A method and system for monitoring above-ground biomass of sparse shrubs in arid regions based on UAV-LAS

Abstract

The present application relates to the technical field of unmanned aerial vehicle ecological remote sensing monitoring, and particularly relates to a method and system for monitoring aboveground biomass of sparse shrubs in arid regions based on UAV-LAS, which adopts a statistical outlier removal algorithm to complete point cloud denoising, realizes ground point classification through a self-adaptive filtering strategy combining quadratic surface fitting and improved progressive triangulated irregular network, constructs a surface model and completes point cloud elevation normalization, completes single tree segmentation in combination with a double-mode strategy of point cloud seed point segmentation and crown height model watershed segmentation, extracts vegetation structure parameters, and calculates aboveground biomass in combination with an allometric growth model. The present application effectively improves the extraction precision of vegetation parameters in sandy land scenes, controls errors from the data source end, and is suitable for large-scale high-precision monitoring of sparse shrubs in arid regions.

Description

Technical Field

[0001] This invention relates to the field of ecological remote sensing monitoring technology for unmanned aerial vehicles (UAVs), and in particular to a method and system for monitoring aboveground biomass of sparse shrubs in arid areas based on UAV-LAS. Background Technology

[0002] Arid and semi-arid regions are an important part of the global terrestrial ecosystem. Sparse shrubs, as the dominant vegetation type in these regions, play a crucial role in windbreak and sand fixation, soil water conservation, biodiversity maintenance, and regional carbon sequestration regulation. Aboveground biomass is a core indicator for quantifying vegetation growth, assessing ecosystem quality, and evaluating the effectiveness of desertification control. Achieving high-precision, non-destructive, and large-scale monitoring of aboveground biomass of sparse shrubs in arid regions has significant scientific value and practical implications for ecological protection, restoration, and management in these areas.

[0003] Current technologies for monitoring aboveground vegetation biomass mainly include two categories: ground surveys and remote sensing inversion. While traditional ground quadrat harvesting methods offer high accuracy, they suffer from drawbacks such as high labor costs, long operation cycles, and significant damage to vegetation. Furthermore, arid regions typically involve vast areas, complex terrain, and poor accessibility, making large-scale, efficient monitoring difficult with ground survey methods. Satellite remote sensing technology can achieve wide-area coverage, but its spatial resolution limitations make it difficult to accurately identify low-lying, sparsely distributed shrubs. Inversion results are also susceptible to interference from large areas of bare soil, resulting in overall accuracy that falls short of the requirements for refined monitoring.

[0004] In recent years, UAV low-altitude remote sensing technology has developed rapidly. UAVs equipped with lidar payloads can acquire three-dimensional point cloud data with centimeter-level resolution, accurately depicting the vertical structure and morphological characteristics of vegetation. This technology offers advantages such as high flexibility, high precision, and minimal weather-related impact, providing a new technical approach for monitoring the biomass of sparse shrubs in arid regions. Several existing technologies have already developed solutions for applying UAV lidar to shrub biomass monitoring.

[0005] For example, Chinese invention patent application CN117347974A discloses a method and device for estimating aboveground biomass of shrub vegetation. It constructs a multi-scale upscaling estimation system integrating ground measurement, UAV monitoring, and satellite inversion. By using object-oriented segmentation of UAV orthophotos to identify the contours and locations of individual shrubs, it matches corresponding lidar point clouds to extract vegetation height and density features. A biomass estimation model is then constructed using a random forest algorithm, ultimately achieving large-scale regional inversion of shrub biomass. While this approach improves the efficiency of large-scale shrub biomass monitoring to some extent, it still has significant technical limitations in practical applications. The preprocessing of lidar point clouds only employs generalized denoising, classification, and normalization processes, without tailoring algorithms to the specific characteristics of arid sandy terrain, high proportion of bare land, and small, dispersed shrub crowns. The accuracy of ground point classification and individual tree segmentation is significantly limited by general parameters. Errors in the extraction of front-end vegetation structure parameters directly transmit to the back-end biomass inversion results, hindering further improvement in monitoring accuracy.

[0006] For example, Chinese invention patent application CN121937871A discloses a method for estimating the biomass of Haloxylon ammodendron forests based on collaborative inversion of multi-source data from UAVs. This method integrates multispectral vegetation index features from UAVs with statistical features of lidar altitude, selects the optimal feature subset by ranking feature importance, and constructs a high-precision biomass inversion model using an extreme gradient boosting algorithm, effectively improving the inversion accuracy of Haloxylon ammodendron biomass in desert forests. However, the core innovation of this scheme is entirely concentrated in the backend feature engineering and machine learning model optimization stages. Preprocessing of lidar point clouds, ground point classification, and individual tree segmentation are only briefly mentioned as general preliminary steps. No dedicated filtering and segmentation technology architecture is designed for the point cloud features of sparse shrubs in arid regions. The frontend point cloud processing heavily relies on the default parameters of general software, making it difficult to adapt to the refined 3D structure extraction requirements of low, sparse shrubs, and failing to address the fundamental problem of vegetation parameter extraction errors from the data source.

[0007] In summary, existing shrub biomass monitoring technologies in arid regions generally exhibit a bias towards "emphasizing back-end inversion model optimization while neglecting front-end point cloud processing and adaptation." Given the unique characteristics of sparse shrubs in arid regions—small canopy size, scattered distribution, and subtle sandy terrain—there is a lack of dedicated combined filtering and dual-mode single-tree segmentation technologies. This results in insufficient accuracy in extracting vegetation structure parameters, becoming a core technical bottleneck restricting the improvement of biomass monitoring accuracy. Therefore, developing a refined lidar point cloud processing method for sparse shrubs in arid regions to improve the accuracy of vegetation structure parameter extraction from the data source, thereby achieving high-precision monitoring of aboveground biomass, is a pressing technical problem to be solved in this field. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for monitoring aboveground biomass of sparse shrubs in arid regions based on UAV-LAS, in order to solve the problem mentioned in the background art that the lack of dedicated combined filtering and dual-mode single-tree segmentation technology leads to insufficient accuracy in extracting vegetation structure parameters, which is a problem that restricts the improvement of biomass monitoring accuracy.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring aboveground biomass of sparse shrubs in arid regions based on UAV-LAS, comprising acquiring lidar point cloud data of the target area through an airborne lidar system of an unmanned aerial vehicle (UAV), performing noise reduction preprocessing, ground point classification, elevation normalization processing, and individual tree segmentation on the point cloud data in sequence, extracting vegetation structure parameters, and calculating the aboveground biomass, specifically including the following steps:

[0010] S1 collects point cloud data and multispectral image data from UAV-borne lidar in the target area, and completes the flight strip stitching of the original point cloud and radiometric correction preprocessing of the multispectral image.

[0011] S2 uses a statistical outlier removal algorithm to denoise the lidar point cloud data. Based on the neighborhood point density analysis, the maximum distance threshold is calculated, and high-order and low-order outlier noise points are removed to obtain the denoised point cloud data.

[0012] S3 employs a filtering strategy combining quadratic surface fitting and an improved progressive densification triangulation algorithm for ground point classification: For gently sloping sandy areas, quadratic surface fitting is used, fitting a quadratic surface based on the neighborhood point set, and filtering ground points through vertical residual thresholds and normal vector angle thresholds; for undulating dune areas, an improved progressive densification triangulation algorithm is used, constructing an initial triangulation and iteratively densifying and determining ground points, and fusing the classification results of the two types of areas to obtain a ground point set.

[0013] S4 performs spatial interpolation on the ground point set based on the inverse distance weighted interpolation method to construct a digital elevation model, generates a digital surface model by combining the original point cloud data, and obtains the canopy height model through difference calculation;

[0014] S5 performs elevation normalization processing on the denoised point cloud data based on the digital elevation model, subtracting the ground elevation value at the corresponding location from the elevation value of each point to eliminate the interference of terrain undulation on vegetation height calculation, and obtains normalized canopy point cloud data.

[0015] S6 employs a dual-mode strategy combining point cloud seed point segmentation and canopy height model watershed segmentation for single tree segmentation: seed point growth segmentation is performed on the normalized canopy point cloud to obtain the first single tree segmentation result, and watershed segmentation is performed on the canopy height model to obtain the second single tree segmentation result. The overlap of the two types of segmentation results is checked and the boundary is corrected to obtain the final spatial boundary of the single shrub.

[0016] S7 extracts vegetation structure parameters of a single shrub based on the single tree segmentation results. The vegetation structure parameters include at least tree height, crown diameter, crown area, and crown volume.

[0017] S8 inputs the vegetation structure parameters into the arid zone sparse shrub allometric growth model calibrated based on ground measurement data of the target area, and calculates the aboveground biomass of a single shrub and the entire area in the target area.

[0018] Preferably, in step S2, the execution rules of the statistical outlier removal algorithm are as follows: set the number of neighboring points to 8 to 12, calculate the average distance between each data point and all points in its neighborhood, and then calculate the median and standard deviation of the average distance between all data points. Use "median + standard deviation multiple × standard deviation" as the maximum distance threshold, and determine points whose average distance exceeds the threshold as noise points and remove them. The standard deviation multiple is 1 to 2.

[0019] Preferably, in step S3, the combined filtering strategy adopts a terrain adaptive switching mechanism: the terrain slope data of the target area is pre-calculated, a slope threshold is set, the quadratic surface fitting method is automatically used for gentle sandy land with a slope less than the threshold, the improved progressive densification triangular mesh algorithm is automatically used for undulating dune areas with a slope greater than or equal to the threshold, and a human-computer interaction correction step is introduced for local complex areas where the classification accuracy is not up to standard, the local classification results are manually adjusted, and finally the complete ground point set is obtained by fusion.

[0020] Preferably, in step S3, the execution steps of the improved progressive encryption triangulation algorithm are as follows:

[0021] S31 performs gridding on the point cloud of the target area, selects the lowest point in each grid as the initial seed point, and sets the grid size to 1 to 2 times the average crown width of the sparse shrubs in the target area.

[0022] S32 uses initial seed points to construct an initial irregular triangular mesh;

[0023] S33 Iterates through all points to be classified, queries the triangle into which the horizontal projection of a point falls, calculates the vertical distance d from the point to the triangle plane, and the maximum value θ of the angle between the line connecting the point to the three vertices of the triangle and the triangle plane.

[0024] S34 sets the distance threshold T that adaptively decays with each iteration. d With angle threshold T θ First iteration T d Set to 0.3m~0.5m, T θ Set to 8°~12°, T after each iteration d Multiply by the attenuation factor of 0.7 to 0.9 until T. dThe iteration terminates when the vertical distance d decreases to 0.05m; if the vertical distance d is less than the current iteration's T... d And the maximum angle θ is less than T in the current iteration. θ If the point is identified as a ground point, it is added to the triangulation network.

[0025] S35 Repeat steps S33 to S34 until all eligible ground points are classified.

[0026] Preferably, in step S4, the parameters of the inverse distance weighted interpolation method are set as follows: the distance power parameter is 2, the sampling points participating in the interpolation are selected by the k nearest neighbor search method, the k value is set to 10 to 15, and each grid cell is weighted by the k ground points closest to its center.

[0027] Preferably, in step S6, the specific rules for the overlap verification and boundary correction are as follows: calculate the boundary overlap of the same shrub in the two types of segmentation results. If the overlap is greater than or equal to a preset threshold, take the average boundary of the two types of results as the final boundary. If the overlap is less than the preset threshold, take the point cloud seed point segmentation result as the benchmark and correct the watershed segmentation result by adjusting the Gaussian smoothing factor to eliminate oversegmentation or undersegmentation errors. The adjustment rule of the Gaussian smoothing factor is: decrease the smoothing factor when undersegmentation occurs and increase the smoothing factor when oversegmentation occurs.

[0028] Preferably, in step S7, the extraction rules for the vegetation structure parameters are as follows: the tree height is obtained by using the maximum elevation method to extract the maximum elevation value in the point cloud of a single shrub as the tree height; the crown boundary is constructed based on the α-shape model to generate a closed contour that fits the crown shape; the crown volume is obtained by using the 3D convex hull algorithm to generate the minimum convex polyhedron that surrounds the crown point cloud and calculate its geometric volume; the vegetation structure parameters also include the crown density, that is, the number of point clouds per unit volume.

[0029] Preferably, step S8 further includes background correction of the biomass calculation results using multispectral image data: calculating the normalized vegetation index and vegetation cover of the target area based on multispectral images, eliminating the interference of bare soil background on the extraction of vegetation structure parameters, optimizing the input parameters of the allometric growth model, and improving the accuracy of biomass calculation.

[0030] A UAV-LAS-based aboveground biomass monitoring system for sparse shrubs in arid regions includes a data acquisition module, a point cloud denoising module, a single tree segmentation module, a parameter extraction module, and a biomass calculation module. It also includes a combined filtering classification module, a land surface model construction module, and a point cloud normalization module.

[0031] The combined filtering classification module has two built-in filtering algorithms: quadratic surface fitting and improved progressive densification triangulation. It can switch filtering strategies according to the terrain features of the target area to process point cloud data of flat sandy areas and undulating sand dune areas respectively, and then output ground point sets.

[0032] The surface model construction module constructs a digital elevation model and a digital surface model based on the inverse distance weighted interpolation method, and calculates and generates a canopy height model.

[0033] The point cloud normalization module performs elevation normalization processing on the point cloud data based on the digital elevation model, eliminates the influence of terrain undulation on vegetation height calculation, and outputs normalized canopy point cloud data.

[0034] The single-tree segmentation module is a dual-mode single-tree segmentation module, which can simultaneously perform point cloud seed point segmentation and canopy height model watershed segmentation, and perform overlap verification and boundary correction on the two types of segmentation results, outputting the spatial boundary of a single shrub.

[0035] Preferably, it also includes a multispectral processing module and a human-computer interaction correction module; the multispectral processing module is used to calculate the normalized vegetation index and vegetation coverage, and to perform background correction on the biomass calculation results; the human-computer interaction correction module is used to manually adjust the ground point classification results and single tree segmentation results in local complex areas to optimize the final monitoring accuracy.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] 1. For the terrain features of alternating flat sandy areas and undulating dunes in arid regions, an adaptive filtering strategy combining quadratic surface fitting and an improved progressively denser triangulation network algorithm is adopted. In flat sandy areas, surface fitting achieves efficient ground point identification, while in undulating dune areas, progressively denser triangulation networks adapt to micro-topographic variations. Simultaneously, local human-computer interaction is used to correct and supplement the classification results in complex areas. This scheme can adapt to the diverse sandy terrain features of arid regions, effectively reducing the probability of misclassification of ground points in areas with high bare land ratios and slight undulations using general single filtering algorithms, improving the accuracy of digital elevation model construction, and providing a stable topographic benchmark for subsequent vegetation height normalization and structural parameter extraction.

[0038] 2. A dual-mode segmentation strategy combining point cloud seed point segmentation and canopy height model watershed segmentation is adopted. This strategy integrates the 3D structural information of the point cloud with the 2D morphological features of the canopy height model. The segmentation results are cross-validated through overlap verification and boundary correction mechanisms, and the Gaussian smoothing factor can be adaptively adjusted according to the error type. This scheme effectively improves the oversegmentation and undersegmentation problems of single algorithms for sparse shrubs in arid areas due to their small canopy width, low plant height, and scattered distribution. It improves the recognition recall rate of individual shrubs and the accuracy of canopy boundary extraction, and can stably capture the 3D morphological features of scattered, low-growing shrubs, avoiding the accuracy degradation of general forest single-tree segmentation algorithms in sparse shrub scenarios.

[0039] 3. This invention prioritizes the point cloud preprocessing and vegetation structure parameter extraction stages, ensuring accurate extraction of core structural parameters such as tree height, crown diameter, crown area, and crown volume through high-precision ground point classification and individual tree segmentation. Furthermore, it utilizes a specific allometric growth model for sparse shrubs in arid regions to calculate aboveground biomass. Compared to solutions relying on backend machine learning models to correct errors, this invention fundamentally reduces the transmission effect of frontend data errors. The biomass inversion results have clear physical meaning and strong interpretability, avoiding the risk of model overfitting in small sample scenarios, and offering superior stability and versatility in monitoring results.

[0040] 4. A hybrid processing flow is adopted, primarily using adaptive algorithm processing supplemented by human-computer interaction correction. The flat sandy areas, which constitute the largest proportion, can achieve fully automated point cloud processing, with manual correction only applied to localized complex terrain and edge areas. This approach ensures both monitoring accuracy and processing efficiency, enabling efficient monitoring of vegetation biomass in large-scale sample areas of arid regions. This solution fills the technical gap between insufficient spatial resolution of satellite remote sensing and the low efficiency and destructive nature of ground surveys, providing practical and high-precision technical support for assessing the effectiveness of desertification control in arid regions, monitoring ecological restoration, and calculating regional carbon sequestration.

[0041] 5. By combining synchronously acquired UAV multispectral imagery, the bare soil background is corrected for vegetation point cloud identification and biomass calculation using the normalized vegetation index and vegetation cover calculation results. This further eliminates interference from large areas of bare land on the extraction of sparse shrub parameters without introducing complex multi-source feature fusion modeling. The overall solution has a clear data processing chain, requires no additional complex modeling steps, and can be directly adapted to the technical framework of existing mainstream LiDAR data processing software, making engineering implementation easier. Attached Figure Description

[0042] Figure 1 This is a flowchart of the process for extracting sparse shrub vegetation structure information in arid areas based on UAV lidar according to the present invention.

[0043] Figure 2This is a schematic diagram of noise points in the point cloud data of the present invention, showing the high-order and low-order gross errors.

[0044] Figure 3 This is a comparison of the point cloud data denoising process before and after the present invention.

[0045] Figure 4 This is a schematic diagram illustrating the iterative encryption principle of the improved progressive encryption triangular mesh algorithm of this invention.

[0046] Figure 5 This is a comparison diagram of the original point cloud RGB display and the ground point classification results of this invention;

[0047] Figure 6 This is a comparison chart showing the effects of the original point cloud data and the digital elevation model of this invention;

[0048] Figure 7 This is a comparison diagram of the effects of the digital elevation model and the digital surface model of this invention;

[0049] Figure 8 This is a schematic diagram illustrating the relationship between the digital surface model, digital elevation model, and canopy height model of the present invention.

[0050] Figure 9 This is a comparison chart showing the effect of the original point cloud elevation and the normalized point cloud elevation of the present invention;

[0051] Figure 10 This is a schematic diagram illustrating the principle of single-tree segmentation for point cloud seed point growth in this invention.

[0052] Figure 11 This is a schematic diagram of the watershed segmentation process in the canopy height model of the present invention;

[0053] Figure 12 This is a comparison diagram of the original point cloud RGB display and the single-tree extraction results of this invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] This embodiment is applied to the monitoring of aboveground biomass of sparse shrubs in the arid desert region of Alxa League. The monitoring targets are typical sparse shrub communities such as Artemisia arenaria and Salix psammophila in this region. This embodiment uses a rotary-wing unmanned aerial vehicle (UAV) platform equipped with an airborne lidar system (UAV-LAS) and a multispectral camera to conduct low-altitude remote sensing monitoring. The flight altitude is set at 100m. LiDAR point cloud data with a point cloud density of no less than 50 points / m² and multispectral image data with a spatial resolution better than 10cm are acquired. The overall processing flow is as follows: Figure 1 As shown.

[0056] Step 1: Raw Data Acquisition and Preprocessing

[0057] Simultaneous acquisition of lidar point cloud data and multispectral image data of the target study area using drones:

[0058] The lidar point cloud data was spliced ​​with flight strips and converted to coordinate system 1, and then converted to the CGCS2000 plane coordinate system to form an original point cloud dataset covering the entire study area.

[0059] Radiometric calibration, geometric correction, and mosaicking of multispectral image data are performed to generate orthophoto products of the study area, which are used for subsequent vegetation index calculation and background correction.

[0060] Step 2: Point cloud denoising based on statistical outlier removal algorithm

[0061] The Statistical Outlier Removal (SOR) algorithm was used to denoise the original point cloud data, removing noise points generated during data acquisition. Noise points mainly include gross errors at higher altitudes caused by low-flying objects such as birds, and gross errors at lower altitudes caused by multipath errors and systematic errors. The distribution characteristics of these two types of noise points are as follows: Figure 2 As shown.

[0062] In this embodiment, the parameters of the SOR algorithm are set as follows: the number of neighboring points is set to 10, and the standard deviation factor is set to 1. The algorithm execution process is as follows: for each data point, search for its 10 neighboring points and calculate the average distance between the point and all neighboring points; calculate the median (meanD) and standard deviation (S) of the average distances of all points, and calculate the maximum distance threshold according to the formula MaxD = meanD + 1 × S; points whose average distance exceeds MaxD are identified as noise points and removed. A comparison of the point cloud effects before and after denoising is provided. Figure 3 As shown, after denoising, discrete noise points in the point cloud data are effectively removed, and the data quality of the effective point cloud is significantly improved.

[0063] Step 3: Ground point classification using terrain-adaptive combined filtering

[0064] An adaptive filtering strategy combining quadratic surface fitting and an improved progressively denser triangulation network algorithm is used for ground point classification. The filtering algorithm is automatically switched according to the terrain features of the study area. The specific process is as follows:

[0065] 3.1 Terrain Zoning and Algorithm Adaptive Switching

[0066] Based on the terrain slope of the original point cloud computing research area, a slope threshold of 3° is set in advance: for gentle sandy areas with a slope of less than 3°, the quadratic surface fitting method is automatically used for ground point classification; for undulating dune areas with a slope of 3° or greater, the improved progressive densification triangulation algorithm is automatically used for ground point classification.

[0067] 3.2 Ground point classification using quadratic surface fitting method

[0068] For gently sloping sandy areas, a neighborhood set of points is selected for each point to be classified, and a quadratic surface model is fitted using the least squares method. The vertical residual from the point to be classified to the fitted surface is calculated, and the surface normal vector is extracted and its angle with the vertical direction is calculated. In this embodiment, the vertical residual threshold is set to 0.15m, and the normal vector angle threshold is set to 10°. If the vertical residual of the point to be classified is less than the threshold and the normal vector angle is less than 10°, it is determined to be a ground point.

[0069] 3.3 Improved Progressive Encryption Triangulated Network Algorithm for Ground Point Classification

[0070] For undulating dune areas, ground point separation is achieved by constructing a triangular network and iteratively refining it. The principle of iterative refinement is as follows: Figure 4 As shown, the specific steps are as follows:

[0071] (1) Gridding and selection of initial seed points: The grid size is set according to the size of the land features in the study area, and the lowest elevation point in each grid is selected as the initial seed point;

[0072] (2) Initial triangular network construction: Construct an initial irregular triangular network using all initial seed points;

[0073] (3) Iterative encryption classification: Traverse all points to be classified, query the triangle into which the horizontal projection of the point falls, calculate the vertical distance d from the point to the triangle plane, and the maximum value of the angle between the point and the line connecting the three vertices of the triangle and the triangle plane; if the vertical distance d is less than the preset iterative distance threshold and the maximum angle is less than the preset iterative angle threshold, then the point is determined to be a ground point and added to the triangulation network.

[0074] (4) Repeat the above classification process until all ground points that meet the conditions are identified.

[0075] 3.4 Classification Result Fusion and Human-Computer Interaction Correction

[0076] The ground point classification results from two different algorithms are combined to form a preliminary ground point set. For areas with complex terrain and poor automatic classification performance, a human-computer interaction correction step is introduced to manually adjust the local classification results, ultimately obtaining a high-precision ground point set. A comparison between the original RGB display of the point cloud and the ground point classification results is shown below. Figure 5 As shown.

[0077] Step 4: Construction of a surface model based on inverse distance weighted interpolation

[0078] Based on the ground point set obtained from the classification, a digital elevation model (DEM) is constructed using the inverse distance weighted interpolation method. At the same time, a digital surface model (DSM) is generated based on the original point cloud data, and a canopy height model (CHM) is obtained through interpolation.

[0079] In this embodiment, the parameters for inverse distance weighted interpolation are set as follows: the distance power parameter p is set to 2; the sampling points are selected using a k-nearest neighbor search method, with k set to 12, meaning that each grid cell is weighted by the 12 ground points closest to its center. The interpolated DEM can accurately characterize the exposed surface morphology of the study area. A comparison of the point cloud data and the DEM results is shown below. Figure 6 As shown; a comparison of the effects of DEM and DSM. Figure 7 As shown; CHM is calculated by subtracting DEM from DSM, and the relationship among the three is as follows: Figure 8 As shown, CHM can intuitively reflect the relative height distribution of the vegetation canopy.

[0080] Step 5: Point cloud elevation normalization processing

[0081] Based on the completed DEM, the denoised raw point cloud data is subjected to elevation normalization: the ground elevation value corresponding to each point is calculated, and the ground elevation value at the corresponding location is subtracted from the original absolute elevation value of the point to make the elevation of the ground point zero, thus obtaining normalized canopy point cloud data.

[0082] Normalization effectively eliminates the interference of terrain undulations on the analysis of vegetation vertical structure, unifying all point clouds to the same ground reference plane. This separates vegetation height information from terrain influence, accurately reflecting the vertical structural characteristics of the vegetation canopy and providing an optimized data foundation for subsequent individual tree segmentation. A comparison of the original point cloud elevation and the normalized point cloud elevation is shown below. Figure 9 As shown.

[0083] Step 6: Dual-mode single-wood splitting processing

[0084] A dual-mode strategy combining point cloud seed point segmentation and canopy height model watershed segmentation is adopted for single-tree segmentation. The specific process is as follows:

[0085] 6.1 Point Cloud Seed Point Growth Segmentation

[0086] Seed point growth segmentation is performed on the normalized canopy point cloud, and the segmentation principle is as follows: Figure 10 As shown: The global elevation maxima in the point cloud are selected as the initial seed point (tree vertex). Using the crown radius threshold, determined based on measured ground data of the target area, as the spacing threshold, growth clustering is performed on points with lower elevations. If the horizontal distance between an adjacent point and the current seed point is less than the threshold, it is added to the point cloud set of the current shrub. If the distance between an adjacent point and the current seed point is greater than the threshold, that point is marked as a new seed point, corresponding to another shrub. This process is iterated to complete the clustering of all points, yielding the first shrub segmentation result.

[0087] In this embodiment, the critical value is set at 1.5m based on the average crown width of shrubs in the study area.

[0088] 6.2 CHM watershed segmentation

[0089] Watershed segmentation is performed on the canopy height model, and the segmentation principle is as follows: Figure 11 As shown: The elevation of CHM is regarded as the terrain, with high points corresponding to mountain peaks and low points corresponding to valleys; the water injection process from bottom to top is simulated, and dividing barriers are constructed at the confluence of water flows in different valleys to divide the boundaries of individual shrubs, thus obtaining the second single tree segmentation result.

[0090] Before watershed segmentation, a Gaussian low-pass filter is applied to CHM to eliminate false watershed lines caused by minor noise dips. The Gaussian smoothing factor σ controls the width of the filtering window. When the watershed result is found to be over-segmented (too many fragments) in the 6.3 verification, σ is increased to enhance smoothing and merge false segments. When under-segmentation is found (two plants are stuck together), σ is decreased to retain more detailed segmentation lines.

[0091] 6.3 Verification and Correction of Segmentation Results

[0092] Overlap verification and boundary correction were performed on the two types of segmentation results: The overlap rate of the boundary of the same shrub in the two types of segmentation results was calculated. If the overlap rate was greater than or equal to 85%, the average boundary of the two types of results was taken as the final boundary; if the overlap rate was less than 85%, the smoothing parameters of the watershed segmentation were adjusted based on the seed point segmentation result of the point cloud to eliminate over-segmentation or under-segmentation errors, and finally obtain the accurate spatial boundary of a single shrub. A comparison between the original RGB display of the point cloud and the single-shrub extraction result is shown below. Figure 12 As shown.

[0093] Step 7: Extraction of vegetation structure parameters

[0094] Based on the final individual tree segmentation results, the core vegetation structure parameters of a single shrub are extracted. The specific extraction rules are as follows:

[0095] Tree height: The maximum elevation value in the normalized point cloud of a single shrub is extracted using the maximum elevation method and taken as the tree height of that shrub.

[0096] Crown boundary and crown area: Based on the α-shape model, a closed outline of the tree crown is constructed to obtain the crown boundary that fits the shape of the tree crown. The area of ​​the boundary polygon is calculated as the crown area.

[0097] Crown volume: Using the 3D convex hull algorithm, the minimum convex polyhedron that encloses all the point clouds of the crown of a single shrub is generated, and its geometric volume is calculated as the crown volume;

[0098] Canopy density: The total number of point clouds per shrub is counted and divided by the canopy volume to obtain the number of point clouds per unit volume, i.e., the canopy density.

[0099] Step 8: Aboveground biomass calculation and background correction

[0100] The extracted vegetation structure parameters, such as tree height, crown area, and crown volume, are input into the allometric growth model of aboveground biomass of sparse shrubs in arid areas, which is calibrated based on ground measurement data of the target area. The aboveground biomass of a single shrub is calculated, and the aboveground biomass of the entire study area is obtained by summing them up.

[0101] Meanwhile, background correction was performed using synchronously acquired multispectral image data: the normalized difference vegetation index (NDVI) and vegetation cover of the study area were calculated based on multispectral orthophotos, the interference of bare soil background on vegetation point cloud identification was eliminated, the input parameters of the allometric growth model were optimized, and the accuracy of biomass calculation was further improved.

[0102] Implementation effect verification

[0103] In this embodiment, 20 ground-measured quadrats in the study area were selected for accuracy verification. The results showed that the ground point classification accuracy of this embodiment reached 94.2%, the single tree identification recall rate reached 91.7%, the determination coefficient R² of aboveground biomass inversion reached 0.88, and the root mean square error was reduced by 13.5% compared with the general single filter + single segmentation algorithm scheme. It can effectively meet the high-precision monitoring needs of aboveground biomass of sparse shrubs in arid areas.

[0104] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring aboveground biomass of sparse shrubs in arid regions based on UAV-LAS, comprising acquiring lidar point cloud data of the target area using an airborne lidar system on a UAV, sequentially performing noise reduction preprocessing, ground point classification, elevation normalization processing, and individual tree segmentation on the point cloud data, extracting vegetation structure parameters, and calculating aboveground biomass, characterized in that... Specifically, the steps include the following: S1 collects point cloud data and multispectral image data from UAV-borne lidar in the target area, and completes the flight strip stitching of the original point cloud and radiometric correction preprocessing of the multispectral image. S2 uses a statistical outlier removal algorithm to denoise the lidar point cloud data. Based on the neighborhood point density analysis, the maximum distance threshold is calculated, and high-order and low-order outlier noise points are removed to obtain the denoised point cloud data. S3 employs a filtering strategy combining quadratic surface fitting and an improved progressive densification triangulation algorithm for ground point classification: For gently sloping sandy areas, quadratic surface fitting is used, fitting a quadratic surface based on the neighborhood point set, and filtering ground points through vertical residual thresholds and normal vector angle thresholds; for undulating dune areas, an improved progressive densification triangulation algorithm is used, constructing an initial triangulation and iteratively densifying and determining ground points, and fusing the classification results of the two types of areas to obtain a ground point set. S4 performs spatial interpolation on the ground point set based on the inverse distance weighted interpolation method to construct a digital elevation model, generates a digital surface model by combining the original point cloud data, and obtains the canopy height model through difference calculation; S5 performs elevation normalization processing on the denoised point cloud data based on the digital elevation model, subtracting the ground elevation value at the corresponding location from the elevation value of each point to eliminate the interference of terrain undulation on vegetation height calculation, and obtains normalized canopy point cloud data. S6 employs a dual-mode strategy combining point cloud seed point segmentation and canopy height model watershed segmentation for single tree segmentation: seed point growth segmentation is performed on the normalized canopy point cloud to obtain the first single tree segmentation result, and watershed segmentation is performed on the canopy height model to obtain the second single tree segmentation result. The overlap of the two types of segmentation results is checked and the boundary is corrected to obtain the final spatial boundary of the single shrub. S7 extracts vegetation structure parameters of a single shrub based on the single tree segmentation results. The vegetation structure parameters include at least tree height, crown diameter, crown area, and crown volume. S8 inputs the vegetation structure parameters into the arid zone sparse shrub allometric growth model calibrated based on ground measurement data of the target area, and calculates the aboveground biomass of a single shrub and the entire area in the target area.

2. The method for monitoring aboveground biomass of sparse shrubs in arid regions based on UAV-LAS according to claim 1, characterized in that, In step S2, the execution rules of the statistical outlier removal algorithm are as follows: set the number of neighborhood points to 8 to 12, calculate the average distance between each data point and all points in its neighborhood, and then calculate the median and standard deviation of the average distance between all data points. Use "median + standard deviation multiple × standard deviation" as the maximum distance threshold, and determine points whose average distance exceeds the threshold as noise points and remove them. The standard deviation multiple is 1 to 2.

3. The method for monitoring aboveground biomass of sparse shrubs in arid regions based on UAV-LAS according to claim 1, characterized in that, In step S3, the combined filtering strategy adopts a terrain adaptive switching mechanism: the terrain slope data of the target area is pre-calculated, a slope threshold is set, the quadratic surface fitting method is automatically used for gentle sandy land with a slope less than the threshold, the improved progressive densification triangular mesh algorithm is automatically used for undulating dune areas with a slope greater than or equal to the threshold, and the human-computer interaction correction link is introduced for local complex areas where the classification accuracy is not up to standard, and the local classification results are manually adjusted, and finally the complete ground point set is obtained by fusion.

4. The method for monitoring aboveground biomass of sparse shrubs in arid regions based on UAV-LAS according to claim 3, characterized in that, In step S3, the execution steps of the improved progressive encryption triangulation algorithm are as follows: S31 performs gridding on the point cloud of the target area, selects the lowest point in each grid as the initial seed point, and sets the grid size to 1 to 2 times the average crown width of the sparse shrubs in the target area. S32 uses initial seed points to construct an initial irregular triangular mesh; S33 Iterates through all points to be classified, queries the triangle into which the horizontal projection of a point falls, calculates the vertical distance d from the point to the triangle plane, and the maximum value θ of the angle between the line connecting the point to the three vertices of the triangle and the triangle plane. S34 sets the distance threshold T that adaptively decays with each iteration. d With angle threshold T θ First iteration T d Set to 0.3m~0.5m, T θ Set to 8°~12°, T after each iteration d Multiply by the attenuation factor of 0.7 to 0.9, until T. d The iteration terminates when the vertical distance d decreases to 0.05m; if the vertical distance d is less than the current iteration's T... d And the maximum angle θ is less than T in the current iteration. θ If the point is identified as a ground point, it is added to the triangulation network. S35 Repeat steps S33 to S34 until all eligible ground points are classified.

5. The method for monitoring aboveground biomass of sparse shrubs in arid regions based on UAV-LAS according to claim 1, characterized in that, In step S4, the parameters of the inverse distance weighted interpolation method are set as follows: the distance power parameter is 2, the sampling points participating in the interpolation are selected by the k nearest neighbor search method, the k value is set to 10 to 15, and each grid cell is weighted by the k ground points closest to its center.

6. The method for monitoring aboveground biomass of sparse shrubs in arid regions based on UAV-LAS according to claim 1, characterized in that, In step S6, the specific rules for the overlap verification and boundary correction are as follows: calculate the boundary overlap of the same shrub in the two types of segmentation results. If the overlap is greater than or equal to a preset threshold, take the average boundary of the two types of results as the final boundary. If the overlap is less than the preset threshold, take the point cloud seed point segmentation result as the benchmark and correct the watershed segmentation result by adjusting the Gaussian smoothing factor to eliminate oversegmentation or undersegmentation errors. The adjustment rule of the Gaussian smoothing factor is: reduce the smoothing factor when undersegmentation occurs and increase the smoothing factor when oversegmentation occurs.

7. A method for monitoring aboveground biomass of sparse shrubs in arid regions based on UAV-LAS, as described in claim 1, is characterized in that... In step S7, the extraction rules for the vegetation structure parameters are as follows: the tree height is obtained by using the maximum elevation method, extracting the maximum elevation value in the point cloud of a single shrub as the tree height; the crown boundary is constructed based on the α-shape model to generate a closed contour that fits the crown shape; the crown volume is obtained by using the 3D convex hull algorithm to generate the minimum convex polyhedron that surrounds the crown point cloud and calculate its geometric volume; the vegetation structure parameters also include crown density, that is, the number of point clouds per unit volume.

8. A method for monitoring aboveground biomass of sparse shrubs in arid regions based on UAV-LAS, as described in claim 1, is characterized in that... Step S8 also includes background correction of biomass calculation results using multispectral image data: the normalized vegetation index and vegetation cover of the target area are calculated based on multispectral images, the interference of bare soil background on the extraction of vegetation structure parameters is eliminated, the input parameters of the allometric growth model are optimized, and the accuracy of biomass calculation is improved.

9. A UAV-LAS-based aboveground biomass monitoring system for sparse shrubs in arid regions, as described in any one of claims 1-8, comprising a data acquisition module, a point cloud denoising module, a single-tree segmentation module, a parameter extraction module, and a biomass calculation module, characterized in that: It also includes a combined filtering classification module, a surface model construction module, and a point cloud normalization module; The combined filtering classification module has two built-in filtering algorithms: quadratic surface fitting and improved progressive densification triangulation. It can switch filtering strategies according to the terrain features of the target area to process point cloud data of flat sandy areas and undulating sand dune areas respectively, and then output ground point sets. The surface model construction module constructs a digital elevation model and a digital surface model based on the inverse distance weighted interpolation method, and calculates and generates a canopy height model. The point cloud normalization module performs elevation normalization processing on the point cloud data based on the digital elevation model, eliminates the influence of terrain undulation on vegetation height calculation, and outputs normalized canopy point cloud data. The single-tree segmentation module is a dual-mode single-tree segmentation module, which can simultaneously perform point cloud seed point segmentation and canopy height model watershed segmentation, and perform overlap verification and boundary correction on the two types of segmentation results, outputting the spatial boundary of a single shrub.

10. A UAV-LAS-based aboveground biomass monitoring system for sparse shrubs in arid regions according to claim 9, characterized in that, It also includes a multispectral processing module and a human-computer interaction correction module; the multispectral processing module is used to calculate the normalized vegetation index and vegetation coverage, and to perform background correction on the biomass calculation results; the human-computer interaction correction module is used to manually adjust the ground point classification results and single tree segmentation results in local complex areas to optimize the final monitoring accuracy.

Images