A forest parameter estimation method based on a real-time target detection network

Abstract

The application discloses a forest parameter estimation method based on a real-time target detection network, and mainly relates to the field of forest parameter estimation. The method comprises the following steps: S1. sparse forest point cloud data is acquired, and grid block is carried out; S2. data pre-processing is carried out; S3. a forest point cloud density map is generated, a single tree crown image target detection training set and a test set are constructed; S4. a real-time target detection model is constructed, and a network structure is obtained; S5. the point cloud density map is detected, the contour region of the single tree crown is acquired, and single tree point cloud is acquired; and S6. the average tree height of the single tree and the average diameter at breast height of the single tree are estimated. The method has the beneficial effects that the forest point cloud is used as input data, real-time image target detection is carried out by being converted into a density map, the method of acquiring single tree point cloud through image target detection is used to efficiently estimate forest parameter information, high real-time performance can be achieved, and the tree crown can be effectively recognized in a dense forest scene or a sparse forest scene, and the purpose of extracting single tree point cloud is achieved.

Description

Technical Field

[0001] This invention relates to the field of forest parameter estimation, specifically a forest parameter estimation method based on a real-time target detection network. Background Technology

[0002] Forests are among the most important ecosystems on Earth. Surveying forest resources provides crucial forest parameters, enabling accurate detection and assessment of changes in forest resources and offering scientific guidance for forestry policy decisions. Traditional methods for obtaining forest parameters require significant manpower and resources for on-site measurements. However, with the development of lidar technology, airborne lidar can quickly and efficiently acquire large-scale three-dimensional forest data, providing a new method for forest parameter estimation.

[0003] For estimating forest parameters, calculations based on point clouds of individual trees in the forest offer the highest accuracy. Currently, traditional methods for obtaining individual tree point clouds from forest point clouds include watershed-based methods, local maximum-based methods, and region growing methods. While these methods effectively acquire individual tree point cloud information, they suffer from high time complexity, requiring calculation of spatial geometric relationships between points. Furthermore, traditional methods lack robustness to different tree species.

[0004] In forest parameter estimation using deep learning, both point cloud-based and image-based methods can accurately acquire point clouds of individual trees to achieve the goal of forest parameter estimation. Point cloud-based deep learning methods can be divided into point cloud object detection and point cloud instance segmentation. Currently, point cloud instance segmentation methods are still immature and cannot be effectively used in outdoor scenes. Point cloud object detection methods can be effectively applied to point cloud tasks in large scenes, but they cannot achieve real-time segmentation across the entire scene. Image-based deep learning methods can also be divided into image-based object instance segmentation methods and image-based object detection methods. Image instance segmentation methods can effectively acquire tree canopy outlines, and some network models can achieve a certain level of real-time performance. However, in some forest point cloud scenarios, the canopy closure is high, and image instance segmentation models cannot perform effectively. Summary of the Invention

[0005] The purpose of this invention is to provide a forest parameter estimation method based on a real-time target detection network. It takes forest point cloud as input data, converts it into density map for real-time image target detection, and uses image target detection to obtain individual tree point clouds to efficiently estimate forest parameter information. It can achieve high real-time performance and can effectively identify tree canopies in both dense and sparse forest scenes, thus achieving the purpose of extracting individual tree point clouds.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] A forest parameter estimation method based on a real-time object detection network includes the following steps:

[0008] S1. Use airborne lidar to collect sample plot data, obtain sparse forest point cloud data, and use a watershed-based method to divide the large scene point cloud into grid blocks;

[0009] S2. Preprocess the original forest point cloud data in the small area grid of step S1, including point cloud denoising and removal of ground points;

[0010] S3. Use the forest point cloud data preprocessed in step S2 to generate a forest point cloud density map, construct a training set and a test set for target detection of single tree canopy images using the density map, and label them using software;

[0011] S4. Construct a real-time object detection model based on YOLOv5, optimize it using the CBAM attention mechanism, and use the Recursive-FPN feature fusion layer to solve the problem of inconsistent tree canopy scale in different data, thus obtaining the network structure YOLOv5-RT with real-time detection capability;

[0012] S5. Use the YOLOv5-RT model trained in step S4 to detect the forest point cloud density map to be detected, obtain the outline region of the individual tree crown, and use the box region to obtain the individual tree point cloud.

[0013] S6. Estimate the average tree height and average diameter at breast height (DBH) of individual trees in the sample plot based on all the individual tree point clouds obtained in step S5.

[0014] Furthermore, the grid partitioning method in step S1 includes:

[0015] 1.1 The watershed algorithm is used to coarsely segment the obtained original forest point cloud data to obtain low-precision single tree point cloud data, and the centroid of each single tree point cloud is calculated.

[0016] 1.2 Divide the entire area of ​​the original forest point cloud data into a 100m*100m grid, and use the position of the centroid to determine the grid to which the corresponding point cloud belongs.

[0017] Furthermore, in step 1.2, the large-scene forest point cloud is divided into small-scene point clouds along the edge of the canopy to ensure the integrity of the canopy point cloud and to construct effective training data.

[0018] Furthermore, in step S2, the preprocessing of the small-area forest point cloud includes:

[0019] 2.1, Use the progressive triangulation encryption filtering algorithm to remove ground points in the forest point cloud;

[0020] 2.2, Voxel filtering algorithm is used to remove noise generated during radar scanning;

[0021] Furthermore, in step S3, the generation of the density map includes:

[0022] 3.1 Voxelize the forest point cloud on the xy plane. There is no limit to the height of the voxel block on the z-axis, and the area of ​​the voxel block on the xy plane is 0.3m*0.3m.

[0023] 3.2 Count the number of points in each voxel block, calculate the extreme values ​​of the number of point clouds in all voxel blocks, discretize the numerical range between the minimum and maximum values, and project a color scale from blue to red. Assign colors to the voxel blocks according to the color scale to obtain the density map of the forest scene.

[0024] Furthermore, step S4 also includes:

[0025] 4.1 In the feature extraction part of the YOLOv5-RT network structure, taking into account the characteristic that the center of the tree canopy is more obvious in the density map, the CBAM attention mechanism is added to the last two layers of ResNet;

[0026] 4.2 In the multi-level feature fusion part of the YOLOv5-RT network structure, Recursive-FPN is used to enhance the original FPN to address the inconsistency of the canopy scale in the density map.

[0027] Furthermore, in step 4.1, the CBAM convolutional attention mechanism described consists of a channel attention mechanism and a spatial attention mechanism, the specific principle of which is as follows:

[0028]

[0029]

[0030] In the formula, Q is the input feature map, Q' is the feature information obtained by weighting Q through channel attention; Q” is the feature information obtained by weighting Q' through spatial attention mechanism; M c (Q) and M s (Q') represent the weight information generated by the channel attention mechanism and the spatial attention mechanism, respectively; This indicates bitwise multiplication between feature maps.

[0031] Furthermore, in step 4.2, the characteristics of Recursive-FPN described are as follows:

[0032] The features of the traditional FPN are fused and then fed back into the backbone network for multiple iterations to achieve effective utilization of feature map information.

[0033] Furthermore, for obtaining the parameter of average tree height in step S6, the elevation difference is calculated using the highest and lowest points of the tree point cloud. The specific formula is as follows:

[0034] T h =z(P top )-z(P ground (2)

[0035] In the formula z(P) top z(P) represents the highest point of the treetop elevation in a single tree point cloud. ground () indicates the lowest point of elevation of a single log;

[0036] The estimation of the average diameter at breast height (DBH) of a single tree in step S6 is obtained by performing point cloud circle fitting on slices of the obtained single tree point cloud at a distance of 1.8m-2m.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] This invention utilizes a deep learning-based real-time image target detection algorithm to estimate forest parameters, efficiently acquiring forest tree height and individual tree diameter at breast height (DBH), demonstrating high application value. To address the low detection efficiency of forest point cloud density maps in high-canopy-closure scenarios, a CBAM attention mechanism is employed to mitigate the loss of canopy information in deep feature maps by the YOLOv5 feature extraction network. Furthermore, to address the issue of varying forest point cloud density maps across different scene scales and canopy sizes for different tree types, PANet is used to enhance the feature fusion layer, preventing the loss of small target canopies. Experiments show that the improved YOLOv5-RT exhibits better recognition performance, acquiring higher-quality individual tree point cloud data to achieve high-accuracy forest parameter estimation. Attached Figure Description

[0039] Appendix Figure 1 This is a schematic diagram of the cloud data of the sample location in this invention.

[0040] Appendix Figure 2 This is a density map of the cloud data at the sample location in this invention.

[0041] Appendix Figure 3 This is a schematic diagram of the overall network structure of YOLOv5-RT of the present invention.

[0042] Appendix Figure 4 This is a schematic diagram of the CBAM attention mechanism of the present invention.

[0043] Appendix Figure 5This is a schematic diagram of the Recursive-FPN feature pyramid structure of the present invention.

[0044] Appendix Figure 6 This is a graph showing the detection results of YOLOv5-RT on the test set according to the present invention.

[0045] Appendix Figure 7 This is the result image of the extracted single-tree point cloud of the present invention. Detailed Implementation

[0046] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0047] Example: A forest parameter estimation method based on a real-time object detection network, comprising the following steps:

[0048] S1. Use airborne lidar to collect sample plot data, obtain sparse forest point cloud data, and use a watershed-based method to divide the large scene point cloud into grid blocks.

[0049] In this embodiment, an unmanned aerial vehicle (UAV) equipped with an airborne LiDAR is used to collect various types of forest point cloud data. The acquired large-scale forest point cloud data is then preprocessed to generate small-scale forest point cloud data. Specifically:

[0050] 1.1 The watershed algorithm is used to coarsely segment the obtained original forest point cloud data to obtain low-precision individual tree point cloud data, and the centroid of each individual tree point cloud is calculated.

[0051] First, airborne lidar equipment is used to scan forest area data, and the x, y, and z data in the scanned data are obtained through software for further processing.

[0052] Secondly, the forest point cloud data is transformed into a canopy height model, and the watershed algorithm is used for grayscale image segmentation. The low-precision single tree point cloud in the scene is extracted from the segmented image region.

[0053] Finally, the centroid of each instance segmented by the watershed algorithm is calculated and projected onto a two-dimensional plane for post-processing.

[0054] 1.2 Divide the entire area of ​​the original point cloud data into a 100m*100m grid.

[0055] Using the position of the centroid to determine the corresponding mesh of the point cloud solves the problem of overly coarse edge cutting caused by directly using meshes to divide large scenes. By dividing the large forest point cloud into smaller scene point clouds along the canopy edges, the integrity of the canopy point cloud is maintained, thus achieving effective training data construction. The final result is a small-area forest scene point cloud, such as... Figure 1 As shown.

[0056] S2. Preprocess the original forest point cloud data in the small area grid, including point cloud denoising and removal of ground points.

[0057] Preprocessing of point clouds in small forest areas mainly includes the following two parts:

[0058] First, a progressive triangulation filtering algorithm is used to remove ground points from the forest point cloud. A sparse triangular mesh is generated using seed points, and the mesh is then iterated layer by layer to achieve the goal of removing both ground and non-ground points from the entire point cloud.

[0059] Secondly, a voxel filtering algorithm is used to remove noise generated during radar scanning. This is achieved by dividing the entire scene's point cloud into small cubes, counting the number of points in each cube, and retaining cubes with a number greater than a certain threshold. This method preserves the detailed features of the forest point cloud scene without affecting the density map generation.

[0060] S3. Use the preprocessed forest point cloud data to generate a forest point cloud density map, construct a training set and a test set for target detection of single tree canopy images using the density map, and label them using software.

[0061] In this embodiment, a 3D point cloud is used as the raw data, and a tree point cloud density map is generated by using the density of local point regions as image features. Specifically, the steps include:

[0062] 3.1 The scene point cloud is rasterized on the xy horizontal plane. The rasterized grid blocks have no height limit on the axis. The area of ​​the grid block on the xy horizontal plane is 0.3m*0.3m, which ensures that the generated density map has a high resolution while retaining more detailed features.

[0063] 3.2 The generated raster blocks are traversed, and the number of point clouds in each raster is counted. After traversal, the extreme values ​​of all raster blocks are counted. The interval between the maximum and minimum values ​​is discretized into 256 parts and mapped onto a color scale from blue to red to obtain more color feature information. Simultaneously, the RGB average value is calculated for each discrete color scale interval as a color representative. Then, the discretized color scale is assigned to the raster blocks to obtain the density map of the entire forest point cloud scene, which is used as the processing object for the object detection algorithm, such as... Figure 2 As shown.

[0064] Finally, the forest point cloud density map was labeled using the LableMe software, with each tree canopy being the object of labeling. The characteristic of the label is that the center of the tree canopy is reddish and the edge of the tree canopy is blued. This is a visualization effect caused by the distribution characteristics of the points in the single tree point cloud.

[0065] S4. Construct a real-time object detection model based on YOLOv5, optimize it using the CBAM attention mechanism, and use the Recursive-FPN feature fusion layer to solve the problem of inconsistent tree canopy scale in different data, thus obtaining the network structure YOLOv5-RT with real-time detection capability.

[0066] First, a basic YOLOv5 tree point cloud density map recognition model is constructed, mainly including:

[0067] The backbone network module uses a multi-layer convolutional neural network to extract multi-level image features from the input tree point cloud density map. Generally, ResNet101 can meet the feature extraction requirements.

[0068] The neck network module fuses the multi-level features extracted by the backbone network module to mitigate the problem of information loss in deep feature maps. The original model uses PANet for multi-level feature fusion.

[0069] The output end performs target detection on the fused image features to determine the location of the target region in the tree point cloud density map, and finally obtains a series of candidate proposal boxes as output for subsequent operations.

[0070] To address the practical problems encountered in tree point cloud density map object detection tasks, this paper proposes an improvement strategy for the original YOLOv5 model, using both of these approaches. The model structure is as follows: Figure 3 As shown, it specifically includes:

[0071] 4.1 In the feature extraction part of the YOLOv5-RT network structure, taking into account the fact that the center of the tree canopy is more obvious in the density map, the CBAM attention mechanism is added to the last two layers of ResNet.

[0072] The CBAM convolutional attention mechanism consists of channel attention and spatial attention mechanisms. Compared to the traditional SENet channel attention mechanism, it can achieve better feature extraction results. Its structural modules are as follows: Figure 4 As shown.

[0073] from Figure 4As can be seen, the output of the feature extraction network is first processed by the channel attention module to obtain a weighted result of the feature map, and then processed by a spatial attention module to obtain the final weighted feature result. Specifically, the feature map first undergoes global max pooling and global average pooling, and then passes through an MLP. The features after the two processing steps are weighted pixel by pixel, and a sigmoid activation function is used to generate the channel attention feature map. Finally, the channel attention feature map is multiplied pixel by pixel with the input features to generate the input features needed by the spatial attention module. The specific formula is as follows:

[0074] M c (F)=δ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0075] Regarding the spatial attention mechanism of CBAM, the input feature channels are compressed, and average pooling and max pooling are performed. Max pooling is used to obtain the maximum value on each feature map channel, and the number of extractions is equal to the height multiplied by the width of the feature map. Average pooling is used to obtain the average value on each feature map channel, and the number of extractions is also equal to the height multiplied by the width of the feature map. The specific formulas are as follows:

[0076] M s (F)=δ(f 7×7 ([AvgPool(F);MaxPool(F)]))

[0077] 4.2 In the multi-level feature fusion part of the YOLOv5-RT network structure, Recursive-FPN is used to enhance the original FPN to address the inconsistency of the canopy scale in the density map.

[0078] Recursive-FPN builds upon basic FPN by employing a recurrent structure. It fuses feature maps from top to bottom before returning them to the input feature map for further fusion. This multi-loop structure fully utilizes the information from the top-level feature maps and effectively avoids information loss, making it highly effective in tree canopy detection tasks. Its model structure is as follows: Figure 5 As shown.

[0079] S5. Use the trained YOLOv5-RT model to detect the forest point cloud density map to be detected, obtain the outline region of the individual tree crown, and use the box region to obtain the individual tree point cloud.

[0080] The processed density map is input into the trained model for inference, which yields several bounding boxes. Each bounding box represents a local region of the tree canopy. The detection results are as follows: Figure 6As shown, point cloud data for each tree can be obtained through 2D to 3D conversion, and the obtained point cloud results are as follows. Figure 7 As shown.

[0081] S6. Estimate the average tree height and average diameter at breast height (DBH) of individual trees in the sample plot based on the obtained point clouds of all individual trees.

[0082] To obtain the parameter of average tree height, the elevation difference is calculated using the highest and lowest points of the tree point cloud. The specific formula is as follows.

[0083] T h =z(P top )-z(P ground (2)

[0084] z(P top z(P) represents the highest point of the treetop elevation in a single tree point cloud. ground () indicates the lowest point of elevation of a single log;

[0085] The average diameter at breast height (DBH) of a single tree is estimated by performing point cloud circle fitting on slices of the obtained single tree point cloud at a distance of 1.8m-2m.

[0086] Matters not covered in this invention are common knowledge.

[0087] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A forest parameter estimation method based on a real-time object detection network, characterized in that: The method comprises the following steps: S1. Collecting sample plot data using an airborne laser radar to obtain sparse forest point cloud data, and using a watershed-based method to divide the large scene point cloud into grids; S2. Preprocessing the original forest point cloud data in the small area grid of step S1, including point cloud denoising and ground point removal; S3. Generating a forest point cloud density map using the preprocessed forest point cloud data of step S2, constructing a single tree crown image target detection training set and test set through the density map, and labeling using software; In step S3, the generation of the density map comprises: 3.

1. Voxelizing the forest point cloud in the xy plane, with no height restriction on the voxel block in the z axis, and the area size in the xy plane being 0.3m*0.3m; 3.

2. Counting the number of points in each voxel block, calculating the extreme value of the number of point clouds in all voxel blocks, discretizing the numerical interval between the minimum value and the maximum value, and projecting the color scale from blue to red, and assigning colors to the voxel blocks according to the color scale to obtain a density map of the forest scene; S4. Constructing a real-time target detection model based on YOLOv5 and optimizing it using CBAM attention mechanism, using Recursive-FPN feature fusion layer to solve the problem of inconsistent tree crown scales in different data, and obtaining a network structure YOLOv5-RT with real-time detection capability; Step S4 further comprises: 4.

1. In the feature extraction part of the YOLOv5-RT network structure, in view of the feature that the crown center in the density map is relatively obvious, CBAM attention mechanism is added in the last two layers of ResNet; 4.

2. In the multi-level feature fusion part of the YOLOv5-RT network structure, in view of the feature that the crown scale in the density map is inconsistent, Recursive-FPN is used to strengthen the original FPN; S5. Using the trained YOLOv5-RT model in step S4 to detect the forest point cloud density map to be detected, obtaining the outline area of the single tree crown, and using the frame area to obtain the single tree point cloud; S6. Estimating the two forest parameters of single tree average height and single tree average diameter in the sample plot according to all the single tree point clouds obtained in step S5. 2.The forest parameter estimation method based on real-time target detection network according to claim 1, wherein: The grid division method in step S1 comprises: 1.

1. Using the watershed algorithm to perform coarse segmentation on the obtained original forest point cloud data to obtain low-precision single tree point cloud data, and calculating the centroid point of each single tree point cloud; 1.

2. Dividing the whole area of the original forest point cloud data into 100m*100m grids, and using the position of the centroid point to determine the grid to which the corresponding point cloud belongs. 3.The forest parameter estimation method based on real-time object detection network according to claim 2, wherein: In step 1.2, the large scene forest point cloud is divided into small scene point clouds along the crown edge to ensure the integrity of the crown point cloud, so as to achieve the construction of effective training data. 4.The forest parameter estimation method based on real-time target detection network according to claim 1, wherein: In step S2, the preprocessing of the small area forest point cloud comprises: 2.

1. Using the progressive triangulation encryption filtering algorithm to remove the ground points in the forest point cloud; 2.

2. Using the voxel filtering algorithm to remove the noise points generated during radar scanning. 5.The forest parameter estimation method based on real-time target detection network according to claim 1, wherein: In step 4.1, the CBAM convolutional attention mechanism described is composed of channel attention mechanism and spatial attention mechanism, and the specific principle is: In the formula, Q is the input feature map, Q' is the feature information obtained after Q is weighted by channel attention, and Q" is the feature information obtained after Q' is weighted by spatial attention mechanism. In step 4.2, the Recursive-FPN feature is characterized by: M c (Q) and M s (Q') are weight information generated by the channel attention mechanism and the spatial attention mechanism, respectively; denotes a bitwise multiplication between feature maps. 6.The forest parameter estimation method based on real-time object detection network according to claim 1, wherein: The fused features of the traditional FPN are input to the backbone network again for multiple cycles to effectively utilize the feature map information. For the acquisition of the average single tree height parameter in step S6, the highest point of the single tree point cloud and the lowest point of the single tree point cloud are used to calculate the height difference, and the specific formula is as follows, 7.The forest parameter estimation method based on real-time object detection network according to claim 1, wherein: For the estimation of the average single tree diameter in step S6, the point cloud circle fitting is performed on the slice of the single tree point cloud at 1.8m-2m. T h = z(P top ) - z(P ground ) (2) z(P top ) represents the highest point of the tree top height of the single tree point cloud, and z(P ground ) represents the lowest point of the location height of the single tree. ​

Images