A forest single-piece digital twin data fusion processing technology and method

Abstract

This invention belongs to the field of digital information technology, specifically a technology and method for fusion and processing digital twin data of individual trees. Addressing the problems of existing methods that struggle to balance individual tree accuracy with regional coverage, poor data fusion adaptability, inaccurate growth model predictions, and a lack of dynamic twin monitoring capabilities, this invention proposes the following solution: It utilizes dual-source data collaboration between satellite LiDAR and UAV LiDAR, combined with digital twin technology to achieve accurate inversion of individual tree volume and construction of dynamic growth models. Dual-source adaptive registration and enhancement algorithms balance UAV individual tree accuracy with satellite regional coverage capabilities, resolving the contradiction between accuracy and range. A "binary volumetric + random forest" model couples the environment and competing factors, improving accuracy. Incremental fusion and dynamic growth models enable efficient twin updates and growth prediction, supporting full life-cycle monitoring and significantly improving the intelligence and efficiency of forestry resource management.

Description

Technical Field

[0001] This invention relates to the field of digital information technology, and in particular to a technology and method for fusion processing of digital twin data of individual trees. Background Technology

[0002] Monitoring the volume of individual trees and predicting their growth are core requirements for forestry resource management and carbon sequestration. LiDAR technology, with its high-precision three-dimensional imaging capabilities, has been widely applied in this field. However, existing technologies still have significant limitations, as detailed below:

[0003] For example, Chinese Patent Application No. 201811224238.7 discloses a forest information measurement system and data processing method based on a two-dimensional laser scanner, including a processing system, a ground measurement system, and an unmanned aerial vehicle (UAV) measurement system. The system includes: acquiring GPS data, IMU data, laser scan data, and time information from the ground measurement system and the UAV measurement system; acquiring GPS data from a GPS reference base station; converting all GPS data to GPB format and processing the converted GPS data to obtain differential global positioning system (GPS) data; integrating the GPS data, IMU data, and laser scan data into the same coordinate system; using a coordinate rotation matrix to match all laser scan data to UTM coordinates, establishing three-dimensional point cloud coordinate information; and thus obtaining complete three-dimensional point cloud information of forest trees. This invention can adapt to the point cloud information measurement of tall trees and trees in forests with high canopy density, greatly saving manpower and achieving high measurement accuracy.

[0004] Another example is single-drone LiDAR monitoring technology. This technology has high accuracy at the tree-scale, but its coverage is limited, data collection costs are high, it cannot achieve efficient monitoring of regional forest stands, and it lacks support for spatial trends at the stand scale, making the volume inversion susceptible to local environmental interference.

[0005] Another example is single-satellite LiDAR monitoring technology. It has a wide coverage area and can quickly obtain macroscopic parameters of forest stands, but the point cloud is sparse and the resolution of individual trees is low, making it difficult to accurately segment individual trees and extract microscopic parameters such as diameter at breast height. The error in the inversion of individual tree volume usually exceeds 20%.

[0006] Traditional LiDAR data fusion and volume inversion methods. Existing fusion methods are mostly simple data overlays, lacking adaptive registration and enhancement algorithms, and exhibiting poor synergy between two-source data. Volume inversion methods often rely on single volumetric or machine learning models, failing to couple environmental and competing factors, resulting in low prediction accuracy of growth models and the inability to construct dynamic digital twins for full lifecycle monitoring.

[0007] In summary, existing technologies suffer from technical problems such as difficulty in achieving both single-tree accuracy and regional coverage, poor data fusion adaptability, inaccurate growth model predictions, and lack of dynamic twin monitoring capabilities. There is an urgent need for an integrated technical solution that can achieve efficient collaboration, accurate inversion, and dynamic simulation of dual-source LiDAR. Summary of the Invention

[0008] Given the problems in the background technology, such as difficulty in achieving both single-tree accuracy and regional coverage, poor data fusion adaptability, inaccurate growth model prediction, and lack of dynamic twin monitoring capabilities, this invention proposes a digital twin data fusion processing technology and method for single-tree forest plots.

[0009] This invention proposes a digital twin data fusion processing technology and method for individual forest trees. Based on the collaborative use of dual-source data from satellite LiDAR and UAV LiDAR, and combined with digital twin technology, it achieves accurate inversion of individual tree volume and construction of a dynamic growth model. The entire process sequentially includes six major stages: data acquisition and adaptation, dual-source data fusion preprocessing, individual tree segmentation and structural parameter extraction, digital twin construction and volume inversion, time-series growth model construction, and dynamic updating of the twin. Specifically, it includes the following steps:

[0010] S1: Data Acquisition and Adaptation: Using UAV LiDAR as the core data source at the single-tree scale and satellite LiDAR as the auxiliary data source at the stand scale, ground-based measured data is collected simultaneously as a calibration and verification benchmark. Spatiotemporal adaptation rules between dual-source LiDAR data and ground-based measured data are established to unify data benchmarks and formats.

[0011] S2: Dual-source data fusion preprocessing: Noise filtering, ground point removal, and data augmentation are performed on the dual-source LiDAR data respectively. Spatiotemporal alignment of the dual-source data is achieved through coarse and fine registration. Then, the data is standardized to provide a foundation for subsequent fusion analysis.

[0012] S3: Tree segmentation and structural parameter extraction: Based on UAV LiDAR data, a multi-algorithm fusion strategy is used to achieve accurate tree segmentation. Key structural parameters and auxiliary parameters such as tree height, diameter at breast height, and crown width are extracted from the segmented tree point cloud clusters.

[0013] S4: Digital Twin Construction and Volume Inversion: Construct a three-in-one digital twin of a single tree, consisting of "geometric model + attribute information + physical rules". Use a model that combines traditional volumetric data with machine learning algorithms to invert the volume of a single tree and calibrate it with ground-based measured data.

[0014] S5: Construction of time-series growth model: Based on multi-period dual-source LiDAR data and ground-measured time-series data, cross-period single-tree tracking is realized, and a dynamic growth model is constructed by coupling environmental factors and competitive factors to characterize the growth law of single-tree volume.

[0015] S6: Dynamic Update and Application of Digital Twins: An incremental fusion algorithm is used to dynamically update digital twins of individual trees. Combined with a growth model to predict individual tree growth trends, satellite LiDAR data is used to extrapolate small-area models over a large area, constructing an integrated digital twin application system. Through dual-source LiDAR collaboration and digital twin fusion technology, four core beneficial effects are achieved: First, the dual-source adaptive registration and enhancement algorithm balances the accuracy of individual trees from UAVs with the coverage capability of satellite regions, achieving a registration error ≤0.03m and resolving the contradiction between accuracy and range. Second, the "binary volumetric + random forest" model couples the environment and competing factors, achieving a volume inversion error ≤8%, improving accuracy by 10-15 percentage points compared to traditional methods. Third, incremental fusion and dynamic growth models enable efficient twin updates and growth prediction, supporting full life-cycle monitoring. Fourth, the integrated system adapts to multiple forestry scenarios, accommodating needs such as precision tending and carbon sequestration, significantly improving the intelligence and efficiency of forestry resource management.

[0016] Preferably, in S1, the UAV LiDAR uses a 905nm or 1550nm laser wavelength, a scanning frequency of 100-200kHz, a point cloud echo count of ≥3 times, and a point cloud density controlled at 20-50 points / ㎡. The density is adaptively adjusted according to the size of the single tree crown, with 30-50 points / ㎡ when the crown width is ≤2m. The flight altitude is 50-100m, and the overlap rate of the flight direction and the side direction is ≥80%. The matching RTK / PPK positioning module achieves a planar accuracy of ±3cm and a vertical accuracy of ±5cm.

[0017] The 1550nm wavelength was prioritized to enhance canopy penetration and reduce interference from foliage shading. The trunk point cloud acquisition rate was ≥85%. Satellite LiDAR data were selected from either GEDI Level 2A or ICESat-2 ATL 08. The effective spot signal-to-noise ratio of GEDI Level 2A was ≥3, and the spot diameter was ~25m. The ground spot land cover type code of ICESat-2 ATL 08 was 1, and the elevation accuracy was ±0.3m. Cloud shadows and atmospheric scattering anomalies were removed using an adaptive thresholding method. The anomaly threshold for GEDI was ±3m of the deviation between the spot elevation and the mean of the surrounding 5 spots, and for ICESat-2 it was ±2m. The average tree height at the stand scale, the canopy height model CHM, and the spatial distribution trend were extracted.

[0018] Ground-based measured data were collected from 20 to 50 standard trees, covering small diameter class (DBH < 10cm), medium diameter class (10cm ≤ DBH ≤ 30cm), and large diameter class (DBH > 30cm), with each diameter class accounting for no less than 20%. Measured parameters included DBH accuracy ±1mm, tree height accuracy ±10cm, crown width accuracy ±5cm, tree age, felled timber volume, and site environmental parameters, including slope, aspect, altitude, and soil organic matter content.

[0019] Spatiotemporal adaptation employs the NTP time synchronization protocol to control timestamp errors to ≤±10min, uses the CGCS2000 coordinate system and the 1985 National Height Standard as the elevation datum, and eliminates spatial offsets from different data sources through a seven-parameter coordinate transformation model with a transformation accuracy of ≤±2cm. By specifying adaptation parameters such as UAV LiDAR wavelength and point cloud density, the 1550nm wavelength is preferentially selected to improve the tree trunk point cloud acquisition rate to ≥85%. Combined with the seven-parameter coordinate transformation, spatial offset elimination of ≤±2cm is achieved, solving the problems of no unified standard for data acquisition and poor consistency between dual-source data in existing technologies. This provides a high-precision, standardized data source for subsequent fusion preprocessing, laying the foundation for accurate single-tree analysis.

[0020] Preferably, in S2, the specific method for dual-source data fusion preprocessing is as follows: the UAV LiDAR point cloud denoising adopts a combination algorithm of "statistical filtering + radius filtering". The statistical filtering sets the number of neighboring points to 16 and the standard deviation threshold to 2.5σ, where σ is the standard deviation of the elevation of the neighboring points, and removes discrete noise points.

[0021] The radius filter is set with a search radius of 0.3m and a minimum number of neighboring points of 8 to remove isolated points;

[0022] Ground point removal employs an improved CSF algorithm that incorporates adaptive terrain slope parameters. The terrain slope is calculated using a 3×3 sliding window, dividing the terrain into gentle slopes with a slope of <15°, gentle slopes with a slope of 15°≤30°, and steep slopes with a slope of >30°. The corresponding filter window sizes are set to 3×3, 5×5, and 7×7 pixels, respectively. The number of iterations is 10–15, and the ground point judgment confidence threshold is 0.85. Suspicious points are then verified a second time using DEM interpolation results. The accuracy of ground point removal is improved by more than 12% compared to the traditional CSF algorithm.

[0023] Based on the non-ground points after removing ground points, a DEM is generated by Kriging interpolation. The Kriging interpolation semivariogram function is a spherical model with an interpolation resolution of 0.5m. The point cloud elevation is normalized with a normalized elevation error ≤2cm.

[0024] Satellite LiDAR data was augmented using a co-kriging interpolation algorithm, with high-precision CHM generated by UAV LiDAR as the covariate. The interpolation accuracy RMSE ≤ 0.5m was achieved, filling in the blank areas of the satellite point cloud.

[0025] Dual-source data registration first involves coarse registration for localization: After the CHM data is denoised by Gaussian filtering with σ=0.8, a 3×3 to 5×5 adaptive window is used to extract the local maximum value as the treetop position. The canopy outline is extracted using the Canny edge detection algorithm. Coarse registration is achieved based on the correspondence between the treetop position and the canopy outline, and the coarse registration error is controlled within 0.5m.

[0026] Then, the improved FPFH-ICP algorithm is used for fine registration, and a tree top feature point set is constructed. The neighborhood radius of the FPFH descriptor is 0.8m, the number of sampling points is 50, and the tree top feature point matching result is used as the initial value of ICP iteration. The maximum number of iterations is set to 50 and the iteration stop threshold is 0.001m. The registration efficiency is improved by more than 30%, and the final spatial registration error of the dual-source data is ≤0.03m.

[0027] Data standardization employs a differentiated strategy: the geometric parameters of tree height and crown width are unified to the [0,1] interval for min-max normalization; spectral and physiological parameters are standardized using Z-score with a mean of 0 and a standard deviation of 1 to ensure consistency in the fusion analysis. The improved CSF algorithm, through terrain-adaptive parameters and secondary verification, increases the accuracy of ground point removal by more than 12%. The improved FPFH-ICP algorithm compresses the registration accuracy to ≤0.03m and improves efficiency by 30%. Collaborative Kriging interpolation fills the gaps in satellite point clouds, effectively solving the defects of traditional fusion algorithms such as poor terrain adaptation, easy registration getting trapped in local optima, and low utilization of sparse point clouds, significantly improving the accuracy and efficiency of dual-source data fusion.

[0028] Preferably, in S3, the single-tree segmentation strategy fused with multiple algorithms is a hybrid algorithm of "CHM peak detection - multi-feature clustering - hyperspectral auxiliary correction", the specific process of which is as follows:

[0029] CHM peak detection: The normalized CHM is smoothed by Gaussian filtering σ=0.8. An adaptive search window of 3×3 to 5×5 is used. A 5×5 window is used in the dense canopy area. The local maximum value is extracted as the candidate point of the tree top. The tree top elevation threshold is set to be more than 0.5m higher than the surrounding pixels. False tree tops formed by dwarf shrubs and branches are removed.

[0030] Multi-feature clustering: Using the treetop as the seed point, the DBSCAN clustering model is constructed by fusing three features: spatial distance of the point cloud, normal vector, and elevation difference. The spatial distance ε is adaptively adjusted according to the average crown width of a single tree, with ε = 0.3 to 0.5 times the average crown width, ranging from 0.6 to 1.2 m. The minimum number of cluster points MinPts = 15, the normal vector threshold is 0.3, the angle with the ground normal vector is ≤18°, and the elevation difference threshold is 0.8 m to segment the point cloud clusters of a single tree.

[0031] Hyperspectral-assisted correction: For canopy overlap areas with canopy closure ≥ 0.7, UAV hyperspectral imagery was introduced, with a band range of 400–1000 nm and a spectral resolution ≤ 5 nm. Six core vegetation indices, including NDVI, EVI, red edge position, and green peak reflectance, were extracted. A tree species classification model was constructed using support vector machine (SVM), with a classification accuracy of ≥ 92%. A spectral feature library for different tree species was established.

[0032] The spectral cosine similarity of adjacent tree point cloud clusters is calculated as cosθ = (A·B) / (|A|×|B|), where A and B are spectral feature vectors. A dynamic similarity threshold is set from 0.3 to 0.5, with a higher threshold for the same tree species and a lower threshold for different tree species. When the threshold is lower than the threshold, the segmentation line is jointly corrected based on the spectral boundary and the spatial distance of the point cloud. The final tree segmentation accuracy is ≥90%, the false negative rate is ≤5%, and the false positive rate is ≤3%.

[0033] Key structural parameter extraction: Tree height is the difference between the maximum elevation of a single tree point cloud and the corresponding DEM elevation, with an accuracy of ±5cm;

[0034] Crown width was obtained by fitting a crown projection ellipse, calculating the major axis CWmax and minor axis CWmin, and taking the average value CW = (CWmax + CWmin) / 2, with an accuracy of ±10cm. Diameter at breast height (DBH) was obtained by fitting point clouds of the trunk at three sections: 1.2m, 1.3m, and 1.4m. Each section used a least-squares cylinder fitting algorithm, with the objective function minΣ[(x-x0)² + (y-y0)² - r²]², where (x0, y0) are the coordinates of the cylinder center, and r is the radius. The fitting residual was controlled within ±0.02m, and the average radius of the three sections was used as the baseline. To obtain the final diameter at breast height (DBH) value, auxiliary parameters were simultaneously extracted, including trunk straightness fitting cylinder axis deviation ≤0.1m / m and canopy density point cloud number ≥5 points / cm³. The "CHM peak detection-multi-feature clustering-hyperspectral auxiliary correction" algorithm achieved a segmentation accuracy of ≥90% for regions with canopy closure ≥0.7. The three-section cylinder fitting combined with outlier removal ensured DBH accuracy within ±3mm, solving the problems of high missed and false detection rates and large errors in microscopic parameter extraction in existing technologies. This provides accurate structural data for volume inversion and twin construction.

[0035] Preferably, in step S4, the single-wood digital twin is constructed in layers using "geometric model + attribute information + physical rules":

[0036] Geometric model: The trunk part is generated as a parametric model based on the cylinder fitting results, divided into segments of 0.5m each, preserving the texture details of the trunk;

[0037] The outer contour of the tree canopy was extracted using the α-shapes algorithm with an α value of 0.8. A three-dimensional mesh model was generated by Poisson reconstruction and Laplacian smoothing. The texture was optimized by upsampling of the isodense point cloud. The number of mesh faces was controlled between 50,000 and 100,000 to balance accuracy and storage cost.

[0038] Attribute information: The structural parameters, ground-based measured data, dual-source LiDAR fusion data and geometric model are bound by a unique identifier UUID. A key-value pair storage structure is used to realize real-time mapping of attributes and geometry. Blockchain technology is used to record data acquisition time, processing algorithm and calibration basis. Hash value verification is used to ensure that the data is tamper-proof.

[0039] Physical rules: Embedded with forest growth mechanisms and environmental response rules, supporting dynamic updates and simulations. Volume inversion employs a combined model of "binary volumetric model + random forest".

[0040] Binary volumetric fitting: Based on ground-measured standard tree data, fitting parameters are applied according to tree species. The general expression is V = a × DBHᵇ × Hᶜ, where V is the volume in m³. For Chinese fir, a = 0.000052–0.000078, b = 1.85–1.95, c = 0.85–0.95; for Masson pine, a = 0.000043–0.000069, b = 1.92–2.08, c = 0.75–0.85; for broadleaf trees, a = 0.000061–0.000087, b = 1.78–1.88, c = 0.92–1.02. The goodness of fit is R² ≥ 0.88. Random forest correction:

[0041] Input features include binary volumetric output values, average tree height of LiDAR forest stands, site environment parameters, and trunk straightness. The model is set with 100 decision trees, a maximum depth of 15, a feature sampling ratio of 0.7, and uses 5-fold cross-validation to optimize hyperparameters.

[0042] Weight calibration: Weights ω1 (0.6~0.7) and ω2 (0.3~0.4) are fitted using the least squares method. The calibration formula is V_calibration = ω1 × V_volume formula + ω2 × V_random forest. After calibration, the volume inversion error is ≤8% and the root mean square error (RMSE) is ≤10%.

[0043] Multi-species mixed forests use hyperspectral-LiDAR joint classification to pre-determine tree species and automatically call the corresponding parameter library, reducing the inversion error by 3-5 percentage points compared to a single model. The hierarchical twin construction balances accuracy and storage cost, and blockchain technology ensures data traceability. The "binary volumetric + random forest" model reduces the inversion error of mixed forests by 3-5 percentage points, solving the problems of poor adaptability, low data reliability, and insufficient accuracy of volume inversion in traditional models, and achieving accurate inversion of single tree volume and secure and controllable twin data.

[0044] Preferably, in step S5, the cross-period single-tree tracking adopts a three-dimensional matching mechanism of "tree top position + trunk structure + spectral characteristics":

[0045] The treetop location is matched with an Euclidean distance threshold of 0.5m, the trunk structure is matched with a similarity of ≥0.8 to the radius of the fitted cylinder, and the spectral features are matched with a cosine similarity of ≥0.6. If two or more conditions are met, it is considered to be the same single tree, and the tracking accuracy is ≥88%. For fallen or felled trees, they are automatically marked and removed, and twins of newly added trees are constructed.

[0046] The competition factor adopts an improved competition index. The calculation logic is as follows: with the target tree as the center, the initial competition area radius is set according to the average crown width of the tree species × 1.5, which is 3-5m for coniferous trees and 5-8m for broad-leaved trees. Then, it is adjusted according to the tree density in the area. When the density is ≥5 trees / 100㎡, it is reduced by 20% and when the density is <2 trees / 100㎡, it is increased by 20%.

[0047] Competition index formula CI=Σ[DBHᵢ / (DBH_target×Dᵢ)]×(CWᵢ / CW_target), where DBHᵢ is the diameter at breast height of the i-th adjacent tree, Dᵢ is the horizontal distance, and CWᵢ is the crown width of the adjacent tree, quantifying the interference intensity of adjacent trees;

[0048] The dynamic growth model is an improved Richards model, with the expression V(t) = K(θ) × (1 - e^(-b(θ) × t))^c, where t is the tree age in years, θ = [T, P, L, CI, S, N] is the vector of environmental and competitive factors, T is the annual average temperature in °C, P is the annual precipitation in mm, L is the duration of sunlight in h, S is the slope, and N is the soil organic matter content.

[0049] c is a shape parameter, 1.2-1.3 for coniferous trees and 1.4-1.5 for broad-leaved trees, and remains constant.

[0050] The upper limit of K(θ) accumulation growth and the growth rate parameter b(θ) are coupled by a multiple linear regression factor, expressed as:

[0051] K(θ)=k0+k1×T+k2×P+k3×CI+k4×S+k5×N, b(θ)=b0+b1×T+b2×P+b3×L+b4×CI;

[0052] The Levenberg-Marquardt algorithm was used to fit the coefficients, with ≤100 iterations and a convergence threshold of 1e-6. The model fit R² ≥0.85 and the prediction accuracy ≥85%, which is 10-15 percentage points higher than the traditional Richards model that only considers tree age. The three-dimensional matching mechanism makes the accuracy of cross-period single tree tracking ≥88%. The improved Richards model, which couples competition factors and multiple environmental factors, improves the prediction accuracy by 10-15 percentage points compared with the traditional model. It solves the shortcomings of existing growth models that only consider tree age, ignore competition and environmental influences, and have large cross-period tracking errors, and achieves accurate characterization and prediction of single tree growth patterns.

[0053] Preferably, in step S6, the incremental fusion algorithm adopts the process of "change region detection - local data update - global consistency verification":

[0054] Change area detection: The Euclidean distance difference method was used to compare point clouds from multiple periods. The distance threshold for the trunk area was 0.05m and for the crown area it was 0.2m. Areas exceeding the threshold were identified as change areas, including areas with increased diameter at breast height, branch extension, crown expansion, and pest and disease stress.

[0055] Local data update: The trunk area is updated incrementally using cylindrical fitting, and the canopy area is locally redrawn with an update resolution of 0.1m, only processing data in the changed areas;

[0056] Global consistency verification: The local update model and the global model are registered using the ICP algorithm, and the verification error is ≤0.03m, ensuring the integrity of the model;

[0057] The data update cycle is adjusted according to the growth stage: 1-3 months for young forests and 6-12 months for mature forests. Ground IoT sensor data is smoothed by Kalman filtering to remove sudden interference and abnormal data.

[0058] Establish a status feedback mechanism. If the deviation between the model prediction and the measured value exceeds 10%, the parameter calibration process will be automatically triggered to adjust the growth model coefficients.

[0059] Large-scale extrapolation is based on the UAV LiDAR single-tree model, establishing a linear correlation model between satellite LiDAR forest stand parameters and single-tree volume. The Kriging spatial interpolation algorithm is used to extrapolate to the regional forest stand level. Multiple 1km×1km verification sample areas are selected for verification. The extrapolation error is ≤15% to be considered qualified. A 50m×50m resolution volume distribution map is generated, which supports docking with the forestry resource management system. The incremental fusion algorithm improves efficiency by more than 60% compared with the full update. Kalman filtering and state feedback mechanism ensure the consistency between twins and physical trees. The Kriging extrapolation error is ≤15%, which solves the problems of high computational cost, easy deviation of model from reality, and lack of regional monitoring in traditional update. It realizes the synergy between precise single-tree management and macro-management of forest stands.

[0060] Preferably, the core optimization of the improved FPFH-ICP algorithm is the initial value constraint of the tree top feature point, which avoids the ICP algorithm from getting trapped in a local optimum.

[0061] Compared to the traditional FPFH-ICP algorithm, the registration time is reduced by more than 30%, and the average distance deviation of the registration accuracy is improved from 0.05m to within 0.03m. It is suitable for the dual-source data alignment requirements of complex scenarios such as mountains and closed forests. By constraining the initial value of treetop feature points, it avoids the local optimum trap of the ICP algorithm, and the average distance deviation of the registration is reduced from 0.05m to within 0.03m. It is specifically adapted to complex scenarios such as mountains and closed forests, and solves the problem that existing registration algorithms cannot balance accuracy and efficiency in complex forestry scenarios, thus enhancing the stability of dual-source data alignment.

[0062] Preferably, the hyperspectral-assisted correction mechanism adds a spectral feature temporal stability verification: by comparing multiple periods of hyperspectral data, the reflectance variation coefficient of each band is calculated, unstable bands with a variation coefficient > 0.2 are eliminated, and 6 to 8 core stable bands are retained for feature extraction;

[0063] The diameter at breast height (DBH) was extracted using three cross-section fittings at 1.2m, 1.3m, and 1.4m. Abnormal cross-sectional radius values ​​were eliminated using the 3σ criterion. If the deviation exceeded the mean ±3σ, and the residual of a single cross-section fitting was >0.02m, 1.1m and 1.5m cross-section fittings were added to ensure that the DBH extraction accuracy was within ±3mm. The spectral temporal stability was verified to eliminate unstable bands. The supplementary cross-section fitting strategy further improved the DBH accuracy, solving the problem that hyperspectral features are easily affected by environmental interference and that trunk curvature leads to abnormal DBH extraction. This provides more reliable support for single-tree segmentation correction and parameter extraction.

[0064] Preferably, the technology and method for fusion processing of single-tree digital twin data also includes a single-tree digital twin quality assessment module and a multi-scenario application mechanism:

[0065] The quality assessment covers geometric accuracy, attribute accuracy, and fusion efficiency. Geometric accuracy is compared with the measured contour using the CloudCompare tool. The average distance deviation of the point cloud is ≤0.05m. The measurement errors of tree height, diameter at breast height, and crown width are controlled within ±5cm, ±3mm, and ±10cm, respectively.

[0066] The accuracy of attributes was verified using 30–50 independent standard trees. The mean absolute percentage error (MAPE) of volume inversion was ≤8%, and the prediction accuracy of the growth model was ≥85%.

[0067] The fusion efficiency requirements are: single tree modeling time ≤ 10 min / tree, single-threaded CPU, main frequency 3.0 GHz, multi-period data update delay ≤ 1 h, single batch processing capacity ≥ 1000 single trees, a three-dimensional quality assessment system is established to clarify accuracy and efficiency indicators, adapt to multiple scenarios such as carbon sequestration accounting and ancient tree protection, carbon sequestration accounting error ≤ 10%, and ancient tree early warning accuracy rate ≥ 80%, which solves the problems of lack of unified quality standards and poor scenario adaptability of existing technologies, and promotes the engineering implementation and large-scale application of technical solutions.

[0068] The beneficial effects of this invention are:

[0069] By integrating dual-source LiDAR and digital twin technologies, four core benefits are achieved: First, the dual-source adaptive registration and enhancement algorithm balances the accuracy of individual trees from UAVs with the regional coverage capability of satellites, achieving a registration error of ≤0.03m and resolving the contradiction between accuracy and range. Second, the "binary volumetric + random forest" model couples the environment and competing factors, achieving a volume inversion error of ≤8%, which is 10-15 percentage points higher than traditional methods. Third, the incremental fusion and dynamic growth model enables efficient twin updates and growth prediction, supporting full life-cycle monitoring. Fourth, the integrated system adapts to multiple forestry scenarios, taking into account the needs of precision tending, carbon sequestration, and other aspects, significantly improving the intelligence and efficiency of forestry resource management. Attached Figure Description

[0070] Figure 1 This is a flowchart illustrating the workflow proposed in this invention. Detailed Implementation

[0071] The present invention will be further explained below with reference to specific embodiments.

[0072] Reference Figure 1 , Example

[0073] This embodiment proposes a data fusion processing technology and method for single-tree digital twins. Based on the collaborative processing of dual-source data from satellite LiDAR and UAV LiDAR, and combined with digital twin technology, it achieves accurate inversion of single-tree volume and construction of dynamic growth models. The entire process includes six major stages: data acquisition and adaptation, dual-source data fusion preprocessing, single-tree segmentation and structural parameter extraction, digital twin construction and volume inversion, time-series growth model construction, and dynamic updating of the twin. Specifically, it includes the following steps:

[0074] S1: Data Acquisition and Adaptation: Using UAV LiDAR as the core data source at the single-tree scale and satellite LiDAR as the auxiliary data source at the stand scale, ground-based measured data is collected simultaneously as a calibration and verification benchmark. Spatiotemporal adaptation rules for dual-source LiDAR data and ground-based measured data are established to unify data benchmarks and formats. The UAV LiDAR uses a 905nm or 1550nm laser wavelength, a scanning frequency of 100-200kHz, ≥3 times of point cloud echoes, and a point cloud density controlled at 20-50 points / ㎡. The density is adaptively adjusted according to the size of the single-tree crown. When the crown width is ≤2m, the density is 30-50 points / ㎡. The flight altitude is 50-100m, and the overlap rate of the flight direction and the side direction is ≥80%. The matching RTK / PPK positioning module achieves a planar accuracy of ±3cm and a vertical accuracy of ±5cm.

[0075] The 1550nm wavelength was prioritized to enhance canopy penetration and reduce interference from foliage shading. The trunk point cloud acquisition rate was ≥85%. Satellite LiDAR data were selected from either GEDI Level 2A or ICESat-2 ATL 08. The effective spot signal-to-noise ratio of GEDI Level 2A was ≥3, and the spot diameter was ~25m. The ground spot land cover type code of ICESat-2 ATL 08 was 1, and the elevation accuracy was ±0.3m. Cloud shadows and atmospheric scattering anomalies were removed using an adaptive thresholding method. The anomaly threshold for GEDI was ±3m of the deviation between the spot elevation and the mean of the surrounding 5 spots, and for ICESat-2 it was ±2m. The average tree height at the stand scale, the canopy height model CHM, and the spatial distribution trend were extracted.

[0076] Ground-based measured data were collected from 20 to 50 standard trees, covering small diameter class (DBH < 10cm), medium diameter class (10cm ≤ DBH ≤ 30cm), and large diameter class (DBH > 30cm), with each diameter class accounting for no less than 20%. Measured parameters included DBH accuracy ±1mm, tree height accuracy ±10cm, crown width accuracy ±5cm, tree age, felled timber volume, and site environmental parameters, including slope, aspect, altitude, and soil organic matter content.

[0077] Spatiotemporal adaptation employs the NTP time synchronization protocol to control timestamp errors ≤ ±10 min, uses the CGCS2000 coordinate system and the 1985 National Height Standard as the elevation datum, and eliminates spatial offsets from different data sources through a seven-parameter coordinate transformation model with a transformation accuracy ≤ ±2 cm. By specifying adaptation parameters such as UAV LiDAR wavelength and point cloud density, the 1550 nm wavelength is preferentially selected to improve the tree trunk point cloud acquisition rate to ≥ 85%. Combined with the seven-parameter coordinate transformation, spatial offset elimination of ≤ ±2 cm is achieved, solving the problems of no unified standard for data acquisition and poor consistency between dual-source data in existing technologies. This provides a high-precision, standardized data source for subsequent fusion preprocessing and lays the foundation for accurate single-tree analysis.

[0078] S2: Dual-source data fusion preprocessing: Noise filtering, ground point removal, and data augmentation are performed on the dual-source LiDAR data respectively. Spatiotemporal alignment of the dual-source data is achieved through coarse and fine registration. Then, the data is standardized to provide a foundation for subsequent fusion analysis. The specific method of dual-source data fusion preprocessing is as follows: The UAV LiDAR point cloud denoising adopts a combination algorithm of "statistical filtering + radius filtering". The statistical filtering sets the number of neighboring points to 16 and the standard deviation threshold to 2.5σ, where σ is the standard deviation of the elevation of the neighboring points, and removes discrete noise points.

[0079] The radius filter is set with a search radius of 0.3m and a minimum number of neighboring points of 8 to remove isolated points;

[0080] Ground point removal employs an improved CSF algorithm that incorporates adaptive terrain slope parameters. The terrain slope is calculated using a 3×3 sliding window, dividing the terrain into gentle slopes with a slope of <15°, gentle slopes with a slope of 15°≤30°, and steep slopes with a slope of >30°. The corresponding filter window sizes are set to 3×3, 5×5, and 7×7 pixels, respectively. The number of iterations is 10–15, and the ground point judgment confidence threshold is 0.85. Suspicious points are then verified a second time using DEM interpolation results. The accuracy of ground point removal is improved by more than 12% compared to the traditional CSF algorithm.

[0081] Based on the non-ground points after removing ground points, a DEM is generated by Kriging interpolation. The Kriging interpolation semivariogram function is a spherical model with an interpolation resolution of 0.5m. The point cloud elevation is normalized with a normalized elevation error ≤2cm.

[0082] Satellite LiDAR data was augmented using a co-kriging interpolation algorithm, with high-precision CHM generated by UAV LiDAR as the covariate. The interpolation accuracy RMSE ≤ 0.5m was achieved, filling in the blank areas of the satellite point cloud.

[0083] Dual-source data registration first involves coarse registration for localization: After the CHM data is denoised by Gaussian filtering with σ=0.8, a 3×3 to 5×5 adaptive window is used to extract the local maximum value as the treetop position. The canopy outline is extracted using the Canny edge detection algorithm. Coarse registration is achieved based on the correspondence between the treetop position and the canopy outline, and the coarse registration error is controlled within 0.5m.

[0084] Then, the improved FPFH-ICP algorithm is used for fine registration, and a tree top feature point set is constructed. The neighborhood radius of the FPFH descriptor is 0.8m, the number of sampling points is 50, and the tree top feature point matching result is used as the initial value of ICP iteration. The maximum number of iterations is set to 50 and the iteration stop threshold is 0.001m. The registration efficiency is improved by more than 30%, and the final spatial registration error of the dual-source data is ≤0.03m.

[0085] Data standardization employs a differentiated strategy: the geometric parameters of tree height and crown width are unified to the [0,1] interval min-max normalization, and spectral and physiological parameters are standardized using Z-score with a mean of 0 and a standard deviation of 1 to ensure consistency in the fusion analysis. The improved CSF algorithm improves the accuracy of ground point removal by more than 12% through terrain adaptive parameters and secondary verification. The improved FPFH-ICP algorithm compresses the registration accuracy to ≤0.03m and improves efficiency by 30%. Kriging interpolation is used to fill the gaps in satellite point clouds, effectively solving the defects of traditional fusion algorithms such as poor terrain adaptation, easy registration getting trapped in local optima, and low utilization of sparse point clouds, and significantly improving the accuracy and efficiency of dual-source data fusion.

[0086] The core optimization of the improved FPFH-ICP algorithm is the initial value constraint of the tree top feature point, which avoids the ICP algorithm from getting trapped in local optima.

[0087] Compared to the traditional FPFH-ICP algorithm, the registration time is reduced by more than 30%, and the average distance deviation of the registration accuracy is improved from 0.05m to within 0.03m. It is suitable for the dual-source data alignment requirements of complex scenarios such as mountains and closed forests. By constraining the initial value of treetop feature points, it avoids the local optimum trap of the ICP algorithm, and the average distance deviation of the registration is reduced from 0.05m to within 0.03m. It is specifically adapted to complex scenarios such as mountains and closed forests, and solves the problem that existing registration algorithms cannot balance accuracy and efficiency in complex forestry scenarios, thus enhancing the stability of dual-source data alignment.

[0088] S3: Individual Tree Segmentation and Structural Parameter Extraction: Based on UAV LiDAR data, a multi-algorithm fusion strategy is employed to achieve accurate individual tree segmentation. Key structural parameters and auxiliary parameters, including tree height, diameter at breast height (DBH), and crown width, are extracted from the segmented individual tree point cloud clusters. The multi-algorithm fusion strategy is a hybrid algorithm combining "CHM peak detection - multi-feature clustering - hyperspectral auxiliary correction." The specific process is as follows:

[0089] CHM peak detection: The normalized CHM is smoothed by Gaussian filtering σ=0.8. An adaptive search window of 3×3 to 5×5 is used. A 5×5 window is used in the dense canopy area. The local maximum value is extracted as the candidate point of the tree top. The tree top elevation threshold is set to be more than 0.5m higher than the surrounding pixels. False tree tops formed by dwarf shrubs and branches are removed.

[0090] Multi-feature clustering: Using the treetop as the seed point, the DBSCAN clustering model is constructed by fusing three features: spatial distance of the point cloud, normal vector, and elevation difference. The spatial distance ε is adaptively adjusted according to the average crown width of a single tree, with ε = 0.3 to 0.5 times the average crown width, ranging from 0.6 to 1.2 m. The minimum number of cluster points MinPts = 15, the normal vector threshold is 0.3, the angle with the ground normal vector is ≤18°, and the elevation difference threshold is 0.8 m to segment the point cloud clusters of a single tree.

[0091] Hyperspectral-assisted correction: For canopy overlap areas with canopy closure ≥ 0.7, UAV hyperspectral imagery was introduced, with a band range of 400–1000 nm and a spectral resolution ≤ 5 nm. Six core vegetation indices, including NDVI, EVI, red edge position, and green peak reflectance, were extracted. A tree species classification model was constructed using support vector machine (SVM), with a classification accuracy of ≥ 92%. A spectral feature library for different tree species was established.

[0092] The spectral cosine similarity of adjacent tree point cloud clusters is calculated as cosθ = (A·B) / (|A|×|B|), where A and B are spectral feature vectors. A dynamic similarity threshold is set from 0.3 to 0.5, with a higher threshold for the same tree species and a lower threshold for different tree species. When the threshold is lower than the threshold, the segmentation line is jointly corrected based on the spectral boundary and the spatial distance of the point cloud. The final tree segmentation accuracy is ≥90%, the false negative rate is ≤5%, and the false positive rate is ≤3%.

[0093] Key structural parameter extraction: Tree height is the difference between the maximum elevation of a single tree point cloud and the corresponding DEM elevation, with an accuracy of ±5cm;

[0094] Crown width was obtained by fitting a crown projection ellipse, calculating the major axis CWmax and minor axis CWmin, and taking the average value CW = (CWmax + CWmin) / 2, with an accuracy of ±10cm. Diameter at breast height (DBH) was obtained by fitting point clouds of the trunk at three sections: 1.2m, 1.3m, and 1.4m. Each section used a least-squares cylinder fitting algorithm, with the objective function minΣ[(x-x0)² + (y-y0)² - r²]², where (x0, y0) are the coordinates of the cylinder center, and r is the radius. The fitting residual was controlled within ±0.02m, and the average radius of the three sections was used as the baseline. To obtain the final diameter at breast height (DBH) value, auxiliary parameters were simultaneously extracted, including trunk straightness fitting cylinder axis deviation ≤0.1m / m and canopy density point cloud number ≥5 points / cm³. The "CHM peak detection-multi-feature clustering-hyperspectral auxiliary correction" algorithm achieved a segmentation accuracy of ≥90% for regions with canopy closure ≥0.7. The three-section cylinder fitting combined with outlier removal ensured DBH accuracy within ±3mm, solving the problems of high missed and false detection rates and large errors in microscopic parameter extraction in existing technologies. This provides accurate structural data for volume inversion and twin construction.

[0095] The hyperspectral-assisted correction mechanism adds a spectral feature temporal stability verification: by comparing multiple periods of hyperspectral data, the reflectance variation coefficient of each band is calculated, unstable bands with a variation coefficient > 0.2 are removed, and 6 to 8 core stable bands are retained for feature extraction; for diameter at breast height (DBH) extraction, three-section fitting at 1.2m, 1.3m, and 1.4m is used, and abnormal section radius values ​​are removed by the 3σ criterion. If the deviation exceeds the mean ± 3σ, and the residual of a single section fitting is > 0.02m, 1.1m and 1.5m section fittings are added to ensure that the DBH extraction accuracy is within ± 3mm. The spectral temporal stability verification removes unstable bands, and the supplementary section fitting strategy further improves the DBH accuracy. This solves the problem that hyperspectral features are easily affected by environmental interference and trunk bending, which leads to abnormal DBH extraction, and provides more reliable support for single-tree segmentation correction and parameter extraction.

[0096] S4: Digital Twin Construction and Volume Inversion: Constructing a three-in-one digital twin of a single tree, consisting of "geometric model + attribute information + physical rules," using a model that combines traditional volumetric methods with machine learning algorithms to invert the volume of a single tree and calibrate it using ground-based measured data. The digital twin of a single tree is constructed in a layered manner using "geometric model + attribute information + physical rules."

[0097] Geometric model: The trunk part is generated as a parametric model based on the cylinder fitting results, divided into segments of 0.5m each, preserving the texture details of the trunk;

[0098] The outer contour of the tree canopy was extracted using the α-shapes algorithm with an α value of 0.8. A three-dimensional mesh model was generated by Poisson reconstruction and Laplacian smoothing. The texture was optimized by upsampling of the isodense point cloud. The number of mesh faces was controlled between 50,000 and 100,000 to balance accuracy and storage cost.

[0099] Attribute information: The structural parameters, ground-based measured data, dual-source LiDAR fusion data and geometric model are bound by a unique identifier UUID. A key-value pair storage structure is used to realize real-time mapping of attributes and geometry. Blockchain technology is used to record data acquisition time, processing algorithm and calibration basis. Hash value verification is used to ensure that the data is tamper-proof.

[0100] Physical rules: Embedded with forest growth mechanisms and environmental response rules, supporting dynamic updates and simulations. Volume inversion employs a combined model of "binary volumetric model + random forest".

[0101] Binary volumetric fitting: Based on ground-measured standard tree data, fitting parameters are applied according to tree species. The general expression is V = a × DBHᵇ × Hᶜ, where V is the volume in m³. For Chinese fir, a = 0.000052–0.000078, b = 1.85–1.95, c = 0.85–0.95; for Masson pine, a = 0.000043–0.000069, b = 1.92–2.08, c = 0.75–0.85; for broadleaf trees, a = 0.000061–0.000087, b = 1.78–1.88, c = 0.92–1.02. The goodness of fit is R² ≥ 0.88. Random forest correction:

[0102] Input features include binary volumetric output values, average tree height of LiDAR forest stands, site environment parameters, and trunk straightness. The model is set with 100 decision trees, a maximum depth of 15, a feature sampling ratio of 0.7, and uses 5-fold cross-validation to optimize hyperparameters.

[0103] Weight calibration: Weights ω1 (0.6~0.7) and ω2 (0.3~0.4) are fitted using the least squares method. The calibration formula is V_calibration = ω1 × V_volume formula + ω2 × V_random forest. After calibration, the volume inversion error is ≤8% and the root mean square error (RMSE) is ≤10%.

[0104] Multi-species mixed forests are classified by hyperspectral-LiDAR joint classification to determine tree species in advance and automatically call the corresponding parameter library, which reduces the inversion error by 3 to 5 percentage points compared with a single model.

[0105] S5: Temporal Growth Model Construction: Based on multi-period dual-source LiDAR data and ground-measured time-series data, it achieves cross-period single-tree tracking, couples environmental factors and competing factors to construct a dynamic growth model, characterizes the growth law of single-tree volume, and constructs hierarchical twins to balance accuracy and storage cost. Blockchain technology ensures data traceability. In S5, cross-period single-tree tracking adopts a three-dimensional matching mechanism of "tree top position + trunk structure + spectral features".

[0106] The treetop location is matched with an Euclidean distance threshold of 0.5m, the trunk structure is matched with a similarity of ≥0.8 to the radius of the fitted cylinder, and the spectral features are matched with a cosine similarity of ≥0.6. If two or more conditions are met, it is considered to be the same single tree, and the tracking accuracy is ≥88%. For fallen or felled trees, they are automatically marked and removed, and twins of newly added trees are constructed.

[0107] The competition factor adopts an improved competition index. The calculation logic is as follows: with the target tree as the center, the initial competition area radius is set according to the average crown width of the tree species × 1.5, which is 3-5m for coniferous trees and 5-8m for broad-leaved trees. Then, it is adjusted according to the tree density in the area. When the density is ≥5 trees / 100㎡, it is reduced by 20% and when the density is <2 trees / 100㎡, it is increased by 20%.

[0108] Competition index formula CI=Σ[DBHᵢ / (DBH_target×Dᵢ)]×(CWᵢ / CW_target), where DBHᵢ is the diameter at breast height of the i-th adjacent tree, Dᵢ is the horizontal distance, and CWᵢ is the crown width of the adjacent tree, quantifying the interference intensity of adjacent trees;

[0109] The dynamic growth model is an improved Richards model, with the expression V(t) = K(θ) × (1 - e^(-b(θ) × t))^c, where t is the tree age in years, θ = [T, P, L, CI, S, N] is the vector of environmental and competitive factors, T is the annual average temperature in °C, P is the annual precipitation in mm, L is the duration of sunlight in h, S is the slope, and N is the soil organic matter content.

[0110] c is a shape parameter, 1.2-1.3 for coniferous trees and 1.4-1.5 for broad-leaved trees, and remains constant.

[0111] The upper limit of K(θ) accumulation growth and the growth rate parameter b(θ) are coupled by a multiple linear regression factor, expressed as:

[0112] K(θ)=k0+k1×T+k2×P+k3×CI+k4×S+k5×N, b(θ)=b0+b1×T+b2×P+b3×L+b4×CI;

[0113] The Levenberg-Marquardt algorithm was used to fit the coefficients, with ≤100 iterations and a convergence threshold of 1e-6. The model fit R² ≥0.85 and the prediction accuracy ≥85%. Compared with the traditional Richards model, which only considers tree age, the prediction accuracy is improved by 10-15 percentage points. The "binary volume + random forest" model reduces the inversion error of mixed forests by 3-5 percentage points, solving the problems of poor adaptability, low data reliability, and insufficient volume inversion accuracy of traditional models. It achieves accurate inversion of single tree volume and safe and controllable twin data.

[0114] S6: Dynamic Update and Application of Digital Twins: An incremental fusion algorithm is used to dynamically update the digital twins of individual trees. Combined with a growth model to predict the growth trend of individual trees, satellite LiDAR data is used to extrapolate small-area models over a large area, constructing an integrated digital twin application system. The incremental fusion algorithm adopts a process of "change area detection – local data update – global consistency verification."

[0115] Change area detection: The Euclidean distance difference method was used to compare point clouds from multiple periods. The distance threshold for the trunk area was 0.05m and for the crown area it was 0.2m. Areas exceeding the threshold were identified as change areas, including areas with increased diameter at breast height, branch extension, crown expansion, and pest and disease stress.

[0116] Local data update: The trunk area is updated incrementally using cylindrical fitting, and the canopy area is locally redrawn with an update resolution of 0.1m, only processing data in the changed areas;

[0117] Global consistency verification: The local update model and the global model are registered using the ICP algorithm, and the verification error is ≤0.03m, ensuring the integrity of the model;

[0118] The data update cycle is adjusted according to the growth stage: 1-3 months for young forests and 6-12 months for mature forests. Ground IoT sensor data is smoothed by Kalman filtering to remove sudden interference and abnormal data.

[0119] Establish a status feedback mechanism. If the deviation between the model prediction and the measured value exceeds 10%, the parameter calibration process will be automatically triggered to adjust the growth model coefficients.

[0120] Large-scale extrapolation is based on the UAV LiDAR single-tree model, establishing a linear correlation model between satellite LiDAR forest stand parameters and single-tree volume. The Kriging spatial interpolation algorithm is used to extrapolate to the regional forest stand level. Multiple 1km×1km verification sample areas are selected for verification. The extrapolation error is ≤15% to be considered qualified. A 50m×50m resolution volume distribution map is generated, which supports the connection with the forestry resource management system. The incremental fusion algorithm improves the efficiency by more than 60% compared with the full update. Kalman filtering and state feedback mechanism ensure the consistency between twins and physical trees. The Kriging extrapolation error is ≤15%. It solves the problems of high computational cost, easy deviation of model from reality, and lack of regional monitoring in traditional update, and realizes the synergy between precise single-tree management and macro-management of forest stands.

[0121] The technology and method for fusion processing of single-tree digital twin data also include a single-tree digital twin quality assessment module and a multi-scenario application mechanism:

[0122] The quality assessment covers geometric accuracy, attribute accuracy, and fusion efficiency. Geometric accuracy is compared with the measured contour using the CloudCompare tool. The average distance deviation of the point cloud is ≤0.05m. The measurement errors of tree height, diameter at breast height, and crown width are controlled within ±5cm, ±3mm, and ±10cm, respectively.

[0123] The accuracy of attributes was verified using 30–50 independent standard trees. The mean absolute percentage error (MAPE) of volume inversion was ≤8%, and the prediction accuracy of the growth model was ≥85%.

[0124] The fusion efficiency requirements are: single tree modeling time ≤ 10 min / tree, single-threaded CPU, main frequency 3.0 GHz, multi-period data update latency ≤ 1 h, single batch processing capacity ≥ 1000 single trees, a three-dimensional quality assessment system is established to clarify accuracy and efficiency indicators, adapt to multiple scenarios such as carbon sequestration accounting and ancient tree protection, carbon sequestration accounting error ≤ 10%, ancient tree early warning accuracy rate ≥ 80%, solving the problems of lack of unified quality standards and poor scenario adaptability of existing technologies, and promoting the engineering implementation and large-scale application of technical solutions;

[0125] By integrating dual-source LiDAR and digital twin technologies, four core benefits are achieved: First, the dual-source adaptive registration and enhancement algorithm balances the accuracy of individual trees from UAVs with the regional coverage capability of satellites, achieving a registration error of ≤0.03m and resolving the contradiction between accuracy and range. Second, the "binary volumetric + random forest" model couples the environment and competing factors, achieving a volume inversion error of ≤8%, which is 10-15 percentage points higher than traditional methods. Third, the incremental fusion and dynamic growth model enables efficient twin updates and growth prediction, supporting full life-cycle monitoring. Fourth, the integrated system adapts to multiple forestry scenarios, taking into account the needs of precision tending, carbon sequestration, and other aspects, significantly improving the intelligence and efficiency of forestry resource management.

[0126] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A technology and method for fusion processing of digital twin data of single forest trees, characterized in that, Based on the collaborative use of satellite LiDAR and UAV LiDAR data, and combined with digital twin technology, this method achieves accurate inversion of single-tree volume and construction of a dynamic growth model. The entire process includes six major stages: data acquisition and adaptation, dual-source data fusion preprocessing, single-tree segmentation and structural parameter extraction, digital twin construction and volume inversion, time-series growth model construction, and dynamic updating of the twin. Specifically, it includes the following steps: S1: Data Acquisition and Adaptation: Using UAV LiDAR as the core data source at the single-tree scale and satellite LiDAR as the auxiliary data source at the stand scale, ground-based measured data is collected simultaneously as a calibration and verification benchmark. Spatiotemporal adaptation rules between dual-source LiDAR data and ground-based measured data are established to unify data benchmarks and formats. S2: Dual-source data fusion preprocessing: Noise filtering, ground point removal, and data augmentation are performed on the dual-source LiDAR data respectively. Spatiotemporal alignment of the dual-source data is achieved through coarse and fine registration. Then, the data is standardized to provide a foundation for subsequent fusion analysis. S3: Tree segmentation and structural parameter extraction: Based on UAV LiDAR data, a multi-algorithm fusion strategy is used to achieve accurate tree segmentation. Key structural parameters and auxiliary parameters such as tree height, diameter at breast height, and crown width are extracted from the segmented tree point cloud clusters. S4: Digital Twin Construction and Volume Inversion: Construct a three-in-one digital twin of a single tree, consisting of "geometric model + attribute information + physical rules". Use a model that combines traditional volumetric data with machine learning algorithms to invert the volume of a single tree and calibrate it with ground-based measured data. S5: Construction of time-series growth model: Based on multi-period dual-source LiDAR data and ground-measured time-series data, cross-period single-tree tracking is realized, and a dynamic growth model is constructed by coupling environmental factors and competitive factors to characterize the growth law of single-tree volume. S6: Dynamic Update and Application of Digital Twins: The incremental fusion algorithm is used to realize the dynamic update of digital twins of individual trees. Combined with the growth model, the growth trend of individual trees is predicted. Satellite LiDAR data is used to realize the large-scale extrapolation of small-area models and build an integrated digital twin application system.

2. The forest single-piece digital twin data fusion processing technology and method according to claim 1, characterized in that, In S1, the UAV LiDAR uses a 905nm or 1550nm laser wavelength, a scanning frequency of 100-200kHz, a point cloud echo count of ≥3 times, and a point cloud density controlled at 20-50 points / ㎡. The density is adaptively adjusted according to the size of the single tree crown. When the crown width is ≤2m, the density is 30-50 points / ㎡. The flight altitude is 50-100m, and the overlap rate of the flight direction and the side direction is ≥80%. The matching RTK / PPK positioning module achieves a planar accuracy of ±3cm and a vertical accuracy of ±5cm. The 1550nm wavelength was prioritized to enhance canopy penetration and reduce interference from foliage shading. The trunk point cloud acquisition rate was ≥85%. Satellite LiDAR data were selected from either GEDI Level 2A or ICESat-2 ATL 08. The effective spot signal-to-noise ratio of GEDI Level 2A was ≥3, and the spot diameter was ~25m. The ground spot land cover type code of ICESat-2 ATL 08 was 1, and the elevation accuracy was ±0.3m. Cloud shadows and atmospheric scattering anomalies were removed using an adaptive thresholding method. The anomaly threshold for GEDI was ±3m of the deviation between the spot elevation and the mean of the surrounding 5 spots, and for ICESat-2 it was ±2m. The average tree height at the stand scale, the canopy height model CHM, and the spatial distribution trend were extracted. Ground-based measured data were collected from 20 to 50 standard trees, covering small diameter class (DBH < 10cm), medium diameter class (10cm ≤ DBH ≤ 30cm), and large diameter class (DBH > 30cm), with each diameter class accounting for no less than 20%. Measured parameters included DBH accuracy ±1mm, tree height accuracy ±10cm, crown width accuracy ±5cm, tree age, felled timber volume, and site environmental parameters, including slope, aspect, altitude, and soil organic matter content. The spatiotemporal adaptation uses the NTP time synchronization protocol to control the timestamp error to ≤±10min. The unified coordinate system is CGCS2000 and the elevation datum is the 1985 National Elevation. The spatial offset of different data sources is eliminated through a seven-parameter coordinate transformation model, and the transformation accuracy is ≤±2cm.

3. The forest single-piece digital twin data fusion processing technology and method according to claim 1, characterized in that, In S2, the specific method for dual-source data fusion preprocessing is as follows: the UAV LiDAR point cloud denoising adopts a combination algorithm of "statistical filtering + radius filtering". The statistical filtering sets the number of neighboring points to 16 and the standard deviation threshold to 2.5σ, where σ is the standard deviation of the elevation of the neighboring points, and removes discrete noise points. The radius filter is set with a search radius of 0.3m and a minimum number of neighboring points of 8 to remove isolated points; Ground point removal employs an improved CSF algorithm that incorporates adaptive terrain slope parameters. The terrain slope is calculated using a 3×3 sliding window, dividing the terrain into gentle slopes with a slope of <15°, gentle slopes with a slope of 15°≤30°, and steep slopes with a slope of >30°. The corresponding filter window sizes are set to 3×3, 5×5, and 7×7 pixels, respectively. The number of iterations is 10–15, and the ground point judgment confidence threshold is 0.

85. Suspicious points are then verified a second time using DEM interpolation results. The accuracy of ground point removal is improved by more than 12% compared to the traditional CSF algorithm. Based on the non-ground points after removing ground points, a DEM is generated by Kriging interpolation. The Kriging interpolation semivariogram function is a spherical model with an interpolation resolution of 0.5m. The point cloud elevation is normalized with a normalized elevation error ≤2cm. Satellite LiDAR data was augmented using a co-kriging interpolation algorithm, with high-precision CHM generated by UAV LiDAR as the covariate. The interpolation accuracy RMSE ≤ 0.5m was achieved, filling in the blank areas of the satellite point cloud. Dual-source data registration first involves coarse registration for localization: After the CHM data is denoised by Gaussian filtering with σ=0.8, a 3×3 to 5×5 adaptive window is used to extract the local maximum value as the treetop position. The canopy outline is extracted using the Canny edge detection algorithm. Coarse registration is achieved based on the correspondence between the treetop position and the canopy outline, and the coarse registration error is controlled within 0.5m. Then, the improved FPFH-ICP algorithm is used for fine registration, and a tree top feature point set is constructed. The neighborhood radius of the FPFH descriptor is 0.8m, the number of sampling points is 50, and the tree top feature point matching result is used as the initial value of ICP iteration. The maximum number of iterations is set to 50 and the iteration stop threshold is 0.001m. The registration efficiency is improved by more than 30%, and the final spatial registration error of the dual-source data is ≤0.03m. Data standardization employs a differentiated strategy: the geometric parameters of tree height and crown width are standardized to the [0,1] interval min-max normalization, and spectral and physiological parameters are standardized using Z-score with a mean of 0 and a standard deviation of 1 to ensure consistency in the fusion analysis.

4. The forest single-piece digital twin data fusion processing technology and method according to claim 1, characterized in that, In S3, the single-tree segmentation strategy fused with multiple algorithms is a hybrid algorithm of "CHM peak detection - multi-feature clustering - hyperspectral auxiliary correction", the specific process of which is as follows: CHM peak detection: The normalized CHM is smoothed by Gaussian filtering σ=0.

8. An adaptive search window of 3×3 to 5×5 is used. A 5×5 window is used in the dense canopy area. The local maximum value is extracted as the candidate point of the tree top. The tree top elevation threshold is set to be more than 0.5m higher than the surrounding pixels. False tree tops formed by dwarf shrubs and branches are removed. Multi-feature clustering: Using the treetop as the seed point, the DBSCAN clustering model is constructed by fusing three features: spatial distance of the point cloud, normal vector, and elevation difference. The spatial distance ε is adaptively adjusted according to the average crown width of a single tree, with ε = 0.3 to 0.5 times the average crown width, ranging from 0.6 to 1.2 m. The minimum number of cluster points MinPts = 15, the normal vector threshold is 0.3, the angle with the ground normal vector is ≤18°, and the elevation difference threshold is 0.8 m to segment the point cloud clusters of a single tree. Hyperspectral-assisted correction: For canopy overlap areas with canopy closure ≥ 0.7, UAV hyperspectral imagery was introduced, with a band range of 400–1000 nm and a spectral resolution ≤ 5 nm. Six core vegetation indices, including NDVI, EVI, red edge position, and green peak reflectance, were extracted. A tree species classification model was constructed using support vector machine (SVM), with a classification accuracy of ≥ 92%. A spectral feature library for different tree species was established. The spectral cosine similarity of adjacent tree point cloud clusters is calculated as cosθ = (A·B) / (|A|×|B|), where A and B are spectral feature vectors. A dynamic similarity threshold is set from 0.3 to 0.5, with a higher threshold for the same tree species and a lower threshold for different tree species. When the threshold is lower than the threshold, the segmentation line is jointly corrected based on the spectral boundary and the spatial distance of the point cloud. The final tree segmentation accuracy is ≥90%, the false negative rate is ≤5%, and the false positive rate is ≤3%. Key structural parameter extraction: Tree height is the difference between the maximum elevation of a single tree point cloud and the corresponding DEM elevation, with an accuracy of ±5cm; Crown width was obtained by fitting the crown projection ellipse, calculating the major axis CWmax and minor axis CWmin, and taking the average value CW = (CWmax + CWmin) / 2, with an accuracy of ±10cm; diameter at breast height (DBH) was obtained by fitting the point cloud of the trunk at three sections of 1.2m, 1.3m, and 1.4m. The least squares cylinder fitting algorithm was used for each section, with the fitting objective function minΣ[(x-x0)²+(y-y0)²-r²]², where (x0, y0) are the coordinates of the cylinder center and r is the radius. The fitting residual was controlled within ±0.02m, and the average of the radii of the three sections was taken as the final DBH value. At the same time, auxiliary parameters were extracted to ensure that the deviation of the cylinder axis of the trunk straightness fitting is ≤0.1m / m and the number of point clouds per unit volume of the crown density is ≥5 points / cm³.

5. The forest single-piece digital twin data fusion processing technology and method according to claim 1, characterized in that, In S4, the single-tree digital twin is constructed in layers using "geometric model + attribute information + physical rules": Geometric model: The trunk part is generated as a parametric model based on the cylinder fitting results, divided into segments of 0.5m each, preserving the texture details of the trunk; The outer contour of the tree canopy was extracted using the α-shapes algorithm with an α value of 0.

8. A three-dimensional mesh model was generated by Poisson reconstruction and Laplacian smoothing. The texture was optimized by upsampling of the isodense point cloud. The number of mesh faces was controlled between 50,000 and 100,000 to balance accuracy and storage cost. Attribute information: The structural parameters, ground-based measured data, dual-source LiDAR fusion data and geometric model are bound by a unique identifier UUID. A key-value pair storage structure is used to realize real-time mapping of attributes and geometry. Blockchain technology is used to record data acquisition time, processing algorithm and calibration basis. Hash value verification is used to ensure that the data is tamper-proof. Physical rules: Embedded with forest growth mechanisms and environmental response rules, supporting dynamic updates and simulations. Volume inversion employs a combined model of "binary volumetric model + random forest": Binary volumetric fitting: Based on ground-measured standard tree data, fitting parameters are applied according to tree species. The general expression is V = a × DBHᵇ × Hᶜ, where V is the volume in m³. For Chinese fir, a = 0.000052–0.000078, b = 1.85–1.95, c = 0.85–0.95; for Masson pine, a = 0.000043–0.000069, b = 1.92–2.08, c = 0.75–0.85; for broadleaf trees, a = 0.000061–0.000087, b = 1.78–1.88, c = 0.92–1.

02. The goodness of fit is R² ≥ 0.

88. Random forest correction: Input features include binary volumetric output values, average tree height of LiDAR forest stands, site environment parameters, and trunk straightness. The model is set with 100 decision trees, a maximum depth of 15, a feature sampling ratio of 0.7, and uses 5-fold cross-validation to optimize hyperparameters. Weight calibration: Weights ω1 (0.6~0.7) and ω2 (0.3~0.4) are fitted using the least squares method. The calibration formula is V_calibration = ω1 × V_volume formula + ω2 × V_random forest. After calibration, the volume inversion error is ≤8% and the root mean square error (RMSE) is ≤10%. Multi-species mixed forests are classified by hyperspectral-LiDAR joint classification to determine tree species in advance and automatically call the corresponding parameter library, which reduces the inversion error by 3 to 5 percentage points compared with a single model.

6. The forest single-piece digital twin data fusion processing technology and method according to claim 1, characterized in that, In S5, the inter-period single-tree tracking adopts a three-dimensional matching mechanism of "tree top position + trunk structure + spectral characteristics": The treetop location is matched with an Euclidean distance threshold of 0.5m, the trunk structure is matched with a similarity of ≥0.8 to the radius of the fitted cylinder, and the spectral features are matched with a cosine similarity of ≥0.

6. If two or more conditions are met, it is considered to be the same single tree, and the tracking accuracy is ≥88%. For fallen or felled trees, they are automatically marked and removed, and twins of newly added trees are constructed. The competition factor adopts an improved competition index. The calculation logic is as follows: with the target tree as the center, the initial competition area radius is set according to the average crown width of the tree species × 1.5, which is 3-5m for coniferous trees and 5-8m for broad-leaved trees. Then, it is adjusted according to the tree density in the area. When the density is ≥5 trees / 100㎡, it is reduced by 20% and when the density is <2 trees / 100㎡, it is increased by 20%. Competition index formula CI=Σ[DBHᵢ / (DBH_target×Dᵢ)]×(CWᵢ / CW_target), where DBHᵢ is the diameter at breast height of the i-th adjacent tree, Dᵢ is the horizontal distance, and CWᵢ is the crown width of the adjacent tree, quantifying the interference intensity of adjacent trees; The dynamic growth model is an improved Richards model, with the expression V(t) = K(θ) × (1 - e^(-b(θ) × t))^c, where t is the tree age in years, θ = [T, P, L, CI, S, N] is the vector of environmental and competitive factors, T is the annual average temperature in °C, P is the annual precipitation in mm, L is the duration of sunlight in h, S is the slope, and N is the soil organic matter content. c is a shape parameter, 1.2-1.3 for coniferous trees and 1.4-1.5 for broad-leaved trees, and remains constant. The upper limit of K(θ) accumulation growth and the growth rate parameter b(θ) are coupled by a multiple linear regression factor, expressed as: K(θ)=k0+k1×T+k2×P+k3×CI+k4×S+k5×N, b(θ)=b0+b1×T+b2×P+b3×L+b4×CI; The Levenberg-Marquardt algorithm was used to fit the coefficients, with ≤100 iterations, a convergence threshold of 1e-6, a model fit R² ≥0.85, and a prediction accuracy ≥85%, which is 10-15 percentage points higher than the traditional Richards model that only considers tree age prediction.

7. The forest single-piece digital twin data fusion processing technology and method according to claim 1, characterized in that, In S6, the incremental fusion algorithm adopts the process of "change region detection - local data update - global consistency verification": Change area detection: The Euclidean distance difference method was used to compare point clouds from multiple periods. The distance threshold for the trunk area was 0.05m and for the crown area it was 0.2m. Areas exceeding the threshold were identified as change areas, including areas with increased diameter at breast height, branch extension, crown expansion, and pest and disease stress. Local data update: The trunk area is updated incrementally using cylindrical fitting, and the canopy area is locally redrawn with an update resolution of 0.1m, only processing data in the changed areas; Global consistency verification: The local update model and the global model are registered using the ICP algorithm, and the verification error is ≤0.03m, ensuring the integrity of the model; The data update cycle is adjusted according to the growth stage: 1-3 months for young forests and 6-12 months for mature forests. Ground IoT sensor data is smoothed by Kalman filtering to remove sudden interference and abnormal data. Establish a status feedback mechanism. If the deviation between the model prediction and the measured value exceeds 10%, the parameter calibration process will be automatically triggered to adjust the growth model coefficients. Large-scale extrapolation is based on the UAV LiDAR single-tree model. A linear correlation model between satellite LiDAR forest stand parameters and single-tree volume is established. The Kriging spatial interpolation algorithm is used to extrapolate to the regional forest stand. Multiple 1km×1km verification sample areas are selected for verification. The extrapolation error is ≤15% to be considered qualified. A 50m×50m resolution volume distribution map is generated, which supports the connection with the forestry resource management system.

8. The forest single-piece digital twin data fusion processing technology and method according to claim 3, characterized in that, The core optimization of the improved FPFH-ICP algorithm is the initial value constraint of the tree top feature point, which avoids the ICP algorithm from getting trapped in a local optimum. Compared to the traditional FPFH-ICP algorithm, the registration time is reduced by more than 30%, and the average distance deviation of the registration accuracy is improved from 0.05m to within 0.03m, making it suitable for the dual-source data alignment needs of complex scenarios such as mountains and dense forests.

9. The forest single-piece digital twin data fusion processing technology and method according to claim 4, characterized in that, The hyperspectral-assisted correction mechanism adds a spectral feature temporal stability verification: by comparing multiple periods of hyperspectral data, the reflectance variation coefficient of each band is calculated, unstable bands with a variation coefficient > 0.2 are removed, and 6 to 8 core stable bands are retained for feature extraction. The diameter at breast height (DBH) was extracted using three cross-section fitting at 1.2m, 1.3m, and 1.4m. Abnormal cross-sectional radius values ​​were eliminated using the 3σ criterion. If the deviation exceeded the mean ±3σ, and the residual of a single cross-section fitting was >0.02m, 1.1m and 1.5m cross-section fittings were added to ensure that the DBH extraction accuracy was within ±3mm.

10. The forest single-piece digital twin data fusion processing technology and method according to claim 1, characterized in that, The technology and method for fusion processing of digital twin data of individual trees also include a quality assessment module for digital twins of individual trees and a multi-scenario application mechanism: The quality assessment covers geometric accuracy, attribute accuracy, and fusion efficiency. Geometric accuracy is compared with the measured contour using the CloudCompare tool. The average distance deviation of the point cloud is ≤0.05m. The measurement errors of tree height, diameter at breast height, and crown width are controlled within ±5cm, ±3mm, and ±10cm, respectively. The accuracy of attributes was verified using 30–50 independent standard trees. The mean absolute percentage error (MAPE) of volume inversion was ≤8%, and the prediction accuracy of the growth model was ≥85%. The fusion efficiency requirements are: single tree modeling time ≤ 10 min / tree, single-threaded CPU, main frequency 3.0 GHz, multi-period data update delay ≤ 1 h, and single batch processing capacity ≥ 1000 single trees.

Images