Forest biomass edge error correction method based on LiDAR and HSI fusion

Abstract

The invention discloses a forest biomass edge error correction method based on LiDAR and HSI fusion, and belongs to the technical field of forest biomass monitoring. The invention aims to solve the problem of constructing an edge error correction system with multi-source data collaborative acquisition, high-precision model driving and automatic process integration. The method comprises the following steps: constructing a single tree segmentation data set based on a forest land multi-source data set: constructing a single tree segmentation training data set fusing a foundation LiDAR point cloud and an unmanned aerial vehicle LiDAR point cloud; constructing an individual tree point cloud three-dimensional segmentation model, and performing individual tree instance segmentation on the forest land three-dimensional point cloud data to obtain point cloud data corresponding to each individual tree; obtaining crown HSI information, constructing a tree species identification model based on the crown HSI information, performing edge tree judgment on the forest single tree subjected to tree species identification, and performing edge error correction and quadrat biomass calculation; and outputting a calibrated quadrat biomass result, wherein the calibrated quadrat biomass result comprises a corrected quadrat biomass distribution diagram, a weight distribution thermodynamic diagram and an error analysis report.

Description

Technical Field

[0001] This invention belongs to the field of forest biomass monitoring technology, specifically relating to a forest biomass edge error correction method based on LiDAR and HSI fusion. Background Technology

[0002] Forest biomass is a core indicator for assessing global carbon sink potential, biodiversity maintenance capacity, and ecosystem service provision levels. Accurate estimation of forest biomass is crucial for climate change response and the formulation of sustainable development strategies. While the development of remote sensing technology has provided technical support for multi-scale monitoring of forest biomass, edge canopy attribution error (ECAE) remains a key bottleneck restricting estimation accuracy.

[0003] Current technologies have the following problems:

[0004] Limited data acquisition dimensions: UAV LiDAR is prone to missing detailed information about the lower canopy and tree trunk, while ground-based LiDAR has limited coverage, making it difficult to meet the needs of both large-scale monitoring and high-precision detail; insufficient fusion of hyperspectral data and LiDAR data, and tree species identification accuracy is greatly affected by spectral noise.

[0005] The edge error correction method is imperfect: the traditional trunk center method determines the quadrat assignment based solely on the trunk position, completely ignoring the overlapping characteristics of the canopy space, which leads to deviations in the estimation of biomass at the quadrat scale; the existing weight allocation model does not incorporate three-dimensional canopy structure parameters, and the correction logic lacks physical mechanism support.

[0006] Poor adaptability of allometric growth equations: The general allometric equations do not take into account tree species specificity and LiDAR three-dimensional structural information, making it difficult to accurately reflect the correlation between individual tree biomass and morphological parameters; the lack of calibration mechanisms between different data sources leads to amplified errors in biomass estimation.

[0007] The system has a low degree of automation: existing calibration tools are mostly single-function modules, lacking full-process integration from data acquisition and processing to modeling and output; uncertainty is not quantified, making it difficult to meet the technical support needs for revising forest resource survey standards. Summary of the Invention

[0008] The problem this invention aims to solve is to construct an edge error correction system that integrates multi-source data collaborative acquisition, high-precision model-driven, and automated processes. It proposes a forest biomass edge error correction method based on the fusion of LiDAR (Light Detection and Ranging) and HSI (Hyperspectral Imaging).

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A forest biomass edge error correction method based on LiDAR and HSI fusion includes the following steps:

[0011] S1. Collect ground-based LiDAR data, UAV LiDAR data, and UAV HSI data, and preprocess them to obtain a multi-source forest dataset containing 900 sample plots;

[0012] S2. Construct a single-tree segmentation dataset based on the forest land multi-source dataset obtained in step S1: Select 100 complete sample plots, manually label the single-tree point cloud, and construct a single-tree segmentation training dataset that integrates ground-based LiDAR point cloud and UAV LiDAR point cloud;

[0013] S3. Construct a 3D segmentation model for single tree point clouds. Train the model based on the single tree segmentation training dataset obtained in step S2. Segment the forest 3D point cloud data into single tree instances to obtain the point cloud data corresponding to each single tree.

[0014] S4. Obtain the crown HSI information of the individual forest trees segmented in step S3, and construct a tree species identification model based on the crown HSI information to obtain the individual forest trees that have been identified by tree species;

[0015] S5. For individual trees in the forest identified in step S4, determine the edge trees, and perform edge error correction and quadrat biomass calculation;

[0016] S6. Output the calibrated quadrat biomass results, including the calibrated quadrat biomass distribution map, weighted heatmap, and error analysis report.

[0017] Furthermore, the specific implementation method of step S1 includes the following steps:

[0018] S1.1. Set up a ground-based LiDAR to perform multi-station scanning of a 300m×300m fixed sample plot, with each station's scanning coverage overlapping by ≥30%, and collect point cloud data of individual tree trunks and the lower part of the canopy; deploy an unmanned aerial vehicle (UAV) platform based on differential positioning principles, flying along a grid flight path, and simultaneously collecting UAV LiDAR point cloud, HSI image data, and Position and Orientation System (POS) data; set the flight path spacing to ≤10m;

[0019] S1.2. Statistical filtering was used to remove noise points from the collected ground-based LiDAR data. The height threshold was set to ±0.2m for ground points and ≤ the highest tree height in the sample plot +1m for canopy points.

[0020] S1.3. The collected UAV LiDAR data was processed using a cloth-based simulation filtering method to separate ground points from vegetation points, generating a digital elevation model (DEM) and a canopy height model (CHM), and extracting the individual tree canopy volume V. UAV Projected area S UAV ;

[0021] S1.4. Atmospheric correction was performed on the acquired HSI image data using the Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) module. The Digital Number (DN) values ​​were converted to reflectance, and the red edge slope and moisture index (NDWI) were extracted. The calculation formula is as follows:

[0022]

[0023] Among them, P 绿 P represents the green band reflectivity. 近红外 Reflectivity in the near-infrared band;

[0024] S1.5. Spatiotemporally align the data obtained in step S1. Establish a Cartesian coordinate system with the lower left corner of the quadrat as the origin. Based on three or more control points of the same name, achieve the unification of coordinates of the three source data through affine transformation (AT). The alignment error is ≤0.1m.

[0025] Furthermore, the specific implementation method of step S3 includes the following steps:

[0026] S3.1. Construct a three-dimensional segmentation model for single-tree point clouds based on deep learning. The model includes a point cloud semantic segmentation sub-network and an instance feature embedding sub-network. The fused ground-based LiDAR point cloud and UAV LiDAR point cloud are used as model inputs. The semantic category label and instance feature vector of each point in the point cloud are predicted respectively.

[0027] S3.2. Based on the instance feature vector and point cloud spatial coordinate information output in step S3.1, the density-aware clustering algorithm is used to divide the point cloud into instances, and the points belonging to the same instance are clustered into the same tree, thereby realizing the three-dimensional segmentation of forest tree point clouds.

[0028] During instance clustering, the clustering distance threshold is adaptively adjusted based on the local height distribution of the point cloud, canopy density, and forest stand structure characteristics to reduce the impact of overlapping areas of adjacent canopies on the segmentation accuracy of individual trees.

[0029] Furthermore, the specific implementation method of step S4 includes the following steps:

[0030] S4.1. Based on the single tree point cloud obtained in step S3, a three-dimensional geometric model of the single tree canopy is constructed using the α-shape algorithm, and the three-dimensional model is projected onto the horizontal plane to obtain the projection area of ​​the single tree canopy;

[0031] S4.2 Based on the projection area of ​​a single tree canopy, the Delaunay triangulation method is used to calculate the projected overlap area (Soverlap) between adjacent single tree canopies. The expression for this area is:

[0032]

[0033] Where A is the canopy projection area of ​​the target quadrat, B is the canopy projection area of ​​the adjacent quadrat, x is the abscissa of the tree projection plane, and y is the ordinate of the tree projection plane; thus, the diameter at breast height (DBH), tree height (H), and canopy volume (V) of a single tree are obtained. ground ;

[0034] S4.3. For the single-tree point cloud identified in step S3, calculate the minimum sample size based on the multinomial distribution algorithm, using the following formula:

[0035]

[0036] Where n is the total number of samples. Let B be the percentage of tree species in category k, B be the chi-square test value of confidence level 1-α / k, K be the total number of tree species categories, and α=0.05 be the significance level. k The expected classification error ratio for the k-th tree species;

[0037] Then, the transform discreteness algorithm is used to evaluate the separability of the samples, as shown in the formula:

[0038]

[0039]

[0040] in, To determine the target transformation dispersion for the two types of tree species, The dispersion of the two tree species is given. , These are the covariance matrices for the two tree species, respectively. , These are the sample mean vectors for the two tree species, respectively. , These represent the degrees of freedom for the corresponding tree species samples;

[0041] S4.4. Feature vectors were constructed based on red edge slope, NDWI, canopy height-to-diameter ratio, and volume dispersion as the training set. An SVM classification model was used for tree species identification through multi-source feature fusion. The SVM classification model was trained using a 5-fold cross-validation method, and the classification results were verified through Kappa coefficient analysis. The formula is as follows:

[0042]

[0043] in, Here, r represents the Kappa coefficients, and r is the number of columns in the error matrix. The number of diagonal pixels in the error matrix. , ...

[0044] S4.5. Store the spectral characteristics of each tree species after classification into the spectral library. For each identified tree species, call the publicly available maturity allometric equation for the corresponding tree species and substitute it with the obtained DBH and H parameters to calculate the biomass.

[0045] Furthermore, the specific implementation method of step S5 includes the following steps:

[0046] S5.1. Perform edge error correction on individual trees in the forest after tree species identification;

[0047] When the distance from the projection center of a single tree's canopy to the boundary of the quadrat is less than or equal to the canopy radius, it is classified as a marginal tree. The weighting coefficient of the marginal tree's canopy in the target quadrat is calculated using the following formula:

[0048]

[0049] Where w is the weighting coefficient, The projected area of ​​the canopy within the target quadrat. This represents the total canopy projection area of ​​a single tree.

[0050] S5.2. Calculate the biomass of the quadrat. For non-marginal trees, calculate the biomass using the general allometric equation. ,for

[0051] Calculate the total biomass of edge trees first. Then, according to the weights, the biomass is allocated to the target quadrat and adjacent quadrats. The biomass allocated to the target quadrat is:

[0052]

[0053] in, Allocate biomass to the target quadrat;

[0054] The consistency correlation coefficient (CCC) and root mean square error (RMSE) are used to evaluate the correction accuracy, as shown in the following formula:

[0055]

[0056]

[0057] Where ρ is the correlation coefficient between the measured and predicted values, σ1 and σ2 are the standard deviations of the measured and predicted biomass, respectively, and μ1 and μ2 are the means of the measured and predicted biomass, respectively. To measure biomass, To predict biomass;

[0058] The consistency correlation coefficient (CCC) and root mean square error (RMSE) are set to evaluate the correction accuracy, with the target CCC ≥ 0.90 and RMSE ≤ 0.3t / tree.

[0059] Uncertainty analysis was performed using Monte Carlo simulation with 1000 iterations. The individual tree segmentation boundaries and allometric equation parameters were randomly perturbed, and the 95% confidence interval for biomass estimation was calculated. :

[0060]

[0061] in, To estimate the mean biomass of quadrats based on edge error correction, The standard deviation of the biomass estimation results obtained from the Monte Carlo simulation.

[0062] Furthermore, step S6 develops an automated correction system based on Python+QGIS, integrating modules for data import, single-tree segmentation, tree species identification, biomass calculation, and error correction, and adopts Dask parallel computing technology to support batch processing.

[0063] The beneficial effects of this invention are:

[0064] This invention presents a forest biomass edge error correction method based on LiDAR and HSI fusion. Through a three-dimensional canopy attribution model, it significantly reduces edge errors compared to traditional methods, thereby significantly improving the accuracy of biomass estimation. It fully leverages the synergistic advantages of LiDAR and hyperspectral imaging, taking into account both structural and spectral characteristics to enhance model robustness and applicability. This provides a technical basis for continuous forest resource inventory and contributes to precise monitoring of forest carbon sequestration and the development of smart forestry. Attached Figure Description

[0065] Figure 1 This is a flowchart of a forest biomass edge error correction method based on LiDAR and HSI fusion as described in this invention;

[0066] Figure 2 This is a structural block diagram of the single-tree point cloud three-dimensional segmentation model of the present invention. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The components of the specific embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations, and the invention may also have other embodiments.

[0068] Therefore, the following detailed description of specific embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected specific embodiments of the invention. All other specific embodiments obtained by those skilled in the art based on these specific embodiments without inventive effort are within the scope of protection of this invention.

[0069] To further understand the invention's content, features, and effects, the following specific embodiments are provided, along with accompanying drawings. Figure 1 and attached Figure 2 Detailed explanation is as follows:

[0070] Example 1:

[0071] A forest biomass edge error correction method based on LiDAR and HSI fusion includes the following steps:

[0072] S1. Collect ground-based LiDAR data, UAV LiDAR data, and UAV hyperspectral data, and preprocess them to obtain a dataset;

[0073] Furthermore, the specific implementation method of step S1 includes the following steps:

[0074] S1.1. Set up a ground-based LiDAR to perform multi-station scanning of a 300m×300m fixed sample plot (containing 900 10m*10m quadrats), with each station's scan coverage overlapping by ≥30%, and collect point cloud data of individual tree trunks and the lower part of the canopy; deploy an unmanned aerial vehicle (UAV) platform based on differential positioning principles, flying along a grid flight path, and simultaneously collecting UAV LiDAR point cloud, HSI image data, and POS (Position and Orientation System) data; set the flight path spacing to ≤10m;

[0075] S1.2. Statistical filtering was used to remove noise points from the collected ground-based LiDAR data. The height threshold was set to ±0.2m for ground points and ≤ the highest tree height in the sample plot +1m for canopy points.

[0076] S1.3. The collected UAV LiDAR data was processed using a cloth-based simulation filtering method to separate ground points from vegetation points, generating a digital elevation model (DEM) and a canopy height model (CHM), and extracting the individual tree canopy volume V. UAV Projected area S UAV ;

[0077] S1.4. The acquired HSI image data is atmospherically corrected using the FLAASH (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes) module. The DN values ​​(Digital Number) are converted to reflectance, and the red edge slope and moisture index NDWI are extracted. The calculation formula is as follows:

[0078]

[0079] Among them, P 绿 P represents the green band reflectivity. 近红外 Reflectivity in the near-infrared band;

[0080] S1.5. Spatiotemporally align the data obtained in step S1. Establish a Cartesian coordinate system with the lower left corner of the quadrat as the origin. Based on three or more control points of the same name, unify the coordinates of the three source data through AT transformation (Affine Transformation). The alignment error is ≤0.1m.

[0081] S2. Construct a segmentation dataset based on the dataset obtained in step S1: Select 100 complete sample plots, manually label the single tree point cloud, and construct a training set that integrates ground-based LiDAR trunk point cloud and UAV LiDAR canopy point cloud;

[0082] S3. Construct a three-dimensional model for dividing individual tree canopies, and train it using the training set obtained in step S2 to achieve the division of individual tree canopies in the forest;

[0083] Furthermore, the specific implementation method of step S3 includes the following steps:

[0084] S3.1. Construct a three-dimensional segmentation model for single-tree point clouds based on deep learning. The model includes a point cloud semantic segmentation sub-network and an instance feature embedding sub-network. The fused ground-based LiDAR point cloud and UAV LiDAR point cloud are used as model inputs. The semantic category label and instance feature vector of each point in the point cloud are predicted respectively.

[0085] S3.2. Based on the instance feature vector and point cloud spatial coordinate information output in step S3.1, the density-aware clustering algorithm is used to divide the point cloud into instances, and the points belonging to the same instance are clustered into the same tree, thereby realizing the three-dimensional segmentation of forest tree point clouds.

[0086] During instance clustering, the clustering distance threshold is adaptively adjusted based on the local height distribution of the point cloud, canopy density, and forest stand structure characteristics to reduce the impact of overlapping areas of adjacent canopies on the segmentation accuracy of individual trees.

[0087] Furthermore, Figure 2 This is a structural diagram of a 3D segmentation model for single-tree point clouds. The point cloud feature encoding module extracts and encodes features from the input 3D point cloud data. The input point cloud includes spatial coordinates and attributes such as intensity and height. This module models the local neighborhood structure and global spatial distribution features of the point cloud, extracting high-dimensional feature representations reflecting the spatial geometric relationships and height distribution characteristics of the point cloud, providing a unified feature foundation for subsequent semantic segmentation and instance differentiation. The semantic segmentation prediction module predicts the semantic category of each point in the point cloud based on the high-dimensional features output by the point cloud feature encoding module, determining whether each point belongs to a tree. The semantic segmentation prediction results are used to filter out non-tree points, reducing the interference of ground points, noise points, and non-vegetation points on the single-tree segmentation results. The instance feature embedding generation module generates a corresponding instance feature vector for each tree point, ensuring high similarity between points belonging to the same tree in the instance feature space, and high distinguishability between points belonging to different trees. Through instance feature embedding, effective differentiation of different single-tree points within spatially adjacent and overlapping canopy regions is achieved. The instance clustering module is used to fuse instance feature vectors, point cloud spatial coordinate information and semantic segmentation results to perform instance clustering on tree points and output the clustering results as single tree point cloud instances.

[0088] S4. For the individual forest trees identified in step S3, perform multi-source feature fusion for tree species identification;

[0089] Furthermore, the specific implementation method of step S4 includes the following steps:

[0090] S4.1. Based on the single tree point cloud obtained in step S3, a three-dimensional geometric model of the single tree canopy is constructed using the α-shape algorithm, and the three-dimensional model is projected onto the horizontal plane to obtain the projection area of ​​the single tree canopy;

[0091] S4.2 Based on the projection area of ​​a single tree canopy, the Delaunay triangulation method is used to calculate the projected overlap area (Soverlap) between adjacent single tree canopies. The expression for this area is:

[0092]

[0093] Where A is the canopy projection area of ​​the target quadrat, B is the canopy projection area of ​​the adjacent quadrat, x is the abscissa of the tree projection plane, and y is the ordinate of the tree projection plane; thus, the diameter at breast height (DBH), tree height (H), and canopy volume (V) of a single tree are obtained. ground ;

[0094] S4.3. For the single-tree point cloud identified in step S3, calculate the minimum sample size based on the multinomial distribution algorithm, using the following formula:

[0095]

[0096] Where n is the total number of samples. Let B be the percentage of tree species in category k, B be the chi-square test value of confidence level 1-α / k, K be the total number of tree species categories, and α=0.05 be the significance level. k The expected classification error ratio for the k-th tree species;

[0097] Then, the transform discreteness algorithm is used to evaluate the separability of the samples, as shown in the formula:

[0098]

[0099]

[0100] in, To determine the target transformation dispersion for the two types of tree species, The dispersion of the two tree species is given. , These are the covariance matrices for the two tree species, respectively. , These are the sample mean vectors for the two tree species, respectively. , These represent the degrees of freedom for the corresponding tree species samples;

[0101] Furthermore, the red edge slope (720-740nm): directly output from the hyperspectral data processing in step S1. Specifically, it calculates the ratio of the reflectance difference between the 720nm and 740nm bands to the wavelength difference in the atmospherically corrected hyperspectral reflectance image, reflecting the difference in chlorophyll content of vegetation, and is a core spectral indicator for tree species identification. The moisture index (NDWI): also derived from the hyperspectral data processing in step S1, reflects leaf moisture content and helps distinguish tree species. The canopy height-to-diameter ratio is calculated from the ratio of the individual tree height to the canopy projected area (Suav) calculated in step S3; the volume dispersion is calculated from the ratio of the standard deviation of the UAV LiDAR canopy volume (Vuav) extracted in step S1 to the mean of the point cloud volume (V subset) of each sub-cluster after individual tree segmentation in step S3.

[0102] S4.4. Feature vectors were constructed based on red edge slope, NDWI, canopy height-to-diameter ratio, and volume dispersion as the training set. An SVM classification model was used for tree species identification through multi-source feature fusion. The SVM classification model was trained using a 5-fold cross-validation method. The classification results of the SVM model were verified through Kappa coefficient analysis. The formula is as follows:

[0103]

[0104] in, Here, r represents the Kappa coefficients, and r is the number of columns in the error matrix. The number of diagonal pixels in the error matrix. , ...

[0105] S4.5. Store the spectral characteristics of each tree species after classification into the spectral library. For each identified tree species, call the publicly available maturity allometric equation for the corresponding tree species and substitute it with the obtained DBH and H parameters to calculate the biomass.

[0106] S5. For individual trees in the forest that have been identified by tree species, perform edge error correction and quadrat biomass calculation;

[0107] Furthermore, the specific implementation method of step S5 includes the following steps:

[0108] S5.1. Perform edge error correction on individual trees in the forest after tree species identification;

[0109] When the distance from the projection center of a single tree's canopy to the boundary of the quadrat is less than or equal to the canopy radius, it is classified as a marginal tree. The weighting coefficient of the marginal tree's canopy in the target quadrat is calculated using the following formula:

[0110]

[0111] Where w is the weighting coefficient, The projected area of ​​the canopy within the target quadrat. This represents the total canopy projection area of ​​a single tree.

[0112] Furthermore, (Projected area of ​​the canopy within the target quadrat) and The total canopy projection area of ​​a single tree comes from steps S1 (UAV LiDAR data processing) and S3 (single tree segmentation and canopy modeling). In step S1, the UAV LiDAR data is filtered to generate a canopy height model (CHM). In step S3, a modified PointNet++ algorithm is used to segment the tree, extracting all point clouds of the tree individually. Then, through Delaunay triangulation or rasterization statistics, the projected area of ​​the entire canopy on the ground is calculated. Simply put, it's the area of ​​the entire shadow cast by the tree's canopy on the ground. Step S1 has already defined the boundary coordinates of the quadrat (e.g., the coordinates of the four corner points of a 10m × 10m quadrat); after segmenting the canopy of a single tree in step S3, an intersection analysis is performed between the canopy projection surface and the quadrat boundary, that is, the area of ​​the entire canopy shadow that falls within the boundary of the target quadrat is calculated. Projected area of ​​the canopy within the target quadrat.

[0113] S5.2. Calculate the biomass of the quadrat. For non-marginal trees, calculate the biomass using the general allometric equation. ,for

[0114] First calculate the total biomass of edge trees. Then, according to the weights, the biomass is allocated to the target quadrat and adjacent quadrats. The biomass allocated to the target quadrat is:

[0115]

[0116] in, Allocate biomass to the target quadrat;

[0117] The consistency correlation coefficient (CCC) and root mean square error (RMSE) are used to evaluate the correction accuracy, as shown in the following formula:

[0118]

[0119]

[0120] Where ρ is the correlation coefficient between the measured and predicted values, σ1 and σ2 are the standard deviations of the measured and predicted biomass, respectively, and μ1 and μ2 are the means of the measured and predicted biomass, respectively. To measure biomass, To predict biomass;

[0121] The consistency correlation coefficient (CCC) and root mean square error (RMSE) are set to evaluate the correction accuracy, with the target CCC ≥ 0.90 and RMSE ≤ 0.3t / tree.

[0122] Uncertainty analysis was performed using Monte Carlo simulation with 1000 iterations. The individual tree segmentation boundaries and allometric equation parameters were randomly perturbed, and the 95% confidence interval for biomass estimation was calculated. :

[0123]

[0124] in, To estimate the mean biomass of quadrats based on edge error correction, The standard deviation of the biomass estimation results obtained from the Monte Carlo simulation.

[0125] S6. Output the results, including the corrected quadrat biomass distribution map, weighted heatmap, and error analysis report.

[0126] Furthermore, step S6 develops an automated correction system based on Python+QGIS, integrating modules for data import, single-tree segmentation, tree species identification, biomass calculation, and error correction, and adopts Dask parallel computing technology to support batch processing.

[0127] Further outputs include: a corrected quadrat biomass distribution map, a weighted heatmap, an error analysis report, and a standardized operation manual. The target margin error is ≤8%.

[0128] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0129] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A forest biomass edge error correction method based on LiDAR and HSI fusion, characterized in that, Includes the following steps: S1. Collect ground-based LiDAR data, UAV LiDAR data, and UAV HSI data, and preprocess them to obtain a multi-source forest dataset containing 900 sample plots; S2. Construct a single-tree segmentation dataset based on the forest land multi-source dataset obtained in step S1: Select 100 complete sample plots, manually label the single-tree point cloud, and construct a single-tree segmentation training dataset that integrates ground-based LiDAR point cloud and UAV LiDAR point cloud; S3. Construct a 3D segmentation model for single tree point clouds. Train the model based on the single tree segmentation training dataset obtained in step S2. Segment the forest 3D point cloud data into single tree instances to obtain the point cloud data corresponding to each single tree. S4. Obtain the crown HSI information of the individual forest trees segmented in step S3, and construct a tree species identification model based on the crown HSI information to obtain the individual forest trees that have been identified by tree species; S5. For individual trees in the forest identified in step S4, determine the edge trees, and perform edge error correction and quadrat biomass calculation; S6. Output the calibrated quadrat biomass results, including the calibrated quadrat biomass distribution map, weighted heatmap, and error analysis report.

2. The forest biomass edge error correction method based on LiDAR and HSI fusion according to claim 1, characterized in that, The specific implementation method of step S1 includes the following steps: S1.

1. Set up a ground-based LiDAR to perform multi-station scanning of a 300m×300m fixed sample plot, with each station's scanning coverage overlapping by ≥30%, and collect point cloud data of individual tree trunks and the lower part of the canopy; deploy an unmanned aerial vehicle (UAV) platform based on differential positioning principles, flying along a grid flight path, and simultaneously collecting UAV LiDAR point cloud, HSI image data, and positioning and orientation system (POS) data; set the flight path spacing to ≤10m; S1.

2. Statistical filtering was used to remove noise points from the collected ground-based LiDAR data. The height threshold was set to ±0.2m for ground points and ≤ the highest tree height in the sample plot +1m for canopy points. S1.

3. The collected UAV LiDAR data was processed using a cloth-based simulation filtering method to separate ground points from vegetation points, generating a digital elevation model (DEM) and a canopy height model (CHM), and extracting the individual tree canopy volume V. UAV Projected area S UAV ; S1.

4. Atmospheric correction was performed on the acquired HSI image data using the FLAASH module (Fast Line Atmospheric Hyperspectral Cube Analysis). The digital quantization (DN) value was converted to reflectance, and the red edge slope and moisture index (NDWI) were extracted. The calculation formula is as follows: Among them, P 绿 P represents the green band reflectivity. 近红外 Reflectivity in the near-infrared band; S1.

5. Spatiotemporally align the data obtained in step S1. Establish a Cartesian coordinate system with the lower left corner of the quadrat as the origin. Based on three or more control points of the same name, achieve coordinate unification of the three source data through affine transformation AT, with an alignment error ≤0.1m.

3. The forest biomass edge error correction method based on LiDAR and HSI fusion according to claim 2, characterized in that, The specific implementation method of step S3 includes the following steps: S3.

1. Construct a 3D segmentation model for single-tree point clouds based on deep learning. The model includes a point cloud semantic segmentation subnetwork and an instance feature embedding subnetwork. The fused ground-based LiDAR point cloud and UAV LiDAR point cloud are used as model inputs. The semantic category label and instance feature vector of each point in the point cloud are predicted respectively. S3.

2. Based on the instance feature vector and point cloud spatial coordinate information output in step S3.1, the density-aware clustering algorithm is used to divide the point cloud into instances, and the points belonging to the same instance are clustered into the same tree, thereby realizing the three-dimensional segmentation of forest tree point clouds. During instance clustering, the clustering distance threshold is adaptively adjusted based on the local height distribution of the point cloud, canopy density, and forest stand structure characteristics to reduce the impact of overlapping areas of adjacent canopies on the segmentation accuracy of individual trees.

4. The forest biomass edge error correction method based on LiDAR and HSI fusion according to claim 3, characterized in that, The specific implementation method of step S4 includes the following steps: S4.

1. Based on the single tree point cloud obtained in step S3, a three-dimensional geometric model of the single tree canopy is constructed using the α-shape algorithm, and the three-dimensional model is projected onto the horizontal plane to obtain the projection area of ​​the single tree canopy; S4.2 Based on the projection area of ​​a single tree canopy, the Delaunay triangulation method is used to calculate the projected overlap area (Soverlap) between adjacent single tree canopies. The expression for this area is: Where A is the canopy projection area of ​​the target quadrat, B is the canopy projection area of ​​the adjacent quadrat, x is the abscissa of the tree projection plane, and y is the ordinate of the tree projection plane; thus, the diameter at breast height (DBH), tree height (H), and canopy volume (V) of a single tree are obtained. ground ; S4.

3. For the single-tree point cloud identified in step S3, calculate the minimum sample size based on the multinomial distribution algorithm, using the following formula: Where n is the total number of samples. Let B be the percentage of tree species in category k, B be the chi-square test value of confidence level 1-α / k, K be the total number of tree species categories, and α=0.05 be the significance level. k The expected classification error ratio for the k-th tree species; Then, the transform discreteness algorithm is used to evaluate the separability of the samples, as shown in the formula: in, To determine the target transformation dispersion for the two types of tree species, The dispersion of the two tree species is given. , These are the covariance matrices for the two tree species, respectively. , These are the sample mean vectors for the two tree species, respectively. , These represent the degrees of freedom for the corresponding tree species samples; S4.

4. Feature vectors were constructed based on red edge slope, NDWI, canopy height-to-diameter ratio, and volume dispersion as the training set. An SVM classification model was used for tree species identification through multi-source feature fusion. The SVM classification model was trained using a 5-fold cross-validation method. The classification results of the SVM model were verified through Kappa coefficient analysis. The formula is as follows: in, Here, r represents the Kappa coefficients, and r is the number of columns in the error matrix. The number of diagonal pixels in the error matrix. , ... S4.

5. Store the spectral characteristics of each tree species after classification into the spectral library. For each identified tree species, call the publicly available maturity allometric equation for the corresponding tree species and substitute it with the obtained DBH and H parameters to calculate the biomass.

5. The forest biomass edge error correction method based on LiDAR and HSI fusion according to claim 4, characterized in that, The specific implementation method of step S5 includes the following steps: S5.

1. Perform edge error correction on individual trees in the forest after tree species identification; When the distance from the projection center of a single tree's canopy to the boundary of the quadrat is less than or equal to the canopy radius, it is classified as a marginal tree. The weighting coefficient of the marginal tree's canopy in the target quadrat is calculated using the following formula: Where w is the weighting coefficient, The projected area of ​​the canopy within the target quadrat. This represents the total canopy projection area of ​​a single tree. S5.

2. Calculate the biomass of the quadrat. For non-marginal trees, calculate the biomass using the general allometric equation. ,for First calculate the total biomass of edge trees. Then, according to the weights, the biomass is allocated to the target quadrat and adjacent quadrats. The biomass allocated to the target quadrat is: in, Allocate biomass to the target quadrat; The consistency correlation coefficient (CCC) and root mean square error (RMSE) are used to evaluate the correction accuracy, as shown in the following formula: Where ρ is the correlation coefficient between the measured and predicted values, σ1 and σ2 are the standard deviations of the measured and predicted biomass, respectively, and μ1 and μ2 are the means of the measured and predicted biomass, respectively. To measure biomass, To predict biomass; The consistency correlation coefficient (CCC) and root mean square error (RMSE) are set to evaluate the correction accuracy, with the target CCC ≥ 0.90 and RMSE ≤ 0.3t / tree. Uncertainty analysis was performed using Monte Carlo simulation with 1000 iterations. The individual tree segmentation boundaries and allometric equation parameters were randomly perturbed, and the 95% confidence interval for biomass estimation was calculated. : in, To estimate the mean biomass of quadrats based on edge error correction, The standard deviation of the biomass estimation results obtained from the Monte Carlo simulation.

6. The forest biomass edge error correction method based on LiDAR and HSI fusion according to claim 5, characterized in that, Step S6 is based on Python+QGIS to develop an automated correction system, which integrates modules for data import, single tree segmentation, tree species identification, biomass calculation, and error correction, and uses Dask parallel computing technology to support batch processing.

Images