Method and system for quantifying the spatial structure of the crown of sand-fixation forest based on ground and unmanned aerial vehicle lidar

Abstract

The application provides a method and system for quantifying the spatial structure of sand-fixing forest canopy based on ground and unmanned aerial vehicle laser radar, and is applied to the technical field of data processing. The application obtains quantified multi-source core data of sand-fixing forest, generates a standardized fusion point cloud data set through point cloud preprocessing and registration correction. The data set is layered at a single tree and stand scale, a three-element correlation node is constructed, and topological information is aggregated to generate a canopy structure feature correlation graph. Then, a neural network model is relied on to extract core features, generate a high-dimensional feature embedding vector, and iterate and optimize through a layered quantification engine combined with actual measurement calibration to construct a quantification basic model. Finally, the accuracy verification and dynamic correction are completed in combination with the measured and growth monitoring data, the multi-index requirements are balanced, and the canopy spatial structure quantification results suitable for the whole monitoring scene of sand-fixing forest are output.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for quantifying the spatial structure of sand-fixing forest canopies based on ground-based and UAV lidar. Background Technology

[0002] In the field of sand-fixing forest ecological monitoring, the canopy spatial structure is the core factor determining the effectiveness of windbreak and sand fixation. Its accurate quantification is crucial for revealing the sand-fixing mechanism and optimizing stand configuration. Currently, existing technologies mainly rely on lidar data acquisition and analysis from a single observation platform. Specifically, single ground-based lidar acquisition can obtain high-density detailed point clouds of the understory, trunks, and canopy gaps, but the observation range is limited, making it difficult to cover the overall stand pattern at the plot scale. Furthermore, it is easily affected by the upper canopy layer, leading to missing data on the upper canopy structure. Single UAV lidar acquisition can quickly obtain continuous surface and spatial pattern information of the canopy at the plot scale, suitable for macroscopic stand structure analysis, but its penetration capability is limited, failing to accurately capture detailed structures under the forest canopy and within the canopy, and easily introducing scale bias.

[0003] Meanwhile, existing technologies still face the following key bottlenecks in the process of multi-source data fusion and structural quantification: Significant differences exist in the observation perspectives and point cloud densities between ground-based and UAV lidar data; a lack of unified spatial references and topographic benchmarks; and an imperfect coarse-to-fine registration and normalization process, resulting in poor comparability and reproducibility of cross-platform data. Simply relying on traditional indicators such as tree height and canopy closure makes it difficult to characterize key features such as canopy void connectivity and vertical stratification differences, failing to fully explain the differentiation in sand-fixing function brought about by different tree species and density configurations. Current quantitative results mostly remain at the structural description level, without establishing a mechanistic link with key windbreak and sand-fixing processes, making it difficult to directly translate into actionable operational decision-making basis such as tree species selection and density optimization. Some quantification methods excessively pursue parameter accuracy, neglecting computational efficiency and multi-scenario data adaptability, thus limiting long-term monitoring and large-scale promotion in complex field environments.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0006] According to one aspect of this application, a method for quantifying the spatial structure of sand-fixing forest canopies based on ground-based and UAV lidar is provided, comprising: acquiring multi-source core data for the entire process of quantifying the spatial structure of sand-fixing forest canopies; performing denoising, filtering, and resampling processing on ground-based and UAV lidar point cloud data based on point cloud preprocessing algorithms, completing accurate registration of point clouds across multiple platforms using a coordinate transformation model, distinguishing different sand-fixing forest community types using a tree species feature recognition model, integrating topographic data into the point cloud correction process, and generating a standardized, integrated fused point cloud dataset for sand-fixing forests; layering the fused point cloud dataset according to stand scale and single tree scale, constructing ternary association nodes including point cloud features, tree entities, and structural parameters in each layer, aggregating point cloud spatial topology information using a neighborhood feature search algorithm, and generating a canopy structure feature association map for the entire scale of sand-fixing forests; A point cloud feature mining neural network is used to model the canopy structure feature correlation map, extracting core features such as point cloud elevation, density, and intersticity, and mining the structural correlation patterns between individual trees and stand scales to generate high-dimensional canopy structure feature embedding vectors. The canopy structure feature embedding vectors are then imported into a hierarchical quantization calculation engine, and iterative optimization is performed using a multi-scale parameter adaptive solution mechanism. A measured data calibration module is introduced, and a correction decision chain is generated through visualization of structural parameter deviations, forming a basic quantitative model of sand-fixing forest structure that integrates ground-based and UAV multi-source data. Based on the quantization basic model, combined with sample plot measured verification data and stand growth dynamic monitoring data, accuracy verification and dynamic correction are performed. A multi-index comprehensive evaluation algorithm is used to balance the requirements of parameter accuracy, computational efficiency, and data adaptability, generating canopy spatial structure quantification results that are suitable for the entire scenario of sand-fixing forest monitoring.

[0007] Another aspect of this application is a system for quantifying the spatial structure of sand-fixing forest canopies based on ground-based and UAV lidar, the system being configured to execute the above-described method for quantifying the spatial structure of sand-fixing forest canopies based on ground-based and UAV lidar by executing the executable instructions.

[0008] According to another aspect of this application, an electronic device includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-described method for quantifying the spatial structure of sand-fixing forest canopies based on ground-based and UAV lidar via executing the executable instructions.

[0009] According to another aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a second processor, implements the above-described method for quantifying the spatial structure of sand-fixing forest canopies based on ground-based and UAV lidar.

[0010] This application presents a method and system for quantifying the spatial structure of sand-fixing forest canopies based on ground-based and UAV lidar. This application achieves accurate quantification of the spatial structure of sand-fixing forest canopies through a complete process of "data acquisition - fusion preprocessing - feature modeling - quantization modeling - accuracy optimization." First, continuous point clouds of the canopy and detailed point clouds of the understory are collected collaboratively from sample plots. After denoising, registration, and topographic normalization, a standardized fusion dataset is generated. Then, the dataset is layered according to the individual tree-stand scale, constructing ternary association nodes and aggregating topological information. High-dimensional features are extracted using a graph convolutional neural network. Finally, through iterative optimization by a layered quantization engine, experimental calibration, and dynamic correction, a quantization result adaptable to all scenarios is output. Core algorithms include ICP registration, CSF filtering, random forest classification, and Kalman filter correction, forming a technical system of "multi-source complementarity, multi-scale integration, and multi-index balance."

[0011] Breaking through the limitations of a single platform, this method achieves seamless integration of the canopy structure through dual-perspective collaborative sampling and precise registration, improving data completeness by over 30% and resolving issues of missing structures and scale bias. A multi-scale quantitative index system is constructed, covering core features such as canopy porosity and vertical stratification, with quantification errors ≤0.2m (length-related) and ≤0.05 (dimensionless), demonstrating significantly superior parameter accuracy compared to traditional methods. Integrating dynamic correction and multi-objective balancing mechanisms, it balances computational efficiency and data adaptability, with a total processing time ≤30 minutes and a community fit rate ≥90%, adapting to complex field monitoring scenarios. The quantitative results are deeply correlated with the sand fixation process, directly providing decision-making basis for tree species configuration and density optimization, driving a technological upgrade from "structural description" to "management support," and possessing strong practicality and promotional value. Attached Figure Description

[0012] Figure 1 This document illustrates a flowchart of a method for quantifying the spatial structure of sand-fixing forest canopies based on ground-based and UAV lidar, according to an embodiment of this application.

[0013] Figure 2 This paper presents a schematic diagram of a system for quantifying the spatial structure of sand-fixing forest canopy based on ground-based and UAV lidar, according to an embodiment of this application. Detailed Implementation

[0014] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0015] In one implementation, Figure 1 The diagram illustrates a flowchart of a method for quantifying the spatial structure of sand-fixing forest canopies based on ground-based and UAV lidar, according to an embodiment of this application.

[0016] S101, acquire multi-source core data of the entire process of quantitative analysis of the spatial structure of the canopy of sand-fixing forests.

[0017] In one implementation, a collaborative acquisition method using UAV lidar and backpack-mounted ground-based lidar is employed to obtain continuous point clouds of the canopy at the sample plot scale and high-density detailed point clouds of the understory / canopy. Simultaneously, supporting tasks such as establishing benchmark constraints, parsing and preprocessing raw data are completed, laying the data foundation for subsequent point cloud fusion and structure quantization. The overall process consists of three key steps: concurrent data acquisition and establishment of benchmark constraints, parsing raw data and stitching point clouds on the same platform, and point cloud preprocessing and ground point classification.

[0018] This step is fundamental to acquiring multi-source core data. Its core is to simultaneously collect point cloud data from both UAVs and backpack-mounted ground-based lidar, while establishing high-precision spatial benchmarks and positioning constraints. This ensures spatial correspondence between the two types of point clouds, balancing the completeness of plot-scale data with the precision of individual tree-scale data. Within the sand-fixing forest plots to be monitored, several evenly distributed, clearly visible, stable, and reproducible locations are selected as benchmark / control points. These serve as spatial references for subsequent point cloud registration. For example, concrete marker stakes are set at the four corners and center of the plot as fixed control points, with coordinate information marked and properly marked in the field to prevent burial by wind and sand or human disturbance.

[0019] Backpack-type LiDAR equipment was used to conduct mobile scanning under the forest canopy, collecting high-density detailed point clouds of the understory, tree trunks, and canopy gaps. During the collection process, real-time high-precision location information was simultaneously acquired via RTK. The walking path adopted a "Z" shape for full coverage, and the end returned to the starting point to form a closed loop, which facilitated subsequent inspection of trajectory drift and stitching consistency. For example, using the Digital Green Soil Backpack-type LiDAR system, a "Z" shaped mobile scanning was performed in the sand-fixing forest sample plot with a row spacing of 5 meters. During the scanning process, RTK centimeter-level positioning was enabled, and the scanning trajectory and point cloud acquisition parameters were recorded throughout. After completion, the device returned to the starting point along the original trajectory to form a closed scanning path.

[0020] Drones equipped with lidar sensors are used to collect canopy point clouds at the sample plot scale, focusing on acquiring continuous point clouds of the upper canopy morphology and the overall pattern of the sample plot. Flight parameters need to be set reasonably according to the height of the sand-fixing forest and the sample plot area. At the same time, differential correction and high-precision positioning support are provided by a base station. For example, using a DJI drone equipped with a digital green soil lidar sensor, for a sea buckthorn sand-fixing forest sample plot with a height of about 3-5 meters, a terrain-following flight altitude of 70 meters, a 75% heading overlap, and a flight speed of 7 m / s are set. The DJID-RTK2 is used as a base station to provide differential correction. The sample plot area is fully covered during the flight, and the flight log and base station status information are saved simultaneously.

[0021] The raw data collected by UAVs and backpack-based ground-based LiDAR are professionally analyzed to extract valid point cloud and attribute information. Point cloud stitching within the same platform is also performed to ensure geometric consistency of the point cloud on a single platform, eliminating spatial deviations between sorties and scan segments. The raw data collected by UAVs and backpack-based ground-based LiDARs are professionally analyzed separately, filtering out invalid data and outputting high-precision point cloud data containing key attributes such as GPS time, echo information, intensity, and 3D coordinates. For example, using LiDAR-compatible data analysis software, the raw radar data collected by UAVs is processed to remove invalid data from abnormal sensor acquisition during flight, outputting a point cloud file with GPS timestamps, echo counts, laser reflection intensity, and WGS84 coordinates. The raw scan data from the backpack-based LiDAR is analyzed to output a high-density detailed point cloud file containing movement trajectory positioning information and 3D coordinates of the point cloud.

[0022] Point cloud data collected by UAVs from different sorties and at different scanning times are striped / stitched together to unify the coordinate system, eliminate translation and rotation deviations between sorties, and ensure the geometric continuity and consistency of UAV point clouds at the plot scale. For example, for data collected by UAVs in two sorties of sand-fixing forest plots, the coordinates of the control points within the plot are used as the reference to stitch and merge the point clouds from the two sorties in the point cloud stitching software, convert them to the local plane coordinate system, and after checking that there are no obvious misalignments at the stitching points, a complete continuous UAV canopy point cloud is output. Based on the movement trajectory and RTK positioning information collected by the backpack radar, the point clouds of different scanning segments are stitched together, and the consistency of the stitched segments of the closed loop is checked to avoid point cloud misalignment or stretching at the boundaries of the sample plots and in the understory area. For example, based on the trajectory information of the backpack radar's "Z"-shaped closed scan, the understory point clouds of each scanning segment are stitched together, with a focus on checking the overlap of the point clouds at the starting and ending points. Areas with slight trajectory drift are locally corrected to ensure that the point cloud morphology of the tree trunk and canopy gaps is not distorted, and a complete high-density detailed point cloud of the understory is output.

[0023] The stitched single-platform point clouds underwent preprocessing such as denoising and filtering to remove invalid information and accurately classify and extract ground points, establishing a reliable ground reference surface to provide a high-quality point cloud data foundation for subsequent multi-source point cloud registration and terrain normalization. Neighborhood consistency smoothing was applied to the stitched UAV and backpack-based ground-based lidar point clouds to reduce the impact of random noise. Simultaneously, outliers in the air and invalid areas outside the sample plots were removed through cropping and statistical analysis. For example, a 3-neighborhood smoothing algorithm was used to optimize the UAV canopy point cloud data, eliminating point cloud spikes caused by random noise from the lidar sensor; point clouds of open areas and roads outside the sample plots were removed through coordinate range cropping, and statistical filtering was used to remove outliers in the air that deviated from the main body of the point cloud. For backpack-based forest understory point clouds, invalid points in the air collected due to sensor movement during scanning, as well as non-sand-fixing forest-related nodules such as rocks and dead trees, were removed.

[0024] Combining the acquisition characteristics of UAV and backpack radar point clouds, different methods such as statistical filtering and radius filtering are used for refined noise reduction to ensure the purity of point cloud data. For example, statistical filtering is used for UAV point clouds to calculate the average distance of each point's neighboring points and remove outliers whose average distance exceeds 3 times the standard deviation. For backpack forest understory point clouds, because the point cloud density is higher and is easily affected by tree trunks and branches, radius filtering is used to set a reasonable radius and the minimum number of neighboring points to remove isolated noise points.

[0025] A professional ground point extraction algorithm is used to perform ground point filtering and classification on the two types of point clouds after denoising, separating ground points from non-ground points and constructing a reliable ground reference surface. For example, the cloth simulation filtering (CSF) algorithm is used for the preprocessed point clouds of UAVs and backpacks. Reasonable cloth resolution, gravity coefficient and other parameters are set to simulate the process of cloth covering the point cloud surface, accurately extract ground points, separate non-ground point clouds such as tree trunks and branches of sand-fixing forests, and output classified point cloud files to provide a ground benchmark for subsequent terrain normalization.

[0026] S102 uses a point cloud preprocessing algorithm to denoise, filter, and resample point cloud data from ground-based and UAV lidar systems. It combines a coordinate transformation model to achieve accurate registration of point clouds across multiple platforms. It uses a tree species feature recognition model to distinguish different sand-fixing forest community types and integrates topographic data into the point cloud correction process to generate a standardized and integrated sand-fixing forest fusion point cloud dataset.

[0027] In one implementation, to address the differences in acquisition characteristics between ground-based and UAV LiDAR point clouds, differentiated professional algorithms are employed for denoising and filtering to eliminate data noise and invalid information. Then, resampling is used to standardize and unify the point cloud density, laying a high-quality data foundation for subsequent registration and fusion. The core algorithms include statistical filtering, radius filtering, and cloth simulation filtering (CSF), with voxel-based resampling employed. For UAV LiDAR point clouds, due to their large acquisition range and susceptibility to aerial noise interference, statistical filtering is used: the mean Euclidean distance of each point's neighboring points is calculated, and a threshold of 3 times the standard deviation is set. Outliers with distances exceeding the threshold are removed, eliminating invalid points caused by birds in flight, sensor noise, etc. For example, for UAV point clouds from sea buckthorn sand-fixing forest plots, with a neighboring point count of 15, the calculated mean neighboring distance is 0.2m. Points exceeding 0.8m are identified as outliers and removed, eliminating aerial noise and isolated noise from the point cloud. For backpack-mounted ground-based lidar point clouds, isolated points are easily generated due to obstruction by tree trunks and rocks during forest understory collection. A radius filtering method is employed: a fixed search radius and a minimum number of neighboring points within that radius are set, and isolated points that do not meet the required number of neighboring points are removed, eliminating noise from forest understory debris and sensor vibration. For example, for backpack point clouds in sand-fixing forests, a search radius of 0.3m and a minimum number of neighboring points of 8 are set. Debris points corresponding to rocks and dead wood, as well as isolated noise points generated during scanning, are removed, while retaining the effective point cloud data of tree trunks and branches.

[0028] For both types of denoised point clouds, the Cloth Simulation Filtering (CSF) method was uniformly used to filter and classify ground points. This algorithm accurately fits the terrain surface by simulating the physical process of cloth covering the point cloud surface, achieving effective separation of ground points from non-ground points such as tree trunks and branches of sand-fixing forests. For example, for point cloud data of mixed sand-fixing forests of Salix psammophila and Caragana korshinskii, the CSF algorithm parameters were set to cloth resolution of 0.5m, gravity coefficient of 1.0, and number of iterations of 500. The settling process of cloth on the point cloud surface was simulated to extract the ground point cloud within the sample plot and separate the non-ground point cloud of trees, providing an accurate ground reference for subsequent terrain normalization. Due to the low density of UAV point clouds (approximately 10 points / m²) and the high density of backpack-based ground point clouds (approximately 50 points / m²), a voxel resampling method was adopted. By constructing a three-dimensional voxel mesh, the two types of point clouds were resampled to achieve standardized and unified point cloud density while preserving the core spatial structural features of the point clouds. For example, a three-dimensional voxel mesh with a side length of 0.2m is constructed, and the point clouds of the UAV and backpack are resampled respectively. A representative point is retained in each voxel (the centroid of all points in the voxel is taken), and the density of the two types of point clouds is unified to 25 points / ㎡. This not only eliminates the influence of the difference in point cloud density on subsequent registration, but also completely preserves the spatial structural information of the canopy and trunk.

[0029] Using a unified spatial geographic coordinate system as a benchmark, a coordinate transformation model is used to normalize the coordinates of point clouds from multiple platforms. A two-stage registration strategy of "coarse registration + fine registration" is then employed to achieve precise spatial alignment between the ground-based and UAV point clouds. The core algorithm includes point-to-point coarse registration based on feature points and fine registration using the Iterative Closest Point (ICP) algorithm. Accuracy verification ensures registration reliability. A seven-parameter Bursa-Wolf coordinate transformation model is constructed to uniformly convert the original acquisition coordinates (such as WGS84 geodetic coordinates and device relative coordinates) of UAV and backpack-based ground-based lidar point clouds into local Cartesian coordinates (Gauss-Kruger projection coordinates), achieving a consistent spatial reference benchmark for the point clouds from multiple platforms. For example, the WGS84 latitude and longitude coordinates of the UAV point cloud and the device relative coordinates of the backpack point cloud are converted into Gauss-Kruger 3-degree Cartesian coordinates (central meridian 108°) using the Bursa-Wolf model, placing the two types of point clouds in the same spatial coordinate system and eliminating registration errors caused by differences in coordinate benchmarks.

[0030] Based on a unified coordinate benchmark, stable feature points within the sample plot (such as control point markers, the base of tall sand-fixing forest tree trunks, and obvious terrain inflection points) are extracted. A point-to-point coarse registration method is used to translate and rotate the foundation and UAV point cloud, quickly eliminating large-scale spatial deviations and laying the foundation for fine registration. For example, in a sea buckthorn sand-fixing forest sample plot, five concrete markers and three terrain inflection points are selected as feature points. The three-dimensional coordinates of these feature points in both types of point clouds are extracted. The registration transformation matrix is ​​solved using the least squares method, and the backpack point cloud is spatially transformed to control the overlap error of feature points between the two types of point clouds within 0.5m, completing the coarse registration. Based on the coarsely registered point cloud, the Iterative Closest Point (ICP) algorithm is used for fine registration. By iteratively finding the closest point pair between the source point cloud (backpack point cloud) and the target point cloud (UAV point cloud), the optimal spatial transformation matrix is ​​solved, gradually reducing the registration error between point clouds and achieving high-precision alignment. For example, using the coarsely registered backpack point cloud as the source point cloud and the UAV point cloud as the target point cloud, the ICP algorithm is set to 200 iterations and a convergence threshold of 0.01m. Through iterative optimization, the average point cloud distance error in the overlapping area of ​​the two types of point clouds is reduced from 0.45m to 0.08m, thus achieving accurate registration.

[0031] The digital surface model (DSM) of the overlapping point cloud region is extracted, the elevation residual is calculated, and typical profile lines are drawn to analyze the continuity of topography and canopy, verifying the reliability of registration. If the residual is too large or the profile lines are distorted, feature points are reselected for coarse registration, and then ICP fine registration is performed again. For example, the DSM elevation residual of the overlapping point cloud region of the sand-fixing forest is calculated, with the mean controlled within 0.1m and the maximum value not exceeding 0.2m. East-west and north-south canopy profile lines of the sample plot are drawn. If the profile lines are continuous without obvious misalignment or distortion, the registration result is determined to be reliable.

[0032] Based on the registered and fused point clouds, point cloud features (such as canopy morphology, point cloud density distribution, trunk texture, elevation features, etc.) of different sand-fixing forest tree species are extracted. A machine learning tree species feature recognition model (random forest classification model) is constructed. Through model training and recognition, the classification of individual tree species and community types in sand-fixing forests is completed, laying the foundation for subsequent targeted structural parameter extraction. Core point cloud features of typical sand-fixing forest tree species (such as sea buckthorn, sand willow, tamarisk, and caragana) are extracted from the registered point clouds. These features include canopy three-dimensional morphological features (canopy width, canopy height, canopy shape coefficient), point cloud density features (canopy upper / lower point cloud density ratio), branch and trunk features (curvature and texture of trunk point cloud), and elevation features (canopy centroid height), forming a tree species feature dataset. For example, point cloud features of sea buckthorn, *Caragana korshinskii*, and *Caragana chinensis* were extracted from the sample plots. Sea buckthorn exhibited small crown width (1.0-1.5m), compact crown shape, and high canopy point cloud density; *Caragana korshinskii* exhibited large crown width (2.0-3.0m), sparse canopy, and high center of gravity; and *Caragana chinensis* exhibited multiple branches and high trunk point cloud curvature. Based on these features, a feature dataset was constructed. Using the extracted tree species point cloud features as input and the tree species types from the field survey as labels, a random forest classification model was constructed. The number of decision trees was set to 100, and the number of feature subsets was set to 5. The dataset was divided into training and validation sets in a 7:3 ratio. Through model training, accurate identification of different tree species was achieved, with a model classification accuracy of no less than 90%. For example, using the point cloud features of sea buckthorn, *Caragana korshinskii*, and *Caragana chinensis* as the training set, the random forest classification model was trained, and the classification accuracy on the validation set reached 92%, meeting the requirements for identifying tree species in sand-fixing forests.

[0033] Based on the identification results of individual tree species, and combined with the composition and distribution characteristics of tree species within the sample plots, community types such as pure forests and mixed forests were classified according to the classification standards for sand-fixing forest communities. The spatial distribution range of different communities was also marked. For example, through model identification, the northern area of ​​the sample plot was dominated by sea buckthorn monoculture and was identified as a pure sea buckthorn forest; the southern area was a mixed distribution of sand willow and tamarisk and was identified as a sand willow-tamarisk mixed forest. The spatial boundaries and ranges of the two communities were marked respectively, completing the community type classification.

[0034] Based on high-precision topographic data of the sample plots, topographic factors such as undulation, slope, and aspect are integrated into the point cloud correction process. A topographic normalization algorithm is used to eliminate the influence of topography on the elevation of the sand-fixing forest point clouds, achieving topographic correction of the point clouds. Simultaneously, local optimization is performed on point clouds in different topographic regions to ensure standardization and consistency. Core data includes a digital terrain model (DTM) and slope and aspect raster data. Combining field-measured topographic control points with ground point clouds extracted using the fabric simulation filtering method, a high-precision digital terrain model (DTM) of the sample plots is constructed. Slope and aspect raster data are calculated based on the DTM, and the topographic data is spatially correlated with the registered point clouds, ensuring that each point cloud point matches the corresponding topographic factors (elevation, slope, and aspect). For example, based on the ground point cloud of the sea buckthorn sand-fixing forest plot, a DTM with a resolution of 0.5m was constructed. The slope of the plot was calculated to be between 3° and 15°, with the slope direction mainly in the northwest direction. The topographic elevation of the DTM was spatially matched with the original elevation of the point cloud points to complete the topographic data fusion.

[0035] A terrain normalization algorithm is employed to subtract the corresponding DTM terrain elevation from the original elevation of each point cloud point, thus obtaining the point cloud's elevation above the ground. This eliminates the influence of terrain undulations on the elevation of the sand-fixing forest canopy and trunks, achieving standardized correction of point cloud elevations. For example, the original elevation of a sea buckthorn canopy point cloud within the sample plot is 1256.8m, while the corresponding DTM terrain elevation is 1255.2m. After terrain normalization, the elevation above the ground for this point is 1.6m, eliminating the influence of terrain protrusions on canopy elevations and making point cloud elevations comparable across different terrain regions. For terrain regions with different slopes and aspects within the sample plot, local optimization is performed on the corrected point clouds: for steep slopes (slope > 10°), neighborhood interpolation is used to supplement point clouds missing due to shading; for shaded / sunny slopes with different lighting conditions, the point cloud density threshold is adjusted to ensure consistent point cloud quality across different terrain regions. For example, in the northwest steep slope area with a slope of 12°-15° in the sample plot, some tree trunks of the *Caragana korshinskii* tree were missing point clouds due to the shading caused by the forest. The neighborhood interpolation method was used to fill in the missing point clouds based on the surrounding effective point clouds. For the point cloud reflection noise caused by strong sunlight on the sunny slope, the statistical filtering method was used again to perform local noise reduction and complete the local correction of the point clouds.

[0036] Multi-source point clouds that have undergone denoising, registration, tree species identification, and terrain correction are integrated and standardized. The attribute labeling, data format, and storage standards of the point clouds are unified, ultimately generating a standardized and integrated point cloud dataset for sand-fixing forests. The UAV point clouds, after all preprocessing and correction, are spatially stitched together with backpack-based ground-based point clouds. The UAV point clouds supplement the continuous canopy morphology and spatial pattern at the plot scale, while the backpack point clouds supplement the details of canopy gaps, trunks, and understory branches and leaves, achieving a seamless point cloud structure from "canopy top" to "canopy bottom." For example, when the UAV canopy point clouds of a mixed forest of *Salix psammophila* and *Caragana korshinskii* are fused with backpack understory point clouds, the UAV point clouds fully present the overall undulation and patch distribution of the plot canopy, while the backpack point clouds supplement the three-dimensional structure of canopy gaps and the fine morphology of tree trunks. The fused point cloud covers the entire plot while preserving detailed information about the understory and canopy.

[0037] The fused point cloud is standardized with attribute annotations. Each point cloud point's attributes include three-dimensional coordinates (Cartesian coordinates, height above ground), tree species type, community type, point cloud source, topographic factors (slope, aspect), and quality label (valid point / missing area), achieving standardized and unified point cloud attributes. For example, each point in the fused point cloud is labeled with Gauss-Kruger plane coordinates and height above ground, along with its tree species (Salix psammophila / Caragana korshinskii), community type (mixed forest), point cloud source (UAV), slope (8°), aspect (northwest), and quality label (valid point), forming a standardized attribute system. The fused point cloud is converted to a common point cloud data format (LAS format), and storage parameters are set according to sand-fixing forest monitoring standards. Simultaneously, a dataset metadata archive is established, recording sample plot information, collection parameters, processing algorithms, accuracy indicators, etc., achieving standardized storage and traceability of the dataset. For example, the fused point cloud is converted to LAS1.4 format, named according to the plot number "GS-L-001", and a metadata archive is established to record the plot location, UAV / backpack collection parameters, CSF filtering, ICP registration, random forest classification and other processing algorithms, as well as accuracy indicators such as point cloud registration error (0.08m) and tree species identification accuracy (92%). Finally, a standardized and integrated sand-fixing forest fused point cloud dataset is generated.

[0038] S103, the fused point cloud dataset is layered according to stand scale and single tree scale. Each layer constructs a ternary association node including point cloud features, tree entities and structural parameters. The spatial topology information of the point cloud is aggregated through the neighborhood feature search algorithm to generate a canopy structure feature association map for the whole scale of sand-fixing forest.

[0039] In one implementation, combining the features of the fused point cloud dataset of sand-fixing forests with the multi-scale requirements for canopy structure quantification, a single-tree-stand scale division rule and a point cloud spatial feature parsing mechanism are introduced to perform scale-layered processing on the fused point cloud dataset, generating dual-scale point cloud layered results for sand-fixing forest stands and single trees. The single-tree scale uses independent tree entities as the dividing unit, defined as the set of all point clouds from the base of the trunk to the top of the canopy, with corresponding point cloud features including individual tree characteristics such as crown shape, trunk shape, and point cloud elevation / density. The stand scale is based on sand-fixing forest plots, divided into several continuous stand units using a regular grid method, defined as the set of point clouds of all tree entities and habitats within each grid, with corresponding point cloud features including canopy undulation, density distribution, and patch pattern, among other group characteristics. Simultaneously, the extraction boundaries and feature dimensions of the dual-scale point clouds are clearly defined.

[0040] A point cloud neighborhood feature analysis algorithm is employed to extract spatial features such as 3D coordinates, point cloud density, normal vector, and curvature from the fused point cloud, providing a feature basis for individual tree segmentation and forest stand division. Addressing the characteristics of sparse canopy and uneven point cloud distribution in sand-fixing forests, the algorithm focuses on analyzing the interstices and vertical stratification features of the canopy point cloud to ensure that the scale division conforms to the structural characteristics of the sand-fixing forest. On the normalized fused point cloud, a point cloud-based individual tree segmentation algorithm (PCS) is used, combining the advantages of region growing and watershed segmentation methods. First, canopy vertices are identified through point cloud elevation clustering. Then, region growing is performed using canopy vertices as seed points, and individual tree segmentation is completed by combining the connectivity constraints of the trunk point cloud, extracting the point cloud set for each independent tree. For example, for the point cloud of the mixed sand-fixing forest of Salix psammophila and Caragana korshinskii, the canopy vertices were identified with an elevation clustering threshold of 1.5m, the regional growth neighborhood radius was set to 0.8m, and the segmentation boundary was constrained by the curvature characteristics of the tree trunk point cloud. The point clouds of 216 independent Salix psammophila and 189 independent Caragana korshinskii in the sample plot were successfully segmented. Each tree point cloud contains complete trunk and canopy point cloud information, forming a layered result of the point cloud at the tree scale.

[0041] A stand-scale regular grid division method was adopted to divide the fused point cloud of the sand-fixing forest sample plot into 20m×20m grids, with each grid serving as a stand unit. All point clouds within each unit (including multiple individual tree point clouds, understory point clouds, and gap point clouds) were extracted, and the spatial location and adjacency relationships of each stand unit were recorded. For example, a 100m×100m sand-fixing forest sample plot fused point cloud was divided into 25 stand units using a 20m×20m grid. Each unit contained several individual tree point clouds of *Salix psammophila* and *Caragana korshinskii*, as well as understory gap point clouds. The complete point cloud set of each unit was extracted to form a stand-scale point cloud stratification result. Spatial correlation and attribute labeling were performed on the individual tree and stand-scale point cloud stratification results. Each individual tree point cloud was labeled with its respective stand unit number, and each stand unit was labeled with the number of individual trees and tree species, generating a dual-scale point cloud stratification result for the sand-fixing forest individual tree-stand, ensuring spatial consistency and feature correlation between the two scales of the point cloud.

[0042] Combining multi-scale structural parameter extraction requirements with point cloud feature association logic, the layered results of dual-scale point clouds are processed to construct ternary association nodes for point cloud features, tree entities, and structural parameters, generating scaled node association matching results. At the tree scale, point cloud features include point cloud elevation, density, gap density, and canopy point cloud normal vector; tree entities are independent sand-fixing forest plants such as *Salix psammophila* and *Caragana korshinskii*, including attributes such as tree species, age, and spatial location; structural parameters are quantifiable structural indicators such as tree height, crown width, crown volume, and trunk diameter. At the stand scale, point cloud features include the average point cloud elevation, density standard deviation, and mean canopy gap density of the stand unit; tree entities are the collection of all trees within the stand unit, including attributes such as tree quantity, species composition, and row / plant spacing; structural parameters are quantifiable structural indicators such as stand canopy closure, average crown height, canopy relief, and patch density.

[0043] For each individual tree point cloud, its core point cloud features are extracted and associated with the corresponding forest entity attributes. Combining this with the extraction standards for individual tree structural parameters in sand-fixing forests, the structural parameters that can be calculated from the corresponding point cloud features of the forest entity are matched, constructing a ternary association node at the individual tree scale. Each node is an association combination of <individual tree point cloud features, individual tree entity, and individual tree structural parameters>. For example, the point cloud of a *Salix psammophila* tree with a diameter at breast height (DBH) of 5cm and a height of 4.2m is extracted, and its point cloud elevation range (0.3-4.2m), canopy point cloud density (30 points / m²), and canopy interstices (0.25) are obtained. These features are then associated with the forest entity attributes of the *Salix psammophila* (tree species: *Salix psammophila*, location: forest stand unit 3-2), and the calculable structural parameters (canopy width, canopy volume, and trunk shape index) are matched to construct a ternary association node at the individual tree scale for the *Salix psammophila*.

[0044] For each forest stand unit's point cloud, its core point cloud features are extracted and associated with the corresponding forest stand tree entity attributes. Combining this with the extraction standards for sand-fixing forest stand structural parameters, the structural parameters that can be calculated from the corresponding point cloud features of the forest stand unit are matched, constructing a stand-scale ternary association node. Each node is an association combination of <forest stand point cloud features, forest stand tree entities, and forest stand structural parameters>. For example, the point cloud features of forest stand unit 3-2 are extracted (average elevation 2.8m, point cloud density standard deviation 8 points / m², average canopy gap 0.3), associated with the unit's tree entity attributes (number of trees: 16, tree species composition: 10 *Salix psammophila* trees and 6 *Caragana korshinskii* trees), and matched with the calculateable structural parameters (canopy closure, average crown height, canopy relief) to construct a stand-scale ternary association node for that forest stand unit.

[0045] The feature-parameter association matching algorithm is adopted. Based on the quantitative calculation relationship between point cloud features and structural parameters, the association strength of the ternary association nodes at the individual tree and stand scales is calculated. Invalid nodes with no direct calculation relationship between features and parameters are eliminated. At the same time, the hierarchical association between individual tree nodes and their respective stand nodes is established, the feature contribution ratio of individual tree nodes in stand nodes is marked, and the scaled node association matching results are generated.

[0046] Based on the spatial topological features of sand-fixing forests, the distribution patterns of multi-scale point clouds, and the logical generation rules for node association construction, a systematic verification of the scaled node association matching results is performed to complete the construction of full-scale ternary association nodes. The construction criteria for ternary association nodes are clarified, including the spatial consistency criterion between point cloud features and tree entities, the feasibility criterion for solving point cloud features and structural parameters, the hierarchical inclusion criterion between single tree nodes and stand nodes, and the feature complementarity criterion between nodes of the same scale. Simultaneously, an error threshold for node association is set (e.g., the spatial positional deviation between point cloud features and tree entities should not exceed 0.5m). Combining the spatial topological features of the fused point cloud of sand-fixing forests (e.g., the neighborhood distribution of trees, canopy connectivity, and spatial distribution of voids), the consistency between the spatial position of tree entities and the spatial distribution of point cloud features in the ternary association nodes is verified. Nodes with spatial positional deviations exceeding the threshold are corrected, and the corresponding point cloud features or tree entities are re-matched. For example, during verification, it was found that the location of a single tree node of a certain *Hedysarum heterotropoides* is forest stand unit 4-1, but its point cloud features all come from forest stand unit 4-2, with a spatial position deviation of 1.2m, which exceeds the error threshold of 0.5m. Therefore, the point cloud features of the *Hedysarum heterotropoides* in forest stand unit 4-1 were re-extracted to correct the ternary association node of the single tree.

[0047] Based on the distribution patterns of point clouds at the individual tree and stand scales in sand-fixing forests (e.g., the density of point clouds in individual trees gradually decreases from the canopy to the trunk, and the density of point clouds in stands shows a patchy distribution along the plot gradient), the point cloud feature values ​​in the nodes are verified to ensure they conform to the distribution patterns of point clouds in sand-fixing forests. Abnormal nodes that significantly deviate from the patterns are investigated, and invalid nodes caused by point cloud noise are removed. For example, if the point cloud density of a certain stand node is labeled as 80 points / m², which is far higher than the normal range of point cloud density in sand-fixing forest stands (15-35 points / m²), it is found to be an anomaly caused by incomplete point cloud denoising. This invalid stand ternary association node is removed and reconstructed.

[0048] Based on the logical rules for node association, the system verifies the completeness of ternary association nodes at the individual tree and stand scales, checking for any missing elements such as point cloud features, tree entities, or structural parameters. It also verifies the completeness of the hierarchical association between individual tree nodes and their respective stand nodes, supplementing nodes with missing elements or relationships. For example, if the verification reveals that the ternary association node for stand unit 2-3 lacks the "patch density" structural parameter, this parameter is supplemented using the point cloud patch features of that stand unit, thus perfecting the stand's ternary association node. The scaled node association matching results, after multi-dimensional verification, correction, and supplementation, are integrated to form a full-scale ternary association node system for sand-fixing forests, encompassing all valid individual tree and stand-scale ternary association nodes, where the hierarchical, spatial, and feature associations between nodes all conform to the rules.

[0049] Based on the spatial distribution characteristics of the fused point cloud and the neighborhood association relationships of forest entities, the corresponding neighborhood search range is automatically matched. An anomaly is marked and the range is corrected based on the point cloud feature-topology information adaptability detection mechanism. Combining the planting patterns and growth characteristics of sand-fixing forests, the neighborhood association relationships of forest entities at the individual tree and stand scales are analyzed. At the individual tree scale, neighborhood associations refer to the spatial location and canopy competition relationship between the target tree and its neighboring trees, with the neighborhood range influenced by the row spacing (typically 1-3m in sand-fixing forests). At the stand scale, neighborhood associations refer to the structural characteristics and patch connectivity relationships between the target stand unit and its neighboring stand units, with the neighborhood range being 1-2 grid units surrounding the target unit. An adaptive neighborhood search algorithm is employed, with full-scale ternary association nodes as the core. At the single-tree scale, the origin is the center of the target tree entity, and a spherical neighborhood search radius (1.5 times the crown width) is automatically matched based on the crown width of the target tree to search the spatial topology information of the surrounding point cloud. At the stand scale, the origin is the geometric center of the target stand unit, and a rectangular neighborhood search range (target unit + 8 surrounding neighboring units) is automatically matched based on the stand unit's grid size to search the spatial topology information of the surrounding stand's point cloud. For example, for a single Salix psammophila tree with a crown width of 2.0m, a spherical neighborhood search radius of 3.0m is automatically matched with its trunk base as the origin; for a 20m×20m stand unit, a rectangular neighborhood search range of 60m×60m (including 8 surrounding neighboring units) is automatically matched.

[0050] A point cloud feature-topology information fit detection mechanism was established. This mechanism calculates the similarity between point cloud features within the matched neighborhood search range and the point cloud features of the target node, as well as the fit between topological information and the neighborhood association of the target forest entity. Fit thresholds were set (similarity ≥ 0.7, fit ≥ 0.6). Cases with fit below the threshold were marked as match anomalies. For these anomalies, the neighborhood search range was corrected based on the actual distribution characteristics of the sand-fixing forest point clouds. At the single-tree scale, a dynamic radius adjustment method was used, increasing or decreasing the spherical neighborhood search radius until the fit reached the threshold. At the stand scale, a grid cell addition / reduction method was used, increasing or decreasing the number of adjacent stand units until the fit reached the threshold. For example, for a single *Salix psammophila* tree with a crown width of 2.0m, the point cloud feature similarity within a 3.0m search radius was 0.58, below the threshold, and was judged as a match anomaly. After adjusting the search radius to 2.5m, the similarity increased to 0.75, and the fit reached 0.68, meeting the fit requirement and completing the single-tree neighborhood search range correction. After matching and correcting the neighborhood search range for all ternary association nodes at the individual tree and stand scales, the optimal neighborhood search range and search type (spherical / rectangular) of each node are recorded to provide a precise search interval for subsequent spatial topology information aggregation.

[0051] According to the preset point cloud feature aggregation rules, the layered point cloud data is associated and fused with the constructed ternary association nodes and the corresponding neighborhood topology information. Spatial topology information is aggregated through a neighborhood feature search algorithm to generate a canopy structure feature association map for the entire scale of sand-fixing forests. The dimensions, methods, and weights of point cloud feature aggregation are clearly defined. Spatially, the point cloud features of individual trees and stand nodes are aggregated with the topological features within the neighborhood search range. In terms of features, statistical quantities such as the mean, variance, and extreme values ​​of core features like point cloud elevation, density, and interstitial density are aggregated. In terms of aggregation methods, a weighted average method and a feature fusion method are used, assigning point cloud density weights to individual tree nodes and spatial distribution weights to stand nodes. Simultaneously, the association method between topological information and ternary association nodes is clearly defined, and neighborhood topology features are labeled in the association attributes of the corresponding nodes. A neighborhood feature search algorithm (combining K-nearest neighbor search and radius search) is employed to extract spatial topological information of the point cloud within the optimal neighborhood search range of each node. This includes spatial connectivity of the point cloud within the neighborhood, neighborhood distribution relationships of tree entities, topological structure of canopy gaps, and neighborhood connectivity of stand patches. Simultaneously, statistical measures of the neighborhood topological features are calculated to quantify the extraction of topological information. For example, within the 2.5m spherical neighborhood of a single *Salix psammophila* tree, the topological information of the point cloud of three *Caragana korshinskii* trees and two *Salix psammophila* trees in the surrounding area is extracted. The calculated average density of the neighborhood point cloud is 25 points / m², the average canopy gap density is 0.28, and the quantified spatial connectivity of the canopy within the neighborhood is 0.72.

[0052] Using layered point cloud data of individual trees and stands as the base layer, full-scale ternary correlation nodes as the core node layer, and extracted neighborhood spatial topology information as the correlation edge layer, the three layers are fused according to preset point cloud feature aggregation rules: the core node layer labels all elements of the ternary correlation nodes, the correlation edge layer labels the neighborhood topology features and correlation strength between nodes, and the base layer labels the spatial distribution and core features of the point cloud, achieving deep fusion of point cloud data, ternary nodes, and topology information. Spatial topology features are then aggregated on the fused multi-source information. At the individual tree scale, the point cloud features of the target individual tree and the neighborhood topology features are aggregated to form a comprehensive feature of the individual tree canopy structure. At the stand scale, the point cloud features of the target stand unit, the feature contributions of internal individual tree nodes, and the topology features of the neighboring stands are aggregated to form a comprehensive feature of the stand canopy structure. Simultaneously, a quantitative transformation relationship between individual tree features and stand features is established to achieve the linkage aggregation of topology information at both scales.

[0053] Using the planar coordinates of the sand-fixing forest plots as a base map, the aggregated canopy structure features at the individual tree and stand scales are visualized and vectorized. The core elements of the ternary associations are presented as nodes, and the neighborhood topological relationships between nodes are represented by weighted lines. Different colors and sizes are used to label the feature differences and association strengths of the nodes, ultimately generating a full-scale canopy structure feature association map of the sand-fixing forest that includes fine-grained individual tree features, overall stand features, dual-scale association patterns, and spatial topological distribution. For example, in the generated canopy structure feature association map of the mixed sand-fixing forest of *Salix psammophila* and *Caragana korshinskii*, circular nodes represent the ternary association elements of individual trees, with node size representing crown width and color representing tree species; square nodes represent the ternary association elements of the stand, with node color representing canopy closure; and arrowed lines represent the neighborhood relationships between nodes, with line thickness representing association strength. This clearly shows the canopy competition relationships of individual trees within the plot, the patchy distribution patterns of the stand, and the dual-scale structural association features.

[0054] S104, based on the point cloud feature mining neural network, models the canopy structure feature association map, extracts the core features of point cloud elevation, density and intersticity, mines the structural association law between individual tree and stand scale, and generates high-dimensional canopy structure feature embedding vector.

[0055] In one implementation, the constructed canopy structure feature association map is combined and matched with preset point cloud feature mining parameters. A point cloud feature mining neural network modeling algorithm and a multi-scale structure association analysis mechanism are introduced to achieve in-depth extraction and pattern focusing of canopy structure features. Based on the quantitative requirements of sand-fixing forest canopy structure, core mining parameters are preset, including point cloud feature sampling step size (0.1m), neighborhood feature aggregation radius (0.5m), single-tree-stand scale association weight coefficient (single-tree scale 0.6, stand scale 0.4), neural network training iterations (500 times), and convergence threshold (0.001). The physical meaning and value basis of the parameters are clarified to ensure that the parameters are adapted to the point cloud distribution characteristics of sand-fixing forests. A parameter-feature adaptability matching algorithm is used to match the preset mining parameters with the node attributes and topological relationships of the canopy structure feature association map one by one. Corresponding mining parameters are assigned to different types of associated nodes (single-tree nodes / stand nodes) to ensure targeted adaptation between parameters and features. For example, for a single tree node of Salix psammophila, the matching neighborhood feature aggregation radius is 0.4m and the sampling step size is 0.08m; for a forest stand unit node, the matching neighborhood feature aggregation radius is 0.6m and the sampling step size is 0.12m, thus achieving differentiated adaptation of parameters.

[0056] Based on the matched parameters and association graph, a point cloud feature mining GCN model is constructed. Nodes in the association graph are used as network input nodes, node attributes (point cloud features, tree entities, structural parameters) are used as input features, and edge weights (neighborhood association strength) are used as feature propagation weights. Deep fusion of node features and neighborhood topological features is achieved through graph convolution operations to extract hidden structural association features. A cross-scale feature attention mechanism is introduced, setting a single-tree-stand scale attention layer in the GCN model. This automatically calculates the association importance weights of single-tree scale features and stand scale features, strengthening the extraction of key association features, suppressing invalid noise features, and focusing on multi-scale structural association patterns. For example, in the association graph modeling of *Salix psammophila*-*Caragana korshinskii* mixed forests, the attention layer identified the association weight between single-tree crown width and stand canopy closure as 0.78, significantly higher than other association features. The model automatically focuses on this core association pattern and deeply extracts the quantitative relationship between the two.

[0057] By aligning with relevant standards for quantifying the canopy structure of sand-fixing forests and a multi-source lidar point cloud database, a feature extraction and analysis model was constructed. Feature extraction weights were calculated using a neural network feature adaptability calculation model, establishing a dynamic correlation mechanism for point cloud features. The model aligns with industry standards such as the "Forest Ecosystem Monitoring Indicator System" and the "Technical Specification for LiDAR Forest Parameter Extraction," clarifying the extraction requirements for core quantitative indicators of sand-fixing forest canopy structure (such as tree height, crown width, canopy closure, PAI, and canopy gap). Simultaneously, a multi-source lidar point cloud database (containing point cloud datasets of typical sand-fixing forest tree species such as sea buckthorn, sand willow, and *Caragana korshinskii*, covering different growth stages and site conditions) was accessed to provide data support for model construction. Based on the aligned standards and database, a multi-scale feature extraction and analysis model was constructed. This model includes sub-models for single-tree feature extraction, stand feature extraction, and cross-scale correlation, respectively extracting features for single-tree nodes, stand nodes, and dual-scale correlation relationships, ensuring that the extraction results conform to industry standards and are data comparable. For example, the single tree feature extraction sub-model extracts 12 core features such as tree height, crown width, and crown volume for sea buckthorn single tree nodes according to standard requirements; the forest stand feature extraction sub-model extracts 8 core features such as canopy closure, average crown height, and patch density for forest stand unit nodes.

[0058] An adaptive weighted kernel function model (Gaussian kernel function type) is adopted as the neural network feature adaptation calculation model. Based on sample data from a multi-source lidar point cloud database, the extraction weight of each feature is calculated, with weight values ​​ranging from 0 to 1. Higher weights indicate a greater contribution of the feature to the quantification of the sand-fixing forest structure. For example, through model calculation, the weight of the point cloud density feature is 0.92, the weight of the canopy gap feature is 0.87, the weight of the single tree crown width feature is 0.85, and the weight of the stand canopy closure feature is 0.81. These high-weight features will be used as the core extraction objects. Based on the calculated feature extraction weights, a dynamic association mechanism for point cloud features is established. By monitoring the changes in association strength during the feature extraction process in real time, the feature weights and association methods are dynamically adjusted to ensure that the feature association always closely matches the actual structure of the sand-fixing forest. For example, when extracting features of a forest stand, it was found that the correlation strength between point cloud elevation and individual tree height dropped from 0.73 to 0.51. The mechanism automatically triggered a weight adjustment, reducing the feature weight of point cloud elevation from 0.78 to 0.62, while strengthening the correlation between point cloud density and tree height to ensure dynamic adaptation of feature association.

[0059] Using feature correlation strength as the core dimension, a multi-dimensional feature correlation matrix is ​​constructed by integrating core feature data of point cloud elevation, density, and interstitial density, as well as structural feature data at the individual tree scale and the stand scale, to extract canopy structure features. Three types of core feature data are collected and organized, including core feature data of point cloud elevation, density, and interstitial density (such as individual tree point cloud elevation sequences, stand point cloud density distribution data, and spatial distribution data of canopy interstitial density), structural feature data at the individual tree scale, and structural feature data at the stand scale (such as canopy closure, average canopy height, canopy undulation, and patch density). All data are standardized (normalized to the [0,1] interval) to eliminate the influence of dimensional differences. The Pearson correlation coefficient method is used to calculate the correlation strength of linearly correlated features, and the mutual information method is used to calculate the correlation strength of nonlinearly correlated features. The final feature correlation strength value (range 0-1, with values ​​closer to 1 indicating stronger correlation) is obtained by combining the two types of results. For example, the Pearson correlation coefficient between point cloud density and crown width of individual *Salix psammophila* trees was 0.83, and the mutual information value between canopy gap and canopy closure of the stand was 0.79, both of which were determined to be strongly correlated features.

[0060] Using the fused feature data as row vectors and the feature association strength as weights, an N×N feature association matrix (where N is the total number of features) is constructed. Matrix elements represent the association strength values ​​between corresponding features, visually presenting the relationships between all features. For example, for sea buckthorn sand-fixing forest data containing 20 features, a 20×20 feature association matrix is ​​constructed. In this matrix, the element value for "point cloud density - single tree crown width" is 0.83, the element value for "canopy gap density - stand canopy closure" is 0.79, and the element value for "single tree height - average stand crown height" is 0.86, clearly demonstrating the strong associations between core features.

[0061] Based on the multidimensional feature correlation matrix, feature combinations with correlation strength higher than a threshold (set to 0.6) are extracted to form a core feature set of the canopy structure of sand-fixing forests. At the same time, the correlation information between features is preserved, providing a foundation for subsequent high-dimensional vector generation. For example, 15 feature combinations with correlation strength higher than 0.6 are extracted from a 20×20 matrix to form a feature set containing 12 core features such as point cloud density, canopy gap, individual tree crown width, and stand canopy closure, fully covering the key structural information at both the individual tree and stand scales.

[0062] Feature association gaps are addressed by optimizing the feature mining model and strengthening multi-scale structural association features through iterative neural network training. This is then integrated with the results of the feature extraction and analysis model to generate a high-dimensional canopy structure feature embedding vector. Based on the multi-dimensional feature association matrix, a gap identification algorithm (a threshold-based sparsity detection algorithm) is used to identify elements in the matrix with an association strength below 0.3, classifying them as feature association gaps. The causes of these gaps are analyzed and marked. For example, the association strength between "single tree trunk diameter and stand patch density" is identified as 0.21, and the association strength between "point cloud gap density and single tree crown shape coefficient" is 0.27, both identified as feature association gaps. Analysis shows this is caused by occlusion of point cloud data in the understory area. A GAN-based feature association gap completion model is constructed, using complete feature association data as training samples. The generator learns the underlying patterns of feature associations and appropriately completes the identified association gaps, making the feature association matrix more complete and coherent. For example, by using a GAN model to fill in the two missing correlations mentioned above, the correlation strength of "single tree trunk diameter - stand patch density" was filled to 0.38, and the correlation strength of "point cloud gap density - single tree crown shape coefficient" was filled to 0.42. The data after filling in the gaps conformed to the correlation patterns of the structural characteristics of sand-fixing forests and showed no obvious anomalies.

[0063] The completed multidimensional feature association matrix was input into the previously constructed GCN model, and iterative training was performed using the Adam optimizer (learning rate set to 0.001). An early stopping mechanism was introduced (training stopped if the validation set loss did not decrease after 10 consecutive iterations) to prevent overfitting and enhance the extraction effect of multi-scale structural association features. After 320 iterations, the validation set loss stopped decreasing, and training ceased. At this point, the model achieved a 91% accuracy rate in extracting single-tree-stand-scale association features, significantly better than in the early training stages. The feature output of the trained GCN model was fused with the results of the multi-scale feature extraction and analysis model. A feature concatenation and dimensionality compression algorithm (Principal Component Analysis, PCA) was used to compress the high-dimensional feature data to a preset dimension (e.g., 128-dimensional, 256-dimensional), generating the final high-dimensional canopy structure feature embedding vector. This vector contains both core feature information and preserves the association patterns between features. The fusion feature data of the mixed forest of Salix psammophila and Caragana korshinskii was compressed to 256 dimensions using the PCA algorithm to generate a 256-dimensional high-dimensional canopy structure feature embedding vector. Each dimension of the vector corresponds to a fused feature information, which can be directly used for subsequent quantization model construction.

[0064] S105 embeds the canopy structure features into a vector and imports it into a hierarchical quantization calculation engine. It uses a multi-scale parameter adaptive solution mechanism for iterative optimization, introduces a measured data calibration module, and generates a correction decision chain through visualization of structural parameter deviations, thus forming a basic quantitative model of sand-fixing forest structure that integrates ground-based and UAV multi-source data.

[0065] In one implementation, the generated high-dimensional canopy structure feature embedding vector is combined and matched with multi-scale quantitative parameters of sand-fixing forests. A hierarchical quantization calculation algorithm and a multi-source point cloud feature fusion mechanism are introduced to achieve dimensionality calculation and parameter transformation of canopy structure features. Based on the canopy structure index system, multi-scale quantitative parameters are defined for individual trees, stands, and plots. Individual tree scale includes tree height, crown width, crown volume, trunk diameter, and crown shape coefficient; stand scale includes canopy closure, average crown height, canopy undulation, patch density, PAI (leaf area index), and FHD (canopy height diversity); plot scale includes overall canopy coverage, number of vertical layers, and dominant canopy height range. The definitions, dimensions, and quantification ranges of each parameter are clearly defined. A feature-parameter mapping matching algorithm is used to establish a mapping relationship between the feature meaning corresponding to each dimension of the high-dimensional feature embedding vector and the multi-scale quantitative parameters. A corresponding high-dimensional feature dimension combination is assigned to each quantitative parameter to ensure that the matching logic conforms to the correlation between the structural features and parameters of the sand-fixing forest. For example, the 3rd to 8th dimensions (corresponding to point cloud elevation statistics) and the 45th to 52nd dimensions (corresponding to single tree canopy morphology) in the 256-dimensional high-dimensional vector are matched with the "single tree height" parameter; the 102nd to 110th dimensions (corresponding to forest stand point cloud density distribution) and the 180th to 185th dimensions (corresponding to forest stand patch characteristics) are matched with the "forest stand canopy closure" parameter.

[0066] A wavelet transform-based hierarchical decomposition algorithm is introduced to perform hierarchical wavelet decomposition of high-dimensional feature embedding vectors at different scales (individual tree, stand, and plot). Low-frequency approximation coefficients and high-frequency detail coefficients at each scale are extracted, achieving dimensionality reduction and scale separation of high-dimensional features while preserving core feature information strongly correlated with quantization parameters at each scale. For example, for a 256-dimensional high-dimensional vector of a *Salix psammophila*-*Hedysarum heterotropoides* mixed forest, a 3-level decomposition is performed using the db4 wavelet basis function. At the individual tree scale, the third-level low-frequency coefficients (corresponding to core individual tree features) are extracted; at the stand scale, the second-level coefficients (corresponding to stand aggregation features) are extracted; and at the plot scale, the first-level coefficients (corresponding to overall plot features) are extracted, completing the hierarchical decomposition of feature dimensions.

[0067] A weighted feature fusion algorithm, combined with a multi-source point cloud feature fusion mechanism, is employed to assign appropriate weights (based on the correlation strength between features and parameters) to the feature coefficients after hierarchical calculation. Through linear fusion and nonlinear mapping, high-dimensional features are transformed into initial values ​​of corresponding multi-scale quantization parameters for sand-fixing forests. For example, for the "stand PAI" parameter, the stand-scale wavelet coefficients are assigned a weight of 0.7, and the interstitial density features after multi-source point cloud fusion are assigned a weight of 0.3. Through weighted summation and Sigmoid mapping, the feature values ​​are transformed into initial PAI values ​​in the range of 0.8-3.2.

[0068] Aligning with the "Technical Specifications for Forest Parameter Extraction by LiDAR" and the "Technical Standards for Ecological Monitoring of Sand-fixing Forests," the extraction accuracy requirements for each quantitative parameter (e.g., single tree height error ≤ 0.2m, forest stand canopy closure error ≤ 0.05), data format requirements, and quality control standards were clarified. At the same time, the compatibility characteristics of ground-based and UAV multi-source data were analyzed, clarifying that UAV point clouds are suitable for calculating macroscopic parameters at the forest stand scale, and ground-based point clouds are suitable for calculating fine parameters at the single tree scale, thus forming a data-parameter compatibility comparison table.

[0069] A multi-scale parameter calculation model was constructed, comprising sub-models for individual tree parameters, stand parameters, and sample plot parameters. Each sub-model employs a different calculation algorithm: the individual tree parameter calculation model uses a random forest regression algorithm (adapted to the nonlinear relationship between fine-grained features of individual trees and parameters); the stand parameter calculation model uses a gradient boosting decision tree (GBDT) algorithm (adapted to the complex correlation of stand aggregation features); and the sample plot parameter calculation model uses a multiple linear regression algorithm (adapted to the linear expression of overall sample plot features). For example, the individual tree parameter calculation model uses wavelet coefficients, point cloud density, and canopy gap density at the individual tree scale as inputs, and uses measured individual tree height and crown width as labels to train a random forest regression model, achieving a goodness-of-fit R² of 0.93; the stand parameter calculation model uses stand-scale feature coefficients and point cloud patch features as inputs to train a GBDT model to calculate parameters such as canopy closure and PAI.

[0070] A particle swarm optimization-support vector regression (PSO-SVR) model was adopted as the adaptive solution accuracy calculation model. With the goals of multi-source data adaptability and parameter extraction accuracy requirements, the solution weights of each sub-model and input features were calculated. The weight values ​​ranged from 0 to 1, with higher weights indicating a greater contribution to solution accuracy. For example, through PSO-SVR model calculation, in the single-tree parameter solution sub-model, the weight of ground-based point cloud features was 0.65, and the weight of UAV point cloud features was 0.35; in the stand parameter solution sub-model, the weight of UAV point cloud features was 0.72, and the weight of ground-based point cloud features was 0.28; in the sample plot parameter solution sub-model, the weight of dual-source point cloud fusion features was 0.80, ensuring that the solution weights adapted to data characteristics and accuracy requirements.

[0071] Based on the calculation weights and multi-source data quality feedback, a dynamic calculation mechanism for canopy structure parameters is established. This mechanism monitors data quality (such as point cloud density and completeness) and calculation errors in real time, and dynamically adjusts the calculation model parameters and feature weights. For example, when the UAV point cloud completeness drops from 95% to 82%, the dynamic calculation mechanism automatically reduces the UAV feature weight in the stand parameter calculation sub-model from 0.72 to 0.58, while increasing the ground-based point cloud feature weight to 0.42, ensuring stable calculation accuracy.

[0072] Using parameter solution fit as the core dimension, this study integrates individual tree-stand-plot scale feature data, multi-source point cloud fusion feature data, and preliminary structural parameter calculation data to construct a multi-dimensional parameter solution correlation matrix. Three types of core data were collected and organized: individual tree-stand-plot scale feature data (feature coefficients after wavelet layered solution), multi-source point cloud fusion feature data (density, gaps, and elevation features after ground-based and UAV point cloud fusion), and preliminary structural parameter calculation data (preliminary parameter values ​​after high-dimensional vector transformation). All data were standardized to eliminate differences in dimensions and numerical ranges. The cosine similarity method was used to calculate the linear fit between the feature data and the preliminary parameter calculation data, while the mean square error method was used to calculate the consistency fit of different data sources for the same parameter. The final parameter solution fit (range 0-1, with values ​​closer to 1 indicating higher fit and more reliable results) was obtained by combining the two sets of results. The cosine similarity between the calculated and measured values ​​of the ground point cloud features for calculating the height of a single tree is 0.94, the mean square error is 0.03, and the overall fit is 0.92. The cosine similarity between the calculated and measured values ​​of the UAV point cloud features is 0.87, the mean square error is 0.08, and the overall fit is 0.85.

[0073] Using multi-scale quantization parameters as rows and multi-source solution data as columns, an M×N dimensional parameter solution correlation matrix is ​​constructed (M is the number of parameters, N is the number of data types). The matrix elements represent the solution fit values ​​between the corresponding parameters and the data, intuitively presenting the support capability of each data for solving different parameters. For the correlation matrix containing 15 core parameters and 6 types of data sources, the fit element value for "stand canopy closure - UAV fusion feature" is 0.91, the fit element value for "single tree crown width - ground point cloud feature" is 0.89, and the fit element value for "number of vertical stratification layers in sample plots - dual-source fusion feature" is 0.86, demonstrating the adaptation relationship between data and parameters.

[0074] The parameter calculation gaps are addressed through a parameter calculation deviation correction model, and the accuracy of structural parameter quantification is enhanced by combining measured data calibration mechanisms. This is integrated with the results of multi-scale parameter calculation models, and a correction decision chain is generated through structural parameter deviation visualization, forming a basic quantitative model of sand-fixing forest structure that integrates ground-based and UAV multi-source data. Based on the multi-dimensional parameter calculation correlation matrix, a gap identification algorithm (threshold-based sparsity detection) is used to identify matrix elements with a fit score below 0.6, which are identified as parameter calculation gaps. The causes of these gaps (such as missing data, insufficient features, and poor algorithm adaptability) are analyzed and marked. For example, the fit score of "single tree trunk diameter - UAV point cloud features" is identified as 0.53, and the fit score of "forest stand FHD - ground-based point cloud features" is 0.57, which are identified as calculation gaps. Analysis shows that the UAV point cloud lacks detailed features of single tree trunk diameter, and the ground-based point cloud lacks aggregated features of forest stand height diversity.

[0075] A Bayesian estimation bias correction model was constructed, using complete fit data from the correlation matrix as prior information. Bayesian inference was used to estimate the solution values ​​and biases at the gaps, correcting the initial parameter values ​​and filling the gaps. For example, for the gap in the "single tree trunk diameter - UAV point cloud characteristics" calculation, using the correlation between single tree height, crown width, and trunk diameter as prior information, Bayesian estimation corrected the initial trunk diameter value from 5.2cm to 4.8cm, reducing the bias from 0.9cm to 0.3cm. For the "stand FHD" gap, the canopy height distribution characteristics from UAV point clouds were integrated, increasing the corrected FHD value from 1.2 to 1.5, which better matches the measured results.

[0076] A K-fold cross-validation calibration method (K=10) was used to construct a calibration mechanism for the measured data. The measured parameter data from the sample plots (such as individual tree height, crown width, stand canopy closure, and PAI measured in the field) were divided into training and validation sets at a ratio of 10:1. The multi-scale parameter calculation model was iteratively calibrated, adjusting model parameters and calculation weights to reduce calculation errors. For example, for 100 measured samples from a seabuckthorn sand-fixing forest, a 10-fold cross-validation was performed. Before calibration, the average error in calculating individual tree height was 0.32 m, which decreased to 0.18 m after calibration; the average error in stand canopy closure was 0.07 before calibration, which decreased to 0.04 after calibration, meeting the standard accuracy requirements.

[0077] A heatmap visualization method is used to present the distribution of solution deviations for each parameter, and a line graph visualization method is used to present the trend of deviation changes with data quality and model parameters. Based on the visualization results, a correction decision chain is generated to clarify the correction strategies under different deviation scenarios (e.g., when the deviation is >0.2m, priority is given to supplementing ground point cloud data; when the deviation is 0.1-0.2m, the model solution weights are adjusted). For example, when generating a heatmap of single tree height solution deviation, it is found that the deviation in the edge area of ​​the sample plot is concentrated in 0.2-0.3m, and the decision chain triggers the correction strategy of "supplementing ground point cloud collection in the edge area + adjusting the number of decision trees in the random forest model"; when the canopy closure deviation is concentrated in the sparse area, the decision chain triggers the correction strategy of "increasing the weight of the gap feature + optimizing the learning rate of the GBDT model".

[0078] The corrected parameter calculation results are integrated with the multi-scale parameter calculation model, dynamic calculation mechanism, and correction decision chain to form a basic quantitative model of sand-fixing forest structure that integrates ground-based and UAV multi-source data. The model includes a parameter calculation module, a deviation correction module, a field measurement calibration module, and a decision output module. It can directly input multi-source point cloud data and output standardized quantitative results of sand-fixing forest structure. For example, the final constructed basic quantitative model of Salix psammophila-Caragana korshinskii mixed forest structure can automatically output 18 quantitative parameters after inputting UAV and ground-based fused point cloud data, such as individual tree height, crown width, trunk diameter, stand canopy closure, PAI, FHD, plot coverage, and number of vertical layers. The average parameter calculation error is ≤0.05 (dimensionless parameters) or ≤0.2m (length-type parameters), which meets the requirements for quantitative analysis of sand-fixing forest structure.

[0079] S106, based on a quantitative basic model combined with field measurement verification data and forest stand growth dynamic monitoring data, performs accuracy verification and dynamic correction. Through a multi-index comprehensive evaluation algorithm, it balances the requirements of parameter accuracy, computational efficiency, and data adaptability, and generates quantitative results of canopy spatial structure that are suitable for monitoring all scenarios of sand-fixing forests.

[0080] In one implementation, field-measured canopy structure parameter data are collected within the sample plot, covering the individual tree scale (tree height, crown width, trunk diameter, etc., with 30-50 trees randomly selected from each species for measurement), the stand scale (canopy closure, PAI, average crown height, etc., with 3-5 replicates per stand unit), and the sample plot scale (overall coverage, number of vertical stratifications, etc., with full plot measurement for verification), ensuring the representativeness and accuracy of the measured data. For example, in a mixed forest sample plot of Salix psammophila and Caragana korshinskii, the tree height (range 2.3-4.8m) and crown width (range 1.2-2.5m) of 216 Salix psammophila trees, the corresponding parameters of 189 Caragana korshinskii trees, and the canopy closure (range 0.4-0.75) and PAI (range 1.2-2.8) of 25 stand units are measured as verification data. The root mean square error (RMSE) is used to calculate the overall error between the quantified and measured values ​​of the parameters, the mean absolute error (MAE) is used to calculate the average deviation, and the coefficient of determination (R²) is used to evaluate the goodness of fit between the quantified and measured values. These three methods are used together to comprehensively verify the accuracy. For example, in verifying the quantified data of a single Salix psammophila tree, the calculated RMSE = 0.19m, MAE = 0.15m, and R² = 0.94, indicating a high good fit between the quantified and measured values. In verifying the quantified data of canopy closure, the RMSE = 0.04, MAE = 0.03, and R² = 0.92, meeting the accuracy requirements.

[0081] A 10-fold cross-validation mechanism was introduced, dividing the measured data into training and validation sets at a 10:1 ratio. This allowed for iterative verification of the stability and generalization ability of the quantification model, identifying the error distribution patterns across different structural types and terrain regions. For example, cross-validation revealed that the RMSE (Real-Time Sequence) of the quantified tree crown width in the edge areas of the sample plots was 0.28m, significantly higher than that in the core areas (RMSE = 0.16m). This indicated insufficient model adaptability in the edge areas, and these areas were marked as key areas for correction.

[0082] Precision level standards are set (R²≥0.9 is Level 1 precision, 0.8≤R²<0.9 is Level 2 precision, 0.7≤R²<0.8 is Level 3 precision, and R²<0.7 is Level 4 precision). Each quantified parameter is labeled with its corresponding precision level and error index, generating quantitative data of forest canopy structure with precision labels. For example, the height of a single *Salix psammophila* tree is labeled as "Level 1 precision (RMSE=0.19m, R²=0.94)," and the canopy width of the stand edge area is labeled as "Level 3 precision (RMSE=0.28m, R²=0.76)."

[0083] Combining the monitoring and evaluation standards for sand-fixing forests with a multi-index comprehensive evaluation strategy, the quantitative data of forest canopy structure with precision markers were processed to complete a multi-dimensional evaluation of parameter accuracy, computational efficiency, and data fit, generating preliminary quantitative results for data optimization. Three core evaluation dimensions were defined: ① Parameter accuracy: RMSE, MAE, and R² were used as core indicators; ② Computational efficiency: single-parameter solution time and total processing time were used as core indicators (single-parameter solution ≤ 0.5s, total processing time ≤ 30min); ③ Data fit: multi-source point cloud data utilization rate and fit rate for different sand-fixing forest community types were used as core indicators (data utilization rate ≥ 85%, community fit rate ≥ 90%). Weights and evaluation thresholds were assigned to each indicator.

[0084] The Analytic Hierarchy Process (AHP) was used to subjectively determine the weights of each dimension (accuracy 0.5, computational efficiency 0.2, data fit 0.3). The entropy weight method was then used to objectively adjust the weights. Combining the measured values ​​of each indicator with the evaluation thresholds, a comprehensive evaluation score (0-100 points) was calculated. A score ≥85 was excellent, 70-84 was good, 60-69 was satisfactory, and <60 was unsatisfactory. For example, the comprehensive evaluation score of the quantitative data on canopy closure of a pure sea buckthorn forest was 88 points (accuracy 45 points, efficiency 18 points, fit 25 points), which was considered excellent; the quantitative data on the single-tree trunk diameter of *Caragana korshinskii* scored 65 points (accuracy 30 points, efficiency 18 points, fit 17 points), which was considered satisfactory, but targeted optimization is needed.

[0085] For quantitative data that fails the comprehensive evaluation or has shortcomings, optimization strategies are formulated (if accuracy is insufficient, model parameters are corrected; if efficiency is too low, the algorithm process is optimized; if adaptability is poor, data features are supplemented) to make preliminary adjustments to the quantitative data and generate preliminary optimized quantitative results. For example, if the individual trunk diameter of *Hedysarum heterotropoides* is not well adapted (community adaptability rate of 82%), the trunk point cloud texture features of *Hedysarum heterotropoides* are supplemented, and after recalculation, the adaptability rate is improved to 93%, generating preliminary optimized trunk diameter quantitative data.

[0086] Based on the requirements of multi-indicator comprehensive evaluation, accuracy verification information, and the generation rules of sand-fixing forest monitoring specifications, the preliminary quantitative results were systematically organized to optimize the data and dynamically correct the quantitative results of canopy spatial structure. The K-means clustering algorithm was used to classify the preliminary optimized data according to sand-fixing forest community type (pure forest / mixed forest) and terrain type (flat land / steep slope). Then, the 3σ criterion outlier removal algorithm was used to remove extreme outliers (such as quantified values ​​exceeding the mean ± 3 standard deviations) from the classified data to ensure data homogeneity. For example, the quantitative data of *Salix psammophila*-*Caragana korshinskii* mixed forest was clustered into two categories: flat land (slope < 10°) and steep slope (slope ≥ 10°). The 3σ criterion was used to remove outliers in the steep slope area with a PAI quantified value > 4.0 (mean 2.5, standard deviation 0.5). A Kalman filter dynamic correction model was constructed, using the classified and organized data as the state vector, the error covariance obtained from accuracy verification as the system noise, and the measured data as the observation vector. Through a prediction-update iterative process, the quantified data was dynamically corrected to reduce cumulative errors. For example, the PAI quantified data of steep slope forest stands was dynamically corrected. After 5 iterations, the RMSE decreased from 0.32 to 0.21, and the corrected PAI value better matched the measured value and the actual growth state of the sand-fixing forest.

[0087] Based on the dynamic characteristics of forest stand growth and the correlation logic between quantification results and accuracy, core correction parameters are automatically generated. Anomalies are identified and parameters are adjusted based on a multi-objective balance detection mechanism. A grey relational analysis algorithm is used to identify core factors affecting quantification accuracy (such as point cloud density, terrain slope, and tree species type). A regression model between these core factors and quantification error is established using a multiple regression algorithm, automatically generating core correction parameters (such as point cloud density correction coefficient and terrain slope compensation coefficient). For example, analysis revealed that for every 10 points / m² decrease in point cloud density, the quantification error for single tree height increases by 0.05m, resulting in a point cloud density correction coefficient (when density < 20 points / m², correction coefficient = 1.0 + (20 - actual density) × 0.005).

[0088] A non-dominated sorting genetic algorithm (NSGA-II) was employed, with parameter accuracy, computational efficiency, and data fit as multi-objective optimization goals. The algorithm detected optimization anomalies in the quantification results (such as excessive pursuit of accuracy leading to low efficiency, or sacrificing accuracy for improved fit). Core correction parameters were adjusted for anomalies. For example, the detection revealed that the quantification data for the vertical stratification of the *Salix psammophila*-*Caragana korshinskii* mixed forest showed an RMSE increase to 0.35m (exceeding the threshold of 0.3m) due to efforts to improve fit. This was marked as an optimization anomaly, and the terrain compensation coefficient was adjusted from 1.2 to 1.05. After correction, the RMSE was 0.27m, maintaining a fit rate of 91%, thus achieving a balance among the multi-objective goals.

[0089] Three core rules are established: ① Data format: The output is uniformly in LAS format point cloud with parameter attribute table and CSV format quantitative parameter table, supporting direct import into GIS software; ② Parameter classification: Quantitative parameters are classified and labeled according to the health level and sand fixation efficiency level of the sand-fixing forest (e.g., canopy closure of 0.6-0.8 is the suitable level); ③ Result visualization: Multi-scale visualization charts of individual trees, stands, and plots are generated (e.g., canopy height heat map, stand structure profile). A weighted fusion algorithm is adopted to weight and fuse the dynamically corrected quantitative data (weight 0.6), core correction parameters (weight 0.2), and model basic parameters (weight 0.2) to ensure information complementarity; the data format, field naming, and parameter units (length unit m, density unit points / m², dimensionless parameters are retained to 2 decimal places) are unified through a format standardization algorithm. For example, the tree height quantification data (corrected mean 3.2m), point cloud density correction coefficient (1.08), and basic model solution parameters of the sea buckthorn pure forest plot are fused together to output the tree height quantification result (3.2m, first-level accuracy, suitable level).

[0090] The integrated output includes multi-scale quantitative parameters, precision indicators, hierarchical annotations, and visualization charts, providing comprehensive quantitative results for various monitoring needs. For example, the output of comprehensive quantitative results for a mixed forest of Salix psammophila and Caragana korshinskii includes detailed parameters for individual trees (tree height, crown width, etc., with precision markers) of 216 Salix psammophila and 189 Caragana korshinskii trees, aggregated parameters for 25 forest stand units (canopy closure, PAI, etc., with hierarchical annotations), overall plot structure parameters (coverage 82%, vertical stratification 3 layers), and a 3D visualization model of the canopy and a heat map of structural differences, making it suitable for comprehensive applications such as daily monitoring, performance evaluation, and management planning.

[0091] In one implementation, such as Figure 2 As shown, this application also provides a device for quantifying the spatial structure of sand-fixing forest canopies based on ground-based and UAV lidar, comprising:

[0092] The multi-source core data acquisition module 201 for sand-fixing forests is used to acquire multi-source core data for the entire process of quantifying the spatial structure of the canopy of sand-fixing forests, including ground-based lidar point cloud data, UAV lidar point cloud data, field measurement and verification data of sample plots, tree species classification archives of sand-fixing forests, and basic topographic and geomorphological data of the operation area.

[0093] The sand-fixing forest integrated point cloud dataset generation module 202 is used to perform noise reduction, filtering and resampling processing on ground-based and UAV lidar point cloud data based on point cloud preprocessing algorithms, complete the accurate registration of point clouds on multiple platforms by combining coordinate transformation models, distinguish different sand-fixing forest community types through tree species feature recognition models, integrate topographic data into the point cloud correction process, and generate a standardized and integrated sand-fixing forest integrated point cloud dataset.

[0094] The module 203 for constructing a full-scale canopy structure feature association map of sand-fixing forests is used to layer the fused point cloud dataset by stand scale and single tree scale. Each layer constructs a ternary association node including point cloud features, tree entities and structural parameters. The spatial topology information of the point cloud is aggregated through the neighborhood feature search algorithm to generate a canopy structure feature association map for the full scale of sand-fixing forests.

[0095] The high-dimensional feature embedding vector generation module 204 for canopy structure is used to model the canopy structure feature association map based on the point cloud feature mining neural network, extract the core features of point cloud elevation, density and intersticity, mine the structural association law between individual tree and stand scale, and generate high-dimensional canopy structure feature embedding vector.

[0096] The sand-fixing forest structure quantitative basic model construction module 205 is used to import the canopy structure features into the hierarchical quantitative calculation engine, adopt a multi-scale parameter adaptive solution mechanism for iterative optimization, introduce a measured data calibration module, and generate a correction decision chain through structural parameter deviation visualization to form a sand-fixing forest structure quantitative basic model that integrates ground-based and UAV multi-source data.

[0097] The module 206 for generating quantitative results of canopy spatial structure of sand-fixing forests is used to verify accuracy and dynamically correct based on the quantitative basic model combined with field measurement verification data and forest stand growth dynamic monitoring data. It balances the requirements of parameter accuracy, computational efficiency and data adaptability through a multi-index comprehensive evaluation algorithm to generate quantitative results of canopy spatial structure that are suitable for the entire monitoring scenario of sand-fixing forests.

[0098] The computer-readable storage medium provided in the above embodiments of this application and the method for quantifying the spatial structure of sand-fixing forest canopy based on ground-based and UAV lidar provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0099] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of the method, system, electronic device, and readable storage medium for evaluating the spatial structure of sand-fixing forest canopies based on ground-based and UAV lidar are basically similar to the embodiments of the method for quantifying the spatial structure of sand-fixing forest canopies based on ground-based and UAV lidar, so the description is relatively simple. Relevant parts can be referred to in the description of the embodiments of the method for quantifying the spatial structure of sand-fixing forest canopies based on ground-based and UAV lidar.

Claims

1. A method for quantifying the spatial structure of the forest canopy of a sand-fixing forest based on ground-based and unmanned aerial vehicle LiDAR, characterized in that, include: Acquire multi-source core data on the entire process of quantifying the spatial structure of sand-fixing forest canopies; Based on the point cloud preprocessing algorithm, the point cloud data of ground-based and UAV lidar are denoised, filtered and resampled. Combined with the coordinate transformation model, the point cloud is accurately registered on multiple platforms. Different sand-fixing forest community types are distinguished by the tree species feature recognition model. The topographic data is integrated into the point cloud correction process to generate a standardized and integrated sand-fixing forest fusion point cloud dataset. The fused point cloud dataset is layered according to stand scale and single tree scale. Each layer constructs a ternary association node including point cloud features, tree entities and structural parameters. The spatial topology information of the point cloud is aggregated by the neighborhood feature search algorithm to generate a canopy structure feature association map for the whole scale of sand-fixing forest. Based on the point cloud feature mining neural network, the canopy structure feature association map is modeled, the core features of point cloud elevation, density and intersticity are extracted, the structural association rules between individual trees and stand scale are explored, and a high-dimensional canopy structure feature embedding vector is generated. The canopy structure features are embedded into a vector and imported into a hierarchical quantization calculation engine. A multi-scale parameter adaptive solution mechanism is used for iterative optimization. A measured data calibration module is introduced. At the same time, a correction decision chain is generated by visualizing the structural parameter deviation, forming a basic model for the quantification of sand-fixing forest structure that integrates ground-based and UAV multi-source data. Based on the quantitative basic model, combined with the field measurement verification data of sample plots and the forest stand growth dynamic monitoring data, the accuracy is verified and dynamically corrected. Through the multi-index comprehensive evaluation algorithm, the requirements of parameter accuracy, calculation efficiency and data adaptability are balanced, and the quantitative results of the canopy spatial structure that are suitable for the whole scenario of sand-fixing forest monitoring are generated.

2. The method as described in claim 1, characterized in that, The fused point cloud dataset is layered according to stand and individual tree scales. Each layer constructs a ternary association node comprising point cloud features, tree entities, and structural parameters. A neighborhood feature search algorithm is used to aggregate the spatial topology information of the point cloud, generating a canopy structure feature association map for the entire scale of sand-fixing forests, including: Combining the characteristics of the fused point cloud dataset of sand-fixing forests with the multi-scale requirement of canopy structure quantification, we introduce the single-tree-stand scale division rule and the point cloud spatial feature parsing mechanism to perform scale layering processing on the fused point cloud dataset, generating the dual-scale point cloud layering results of sand-fixing forest stands and single trees. By combining the requirements for multi-scale structural parameter extraction with the logic of point cloud feature association, the layered results of dual-scale point cloud are processed to complete the construction of ternary association nodes for point cloud features, forest entities, and structural parameters, and to generate scaled node association matching results. Based on the spatial topological characteristics of sand-fixing forests, the distribution patterns of multi-scale point clouds, and the logical generation rules for node association, the scaled node association matching results are systematically verified to complete the construction of full-scale ternary association nodes. Based on the spatial distribution characteristics of the fused point cloud and the neighborhood association relationship of the forest entity, the corresponding neighborhood search range is automatically matched, and the matching anomalies are marked and the range is corrected based on the point cloud feature-topology information adaptability detection mechanism. According to the preset point cloud feature aggregation rules, the layered point cloud data is associated and fused with the constructed ternary association nodes and the corresponding range of neighborhood topology information. The spatial topology information is aggregated through the neighborhood feature search algorithm to generate a canopy structure feature association map for the whole scale of sand-fixing forests.

3. The method as described in claim 2, characterized in that, Based on a point cloud feature mining neural network, a model of the canopy structure feature association map is constructed. Core features such as point cloud elevation, density, and intersticity are extracted, and structural association patterns between individual trees and stands are explored. This generates a high-dimensional canopy structure feature embedding vector, including: The constructed canopy structure feature association map is combined and matched with the preset point cloud feature mining parameters. A point cloud feature mining neural network modeling algorithm and a multi-scale structure association analysis mechanism are introduced to achieve in-depth extraction and pattern focusing of canopy structure features. By connecting with relevant standards for the quantitative analysis of the canopy structure of sand-fixing forests and a multi-source lidar point cloud database, a feature extraction and analysis model is constructed. The feature extraction weights are calculated through a neural network feature adaptability calculation model, and a dynamic correlation mechanism for point cloud features is established. Using feature correlation strength as the core dimension, a multi-dimensional feature correlation matrix is ​​constructed by integrating core feature data of point cloud elevation, density, and interstitial density, as well as structural feature data of individual trees and structural feature data of forest stands, to achieve the extraction of canopy structure features. By optimizing the feature mining model to address feature association gaps, and combining iterative training mechanisms of neural networks to strengthen multi-scale structural association features, the results of the feature extraction and analysis model are integrated to generate high-dimensional canopy structure feature embedding vectors.

4. The method as described in claim 1, characterized in that, The canopy structure features are embedded into a vector and imported into a hierarchical quantization calculation engine. A multi-scale parameter adaptive solution mechanism is used for iterative optimization. A measured data calibration module is introduced, and a correction decision chain is generated through visualization of structural parameter deviations. This forms a basic quantitative model of sand-fixing forest structure that integrates ground-based and UAV multi-source data, including: The generated high-dimensional canopy structure feature embedding vector is combined and matched with the multi-scale quantization parameters of sand-fixing forest. A hierarchical quantization calculation algorithm and a multi-source point cloud feature fusion mechanism are introduced to realize the dimensionality calculation and parameter transformation of canopy structure features. To align the quantitative standards for canopy structure of sand-fixing forests with the requirements for adapting ground-based and UAV multi-source data, a multi-scale parameter calculation model was constructed. The parameter calculation weights were calculated through an adaptive calculation accuracy calculation model, and a dynamic calculation mechanism for canopy structure parameters was established. Using parameter solution fit as the core dimension, a multi-dimensional parameter solution correlation matrix is ​​constructed by integrating single-tree-stand-plot scale feature data, multi-source point cloud fusion feature data, and preliminary structural parameter calculation data. The parameter calculation gap is addressed by using a parameter calculation deviation correction model, and the accuracy of structural parameter quantification is enhanced by combining measured data calibration mechanism. The results are then integrated with the results of the multi-scale parameter calculation model, and a correction decision chain is generated through visualization of structural parameter deviations. This forms a basic quantitative model of sand-fixing forest structure that integrates ground-based and UAV multi-source data.

5. The method as described in claim 4, characterized in that, Based on a quantitative fundamental model combined with field measurement data from sample plots and forest stand growth dynamic monitoring data, accuracy verification and dynamic correction are performed. A multi-index comprehensive evaluation algorithm balances parameter accuracy, computational efficiency, and data adaptability requirements, generating quantitative results of canopy spatial structure suitable for all scenarios of sand-fixing forest monitoring, including: By combining the basic model of the quantitative structure of sand-fixing forests with the field measurement verification data, an accuracy verification algorithm and a dynamic correction and verification mechanism are introduced to perform error analysis and accuracy verification on the preliminary calculation data of the canopy structure, and generate quantitative data of the canopy structure with accuracy markers. By combining the monitoring and evaluation standards for sand-fixing forests with a multi-index comprehensive evaluation strategy, the quantitative data of forest canopy structure with precision markers are processed to complete the multi-dimensional evaluation of parameter accuracy, calculation efficiency, and data adaptability, and generate preliminary quantitative results to optimize the data. Based on the requirements of multi-indicator comprehensive evaluation, accuracy verification information and sand-fixing forest monitoring standard generation rules, the preliminary quantitative results optimization data are systematically organized to complete the dynamic correction of the quantitative results of canopy spatial structure. Based on the dynamic characteristics of forest stand growth and the correlation logic of quantitative results, core correction parameters are automatically generated, and optimization anomalies are marked and parameters are adjusted based on a multi-objective balance detection mechanism. According to the preset quantitative result generation rules, the dynamically corrected accuracy optimization information is linked and integrated with the quantitative basic model and core correction parameters to generate quantitative results of the canopy spatial structure that are suitable for the entire monitoring scenario of sand-fixing forests.

6. A system for quantifying the spatial structure of sand-fixing forest canopies based on ground-based and UAV lidar, characterized in that, The system is configured to execute the method for quantifying the spatial structure of sand-fixing forest canopy based on ground-based and UAV lidar, as described in any one of claims 1 to 5, by executing executable instructions.

7. An electronic device, characterized in that, include: First processor; and memory for storing executable instructions of the first processor; The first processor is configured to execute the method for quantifying the spatial structure of sand-fixing forest canopy based on ground-based and UAV lidar as described in any one of claims 1 to 5 by executing the executable instructions.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the method for quantifying the spatial structure of sand-fixing forest canopy based on ground-based and UAV lidar as described in any one of claims 1 to 5.

Images