A forest volume calculation method based on 3D point cloud data and neural network

Abstract

The present invention discloses a forest stock calculation method based on three-dimensional point cloud data and a neural network. The method comprises preprocessing the point cloud data, including point cloud denoising, filtering and normalization, and removing redundant data such as surrounding ground objects, non-ground points, etc. other than single tree information; using a marker-controlled watershed algorithm based on morphological filtering to mark the center points of tree crowns to achieve single tree segmentation, and extracting single tree parameters such as single tree diameter at breast height, tree height, and crown width; establishing a neural network stock inversion model, which uses single tree parameters as an input layer and forest stock as an output layer, and uses an optimization algorithm to optimize the training process; using the stock inversion model to estimate the forest stock, inputting the diameter at breast height, tree height, and crown width in the sample plot inventory data into the model, and outputting the forest stock estimation result.

Description

Technical Field

[0001] The present invention belongs to the technical field of forest volume inversion, and in particular relates to a forest volume calculation method based on three-dimensional point cloud data and a neural network. Background Art

[0002] As the largest, most complex, and most ecologically complete natural ecosystem on land, forest ecosystems play an irreplaceable role in the global carbon cycle and climate regulation. Entering the 21st century, as ecological and environmental protection has become a core issue in global development, the strategic value of forest resources has become increasingly prominent. Estimating forest resources through scientific surveys has become a key technology for obtaining basic forestry data and supporting decision-making analysis. In the forest resource monitoring system, the accurate estimation of forest stock volume (FSV) has always been a research focus and technical difficulty. FSV, defined as the total volume of tree trunks within a specific area, is not only a core indicator of forest resource surveys, but is also strongly correlated with key ecological parameters such as forest biomass, biodiversity, and carbon storage, making it an important basis for measuring the quality of forest ecosystems.

[0003] Traditional stock volume survey methods (such as the timber volume table method and the standard tree method) face significant technical bottlenecks. On the one hand, they require significant manpower, time, and financial resources, and are limited by spatiotemporal sampling density. On the other hand, existing quantitative models are mostly based on linear parameters such as tree height and diameter at breast height (DBH), which makes it difficult to overcome the nonlinear effects of topographic factors such as altitude, slope, and aspect on stock volume. In particular, they have poor adaptability to tree species in different site conditions. With the advancement of remote sensing technology and intelligent algorithms, forest stock volume estimation is gradually transitioning towards multi-source data fusion and nonlinear modeling. Among them, the BP algorithm, with its powerful nonlinear mapping capabilities, simple data preprocessing process, and high prediction accuracy, has demonstrated significant advantages over traditional linear models, the K-nearest neighbor algorithm, and random forest methods, and is widely used in the field of dynamic forest resource monitoring.

[0004] However, traditional BP algorithms still suffer from inherent flaws in practical applications: the randomness of initial weights and thresholds leads to unstable prediction results, the gradient descent mechanism can easily lead to slow convergence, and the algorithm lacks global search capabilities, often falling into local minima during training. This severely limits the robustness and generalization of the accumulation estimation model. In existing technologies, the key to overcoming current technical bottlenecks is to deeply couple intelligent optimization algorithms with BP algorithms to construct an accumulation inversion model with adaptive parameter adjustment capabilities. Summary of the Invention

[0005] The purpose of the present invention is to provide a forest volume calculation method based on three-dimensional point cloud data and neural network, which can effectively solve the problems existing in the above-mentioned prior art.

[0006] The present invention adopts the following technical solution: a forest volume calculation method based on three-dimensional point cloud data and neural network, which specifically includes the following steps:

[0007] Step (1) preprocessing the point cloud data, including point cloud denoising, filtering and normalization operations, to remove redundant data such as surrounding ground objects and non-ground points other than single tree information;

[0008] Step (2) Mark the center points of the tree crowns using the marker-controlled watershed algorithm based on morphological filtering to achieve single tree segmentation and extract single tree parameters such as single tree diameter at breast height, tree height, and crown width;

[0009] Step (3) establishing a neural network stock volume inversion model, wherein the model uses the single tree diameter at breast height, tree height, and crown width as the input layer, and the forest stock volume as the output layer, and uses an adaptive optimization algorithm to optimize the training process;

[0010] Step (4) estimates the forest volume using the volume inversion model, inputs the single tree parameters obtained by extracting the single tree from the pre-processed point cloud data through single tree segmentation into the model, and outputs the forest volume estimation result.

[0011] Furthermore, the implementation process of step (1) is as follows:

[0012] (1.1) In CloudCompare software, a Euclidean distance-based denoising algorithm was used to perform denoising. Outliers were identified and removed by setting a distance threshold, and significant noise not detected by the algorithm was removed through manual interaction.

[0013] (1.2) Use cloth filtering algorithm to separate ground points and non-ground points in point cloud data;

[0014] (1.3) The denoised point cloud is normalized by subtracting the ground point elevation obtained by filtering to obtain normalized point cloud data without terrain. The Z value of each point cloud represents the vertical height from this point to the ground.

[0015] Furthermore, the implementation process of step (2) is as follows:

[0016] (2.1) Use the gradient operator to calculate the point cloud gradient image, and then perform corrosion, expansion, opening and closing operations on the gradient image to obtain the single tree segmentation result;

[0017] (2.2) The height of a single tree is obtained by subtracting the Z coordinate of the ground point cloud from the maximum Z coordinate of a single vertex in the single tree point cloud data;

[0018] (2.3) The DBH of a single tree was obtained by performing ellipse fitting on the point cloud at DBH height (1.3 m) using the least squares fitting method;

[0019] (2.4) The crown width is calculated by using the convex hull algorithm. The point cloud data of a single tree is projected onto a two-dimensional plane. The two-dimensional convex polygon of the single tree is calculated from the discrete point data of the two-dimensional plane. The convex points on the two-dimensional convex polygon are connected in pairs, and the distance between the convex points is calculated. The maximum distance between the convex points is selected as the crown width of the single tree. At the same time, the rationality of the crown width is further verified by calculating the area of ​​the convex polygon.

[0020] Furthermore, the implementation process of step (3) is as follows:

[0021] (3.1) The BP (Back Propagation) algorithm is used as the basic algorithm of the neural network, with single tree diameter at breast height, tree height, and crown width as the input layer, and forest volume as the output layer;

[0022] The network model based on the BP algorithm is a feedforward neural network that adjusts thresholds and weights through error feedback, gradually self-corrects and optimizes, so that the model's estimated results and expected outputs continue to approach each other. This model mainly adjusts neuron weights and optimizes the network structure through forward and reverse calculations.

[0023] Forward calculation: The input layer is processed layer by layer through the hidden layer and then transmitted to the output layer. The state of each layer of neurons is only related to the neurons in the previous layer.

[0024] Reverse calculation: propagate the error signal backward along the original connection channel, and adjust the weights of each neuron layer by layer according to the error feedback to make the estimation error continuously smaller. When the error no longer decreases with the number of iterations or drops to a certain value, the model is considered to have converged.

[0025] (3.2) Construct an adaptive optimization neural network inversion model, introduce the adaptive optimization mechanism into the BP algorithm, and use a dynamic adaptive global search algorithm to optimize the weights and threshold initial solution space of the BP network;

[0026] The adaptive optimization mechanism calculates the particle fitness variance in real time and adaptively adjusts the search weight factor, making the optimization process dynamically balanced between global exploration and local refinement, solving the problem that traditional BP algorithms are prone to falling into local extreme values.

[0027] Specifically, the BP algorithm weights and thresholds are mapped to particle position parameters, the training sample output error is used as the fitness function, the particle speed and position are iteratively updated, and the optimal initial parameter combination is searched in the weight space;

[0028] (3.3) Model training process;

[0029] First, the adaptive optimization mechanism is used to globally optimize the initial parameters of the BP network, and the search range is converged to the local optimal solution neighborhood; then, the optimized parameters are used as the initial values ​​of the BP network, and the gradient descent algorithm is used for iterative training to construct a forest volume inversion model with adaptive parameter optimization capabilities.

[0030] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0031] 1. The adaptive optimization mechanism introduced effectively solves the problem of traditional BP algorithms being prone to falling into local extremes. By dynamically adjusting the inertia factor and search weight, the model can maintain stable convergence under different forest types and point cloud density conditions.

[0032] 2. High transferability. By mapping 3D point cloud data with plot inventory data, a nonlinear correlation model is established between individual tree parameters such as DBH, tree height, and crown width and plot-scale stock volume. Through the feature transfer learning capabilities of neural networks, stock volume inversion from plots to large-scale areas is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0034] Figure 1 It is a schematic diagram of the process of the present invention;

[0035] Figure 2 This is a diagram of the single tree segmentation result obtained in this embodiment;

[0036] Figure 3 Schematic diagram of the adaptive optimization neural network inversion model;

[0037] Figure 4 This is the forest volume inversion result diagram obtained in this example. DETAILED DESCRIPTION

[0038] To further clarify the objectives and advantages of the present invention, the present invention is described in detail below with reference to embodiments. It should be understood that the following description merely describes a forest volume calculation method based on three-dimensional point cloud data and a neural network, or several specific implementations thereof, and does not strictly limit the scope of protection claimed herein.

[0039] Example: Figure 1 As shown, a forest volume calculation method based on three-dimensional point cloud data and neural network includes the following steps:

[0040] Step (1) preprocessing the point cloud data, including point cloud denoising, filtering and normalization operations, to remove redundant data such as surrounding ground objects and non-ground points other than single tree information;

[0041] Wherein, step (1) includes the following steps:

[0042] (1.1) In the CloudCompare software, a Euclidean distance-based denoising algorithm is used for denoising. Outliers are identified and removed by setting a distance threshold, and significant noise not detected by the algorithm is removed through manual interaction. In a preferred embodiment, the distance threshold can be set to 1.5 times the average point spacing. The k-nearest neighbors (k=10) of each point are searched using a KD tree, the Euclidean distance is calculated, and outliers with a distance greater than the threshold (such as isolated noise points and non-vegetation points such as birds) are marked and deleted. Manual interactive denoising involves manually removing significant noise missed by the algorithm (such as power lines and building remnants) by selecting and deleting them using the CloudCompare "Selection Tool". In this embodiment, both automatic and manual denoising are used to more accurately process point cloud data.

[0043] (1.2) Use cloth filtering algorithm to separate ground points and non-ground points in point cloud data;

[0044] The cloth filtering algorithm flips the point cloud. If a piece of cloth falls from above due to gravity, the final falling cloth can represent the current terrain.

[0045] (1.3) The denoised point cloud is normalized by subtracting the ground point elevation obtained by filtering to obtain normalized point cloud data without terrain. The Z value of each point cloud represents the vertical height from the point to the ground, eliminating the influence of terrain undulation on the extraction of single tree parameters.

[0046] Step (2) uses the marker-controlled watershed algorithm based on morphological filtering to mark the center point of the tree crown, realize single tree segmentation, and extract single tree parameters such as single tree diameter at breast height, tree height, and crown width;

[0047] Wherein, step (2) includes the following steps:

[0048] (2.1) Use the gradient operator to calculate the point cloud gradient image, and then perform corrosion, expansion, opening and closing operations on the gradient image to obtain the single tree segmentation result;

[0049] (2.2) The height of a single tree is obtained by subtracting the Z coordinate of the ground point cloud from the maximum Z coordinate of a single vertex in the single tree point cloud data;

[0050] (2.3) The DBH of a single tree was obtained by performing ellipse fitting on the point cloud at DBH height (1.3 m) using the least squares fitting method;

[0051] (2.4) Crown width: Using the convex hull algorithm, the individual tree point cloud data is projected onto a two-dimensional plane. A two-dimensional convex polygon is calculated from the discrete point data on the two-dimensional plane. The convex points on the two-dimensional convex polygon are connected in pairs, and the spacing between the convex points is calculated. The maximum spacing between the selected convex points is used as the crown width of the individual tree. The rationality of the crown width is further verified by calculating the area of ​​the convex polygon. Data with a ratio of area to square of crown width between 0.5 and 1.0 are considered normal. Otherwise, data are considered abnormal and excluded from the subsequent neural network accumulation inversion model.

[0052] In the pre-processed point cloud data, some single tree point clouds are extracted as follows Figure 2 shown.

[0053] The tree height and diameter at breast height of each plot vegetation were extracted from the extracted single tree point cloud data, and the average values ​​of tree height, diameter at breast height and crown width of each plot were calculated as shown in Table 1.

[0054] Table 1 Vegetation parameters of the sample plot

[0055] Plot number species Average tree height / m Average breast diameter / cm Average crown width / cm 1 bamboo 16.8 9.81 82.9 2 bamboo 15.65 8.17 75.7 3 bamboo 18.46 8.92 82.7 4 Chinese fir 13.18 15.57 124.3 5 Chinese fir 14.54 19.68 165.3 6 Chinese fir 9.78 5.16 29.6 7 Chinese fir 15.48 16.17 139.3 8 mixed coniferous forest 19.21 8.72 83.25 9 mixed coniferous forest 25.17 11.51 140.95 10 mixed coniferous forest 24.84 10.87 132.9 11 mixed coniferous forest 20.93 11.28 117.45 12 mixed coniferous forest 16.75 10.84 92.15 13 Broadleaved mixed forest 29.31 15.89 352.3 14 Broadleaved mixed forest 17.88 11.52 281.7 15 Broadleaved mixed forest 11.59 9.48 246.19 16 Mixed coniferous and broad-leaved forest 8.72 7.55 95.28

[0056] Step (3) establish a neural network volume inversion model, which uses the single tree diameter at breast height, tree height, and crown width as the input layer and the forest volume as the output layer, and uses an adaptive optimization algorithm to optimize the training process;

[0057] Wherein, step (3) includes the following steps:

[0058] (3.1) The BP (Back Propagation) algorithm is used as the basic algorithm of the neural network, with single tree diameter at breast height, tree height, and crown width as the input layer, and forest volume as the output layer;

[0059] The network model based on the BP algorithm is a feedforward neural network that adjusts thresholds and weights through error feedback, gradually self-corrects and optimizes, so that the model's estimated results and expected outputs continue to approach each other. This model mainly adjusts neuron weights and optimizes the network structure through forward and reverse calculations.

[0060] Forward calculation: Data is processed layer by layer through the input layer and then transmitted to the output layer. The state of each neuron is only related to the neurons in the previous layer. The number of neurons in the input layer is 3, corresponding to the tree diameter at breast height, tree height, and crown width. The number of neurons in the hidden layer is 5, with 6, 5, 4, 3, and 2, respectively. The number of neurons in the output layer is 1, corresponding to the forest volume.

[0061] Reverse calculation: propagate the error signal backward along the original connection channel, and adjust the weights of each neuron layer by layer according to the error feedback to make the estimation error continuously smaller. When the error no longer decreases with the number of iterations or drops to a certain value, the model is considered to have converged.

[0062] (3.2) Construct an adaptive optimization neural network inversion model, introduce the adaptive optimization mechanism into the BP algorithm, and use a dynamic adaptive global search algorithm to optimize the weights and threshold initial solution space of the BP network;

[0063] The adaptive optimization neural network inversion model is shown in the figure Figure 3 As shown in the figure, the adaptive optimization mechanism calculates the particle fitness variance in real time and adaptively adjusts the search weight factor, so that the optimization process dynamically balances between global exploration and local refinement, solving the problem that the traditional BP algorithm is prone to falling into local extreme values.

[0064] Specifically, the BP algorithm weights and thresholds are mapped to particle position parameters, the training sample output error is used as the fitness function, the particle speed and position are iteratively updated, and the optimal initial parameter combination is searched in the weight space;

[0065] The principle of adaptive optimization algorithm is:

[0066] A group of particles are randomly initialized in the solution space. After the algorithm starts, the particles begin to move in the solution space according to their own speed and fitness value. During the movement, the flight direction and speed of the particles are constrained by two optimal values, namely the individual optimal value P_best and the population optimal value G_best. The individual optimal value is the optimal fitness value searched by the individual particle so far, and the population optimal value is the optimal fitness value searched by the population of all particles so far. The specific steps of the algorithm are as follows:

[0067] 1) Set the number of particles, initial velocity and position, acceleration constants c1 and c2, inertia factor ω, optimization function and solution space in the adaptive optimization algorithm.

[0068] The designed adaptive optimization algorithm has an initial particle number of 50. Generally speaking, if the particle number is too small, it will fall into the local optimum and cannot find the global optimal solution. Too many ion beams will increase the computational complexity. In this embodiment, the initial particle number is preferably 50. The initial position and velocity are randomly set; the acceleration constants c1=1.5 and c2=1.2; the initial value of the inertia factor ω is calculated by the following three formulas:

[0069]

[0070]

[0071]

[0072] Among them, X is the fitness variance of the particle population; is the average particle fitness of the population.

[0073] 2) The fitness of the particle is calculated by the optimization function and the individual and population optimal values ​​are obtained by comparison. At the same time, the position and fitness of the particle after each iteration are saved in the individual optimal value of the particle, and the population optimal value of all particles is updated.

[0074] 3) In each iteration, the speed and position of each particle are changed by the constraints of individual optimal value, population optimal value and optimization function. i represents the current position of the i-th particle, p i represents the individual optimal value of the i-th particle, p g Represents the current optimal value of the population, and the iteration formula is:

[0075]

[0076]

[0077] Among them, i=1,2,3,…, represents the particle number; d=1,2,3,…, represents the latitude number; the non-negative real inertia factor ω is used to control the number of particles that inherit the current velocity; r1 and r2 are two random numbers between 0 and 1; c1 and c2 are acceleration factors; α is a constraint factor used to control the weight of the velocity.

[0078] In the adaptive optimization algorithm, the inertia factor ω is adjusted based on the particle fitness value. If the particle fitness is high after a certain iteration, its inertia factor is reduced to find the local optimal value; if the particle fitness is low after a certain iteration, its inertia factor is increased to find the global optimal value. The value of ω is adaptively adjusted using the following formula:

[0079]

[0080] Among them, ω min and ω max are the minimum and maximum values ​​of the inertia factor, where ω max Not more than 1.5, ω min Not less than ω s ; Preferably, ω min is 0.3, ω max is 1.2; ω s It is used to correct the size of the inertia factor to avoid premature termination of the algorithm. Its value is set to 0.1. δ is the evolution factor, which represents the relationship between the current number of iterations and the maximum number of iterations. The following formula is used to calculate δ.

[0081]

[0082] Where t is the current iteration number; T is the maximum iteration number; and s is an integer greater than 1.

[0083] 4) Compare the optimal value of the whole system obtained in step 2) after each iteration with the saved optimal value of the whole system. If it is better, update it; otherwise, proceed to the next iteration.

[0084] 5) Compare the optimal value obtained after the iteration with the optimal value set before the algorithm starts. If the optimal value has been reached or the number of iterations reaches the set upper limit, stop the iteration.

[0085] (3.3) Model training process;

[0086] First, the adaptive optimization mechanism is used to globally optimize the initial parameters of the BP algorithm, and the search range is converged to the local optimal solution neighborhood; then the optimized parameters are used as the initial values ​​of the BP algorithm, and the gradient descent algorithm is used for iterative training to construct a forest stock inversion model with adaptive parameter optimization capabilities.

[0087] Step (4) uses the stock volume inversion model to estimate the forest stock volume, inputs the diameter at breast height, tree height, and crown width in the plot inventory data into the model, and outputs the forest stock volume estimation result.

[0088] The accumulation of small classes was calculated based on the cloud data of sample points and the remote sensing inversion model, and the results are shown in Table 2.

[0089] Table 2 Estimation results of plot volume

[0090] Plot number 2017 accumulated m³ / hm² (small class data) 2022 accumulation m³ / hm² (point cloud computing) 1 3.36 4.81 2 2.18 3.96 3 17.32 22.54 4 15.36 20.17 5 9.07 15.79 6 15.79 18.85 7 7.85 10.59 8 3.29 5.12 9 6.98 10.17 10 9.85 17.26 11 6.72 10.05 12 5.29 6.21 13 10.44 20.98 14 7.65 12.65 15 3.38 5.16 16 11.29 18.25

[0091] Overall, from 2017 to 2022, the forest stock in the sample areas increased to varying degrees, with an increase of 17.39% to 100.96%. Figure 4 This is the inversion result of forest stock in 2022.

Claims

1. A method for calculating forest volume based on three-dimensional point cloud data and neural networks, characterized by: The steps include: Step (1) point cloud preprocessing: denoising, filtering and normalizing the acquired three-dimensional point cloud data to remove redundant data irrelevant to the individual tree information; Step (2) Single tree segmentation and parameter extraction: the center point of the crown is marked using the marker-controlled watershed algorithm based on morphological filtering to achieve single tree segmentation and extract the parameters of single tree diameter at breast height, tree height, and crown width; Step (3) building and training a stock volume inversion model, constructing a BP neural network model with single tree diameter at breast height, tree height, and crown width as the input layer and forest volume as the output layer, and using an adaptive optimization algorithm to optimize the initial weights and thresholds of the BP neural network model to improve the model convergence and generalization ability; Step (4) Stock volume estimation: using the stock volume inversion model to estimate the forest stock volume, the diameter at breast height, tree height, and crown width in the plot inventory data are input into the model, and the forest stock volume estimation result is output; The adaptive optimization algorithm in step (3) is an improved particle swarm optimization algorithm, which optimizes the initial parameters of the BP network through the following mechanism: Mechanism 1: Map the weights and thresholds of the BP neural network to the particle position parameters in the particle swarm optimization, and use the output error of the training sample as the fitness function; Mechanism 2: By calculating the fitness variance of the particle population in real time: Where N is the number of particles, f avg is the average particle fitness of the population, dynamically adjusting the search weight factor; i is the particle number, and d is the dimension number of the particle position; Mechanism 3: Adaptively adjust the inertia factor ω based on the particle fitness value: if the particle fitness is high, reduce ω to focus on local refinement; if the fitness is low, increase ω to enhance global search. The specific formula is: oh = oh min +(ω max -oh min )ed+(1-e)ω s Among them, ω min and ω max is the minimum and maximum value of the inertia factor; ω s =0.1 is the correction factor; δ is the evolution factor, t is the current iteration number, T is the maximum iteration number, s is an integer greater than 1; ε = sin(arctanX); Mechanism 4: The particle velocity update formula is: The particle position update formula is: Where α is the constraint factor; c1 and c2 are acceleration constants, and c1 = 1.5, c2 = 1.2; r1, r2 are two random numbers between 0 and 1, and x is used to calculate the value of the acceleration constant. i represents the current position of the i-th particle, p i represents the individual optimal value of the i-th particle, p g Indicates the current optimal value of the population; The training process of the adaptive optimization algorithm in step (3) is as follows: (3.1) Initialize the particle swarm parameters: the number of particles is 50, the initial position and velocity are randomly set, and the initial value of the inertia factor is obtained by calculating the fitness variance; (3.2) By iteratively updating the particle velocity and position, the optimal initial parameter combination is searched in the weight space. When the fitness error no longer decreases with the number of iterations or reaches the set threshold, the global optimization is stopped; (3.3) The optimized weights and thresholds are used as the initial values ​​of the BP algorithm, and local refinement training is performed through the gradient descent algorithm to construct a forest volume inversion model.

2. The forest volume calculation method based on three-dimensional point cloud data and neural network according to claim 1 is characterized in that: The step (1) comprises: (1.1) Denoising was performed using the Euclidean distance-based denoising algorithm in CloudCompare software. Outliers were identified and removed by setting a distance threshold, and manual interaction was used to remove significant noise not detected by the algorithm. (1.2) Use cloth filtering algorithm to separate ground points from non-ground points in point cloud data, and simulate the falling of cloth under gravity to fit the terrain; (1.3) By subtracting the denoised point cloud from the filtered ground point elevation and performing a normalization operation, a normalized point cloud data without terrain is obtained, where the Z value of the point cloud represents the vertical height from this point to the ground.

3. The forest volume calculation method based on three-dimensional point cloud data and neural network according to claim 1 is characterized in that: The step (2) comprises: (2.1) Use the gradient operator to calculate the point cloud gradient image, and perform corrosion, expansion, opening and closing operations on the gradient image to obtain the single tree segmentation result; (2.2) The height of a single tree is obtained by subtracting the Z coordinate of the ground point cloud from the maximum Z coordinate of a single vertex in the single tree point cloud data; (2.3) The DBH of a single tree was obtained by performing ellipse fitting on the point cloud at a DBH height of 1.3 m using the least squares fitting method; (2.4) The crown width of a single tree is calculated by using the convex hull algorithm. The point cloud data of the single tree is projected onto a two-dimensional plane. The two-dimensional convex polygon of the single tree is calculated from the discrete point data of the two-dimensional plane. The convex points on the two-dimensional convex polygon are connected in pairs, and the distance between the convex points is calculated. The maximum value of the distance between the convex points is selected as the crown width of the single tree, and the rationality of the crown width is verified by calculating the area of ​​the convex polygon.

4. The forest volume calculation method based on three-dimensional point cloud data and neural network according to claim 1, characterized in that: In step (3), the BP neural network model adjusts the neuron weights through forward calculation and reverse calculation to optimize the network structure; The forward calculation is processed layer by layer from the input layer to the hidden layer and then transmitted to the output layer. The state of each layer of neurons is only related to the neurons in the previous layer. The reverse calculation is to propagate the error signal backward along the original connection channel, and adjust the weight of each neuron layer by layer according to the error feedback to make the estimation error continuously smaller. When the error no longer decreases with the number of iterations or drops to a certain value, the model is considered to have converged.

5. The method for calculating forest volume based on three-dimensional point cloud data and neural network according to claim 1, characterized in that: The implementation process of step (4) is as follows: The stand tree height, diameter at breast height, and crown width parameters were extracted from the plot inventory data and input into the adaptively optimized BP algorithm to obtain the forest volume estimation results.

Images