane, using a gridded process (usually with a grid size of 0.5m × 0.5m), and calculating the ratio of the number of grids covered by the point cloud to the total number of grids. DBH is extracted by taking a section of the trunk point cloud at a height of 1.3m above the ground, fitting a cylinder using the RANSAC algorithm, and obtaining the diameter value as the DBH. Tree height is obtained by calculating the height difference between the highest point in the single-tree point cloud set and the corresponding ground point, where the ground point is extracted using progressive morphological filtering. Outlier removal and accuracy verification are performed during the acquisition of these parameters to ensure the reliability of the measurement results.
[0038] As a preferred embodiment, the method further includes: Based on the spectral image and the spatial location information, the red light band reflectance and near-infrared light band reflectance of each tree in the area to be identified are obtained; The vegetation index of each tree is calculated based on the red light band reflectance and the near-infrared light band reflectance. Based on the vegetation index and a preset vegetation index threshold, live and dead trees in the area to be identified are identified.
[0039] Furthermore, to effectively provide a basis for tree harvesting, this embodiment uses the red and near-infrared reflectance of each tree in the spectral image to calculate the vegetation index of each tree, such as NDVI and EVI values. It is understood that the vegetation index of dead and living trees differs in their spectral reflectance characteristics. Living trees have strong photosynthesis, therefore their near-infrared reflectance is higher, and correspondingly, their vegetation index is also higher. Dead trees, due to a lack of biological activity, withered or fallen leaves, result in a significantly reduced spectral reflectance and a lower vegetation index. By setting a reasonable vegetation index threshold, living and dead trees can be distinguished.
[0040] In one optional embodiment, the logging strategy after tree species identification adopts a "minimally invasive strip opening + precise replanting and intercropping" approach, minimizing the "wounds" created for forest quality improvement and providing the necessary ecological factors for the growth of planted native broad-leaved trees and precious tree species, thus achieving precise improvement in forest quality. Specifically, the "minimally invasive strip opening" precisely divides the work plots according to the standing tree structure, tree species composition, and site conditions. Strips are opened along contour lines with a layout of "clearing 1.5m and leaving 1.5m," preserving existing native broad-leaved trees. Following the principles of "harvesting bad trees and keeping good ones, harvesting old trees and keeping strong ones, harvesting dense trees and keeping evenly distributed ones," selective logging is used to selectively remove "disease and damaged" trees and infected Masson pine trees, with the selective logging intensity controlled below 40%. The harvested infected Masson pine trees are crushed and sprayed into the strip. Other damaged trees and branches are evenly cut into sections and piled along the strip with the cleared weeds to form fertilizer through composting. This also serves as the edge of the planting strip, reducing longitudinal soil and water erosion and retaining nutrients in the forest. "Precision replanting and intercropping" is based on botany and niche theory, and combines factors such as forest function, spatial morphology, tree species composition, standing tree structure and site conditions. Under the premise of preserving the existing native broad-leaved trees, it optimizes and improves forest quality by precisely replanting and intercropping with high-quality native tree species suitable for local growth through three steps: "determining function, determining space, and determining configuration".
[0041] The tree species identification method provided in this invention preliminarily identifies tree species within the identification area based on spectral features in spectral images, and combines this with spatial location information of each tree provided by laser point cloud data. This not only accurately identifies the tree species of each tree within the identification area, but also precisely identifies the spatial distribution of different tree species, improving the efficiency of tree species identification and enabling rapid identification of tree species over a large area.
[0042] Please see Figure 3 A second aspect of the present invention provides a tree species identification device, comprising: Data acquisition module 101 is used to acquire spectral images and laser point cloud data of the area to be identified; The initial tree species identification module 102 is used to perform semantic segmentation on the spectral image using a preset tree species identification model to obtain the initial tree species identification result of the region to be identified; wherein, the tree species identification model is obtained by training a deep learning model using a training set of spectral images with tree species labels; the tree species labels include Masson pine labels and broad-leaved tree labels; The tree location identification module 103 is used to obtain the spatial location information of each tree in the area to be identified based on the laser point cloud data. The target tree species identification result acquisition module 104 is used to overlay the initial tree species identification result with the spatial location information to obtain the target tree species identification result of the area to be identified.
[0043] As a preferred embodiment, the initial tree species identification module 102 is used to perform semantic segmentation on the spectral image using a preset tree species identification model to obtain the initial tree species identification result of the region to be identified, specifically including: The spectral image is cropped into several cropped images according to a preset image cropping size; Based on the tree species identification model, feature extraction is performed on the cropped image to obtain shallow spectral features and deep semantic features; The deep semantic features are processed by parallel convolution through the hollow pyramid pooling module to obtain multi-scale contextual features. The shallow spectral features are fused with the multi-scale context features, and the tree species probability distribution of each pixel in the fused feature map is calculated to obtain the tree species identification result of the cropped image. The tree species identification results of each cropped image are fused to obtain the initial tree species identification result of the region to be identified.
[0044] As a preferred embodiment, the dilated pyramid pooling module includes a convolution module, three dilated convolution modules, a global pooling module, and a feature fusion module; The initial tree species identification module 102 is used to perform parallel convolution processing on the deep semantic features through the dilated pyramid pooling module to obtain multi-scale contextual features, specifically including: The deep semantic features are processed by parallel convolution of the convolutional module and the three dilated convolutional modules respectively to obtain several branches for feature extraction. The deep semantic features are pooled using the global pooling module to obtain global features. The feature fusion module fuses the global features with the features extracted from several branches to obtain the multi-scale context features.
[0045] As a preferred embodiment, the initial tree species identification module 102 is used to fuse the shallow spectral features with the multi-scale contextual features, and calculate the tree species probability distribution of each pixel in the fused feature map to obtain the tree species identification result of the cropped image, specifically including: The multi-scale context features are upsampled using bilinear interpolation. The shallow spectral features are channel compressed by a 1×1 convolutional layer so that the spatial resolution of the channel-compressed shallow spectral features is consistent with the spatial resolution of the upsampled multi-scale context features. The shallow spectral features after channel compression are fused with the upsampled multi-scale context features and enhanced to obtain a fused feature map. The fused feature map is upsampled using bilinear interpolation to restore its size to the original image size. Calculate the tree species probability distribution of each pixel in the fused feature map to obtain the tree species identification result of the cropped image.
[0046] As a preferred embodiment, the initial tree species identification module 102 is used to fuse the tree species identification results of each of the cropped images to obtain the initial tree species identification result of the region to be identified, specifically including: Obtain the probability distribution of several tree species for each overlapping pixel within the overlapping region between the cropped images; A multiplication fusion strategy is used to fuse several tree species probability distributions for each overlapping pixel to obtain a fused tree species probability distribution for each overlapping pixel. Based on the tree species corresponding to the tree species with the highest probability in the fused tree species probability distribution, the tree species identification result of each overlapping pixel is determined; The initial tree species identification result of the region to be identified is obtained based on the tree species identification result of each overlapping pixel and the tree species identification result of each non-overlapping pixel in each cropped image.
[0047] As a preferred embodiment, the device further includes a tree parameter identification module, used for: Based on the laser point cloud data, each tree in the area to be identified is segmented into individual trees to obtain the structural features of each tree; Based on the structural features, calculate the tree parameters for each tree in the area to be identified; wherein the tree parameters include canopy closure, leaf area index, diameter at breast height (DBH), and tree height.
[0048] As a preferred embodiment, the device further includes a tree condition recognition module, used for: Based on the spectral image and the spatial location information, the red light band reflectance and near-infrared light band reflectance of each tree in the area to be identified are obtained; The vegetation index of each tree is calculated based on the red light band reflectance and the near-infrared light band reflectance. Based on the vegetation index and a preset vegetation index threshold, live and dead trees in the area to be identified are identified.
[0049] The tree species identification device provided in this embodiment of the invention performs preliminary identification of tree species in the area to be identified based on spectral features in spectral images, and combines the spatial location information of each tree provided by laser point cloud data. It can not only accurately identify the tree species of each tree in the area to be identified, but also accurately identify the spatial distribution of different tree species, thereby improving the efficiency of tree species identification and enabling rapid identification of tree species over a large area.
[0050] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the tree species identification method described in any embodiment of the first aspect.
[0051] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.
[0052] The electronic device may include, but is not limited to, a processor and a memory. Those skilled in the art will understand that the schematic diagram is merely an example of an electronic device and does not constitute a limitation on the electronic device. It may include more or fewer components than illustrated, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.
[0053] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the electronic device through various interfaces and lines.
[0054] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.
[0055] A fourth aspect of the present invention provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the tree species identification method described in any embodiment of the first aspect.
[0056] Wherein, if the modules / units integrated in the electronic device 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, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.
[0057] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. A method for identifying tree species, characterized in that, include: Acquire spectral images and laser point cloud data of the region to be identified; A preset tree species identification model is used to perform semantic segmentation on the spectral image to obtain the initial tree species identification result of the region to be identified; wherein, the tree species identification model is obtained by training a deep learning model using a training set of spectral images with tree species labels; the tree species labels include Masson pine labels and broadleaf tree labels; Based on the laser point cloud data, obtain the spatial location information of each tree in the area to be identified; The initial tree species identification result is superimposed with the spatial location information to obtain the target tree species identification result for the area to be identified.
2. The tree species identification method as described in claim 1, characterized in that, The step of using a preset tree species identification model to perform semantic segmentation on the spectral image to obtain the initial tree species identification result for the region to be identified specifically includes: The spectral image is cropped into several cropped images according to a preset image cropping size; Based on the tree species identification model, feature extraction is performed on the cropped image to obtain shallow spectral features and deep semantic features; The deep semantic features are processed by parallel convolution through the hollow pyramid pooling module to obtain multi-scale contextual features. The shallow spectral features are fused with the multi-scale context features, and the tree species probability distribution of each pixel in the fused feature map is calculated to obtain the tree species identification result of the cropped image. The tree species identification results of each cropped image are fused to obtain the initial tree species identification result of the region to be identified.
3. The tree species identification method as described in claim 2, characterized in that, The dilated pyramid pooling module includes a convolution module, three dilated convolution modules, a global pooling module, and a feature fusion module. The process of performing parallel convolution on the deep semantic features using a hollow pyramid pooling module to obtain multi-scale contextual features specifically includes: The deep semantic features are processed by parallel convolution of the convolutional module and the three dilated convolutional modules respectively to obtain several branches for feature extraction. The deep semantic features are pooled using the global pooling module to obtain global features. The feature fusion module fuses the global features with the features extracted from several branches to obtain the multi-scale context features.
4. The tree species identification method as described in claim 2, characterized in that, The process of fusing the shallow spectral features with the multi-scale contextual features and calculating the tree species probability distribution of each pixel in the fused feature map to obtain the tree species identification result of the cropped image specifically includes: The multi-scale context features are upsampled using bilinear interpolation. The shallow spectral features are channel compressed by a 1×1 convolutional layer so that the spatial resolution of the channel-compressed shallow spectral features is consistent with the spatial resolution of the upsampled multi-scale context features. The shallow spectral features after channel compression are fused with the upsampled multi-scale context features and enhanced to obtain a fused feature map. The fused feature map is upsampled using bilinear interpolation to restore its size to the original image size. Calculate the tree species probability distribution of each pixel in the fused feature map to obtain the tree species identification result of the cropped image.
5. The tree species identification method as described in claim 2, characterized in that, The step of fusing the tree species identification results of each of the cropped images to obtain the initial tree species identification result of the region to be identified specifically includes: Obtain the probability distribution of several tree species for each overlapping pixel within the overlapping region between the cropped images; A multiplication fusion strategy is used to fuse several tree species probability distributions for each overlapping pixel to obtain a fused tree species probability distribution for each overlapping pixel. Based on the tree species corresponding to the tree species with the highest probability in the fused tree species probability distribution, the tree species identification result of each overlapping pixel is determined; The initial tree species identification result of the region to be identified is obtained based on the tree species identification result of each overlapping pixel and the tree species identification result of each non-overlapping pixel in each cropped image.
6. The tree species identification method as described in claim 1, characterized in that, The method further includes: Based on the laser point cloud data, each tree in the area to be identified is segmented into individual trees to obtain the structural features of each tree; Based on the structural features, calculate the tree parameters for each tree in the area to be identified; wherein the tree parameters include canopy closure, leaf area index, diameter at breast height (DBH), and tree height.
7. The tree species identification method as described in claim 1, characterized in that, The method further includes: Based on the spectral image and the spatial location information, the red light band reflectance and near-infrared light band reflectance of each tree in the area to be identified are obtained; The vegetation index of each tree is calculated based on the red light band reflectance and the near-infrared light band reflectance. Based on the vegetation index and a preset vegetation index threshold, live and dead trees in the area to be identified are identified.
8. A tree species identification device, characterized in that, include: The data acquisition module is used to acquire spectral images and laser point cloud data of the area to be identified; The initial tree species identification module is used to perform semantic segmentation on the spectral image using a preset tree species identification model to obtain the initial tree species identification result of the region to be identified; wherein, the tree species identification model is obtained by training a deep learning model using a training set of spectral images with tree species labels; the tree species labels include Masson pine labels and broad-leaved tree labels; The tree location identification module is used to obtain the spatial location information of each tree in the area to be identified based on the laser point cloud data. The target tree species identification result acquisition module is used to overlay the initial tree species identification result with the spatial location information to obtain the target tree species identification result of the area to be identified.
9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the tree species identification method according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the tree species identification method according to any one of claims 1 to 7.