Method for retrieving leaf area index of desert shrub based on micro-pixel analysis projection model
By constructing a microvoxel analysis projection model, generating 3D point cloud data using UAVs, and improving the projection function, the accuracy and applicability issues of desert shrub LAI measurement were solved, achieving high-precision, non-destructive LAI inversion.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XINJIANG UNIVERSITY
- Filing Date
- 2026-03-02
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies for measuring the leaf area index (LAI) of desert shrubs are highly destructive, inefficient, inaccurate, and difficult to adapt to the unique morphological characteristics of desert shrubs, resulting in large measurement errors.
A microvoxel analysis-based projection model was adopted. Three-dimensional point cloud data was generated by multi-angle photography from UAVs. SfM and MVS technologies were combined to perform fine branch and leaf separation, construct a microvoxel analysis model, improve the projection function, calculate leaf area density and porosity, and invert LAI.
It achieves high-precision, non-destructive, all-weather LAI inversion, breaking through the bottlenecks of the uniformity assumption and fine structure detection in traditional methods, and improving measurement accuracy and applicability.
Smart Images
Figure CN122134881A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of quantitative remote sensing technology, and in particular to a method for inverting the leaf area index of desert shrubs based on a microvoxel analysis projection model. Background Technology
[0002] In desert ecosystems, leaf area index (LAI) is a core indicator for assessing shrub growth, monitoring density, and reflecting water cycle, carbon sequestration, and ecosystem stability. In environments of extreme drought and fragile ecological balance, accurate LAI monitoring is crucial for early identification of degradation signals and for supporting conservation strategies.
[0003] Currently, the leaf area index (LAI) of desert shrubs is mainly measured using direct harvesting and indirect optical methods. Direct harvesting is highly destructive and inefficient, failing to meet the requirements for large-scale non-destructive methods. While convenient, indirect optical methods are susceptible to weather and light interference and lack detailed leaf structure information. For the unique morphology of desert shrubs (such as Haloxylon ammodendron and Tamarix chinensis) with sparse canopies, large gaps, small assimilating branches (<2mm in diameter), clustered distribution, and a high proportion of wood, the uniform random distribution assumption relied upon by Beer's Law leads to severe distortion of the projection function, resulting in a significant underestimation of LAI. Simultaneously, traditional optical instruments struggle to effectively distinguish between woody and leaf components, causing a significant overestimation of LAI. LiDAR, with its extremely low reflectivity of fine needle-like assimilating branches and sparse point clouds, further amplifies the error. These shortcomings severely limit the accuracy and applicability of existing methods in desert shrub scenarios. Summary of the Invention
[0004] The technical problem to be solved by the present invention is to provide a high-precision, non-destructive inversion method for the leaf area index of desert shrubs based on a microvoxel analysis projection model.
[0005] To address the above problems, the present invention provides a method for inverting the leaf area index of desert shrubs based on a microvoxel analysis projection model, comprising the following steps: S1 acquires multi-angle image data of desert shrubs and constructs 3D point cloud data: Typical desert shrubs were selected as observation targets in the field experimental area. UAVs or ground photography equipment were used to take pictures of the shrubs from multiple angles. After collecting the two-dimensional image sequence from multiple perspectives, three-dimensional reconstruction was carried out using processing software based on structure of motion recovery (SfM) and multi-view stereo vision (MVS) technology to generate high-density three-dimensional point cloud data containing fine spatial structure information of the shrubs. S2 performs preprocessing and fine branch / leaf separation on the raw point cloud data: First, the original point cloud is converted into a unified relative coordinate format and outliers are removed. Then, cKDTree is used to accelerate the query calculation of the local density of the point cloud, extract the connection length feature, and apply an unsupervised Gaussian mixture model (GMM) to fit the connection length distribution and adaptively obtain the segmentation threshold. Finally, the density-based DBSCAN clustering algorithm is combined to accurately divide the point cloud into leaf point cloud and branch point cloud. S3 constructs a microvoxel analysis model based on the separated leaf point cloud and calculates the leaf area density: The leaf point cloud is vertically layered and voxelized according to a preset height interval. For each micro voxel layer, its horizontal projected area and three-dimensional volume are calculated. Then, based on these geometric parameters, the leaf area density (LAD) of each layer is calculated, which is the ratio of the estimated leaf area to the volume occupied by that layer. Improved derivation of the S4 projection function: By integrating the leaf area density and optical path information within the microscopic voxels, the improved projection function is derived. G ( i Calculation formula: In the formula: i The observed zenith angle of the incident light, in degrees; k is the extinction coefficient of the shrub, dimensionless; LAD The leaf area density after polymerization, in units of m³. 2 / m 3 ; H The height of the shrub canopy is in meters (m). S5 calculates canopy porosity using point cloud data: First, the Fibonacci spherical spiral algorithm is used to generate uniformly distributed simulated ray directions, with the initial ray height randomly sampled within the canopy height range. Then, the distance between the ray and the leaf point cloud is calculated; if the distance is less than a set threshold, the light is considered to be blocked; otherwise, the light is considered to have penetrated. Finally, specific zenith angles are statistically analyzed. i The porosity at that angle is the ratio of the number of rays penetrating the canopy in a given direction to the total number of rays. P ( i ); S6 leaf area index inversion: By observing the porosity at different angles P ( i ) and the corresponding improved projection function G ( i Substituting into the following LAI inversion formula, the accurate leaf area index (LAI) of the target shrub can be obtained. The inversion formula is: .
[0006] In step S1, the shooting scheme is set as follows: the azimuth angle is in the range of 0° to 360°; the zenith angle is in the range of 0° to 75°, and multiple gradients are set.
[0007] In step S3, the horizontal projected area is calculated based on the boundary range of the point cloud layer, and a point density adjustment factor is introduced to compensate for the sparsity of the point cloud data; the volume occupied in three-dimensional space is calculated using the minimum bounding box method.
[0008] Compared with the prior art, the present invention has the following advantages: 1. This invention, by constructing a microvoxel analysis projection model (MVAPM) and generating high-precision point clouds through SFM+MVS multi-view photography, achieves fine branch and leaf segmentation and reconstructs a projection function adapted to cluster structures, completely breaking through the bottlenecks of traditional uniformity assumptions and fine structure detection, and achieving high precision (R... 2 With a maximum RMSE of 0.944 and a minimum RMSE of 0.0043, the non-destructive, all-weather inversion provides a practical and efficient new technological approach for desert ecological monitoring.
[0009] 2. It solved the technical problem that traditional measurement methods are difficult to adapt to the unique morphological characteristics of desert shrubs: In existing technologies, both optical instrument measurement methods based on Beer's law and traditional geometric measurement methods typically assume a uniform distribution of the vegetation canopy medium and rely on a pre-defined leaf tilt angle distribution function (such as an ellipsoidal distribution). However, desert shrubs (such as Haloxylon ammodendron and Tamarix chinensis) exhibit a non-uniform structure characterized by clustered assimilated branches, extremely sparse canopies, and large gaps. Existing methods ignore this structural heterogeneity, leading to significant deviations in projection function calculations. This invention constructs a microvoxel analysis model, discretizing the macroscopic irregular canopy into quantifiable microvoxel units. By deriving the projection function using the leaf area density (LAD) integral within the microvoxels, it overcomes the dependence of traditional models on "uniform canopy" and "specific leaf tilt angle distribution," significantly improving the physical realism and accuracy of LAI inversion for complex shrub structures.
[0010] 3. The calculation logic of the traditional projection function G(θ) has been corrected, improving the universality of the inversion model: In the traditional LAI inversion formula, the projection function G ( iTypically, a projection function is a fixed parameter based on empirical values or a function based on an ideal distribution. This invention not only abandons ideal assumptions but also establishes a physical derivation mechanism based on actual point cloud data. By calculating the leaf area density and optical path information within a single micro-voxel, an improved projection function suitable for the current target shrub is directly derived. Experimental data show that the projection function values calculated by this method are in high agreement with measured values at different observation angles (with extremely high correlation coefficients), especially exhibiting stability far exceeding that of traditional models in the mid-to-high zenith angle region. This means that the method of this invention has stronger environmental adaptability and can accurately reflect the light interception characteristics of desert shrubs at different growth stages and morphologies. Attached Figure Description
[0011] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0012] Figure 1 This is a roadmap of the algorithm technology of this invention.
[0013] Figure 2 This is a comparison image of point cloud branch and leaf segmentation using the algorithm in this invention.
[0014] Figure 3 This describes the calculation process of the projection function when point clouds are combined with unit voxels in this invention. Wherein: (a) represents the macroscopic overall voxel; (b) represents the microscopic individual voxel.
[0015] Figure 4 This is a flowchart illustrating the workflow of generating 3D point clouds using COLMAP combined with multi-angle scene photos in an embodiment of the present invention.
[0016] Figure 5 This is a comparison of the segmentation accuracy of point cloud branches and leaves in a desert vegetation scene in an embodiment of the present invention. Wherein: (a)-(d): Value distribution of branch and leaf segmentation accuracy index for tree type 1, tree type 2, tree type 3 and tree type 4.
[0017] Figure 6 This is a comparison of gap probability values in the embodiments of the present invention. Wherein: (a)-(d): gap probability calculated by the algorithm for tree species 1 to tree species 4 and the calculated value by the LESS four-flux model; (e)-(f): gap probability values of the point cloud algorithm for Haloxylon ammodendron and Tamarisk and the gap probability values measured in the field. Detailed Implementation
[0018] like Figure 1 As shown, the method for inverting the leaf area index of desert shrubs based on a microvoxel analysis projection model includes the following steps: S1 acquires multi-angle image data of desert shrubs and constructs 3D point cloud data: Typical desert shrubs were selected as observation targets in the field experimental area, and multi-angle photography was carried out around the shrubs using drones or ground photography equipment. To ensure the integrity of the subsequent 3D reconstruction, the shooting angles were strictly controlled during the shooting process: the azimuth angle was within the range of 0° to 360°; the zenith angle was within the range of 0° to 75° (including key angles such as 15°, 30°, 45°, 60° and 75°), and multiple gradients were set.
[0019] After acquiring multi-view 2D image sequences, 3D reconstruction is performed using processing software (such as COLMAP) based on Structure of Motion (SfM) and Multi-View Stereo Vision (MVS) technologies. This process sequentially involves feature point extraction, image matching, sparse point cloud reconstruction, and dense point cloud reconstruction. Ultimately, high-density 3D point cloud data containing detailed spatial structure information of shrubs is generated.
[0020] S2 performs preprocessing and fine branch / leaf separation on the raw point cloud data: Because the non-photosynthetic xylem (trunk and branches) accounts for a large proportion of desert shrubs, its failure to be removed would significantly affect the accuracy of leaf area index calculation. Therefore, this invention employs a tree structure-guided segmentation process.
[0021] First, the original point cloud is converted to a unified relative coordinate format and outliers are removed. Then, cKDTree is used to accelerate the calculation of local density in the point cloud, extracting connection length features. An unsupervised Gaussian Mixture Model (GMM) is applied to fit the connection length distribution, thus adaptively obtaining the segmentation threshold and overcoming the limitations of manually setting thresholds. Finally, the density-based DBSCAN clustering algorithm is used to refine the preliminary classification results, precisely dividing the point cloud into leaf point clouds (i.e., assimilated branch parts) and branch point clouds (i.e., xylem parts). Figure 2 As shown, the separated leaf point cloud can accurately reflect the photosynthetic area of the shrub, providing a clean data foundation for subsequent calculations.
[0022] S3 constructs a microvoxel analysis model based on the separated leaf point cloud and calculates the leaf area density: To quantify the non-uniform structure within the canopy, this invention discretizes the macroscopic shrub canopy into tiny voxel units in three-dimensional space. Specifically, the leaf point cloud is vertically layered and voxelized at preset height intervals (e.g., 0.1 meters), and for each micro-voxel layer, its horizontal projected area and three-dimensional volume are calculated. Specifically, the horizontal projected area is calculated based on the boundary range of the point cloud layer, and a point density adjustment factor is introduced to compensate for the sparsity of the point cloud data; the three-dimensional volume is calculated using the minimum bounding box method. Then, based on these geometric parameters, the leaf area density (LAD) of each layer is calculated, which is the ratio of the estimated leaf area to the volume occupied by that layer. Through this micro-voxelization process, the complex macroscopic canopy is transformed into a set of quantifiable microscopic parameters.
[0023] Improved derivation of the S4 projection function: Traditional Beer's Law typically relies on an assumed leaf tilt angle distribution function (such as an ellipsoidal distribution), which deviates from the actual tufted structure of desert shrubs. This invention utilizes the aforementioned microvoxel model to re-establish the physical relationship between leaf area projection and light interception from a microscopic perspective. For example... Figure 3 As shown, by integrating the leaf area density and optical path information within microscopic voxels, and based on the principle of radiative transfer, an improved projection function is derived. G ( i Calculation formula: In the formula: i The observed zenith angle of the incident light, in degrees; k is the extinction coefficient of the shrub, dimensionless; LAD The leaf area density after polymerization, in units of m³. 2 / m 3 ; H The height of the shrub canopy is in meters (m).
[0024] This formula no longer relies on the idealized leaf tilt angle assumption, but directly utilizes the spatial distribution information in point cloud data, which can more accurately describe the transmission characteristics of light in complex canopies.
[0025] S5 calculates canopy porosity using point cloud data: This step uses ray tracing technology for simulation. First, the Fibonacci spherical spiral algorithm is used to generate uniformly distributed simulated ray directions, with the ray starting height randomly sampled within the canopy height range. Then, the distance between the ray and the leaf point cloud is calculated; if the distance is less than a set threshold, the ray is considered to be intercepted; otherwise, the ray is considered to have penetrated. Finally, a specific zenith angle is calculated. i The porosity at that angle is the ratio of the number of rays penetrating the canopy in a given direction to the total number of rays.P ( i ).
[0026] S6 leaf area index inversion: By observing the porosity at different angles P ( i ) and the corresponding improved projection function G ( i Substituting into the following LAI inversion formula, the accurate leaf area index (LAI) of the target shrub can be obtained. The inversion formula is: .
[0027] Example This embodiment provides a specific process for operating the present invention in practical applications, aiming to solve the problem of traditional measurement errors caused by sparse shrub canopies and clustered assimilative branches in desert areas.
[0028] The desert shrub leaf area index inversion method based on the microvoxel analysis projection model (MVAPM) includes the following steps: S1 acquires multi-angle image data of desert shrubs and constructs 3D point cloud data: A sample plot of desert shrubs (such as Haloxylon ammodendron or Tamarix chinensis) was selected, and a drone (such as the DJI Air3s series) was used to photograph individual shrubs from multiple angles. To ensure the integrity of the point cloud reconstruction, the shooting scheme was set as follows: within the azimuth range of 0° to 360°, a shooting point was set every 45° or 60°; at the same time, multiple gradients (such as 15°, 30°, 45°, 60°, and 75°) were set in the zenith angle direction. The acquired image data was imported into COLMAP software, and using the Structure for Motion Restoration (SfM) and Multi-View Stereo Vision (MVS) algorithms, through feature extraction, sparse reconstruction, and dense reconstruction steps, high-density 3D point cloud data containing RGB color information was generated, such as... Figure 4 As shown.
[0029] S2 performs preprocessing and fine branch / leaf separation on the raw point cloud data: The original point cloud was converted to a relative coordinate format and outliers were removed using CloudCompare software. Then, the tree-guided segmentation algorithm proposed in this invention was employed: cKDTree was used to accelerate the calculation of local point cloud density and extract connection length features; an unsupervised Gaussian mixture model (GMM) was applied to fit the connection length distribution, adaptively obtaining the segmentation threshold; finally, the DBSCAN density clustering algorithm was combined to accurately segment the point cloud into "branch point cloud" (xylem) and "leaf point cloud" (assimilated branches). This step effectively eliminated the interference of non-photosynthetic components on LAI calculation. In the accuracy verification stage, the segmentation algorithm of this invention was applied to the above point cloud data. The Matthews correlation coefficient (MCC) and overall accuracy (OA) were calculated based on the statistical classification results (true positives TP, false positives FP, true negatives TN, and false negatives FN). Experimental results show that the algorithm of this invention performs particularly well on Tree3 and Tree4, with average MCC values reaching 0.96194 and 0.97636 respectively, and OA values exceeding 0.98. This indicates that the algorithm can identify and separate fine leaf point clouds from branch point clouds with extremely high accuracy. Figure 5 As shown, even for the more complex Tree1 and Tree2, the OA value remains above 0.95.
[0030] S3 constructs a microvoxel analysis model based on the separated leaf point cloud and calculates the leaf area density: The separated leaf point cloud was voxelized vertically at 0.1-meter intervals. For each micro-voxel, its horizontal projected area (introducing a density adjustment factor) and three-dimensional minimum bounding volume were calculated to obtain the local leaf area density (LAD) of each layer.
[0031] Improved derivation of the S4 projection function: Based on the principle of radiative transfer, the improved projection function formula derived in this invention is used. G ( i The calculation is performed using a formula that directly reflects the non-uniform structural characteristics of the current shrub by integrating the extinction coefficient and optical path within the microvoxel, no longer relying on the idealized leaf tilt angle distribution assumption.
[0032] S5 calculates canopy porosity using point cloud data: Using the Fibonacci spherical spiral algorithm to generate uniformly distributed simulated rays, ray tracing simulations were performed on leaf point clouds, and the light penetration ratio at a specific zenith angle was statistically analyzed to obtain the porosity. P ( i ).
[0033] S6 leaf area index inversion: The calculated G ( i )and P ( i By substituting the modified Beer's Law formula, the precise leaf area index (LAI) of the target shrub can be calculated.
[0034] In the LAI inversion accuracy verification stage, the four-component data generated by the LESS radiative transfer model were used as a standard reference. Comparative analysis is as follows: Figure 6 As shown, the LAI values calculated by the MVAPM model in this invention have a very high correlation with the simulated values by the LESS model. Particularly in the Tree 2, Tree 3, and Tree 4 scenarios, the Pearson correlation coefficient (r) ranges from 0.96405 to 0.9722, the coefficient of determination (R²R²R²) is as high as 0.92791 to 0.94403, and the root mean square error (RMSE) is as low as 0.0043 to 0.0057, demonstrating the effectiveness and accuracy of this method in monitoring sparse vegetation in deserts. This also indicates that under ideal simulation conditions, the method proposed in this invention can effectively correct for canopy aggregation effects and provide high-precision LAI estimation.
Claims
1. A method for inverting the leaf area index of desert shrubs based on a microvoxel analysis projection model, comprising the following steps: S1 acquires multi-angle image data of desert shrubs and constructs 3D point cloud data: Typical desert shrubs were selected as observation targets in the field experimental area. Multiple angles of photography were taken around the shrubs using drones or ground photography equipment. After collecting the two-dimensional image sequence from multiple perspectives, three-dimensional reconstruction was carried out using processing software based on motion recovery structure and multi-view stereo vision technology to generate high-density three-dimensional point cloud data containing fine spatial structure information of the shrubs. S2 performs preprocessing and fine branch / leaf separation on the raw point cloud data: First, the original point cloud is converted into a unified relative coordinate format and outliers are removed. Then, cKDTree is used to accelerate the query calculation of the local density of the point cloud, extract the connection length feature, and apply an unsupervised Gaussian mixture model to fit the connection length distribution to adaptively obtain the segmentation threshold. Finally, the density-based DBSCAN clustering algorithm is combined to accurately divide the point cloud into leaf point cloud and branch point cloud. S3 constructs a microvoxel analysis model based on the separated leaf point cloud and calculates the leaf area density: The leaf point cloud is vertically layered and voxelized according to a preset height interval. For each micro voxel layer, its horizontal projected area and three-dimensional volume are calculated. Then, based on these geometric parameters, the leaf area density of each layer is calculated, which is the ratio of the estimated leaf area to the volume occupied by that layer. Improved derivation of the S4 projection function: By integrating the leaf area density and optical path information within the microscopic voxels, the improved projection function is derived. G ( θ Calculation formula: In the formula: θ The observed zenith angle of the incident light, in degrees; k is the extinction coefficient of the shrub, dimensionless; LAD The leaf area density after polymerization, in units of m³. 2 / m 3 ; H The height of the shrub canopy is in meters (m). S5 calculates canopy porosity using point cloud data: First, the Fibonacci spherical spiral algorithm is used to generate uniformly distributed simulated ray directions, with the initial ray height randomly sampled within the canopy height range. Then, the distance between the ray and the leaf point cloud is calculated; if the distance is less than a set threshold, the light is considered to be blocked; otherwise, the light is considered to have penetrated. Finally, specific zenith angles are statistically analyzed. θ The porosity at that angle is the ratio of the number of rays penetrating the canopy in a given direction to the total number of rays. P ( θ ); S6 leaf area index inversion: By observing the porosity at different angles P ( θ ) and the corresponding improved projection function G ( θ Substituting the values into the LAI inversion formula below, the precise leaf area index of the target shrub can be obtained. The inversion formula is: 。 2. The method for inverting the leaf area index of desert shrubs based on a microvoxel analysis projection model as described in claim 1, characterized in that: In step S1, the shooting scheme is set as follows: the azimuth angle is in the range of 0° to 360°; the zenith angle is in the range of 0° to 75°, and multiple gradients are set.
3. The method for inverting the leaf area index of desert shrubs based on a microvoxel analysis projection model as described in claim 1, characterized in that: In step S3, the horizontal projected area is calculated based on the boundary range of the point cloud layer, and a point density adjustment factor is introduced to compensate for the sparsity of the point cloud data; the volume occupied in three-dimensional space is calculated using the minimum bounding box method.