Accurate prediction method of tree diameter at breast height volume based on optimized fuzzy deep network

Abstract

The present invention discloses a method for accurately predicting the diameter at breast height and volume of trees based on an optimized fuzzy deep network, comprising: preprocessing forest site cloud data and segmenting individual trees; obtaining tree parameters of individual trees; establishing a tree parameter prediction network, the tree parameter prediction network comprising a fuzzy deep network and a pigeon flock optimization module, training the tree parameter prediction network, outputting predicted values ​​after training and transmitting them to the pigeon flock optimization module, the pigeon flock optimization module updating the parameters of the fuzzy deep network and completing the search for optimal parameters, the fuzzy deep network completing adaptive training based on the optimal parameters, and establishing a tree parameter prediction model; the present invention develops a tree parameter prediction model, proposes an adaptive algorithm to enhance the generalization ability of the tree parameter prediction model for different tree varieties, embeds an attention mechanism module to enhance the robustness of the network, integrates the pigeon flock optimization algorithm to adjust the parameters of the fuzzy deep network in real time, and further improves the prediction accuracy and learning ability of the model.

Description

Technical Field

[0001] The present invention belongs to the technical field of forest parameter research, and specifically relates to a method for accurately predicting the diameter at breast height and volume of trees based on an optimized fuzzy depth network. Background Art

[0002] Accurate prediction of tree diameter at breast height (DBH) and volume plays a crucial role in forest resource surveys, national strategic timber reserves, and carbon sink assessments. In recent years, methods for predicting tree parameters have been broadly categorized into two types: manual measurement and laser point cloud-based methods. While these two methods differ in data acquisition, they share certain similarities in their approach to building prediction models.

[0003] Forest parameter prediction methods based on individual tree measurements typically establish candidate models to estimate tree parameter values. These models, combined with tree height residuals, establish a tree height-diameter-at-breast-root (DBH) prediction model. Alternatively, stepwise regression and partial least squares methods are used to establish four candidate models to estimate forest volume. Alternatively, nonlinear mixed-effects models are used to predict plantation crown volume, crown surface area, and biomass. Such studies often select a number of commonly used growth models or related extended models as candidate models, screening them using a series of evaluation criteria. However, the type and number of candidate models often vary depending on the research objectives.

[0004] With the rapid development of laser measurement technology and its application in forestry informatization, the method of obtaining tree parameters has shifted from traditional tree-by-tree measurement to the generation of 3D laser point clouds. Laser scanning offers exceptionally high range detection capabilities and stability. Previous work has also combined computer graphics and machine learning theories to process and analyze 3D point clouds of trees. For example, some progress has been made in using the watershed algorithm for crown segmentation of tree point clouds, establishing a tree species identification framework using support vector machines, estimating individual tree diameters at breast height using randomized Hough transforms and octree segmentation, and extracting tree height based on growth direction.

[0005] With the continuous breakthroughs in artificial intelligence (AI) and the maturity of laser point cloud scanning technology, machine learning techniques have emerged in forest point cloud processing, leading to the emergence of several tree parameter prediction models. For example, a forest parameter prediction model proposed in 2020 proposed a hierarchical estimation approach that combines forward models from synthetic aperture radar (SAR), LiDAR, and passive optical systems to generate geometric and electromagnetic models of real forest stands, combining these to estimate forest parameters. A diameter-at-breast (DBH) prediction model proposed in 2021 uses a generalized nonlinear mixed-effects approach, adding site-level random effects to improve model generalization. It also uses variables at the individual tree and laser point cloud levels as predictors, achieving high prediction accuracy. Some of this work directly draws information from forest point clouds and establishes prediction models based on stand-scale parameters. For example, methods such as k-nearest neighbor prediction models, allometric models, prediction models that integrate point cloud features, and limited area growth have been used to predict aboveground biomass across large forest areas in Sweden, study the relationship between DBH and forest biomass in tropical forests, calculate canopy cover in ginkgo plantations, and automatically segment forest stands in Dayekou, Qilian Mountains. Some studies have also performed single-tree segmentation on forest point clouds to obtain single-tree parameters, and then used random forests, support vector machines, and BP neural networks to establish a single-tree DBH prediction model for larch-fir mixed forests, constructed a basal area growth model for Dagangshan artificial forests, and created a tree height prediction model for fir trees in Lexian Forest Farm.

[0006] Although machine learning has made certain achievements in the prediction of forest parameters, the following problems still exist: 1) In actual forests, some individuals are affected by natural disasters or forest competition and become necrotic, oppressed, etc., which become abnormal samples. Most forest parameter prediction models lack a mechanism to judge abnormal samples. The models are easily disturbed by these samples, and their prediction accuracy needs to be improved; 2) Prediction models based on machine learning use trial and error, empirical formulas and other methods to determine model parameters. They cannot make further selections on the number of algorithm parameters and rely on prior experience. There is still room for improvement in the structure and modules of relevant prediction models; 3) Real-time acquisition of forest growth parameters is beneficial to the planting and cultivation of large-scale artificial forests, but there is currently not much related work on the combination of artificial intelligence and airborne lidar for forest parameter prediction. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method for accurately predicting the diameter at breast height and volume of trees based on an optimized fuzzy deep network in response to the shortcomings of the existing technology. This method for accurately predicting the diameter at breast height and volume of trees based on an optimized fuzzy deep network develops a tree parameter prediction model, establishes a nonlinear relationship between tree parameters, proposes an adaptive algorithm to enhance the generalization ability of the tree parameter prediction model for different tree varieties, embeds an attention mechanism module to enhance the robustness of the network, and integrates a pigeon flock optimization algorithm to adjust the parameters of the fuzzy deep network in real time, so as to further improve the prediction accuracy and learning ability of the model.

[0008] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0009] The accurate prediction method of tree diameter at breast height volume based on optimized fuzzy deep network includes:

[0010] Step 1: Obtain forest point cloud data using airborne lidar;

[0011] Step 2: De-noise the point cloud data, filter the point cloud data using the point cloud ground point filtering method, and perform single tree segmentation on the filtered point cloud data;

[0012] Step 3: Based on the segmented point cloud of individual trees, obtain the forest parameters of individual trees, including east-west crown width, north-south crown width, tree height, point cloud density, and crown volume;

[0013] Step 4: Use manual mapping methods to obtain the diameter at breast height and volume of the corresponding trees;

[0014] Step 5: Take the east-west crown width, north-south crown width, tree height, point cloud density, crown volume, diameter at breast height and timber volume of multiple trees as training sample datasets;

[0015] Step 6: Establish a tree parameter prediction network, which includes a fuzzy deep network and a pigeon flock optimization module. The tree parameter prediction network is trained using a training sample data set. The input of the tree parameter prediction network is the tree's east-west crown width, north-south crown width, tree height, point cloud density and crown volume, and the output is the tree's diameter at breast height and volume. After the fuzzy deep network training is completed, the predicted value is output and transmitted to the pigeon flock optimization module. The pigeon flock optimization module updates the parameters of the fuzzy deep network. When the pigeon flock optimization module completes the optimal parameter search of the fuzzy deep network, the fuzzy deep network completes adaptive training based on the optimal parameters to establish the final tree parameter prediction model.

[0016] Step 7: Collect the point cloud data of the forest to be tested and obtain the east-west crown width, north-south crown width, tree height, point cloud density and crown volume of each tree in the forest to be tested according to the methods of steps 2 and 3. Input the east-west crown width, north-south crown width, tree height, point cloud density and crown volume of the trees into the final forest parameter prediction model to obtain the predicted values ​​of the diameter at breast height and volume of the corresponding trees.

[0017] As a further improved technical solution of the present invention, the method for obtaining the forest parameters of individual trees in step 3 is:

[0018] The maximum distance in the east-west direction of the crown point cloud of a single tree is selected as the east-west crown width; the maximum distance in the north-south direction of the crown point cloud of a single tree is selected as the north-south crown width; the vertical distance between the highest point of the crown point cloud of a single tree and the horizontal plane is the tree height; the total number of point clouds of a single tree divided by the projected area of ​​the crown is the point cloud density; the convex hull volume of the crown point cloud of a single tree is calculated, and the convex hull volume is the crown area.

[0019] As a further improved technical solution of the present invention, the method for obtaining the diameter at breast height and volume of the tree in step 4 is:

[0020] At the trunk position of the tree 1.3m above the ground, use a tape measure to obtain the circumference of the trunk, which is the diameter at breast height of the tree;

[0021] The upper diameter of the trunk of the selected tree is The point is the shape point, D is the tree diameter at breast height, and the length from the shape point to the treetop is H. t The cross-sectional area of ​​the diameter at breast height is S D , the tree height is measured as H, then the volume of the tree is:

[0022]

[0023] Where: r is the stem shape index; S vl The volume of trees.

[0024] As a further improved technical solution of the present invention, the fuzzy deep network in step 6 includes an adaptive fuzzy layer, a fuzzy reasoning layer, and an attention-based weight update layer from input to output.

[0025] As a further improved technical solution of the present invention, the calculation process of the adaptive fuzzy layer is:

[0026] Input attribute x u There are k fuzzy subsets in the adaptive fuzzy layer, u∈{1,2,3,4,5}, where x 1 、x 2 、x 3 、x 4 、x 5The input attributes of the adaptive fuzzy layer are the east-west crown width, north-south crown width, tree height, point cloud density and crown area. Each input attribute has k fuzzy subsets. The adaptive fuzzy layer has a total of 5×k fuzzy subsets. The training samples Input attribute x into the adaptive fuzzy layer u The jth fuzzy subset to which it belongs, outputs the membership degree As shown in formula (2):

[0027]

[0028] Where: x u The center of the jth fuzzy subset in the adaptive fuzzy layer is x u The variance of the jth fuzzy subset in the adaptive fuzzy layer is j = 1, ..., k, the total number of fuzzy subsets of the input attributes is k, and k is updated by the pigeon flock optimization module, i = 1, ..., n, n is the number of training samples, that is, the total number of trees in the training sample;

[0029] According to x u Solve for n training samples The local density and distance i=1,…,n,u∈{1,2,3,4,5},local density The calculation formula is:

[0030]

[0031] Where: yes The local density, d u is x u The cutoff distance, u∈{1,2,3,4,5}, i=1,2,…,n, m=1,2,…,n;

[0032] Calculate the local density greater than All training samples and The Euclidean distance of distance As shown in formula (4):

[0033]

[0034] Where: yes The local density and m=1,2,…n; for and The Euclidean distance between them, u∈{1,2,3,4,5}; if The local density of is the largest, then And m≠i;

[0035] calculate As input attribute x u The probability value of the initial center of clustering As shown in formula (5):

[0036]

[0037] Where: max(ρ u ) is ρ u The maximum value of n local densities; min(ρ u ) is ρ u The minimum value of n local densities; max(δ u ) is δ u The maximum value of n distances; min(δ u ) is δ u The minimum value of n distances; ρ u Take n training samples as input attributes x u The local density, including n local densities; δ u Take n training samples as input attributes x u The distance, including n distances;

[0038] Step (a), input attribute x u Among the n training samples according to Sort the values ​​in descending order and take the first k sample points as x u The initial center point of DPKM clustering for n training samples; Step (b), calculate the input attribute x u The Euclidean distance between the n training samples and the k class centers is calculated. According to the minimum distance distribution principle, the training samples are assigned to the class with the nearest center. After the n training samples are assigned once, the mean of each class is calculated, and the mean is used as the new class center and the class center is updated. Repeat the above steps (b) and (c) until the change of the class center is less than the set error. The DPKM clustering is completed. At this time, the center of the k clusters is x u The center of the jth fuzzy subset in the adaptive fuzzy layer j=1,…,k;

[0039] When the input attribute x u The center of the jth fuzzy subset in the adaptive fuzzy layer as well as DPKM cluster After establishment, calculate x according to the following adaptive algorithm uThe variance of the jth fuzzy subset in the adaptive fuzzy layer

[0040] First calculate x u The center of the k fuzzy subset centers As shown in formula (6):

[0041]

[0042] Then solve the clustering Internal samples and The average Euclidean distance As shown in formula (7):

[0043]

[0044] Where: express Internal samples and Euclidean distance; for Internal samples the number of

[0045]

[0046] Where: for The width factor of α is the variance scaling factor, which is updated by the pigeon swarm optimization module. For input attribute x u The maximum distance between the centers of the k fuzzy subsets; u∈{1,2,3,4,5}, j=1,2,…,k.

[0047] As a further improved technical solution of the present invention, the calculation process of the fuzzy reasoning layer is:

[0048] The membership degree is the input of the fuzzy reasoning layer. The product reasoning method is used to establish the fuzzy unit of the fuzzy reasoning layer and calculate the output value of the jth unit of the i-th training sample. As the fuzzy unit output, as shown in formula (9):

[0049]

[0050] Where: is the product of the membership degree of the jth fuzzy subset of the i-th training sample under different input attributes; After normalization, u∈{1,2,3,4,5}, i=1,…,n,j=1,2,…,k, n is the total number of trees in the training sample, and the total number of fuzzy units is k.

[0051] As a further improved technical solution of the present invention, the calculation process of the attention-based weight update layer is:

[0052] The loss function is shown in formula (10):

[0053]

[0054] Where: y 1 is the predicted value of DBH, y 2 is the predicted value of timber volume; is the measured value of DBH, is the measured value of timber volume; Q i is the attention weight of the i-th training sample, i = 1, 2, ..., n;

[0055] The output of the fuzzy inference layer The connection weight between the diameter at breast height is The connection weight between and volume is Then for all training samples:

[0056]

[0057] Where: represents the jth output of the i-th training sample from the fuzzy inference layer value; represents the predicted value of DBH of the i-th training sample, represents the volume prediction value of the i-th training sample; j = 1, 2, ..., k, i = 1, 2, ..., n, n is the total number of training samples, and the total number of fuzzy units is k;

[0058] The training samples are normalized and preprocessed, and then the parameter matrix of the i-th tree in the training sample is Contains 7 measured tree parameters, the tree parameter matrix z i From left to right in the middle are the east-west crown width, north-south crown width, tree height, point cloud density, crown volume, timber volume and diameter at breast height. The average east-west crown width of all training samples is avg 1 The average north-south crown width of all training samples is avg 2 , the average tree height of all training samples is avg 3 , the average point cloud density of all training samples is avg 4 , the average crown volume of all training samples is avg 5 , the average volume of all training samples is avg 6 , the average DBH value of all training samples is avg 7The average value matrix of each tree parameter in the training sample is avg = [avg 1 ,…,avg 7 ], Q i The calculation method of is shown in formula (12):

[0059]

[0060] Where: Represents the tree parameter matrix z i The pth tree parameter from left to right, avg p Represents the pth average value of the average matrix avg from left to right, p∈{1,2,3,4,5,6,7}; cosine similarity Sim(z i ,avg) and Euclidean distance Dist(z i ,avg) represent the intrinsic relationship between the forest parameters of the i-th tree and the average value avg; τ is the attention weight scaling coefficient, which is updated by the pigeon flock optimization module; i = 1, 2, ..., n;

[0061] The attention-based connection weights are continuously updated through back propagation, and the update method is shown in formula (13):

[0062]

[0063] Where: η ranges from 0 to 1, indicating learning efficiency; t is the current number of iterations;

[0064] The initial value of the attention-based connection weight is randomly given and continuously updated through backpropagation. When the number of iterations reaches the maximum, the update is terminated.

[0065] As a further improved technical solution of the present invention, the calculation process of the pigeon flock optimization module is:

[0066] The pigeon flock optimization module is used to optimize the parameters k, α, and τ in the fuzzy deep network. The combination of k, α, and τ is called the model parameter combination δ[k, α, τ]. In the parameter search space, k is 0 to 100 and must be an integer, α is 0 to 30, and τ is 0 to 10. The pigeon flock optimization module is initially composed of L groups of model parameters δ l [k l ,α l ,τ l ](l=1,2,…L), that is, the pigeon group optimization module has L groups of model parameters to optimize and t max iterations;

[0067] The lth group of model parameter combinations in the tth iteration is Its fitness is shown in formula (14):

[0068]

[0069] Where: t is the current iteration number of the pigeon group; l = 1, 2, ..., L; i = 1, 2, ..., n, n is the total number of training samples; is the DBH prediction value of the fuzzy depth network, is the volume prediction value of the fuzzy depth network, is the measured value of DBH, is the measured value of timber volume;

[0070] The search strategy for model parameter combinations is divided into two stages according to the number of iterations. The initial stage is the first stage. When the number of iterations reaches the maximum number of iterations t max When the search reaches 80%, it enters the second stage. The specific search strategy is as follows:

[0071] The calculation process of the first stage of the search strategy for model parameter combinations is shown in formula (15):

[0072]

[0073] Where: is the cosine iteration weight term; t max is the maximum number of iterations of the pigeon group; ε is a very small constant; rand(0,1) is a random number between [0,1]; array is the lth group of model parameter combinations for the tth iteration For arrays Speed; l=1,2,…L;

[0074] When the fitness of the optimal parameter combination of L group model parameter combinations does not change for a long time, the L group model parameter combinations are sorted in descending order according to fitness, and the ones with higher fitness are ranked. The group parameters are subjected to population mutation, and the mutation method is shown in formula (16):

[0075]

[0076] Where: for The updated value of population mutation, Do rounding; if In the range of 0 to 100, otherwise If the value is within the range of 0 to 30, it will be updated; otherwise, it will not be updated. If the value is within the range of 0 to 10, it will be updated; otherwise, it will not be updated. rand(-1,1) is a random number in the range [-1,1];

[0077] The calculation process of the second stage of the model parameter combination search strategy is shown in formula (17):

[0078]

[0079] Where: is the tentative value of the lth group of model parameters at time t; if Then update the lth group of model parameters, that is, Otherwise, no update. l=1,2,…L;

[0080] After each iteration, some model parameter combinations with higher fitness values ​​are discarded and the number of model parameter groups L is updated. When the number of iterations reaches the maximum number of iterations t max Or when only one set of model parameter combinations is left, the iteration ends, and the parameter combination with the lowest fitness among the L sets of model parameter combinations is output, that is, the optimal values ​​of the given parameters k, α, and τ, and the model parameter combination is passed into the fuzzy deep network to complete the training.

[0081] The beneficial effects of the present invention are:

[0082] The present invention proposes an optimized fuzzy deep network to carry out rubber forest tree parameter prediction model, establishes the nonlinear relationship between forest parameters, proposes an adaptive algorithm to enhance the generalization ability of the prediction model to different rubber forest varieties, embeds the attention mechanism module to enhance the robustness of the network, and integrates the pigeon flock optimization algorithm (Pigeon-inspired Optimization) to adjust the parameters of the fuzzy deep network in real time, so as to further enhance the prediction accuracy and learning ability of the model. For rubber trees of different varieties, the rubber forest parameter prediction model of the present invention can accurately invert complex forest parameters and provide quantitative decision-making and data support for the afforestation and growth tending of different varieties of rubber trees.

[0083] Based on a fuzzy deep neural network, this invention can refine complex prediction models. It proposes an adaptive learning algorithm to determine the network structure, combines it with a pigeon flock optimization algorithm to search for optimal parameters, and improves the effectiveness of the adaptive algorithm. It also incorporates an attention mechanism to identify abnormal data in training samples. This forest parameter prediction model further improves the accuracy of prediction results for key parameters in rubber plantations.

[0084] The present invention predicts the diameter at breast height and the volume of individual rubber trees by setting up a forest parameter prediction model, and the partial forest parameters automatically obtained by airborne laser point cloud are predicted. This forest parameter prediction model combines the advantages of fuzzy deep network, attention mechanism and pigeon flock optimization algorithm, can be based on the nonlinear relationship between forest parameters, the weight distribution of abnormal samples, model parameter optimization strategy, realize adaptive model building, can be applicable to the establishment of most complex forest relationships, with good universality and robustness. Based on fuzzy deep network combined with multiple artificial intelligence algorithms, the rubber trees of different varieties are completed automatically according to the training of the model according to the forest parameter prediction model, and is adapted to the prediction of key parameters of the same variety rubber trees in the same forest land, is one of the landing applications of artificial intelligence technology in forestry neighborhood. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 (a) is a schematic diagram of the location of rubber forest plot 1.

[0086] Figure 1 (b) is a schematic diagram of the location of rubber forest plot 2.

[0087] Figure 1 (c) in the figure is a schematic diagram of the location of rubber forest plot 3.

[0088] Figure 1 (d) in the figure is a schematic diagram of Reken 628.

[0089] Figure 1 (e) in the figure is a schematic diagram of Reken 525.

[0090] Figure 1 (f) in the figure is a schematic diagram of Reyan 72059.

[0091] Figure 1 (g) in FIG is a schematic diagram of PR107.

[0092] Figure 2 (a) in the figure is the result of single tree segmentation in rubber forest plot 1.

[0093] Figure 2 (b) in the figure is the result of single tree segmentation in rubber forest plot 2.

[0094] Figure 2 (c) in the figure is the result of single tree segmentation in rubber forest plot 3.

[0095] Figure 3 These are the rubber tree point cloud data of Reken 628, Reken 525, Reyan 72059 and PR107.

[0096] Figure 4 This is the overall framework diagram of the forest parameter prediction network.

[0097] Figure 5 Graph of the attention-based weight update process.

[0098] Figure 6 (a) in the figure is the iterative optimization result diagram of the pigeon flock optimization module for the rubber tree network parameters of Reyan 72059.

[0099] Figure 6 (b) is the iterative optimization result diagram of the pigeon flock optimization module for the Reken 525 rubber tree network parameters.

[0100] Figure 6 (c) in the figure is the iterative optimization result diagram of the pigeon flock optimization module for the Reken 628 rubber tree network parameters.

[0101] Figure 6 (d) is the result of iterative optimization of the PR107 rubber tree network parameters by the pigeon flock optimization module.

[0102] Figure 7 This is a curve diagram showing the fitness of the optimal parameter group of the pigeon swarm optimization module as the number of network training times increases.

[0103] Figure 8 (a) in the figure is an iterative curve graph of the loss value of PR107 rubber tree during the back propagation process.

[0104] Figure 8 (b) is the iterative curve graph of the loss value of the Reyan 72059 rubber tree during the back propagation process.

[0105] Figure 8 (c) in the figure is the iterative curve graph of the loss value of Reken 525 rubber tree during the back propagation process.

[0106] Figure 8 (d) in the figure is the iterative curve graph of the loss value of Reken 628 rubber tree during the back propagation process.

[0107] Figure 9 (a) is a comparison analysis of the predicted and measured values ​​of DBH

[0108] Figure 9 (b) is a comparison analysis chart of the predicted volume value and the measured value. DETAILED DESCRIPTION

[0109] The specific embodiments of the present invention are further described below with reference to the accompanying drawings:

[0110] In recent years, airborne lidar has been widely used in forest resource surveys and parameter inversion, but it is also difficult to obtain complex forest parameters such as diameter at breast height and volume due to visual obstruction. In response to this problem, the present embodiment provides a method for accurately predicting the diameter at breast height and volume of trees based on an optimized fuzzy deep network. First, an optimized fuzzy learning network integrating an attention mechanism module is constructed, and a multi-parameter autonomous optimization module based on a pigeon flock optimization algorithm is added. Secondly, a single plant separation algorithm is used in combination with artificial forest surveys to extract multiple growth parameters of four varieties of rubber trees (Reken 628, Reken 525, Reyan 72059, PR107) from the airborne point clouds of three forests as training sets and bring them into the deep learning network to optimize the training parameters. Finally, the test sets of the four varieties are respectively brought into the trained network to predict the key parameters of the trees and compare and analyze them with the true values. The results show that the comparison results of the predicted and measured values ​​of the four rubber tree diameters at breast height are: RMSE<1.75cm, R 2 >91.42%; the comparison results of the predicted and measured volumes of the four rubber trees all meet the following requirements: RMSE<0.052m 3 , R 2 >90.14%. Compared to traditional backpropagation and radial basis function neural networks, the correlation of forest parameter inversion results obtained by the deep learning network in this paper is 4-9% higher. This embodiment applies the latest artificial intelligence technology to forest airborne laser point clouds to achieve accurate prediction of tree diameter at breast height and volume, which can meet the needs of large-scale rubber plantation parameter inversion and management surveys. The specific steps are described below.

[0111] 1. Materials and data:

[0112] 1.1 Study area and data collection:

[0113] The study area is located in the rubber tree plantation in Danzhou City, northwest of Hainan Island. In this example, three multi-species rubber forest plots were selected from the plantation, as shown in Google Maps. Figure 1As shown in (a), (b), and (c) in the figure. The region features a typical hilly plateau terrain and a tropical monsoon climate. The region receives an average annual precipitation of 1815 mm, with the rainy season (May to October) accounting for over 89% of the total annual rainfall. The average annual temperature is approximately 23°C, which is sufficient for the growth of rubber trees. The rubber tree varieties Reken 628, Reken 525, Reyan 72059, and PR107 found in the three plots are excellent varieties characterized by high and stable yields, strong stress tolerance, and a high survival rate. They are all cultivated on a large scale in Hainan. Reken 628 is highly resistant to cold and wind, making it a high-quality variety with relatively stable yields and wide adaptability. Reken 525 and Reken 523 grow quickly, mature early, and have high yields, making them excellent varieties for both rubber and wood. PR107 has low initial tapping yields, but its high dry rubber content, resistance to harsh conditions, and high-frequency tapping yields increase steadily over time, making it an excellent, high-yielding variety. Therefore, the above four varieties of rubber trees (of different ages) were selected from a multi-variety rubber tree plantation. Figure 1 (d), (e), (f) and (g) in the figure respectively indicate Reken 628, Reken 525, Reyan 72059 and PR107.

[0114] The airborne lidar equipped with the Velodyne HDL-32E lidar sensor can achieve a vertical field of view (FOV) from -30.67° to +10.67°, a 360° horizontal field of view, an operating frequency of 10HZ, a measurement range of 70m, and a measurement accuracy of + / -2cm. The shooting mode of the airborne lidar is set to continuous shooting, and the flying line route is a pre-programmed "back and forth rectangular parallel" route (such as Figure 1 The flight speed, flight altitude, and laser scanning overlap were set to 10 m / s, 30 m (above the take-off position), and 30%, respectively, to ensure the acquisition of complete branch parameters and a clear vertical structure of the rubber tree. The final extracted point cloud was stored in LAS 1.2 format.

[0115] 1.2, training samples and test samples:

[0116] After acquiring the point cloud data of the rubber tree forest through the airborne laser radar, Gaussian filtering is used for denoising, and cloud ground point filtering (CSF) is used to eliminate the unfavorable factors of the terrain. Then, this embodiment adopts the existing double Gaussian filter and energy function minimization single tree segmentation method, which is universal in China's subtropical forests. After experimental verification, it is applicable to rubber forests and has a good segmentation effect at the junction of the crowns. The segmentation results of the three rubber tree plots are represented by different colors, as shown in the following figure. Figure 2 shown. Figure 2 (a) in the figure is the result of single tree segmentation in rubber forest plot 1. Figure 2 (b) in the figure is the result of single tree segmentation in rubber forest plot 2. Figure 2 (c) in the figure is the result of single tree segmentation in rubber forest plot 3.

[0117] There are 1,364 rubber trees in three polyclonal and multi-variety rubber tree plots. Based on the principle of keeping the branches and trunks of individual trees as complete as possible, 813 trees were selected from the plots through manual visual inspection of individual tree point cloud data, including four varieties (about 200 trees of each variety). The specific morphological characteristics of individual trees of different varieties are as follows: Figure 3 shown. Figure 3 Point cloud data of rubber trees of different clone varieties. Figure 3 In the figure, from top to bottom, the single rubber trees in the first and second rows are Reken 628; the single rubber trees in the third and fourth rows are Reken 525; the single rubber trees in the fifth and sixth rows are Reyan 72059; and the single rubber trees in the seventh and eighth rows are PR107.

[0118] Reken 628 has an upright, almost branchless trunk and a small, broom-shaped crown. Its leaves are oval, thick, and shiny, with three separate leaflets. Reken 525 has less skewing, with branches branching at a lower height and at a larger angle. It has a large number of branches and a large, multi-headed crown. Reyan 72059 is more flexible and easily bends, with more drooping branches, a large number of branches at a larger angle, and a large, fan-shaped crown. PR107 is straighter and more wind-resistant, with branches branching at a higher height, fewer branches, and more dichotomous branches. It has a small, broom-shaped crown and oblong leaves with regular, small waves on the margins.

[0119] Based on the segmented point cloud of a single rubber tree, the following method was used to automatically obtain tree parameters, including east-west crown width, north-south crown width, tree height, point cloud density, and crown volume. The method is as follows: The maximum distance in the east-west direction of each tree crown point cloud was selected as the east-west crown width parameter; the north-south crown width parameter was similarly extracted in the north-south direction; the vertical distance between the highest point of the individual tree point cloud and the horizontal plane was the tree height parameter; the total number of individual tree point clouds divided by the projected area of ​​the tree crown was the point cloud density; and the crown volume parameter was calculated using the AlphaShape method.

[0120] Since the forest parameters such as the diameter at breast height and timber volume of rubber trees are not easy to obtain directly from the airborne single plant point cloud, the present embodiment adopts the method of manual mapping to obtain the diameter at breast height and timber volume of rubber trees, and the specific method is as follows. At the trunk position of the rubber tree 1.3m above the ground, the circumference of the trunk is obtained by a tape measure, and then the diameter at breast height parameters are obtained. The tree volume parameters are obtained using the tree measurement method. In the actual measurement, the upper diameter of the rubber trunk is selected to be The point is the shape point, D is the diameter of the rubber tree at breast height, and the length from the shape point to the treetop is H. t , measure the cross-sectional area S of the breast diameter D, measure the height H of the rubber tree and use the following formula to obtain the volume of the rubber tree.

[0121]

[0122] Where: r is the stem shape index; S vl The volume of rubber trees is shown in Table 1. After automatically acquiring parameters from individual tree point clouds and manually mapping, we obtained the tree parameters of approximately 200 trees of each species and divided them into training and test sets. The tree parameters and training samples for the four rubber trees, Reken 628, Reken 525, Reyan 72059, and PR107, are shown in Table 1.

[0123] Table 1 shows the tree parameters and training samples of different rubber tree species in the study plot:

[0124]

[0125] 2. Forest parameter prediction model:

[0126] 2.1. Model overall architecture design:

[0127] Tree parameters such as diameter at breast height (DBH) and volume are not easily acquired from airborne individual tree point clouds. Therefore, it is crucial to establish a tree parameter prediction model to obtain these parameters by identifying corresponding relationships between tree parameters. Multi-species rubber plantations share similar soil and climate conditions. Typically, tree parameters for the same rubber variety are normally distributed around the average value for that variety. However, factors such as rubber tree necrosis and intraspecific competition can cause some individual tree parameters to differ significantly from those of the rest of the population, necessitating a prediction model capable of autonomously identifying these differences. Given the diverse variety of rubber plantations and their varying growth morphologies, parameter prediction requires adaptive learning. Furthermore, parameter optimization in the prediction model significantly impacts prediction performance.

[0128] A comprehensive study of commonly used neural network prediction models found that the attention mechanism effectively enhances the anti-interference ability of the prediction model. Adding the forest parameter attention mechanism improves the robustness of the forest parameter prediction model and increases the accuracy of the prediction model.

[0129] The membership function of the fuzzy subset in the fuzzy deep network often adopts a Gaussian function, which adaptively determines the center and variance of the fuzzy subset, reflecting the learning ability of the fuzzy deep network. Therefore, this embodiment proposes a DPKM algorithm that combines the density peaks clustering algorithm (DPC) with the K-Means algorithm to determine the center of the membership function, and proposes an algorithm for adaptively determining the variance based on the Euclidean distance between the centers. The pigeon-inspired optimization algorithm (PIO) can effectively solve the parameter optimization problem of the fuzzy deep network. Therefore, a forest parameter prediction model that integrates the pigeon-inspired optimization module and the fuzzy deep network is designed.

[0130] The overall framework of the forest parameter prediction network proposed in this embodiment is as follows: Figure 4 As shown in the figure, it is divided into a fuzzy deep network and a pigeon flock optimization module. The pigeon flock optimization module updates the parameters of the fuzzy deep network. After the network completes training, it outputs the predicted value and returns it to the pigeon flock module, which serves as a prerequisite for the pigeon flock module to provide the parameter fitness value solution. After the pigeon flock module completes the search for the optimal parameters of the fuzzy deep network, the fuzzy deep network completes adaptive training based on the optimal parameters and establishes the final forest parameter prediction model. From input to output, the fuzzy deep network consists of an adaptive fuzzy layer, a fuzzy inference layer, and an attention-based weight update layer. 1 、x 2 、x 3 、x 4 、x 5 The input attributes of the adaptive fuzzy layer are the east-west crown width, north-south crown width, tree height, point cloud density, and crown area. Each input attribute has the same number of fuzzy subsets, and the center and variance of the fuzzy subset membership function are determined by the adaptive algorithm. The sample of the input attribute is passed into its own fuzzy subset output membership value x u The corresponding membership degree is u∈{1,2,3,4,5},j=1,…,k,the total number of fuzzy subsets is k. Product reasoning determines the fuzzy unit of the fuzzy reasoning layer, and the membership degree h is passed into the fuzzy unit. One unit corresponds to one output, and the output value is j=1,…,k, the total number of fuzzy units is k. The attention-based weight update layer is passed in, which continuously updates the connection weights through back propagation, and finally outputs the predicted value y after weighted operation. 1 、y 2 , where y 1 is the predicted value of DBH, y 2 is the predicted value of timber volume.

[0131] 2.2, Adaptive fuzzy layer:

[0132] Input attribute x u There are k fuzzy subsets in the adaptive fuzzy layer, u∈{1,2,3,4,5}, so the adaptive fuzzy layer has a total of 5×k fuzzy subsets. Training samples Input attribute x into the adaptive fuzzy layer u The jth fuzzy subset to which it belongs, outputs the membership degree As shown in formula (2).

[0133]

[0134] Where: x u The center and variance of the jth fuzzy subset in the adaptive fuzzy layer are x u The training samples are clustered according to the DPKM algorithm proposed in this embodiment. is the cluster center; according to The Euclidean distance between The sample density within the DPKM cluster to which it belongs, and the variance is adaptively determined j=1,…,k, the total number of fuzzy subsets of the input attributes is k, and k is updated by the pigeon flock optimization module, i=1,…,n, n is the total number of rubber trees in the training sample.

[0135] The DPKM algorithm is based on x u Solve for n training samples The local density and distance And calculated based on the above two Likelihood value as the initial center of the cluster Then clustering is performed. The specific method is as follows, i = 1,…,n, u∈{1,2,3,4,5}.

[0136]

[0137] Where: yes The local density, d u is x u The cutoff distance of , u∈{1,2,3,4,5}, i=1,2,…,n, m=1,2,…,n. Calculate Need to consider x u n training samples and The Euclidean distance of u Sure In the formula (3), the distance between the training samples in the neighborhood is cut off and the training samples in the neighborhood are enlarged. The local density influence reduces the effect of sample points outside the neighborhood; d uThe selection principle is to let x u In the training sample The ratio of the number of plants in the neighborhood to the total number of plants n is 1% to 2%. u The n training samples are sorted in descending order according to the size of the local density. The minimum Euclidean distance between any two samples in the training samples is distance As shown in the following formula.

[0138]

[0139] Where: yes The local density and m=1,2,…n; for and The Euclidean distance between them is u∈{1,2,3,4,5}. In particular, if The local density is the largest, And m≠i.

[0140] when The local density and distance After confirmation, calculate As input attribute x u The likelihood value of the initial center when performing DPKM clustering As shown in formula (5).

[0141]

[0142] Where: max(ρ u )、min(ρ u ) are ρ u The maximum and minimum values ​​of n local densities; max(δ u )、min(δ u ) are δ u The purpose of formula (5) is to convert the local density ρ of different scales into u and distance δ u , normalized to the product of the same scale. ρ u Refers to the local density of the u-th input attribute, and the u-th input attribute has n local densities. u It refers to the distance of the u-th input attribute, and the u-th input attribute has n distances.

[0143] So far, x u Among the n training samples according to Sort the values ​​in descending order and take the first k sample points as x u The initial center point of DPKM clustering for n training samples; secondly, calculate the input attribute x u The Euclidean distance between the n training samples and the k class centers is calculated. According to the minimum distance distribution principle, the training samples are assigned to the class with the nearest center. After all n samples have been assigned once, the mean of each class is calculated, the class center is updated, and the above two steps are repeated (i.e., the second and third steps) until the change in the class center is less than the set error. DPKM clustering is completed, and the center of the k clusters is x. u The fuzzy subset center of j=1,…,k.

[0144] When the input attribute x u The fuzzy subset center of as well as DPKM cluster After establishment, calculate x according to the following adaptive algorithm u The fuzzy subset variance First calculate x u The center of the k fuzzy subset centers As shown in formula (6).

[0145]

[0146] Then solve the clustering Internal samples and The average Euclidean distance Is to calculate the variance The prerequisite is as shown in formula (7).

[0147]

[0148] Where: express Internal samples and Euclidean distance; for Internal samples The number of

[0149]

[0150] Where: for The width factor of α is the variance scaling factor, which is updated by the pigeon swarm optimization module. For input attribute x uThe maximum distance between the centers of the k fuzzy subsets; u∈{1,2,3,4,5}, j=1,2,…,k.

[0151] This layer gives the initial number of fuzzy subsets k and variance scaling coefficient α. When the pigeon swarm optimization module transmits new k and α values, the model structure of the adaptive fuzzy layer also changes accordingly.

[0152] 2.3, Fuzzy Reasoning Layer:

[0153] The fuzzy deep network fuzzifies different input attributes in turn to perform fuzzy reasoning on the complex nonlinear relationship between forest parameters. It is very effective in processing complex models that are difficult to accurately calculate, which makes up for the shortcomings of traditional neural networks.

[0154] The membership degree is the input of the fuzzy reasoning layer. The product reasoning method is used to establish the fuzzy unit of the fuzzy reasoning layer and calculate the output value of the jth unit of the i-th training sample. Output as fuzzy unit.

[0155]

[0156] Where: is the product of the membership degree of the jth fuzzy subset of the i-th training sample under different input attributes; After normalization, u∈{1,2,3,4,5}, i=1,…,n,j=1,2,…,k, n is the total number of rubber trees in the training sample, and the total number of fuzzy units is k.

[0157] 2.4. Attention-based weight update layer:

[0158] The initial value of the attention-based connection weight w is randomly given and continuously updated by back propagation as follows: Figure 5 As shown in , when the number of iterations reaches the maximum, the update is terminated. The value is used as the input of this layer. The predicted value y is output by weighted operation with the connection weight w to realize the defuzzification calculation of the fuzzy depth network.

[0159] The loss function combines the predicted value y and the measured value The attention weight Q is shown in formula (10).

[0160]

[0161] Where: y 1 is the predicted value of DBH, y 2 is the predicted value of timber volume; for The corresponding measured value; Qi is the attention weight of the i-th training sample, i=1,2,…,n. Fitness value The connection weight between the diameter at breast height is The connection weight between and volume is Then for all training samples:

[0162]

[0163] Where: represents the jth output of the i-th training sample from the fuzzy inference layer value; represents the predicted value of DBH of the i-th training sample, represents the volume prediction value of the i-th training sample; j = 1, 2, ..., k, i = 1, 2, ..., n, n is the total number of training samples, and the total number of fuzzy units is k.

[0164] The essence of the attention mechanism is weight distribution. In the loss function, the weight Q is assigned to the rubber tree sample, which has the ability to resist interference with the rubber tree training samples with abnormal growth status. In order to eliminate the influence of the dimensions between different tree parameters, the training samples are normalized preprocessed. Then, the parameter matrix of the i-th tree in the training sample is There are 7 measured tree parameters, namely, east-west crown width, north-south crown width, tree height, point cloud density, crown volume, timber volume, and diameter at breast height. The average value matrix of each tree parameter in the training sample is avg = [avg 1 ,…,avg 7 ], Q i From the training sample z i Get the intrinsic connection between avg.

[0165]

[0166] Where: Represents the tree parameter matrix z i The pth tree parameter from left to right, avg p Represents the pth average value of the average matrix avg from left to right, p∈{1,2,3,4,5,6,7}; cosine similarity Sim(z i ,avg) and Euclidean distance Dist(z i ,avg) represents the intrinsic connection between the parameters of the i-th rubber tree and the average value avg; τ is the attention weight scaling coefficient updated by the pigeon flock optimization module; i=1,2,…,n.

[0167] The connection weight w is continuously updated through back propagation, and the update method is shown in formula (13).

[0168]

[0169] Where: η ranges from 0 < η < 1, indicating learning efficiency; t is the current number of iterations.

[0170] At the same time, given the initial attention weight scaling coefficient τ, each time the pigeon flock optimization module updates the τ value, the forest parameter prediction effect of the fuzzy deep network also changes accordingly.

[0171] 2.5. Pigeon flock optimization module:

[0172] This module optimizes the key parameters k, α, and τ in the fuzzy deep learning network. k is the total number of fuzzy subsets in the adaptive fuzzy layer, α is the variance scaling factor of formula (8), and τ is the attention weight scaling factor τ of formula (12). The combination of k, α, and τ is called the model parameter combination δ[k, α, τ]. In the parameter search space, k is 0 to 100 and must be an integer, α is 0 to 30, and τ is 0 to 10. The pigeon flock optimization module is initially composed of L groups of model parameters δ l [k l ,α l ,τ l ](l=1,2,…L), that is, this module has L groups of model parameters to be optimized and t max iterations, the tth iteration is based on the search strategy of this paper When the iteration is terminated, the optimal parameter combination δ in the L group of model parameters is searched. l , and passed into the fuzzy deep network to complete the training of the fuzzy deep network.

[0173] First, the update of the parameters k and α of the adaptive fuzzy layer will result in the output value of this layer Secondly, the update of the coefficient τ of the attention mechanism will affect the calculation of the attention weight Q, and then affect the loss function, that is, the iteration of the connection weight w in formula (10); Finally, the prediction value of the fuzzy deep learning network will change with The connection weight w is updated and changes accordingly. This embodiment uses the predicted values ​​and measured values ​​output by the completed fuzzy depth training model to calculate the fitness of the model parameter combination, as shown in formula (14). The lower the fitness value, the closer this set of parameters is to the optimal model parameter combination.

[0174]

[0175] Where: The lth group of model parameter combinations for the tth iteration is t is the current iteration number of the pigeon group; l = 1, 2, ..., L; i = 1, 2, ..., n, where n is the total number of training samples; is the predicted value of DBH and volume by fuzzy depth network, For The corresponding measured value.

[0176] The search strategy for model parameter combinations is divided into two stages according to the number of iterations. The initial stage is the first stage, and when the number of iterations reaches 80% of the maximum number of iterations, it enters the second stage. The specific search strategy is shown below.

[0177] In the first phase of the model parameter combination search strategy, this embodiment proposes a cosine iterative weight term and population mutation to help individuals escape local optimal solutions. This allows the fuzzy deep network to complete training by providing the optimal model parameter combination. The addition of the cosine iterative weight term prioritizes global search capabilities in the early stages of iteration, while strengthening local search capabilities in the later stages, meeting actual iteration requirements, as shown below.

[0178]

[0179] Where: is the cosine iteration weight term; t max is the maximum number of iterations of the pigeon group; ε is a very small constant; rand(0,1) is a random number between [0,1]; the value of the lth group of model parameter combinations in the tth iteration is When the value is updated, rounding is required; For arrays Speed; l=1,2,…L;

[0180] In order to further enhance the ability of the PIO algorithm to escape from the local optimal solution, when the fitness of the optimal parameter combination of the L group of model parameter combinations does not change for a long time, it means that the forest parameter prediction model of this embodiment has fallen into a local extreme value. At this time, the L group of model parameter combinations are sorted in descending order according to fitness. l=1,2,…L, with higher fitness The group parameters are subjected to population mutation, and the mutation method is as follows.

[0181]

[0182] Where: for The updated value of population mutation, Do rounding; if In the range of 0 to 100, otherwise according to Update method, If the value is in the range of 0 to 30 or 0 to 10, it will be updated; otherwise, it will not be updated. rand(-1,1) is a random number in the range [-1,1].

[0183] In the second stage of the model parameter combination search strategy, the update strategy of the model parameters is adjusted, assuming that δ c (t) is the center position of all model parameter combinations at time t, and the parameter group flies towards the center position.

[0184]

[0185] Where: is the tentative value of the lth group of model parameters at time t; if Then update the lth group of model parameters Otherwise do not update l=1,2,…L. After each iteration, some model parameter combinations with higher fitness values ​​are discarded and the number of model parameter groups L is updated, so that the better model parameter combinations are retained while ensuring the convergence of the algorithm. When the number of iterations reaches the maximum number of iterations t max Or when only one set of model parameter combinations is left, the iteration ends, and the parameter combination with the lowest fitness among the L sets of model parameter combinations is output, that is, the optimal values ​​of the given parameters k, α, and τ, and the model parameter combination is passed into the fuzzy deep network to complete the training.

[0186] 3. Results and Discussion

[0187] 3.1. Training and testing results of the pigeon flock optimization module:

[0188] The training and testing of the tree parameter prediction model were performed on a Windows 10 64-bit server equipped with an AMD Ryzen 7 4800H CPU @ 2.9 GHz processor and 16 GB RAM. In the tree parameter prediction model constructed in this embodiment, the maximum number of iterations of the weight update layer was set to 200, and the learning efficiency η was set to 8; the total number of model parameter combinations L of the pigeon flock optimization module was set to 32, and the maximum number of iterations t max Set to 50.

[0189] The 32 sets of model parameters in the initial round are randomly selected and evenly distributed in the parameter search space. With the continuous iteration of the pigeon flock module, the values ​​of k, α, and τ in the model parameter combination are continuously updated. The results of the optimal model parameter combination of different varieties of training sets at different stages of iteration are shown in Table 2. The 32 sets of parameters of each variety gradually converge to the optimal array, indicating that the pigeon flock module can adaptively learn the optimal model parameter combination of different varieties of rubber trees, such as Figure 6 shown. Figure 6 (a) in the figure is the iterative optimization result diagram of the pigeon flock optimization module for the rubber tree network parameters of Reyan 72059. Figure 6(b) is the iterative optimization result diagram of the pigeon flock optimization module for the Reken 525 rubber tree network parameters. Figure 6 (c) in the figure is the iterative optimization result diagram of the pigeon flock optimization module for the Reken 628 rubber tree network parameters. Figure 6 (d) is the result of iterative optimization of the PR107 rubber tree network parameters by the pigeon flock optimization module.

[0190] Table 2 shows the optimal results of the model parameter combination of the pigeon flock optimization module at different stages:

[0191]

[0192] In the iterative process of the pigeon group module to find the optimal model parameter combination of the training set, there are 32 sets of model parameters in each round, among which the one with the lowest fitness is the optimal model parameter combination of this round. The fitness curve composed of the optimal model parameter combination at different iteration stages is as follows: Figure 7 As shown. The fitness of the optimal model parameter combination shows a downward trend, indicating that the forest parameter prediction model of this embodiment is a global optimization process. The fitness curves of the training sets of different varieties have a significant decrease in the first 30 Epochs, indicating that the parameters of the forest parameter prediction model are rapidly approaching the optimal array. The model parameters are sequentially passed into the corresponding modules of the fuzzy deep network. On the basis of the fuzzy deep network adaptively constructing the network structure according to the training samples, the correlation coefficients in the neural network are adjusted, so that the optimal parameter group in the initial round can also achieve a good fitness value. After 50 Epochs, the fitness values ​​of the training samples of Reyan 72059, Reken 525, Reken 628, and PR107 converged to 0.025, 0.022, 0.016, and 0.015, respectively, indicating that the forest parameter prediction model constructed in this embodiment has the ability to accurately predict parameters.

[0193] When the pigeon flock optimization module confirms the optimal model parameter combination for different varieties, it is passed to the fuzzy deep network to complete the training of the forest parameter prediction model for each variety. At this time, in the attention-based weight update layer, the loss value E during the training process is as follows: Figure 8 shown. Figure 8 (a) in the figure is an iterative curve graph of the loss value of PR107 rubber tree during the back propagation process. Figure 8 (b) is the iterative curve graph of the loss value of the Reyan 72059 rubber tree during the back propagation process. Figure 8 (c) in the figure is the iterative curve graph of the loss value of Reken 525 rubber tree during the back propagation process. Figure 8(d) in is the iterative curve diagram of the loss value of Reken 628 rubber tree in the back propagation process. In order to improve the training efficiency of the prediction model of the present embodiment, the mini-batch gradient descent (Mini-Batch Gradient Descent) method is adopted in the weight update layer based on attention to carry out back propagation, resulting in local oscillation of the regression loss value. However, with the continuous iteration of the learning process, the loss value E is generally on a downward trend, indicating that the fuzzy depth network of the present embodiment has good convergence. After 100 iterations, the loss value E of PR107, Reyan 72059, Reken 525 and Reken 628 converge to 0.00176, 0.00349, 0.00345, 0.00072 respectively, indicating that the fuzzy depth network constructed by the present embodiment has good forest parameter prediction ability.

[0194] 3.2. Comparison with existing methods:

[0195] Based on the prediction model and traditional method of the present embodiment, the prediction result of rubber tree diameter at breast height, volume parameter is as shown in table 3.BP (Back Propagation) neural network is the method for feedforward neural network based on multilayer, for the definite dependence experience and trial and error of network structure, and activation function has globality, can interfere with each other, therefore easily fall into the problem of local minimum.RBF (Radial Basis Function) neural network and BP are all applicable to nonlinear model establishment, but the local activation function of RBF overcomes the mutual interference problem of BP global activation function, and for new training set, only need hidden layer neuron node number and connection weight to change, learning speed has larger improvement than BP algorithm, convergence is also easier to ensure, therefore RBF easily obtains more excellent result.GRNN (General Regression Neural Network) is a kind of radial basis neural network, compared to traditional radial basis network, between hidden layer and output layer, added summation layer, has more advantages than RBF in aspects such as sample data is less, data instability. However, RBF and GRNN neural networks often determine the network structure through trial and error and empirical formulas, relying on prior experience; at the same time, the above methods lack a mechanism for judging abnormal data in training samples, which reduces the robustness of the neural network. The method of this embodiment is based on a fuzzy deep network, which can accurately predict complex prediction models, proposes an adaptive learning algorithm to determine the network structure, combines the pigeon flock optimization algorithm to search for optimal parameters, improves the effect of the adaptive algorithm, and adds an attention mechanism to judge the abnormal data of the training samples. Table 3 lists the comparative results of the four methods for predicting diameter at breast height and volume. It can be seen from the table that the method of this embodiment has achieved better quantitative results in the three indicators of determination coefficient (R2), root mean square error (RMSE), and mean absolute percentage error (MAPE). It can be seen that the forest parameter prediction model of this embodiment further improves the prediction result accuracy of the key parameters of rubber forests.

[0196] Table 3 shows the prediction results of tree parameters using different methods:

[0197]

[0198] 3.3 Analysis of tree parameter prediction results:

[0199] After the fuzzy deep network establishes the optimal model parameters through the pigeon flock optimization module, the fuzzy deep network adaptively establishes the forest parameter prediction model of each variety based on the training set of different rubber tree varieties. Table 4 gives the actual measured values ​​of the diameter at breast height and timber volume of four rubber trees, Reken 628, Reken 525, Reyan 72059, and PR107, and the predicted values ​​of this embodiment. At the same time, by comparing the index R 2 , RMSE and MAPE quantitatively analyze the effectiveness of the method in this embodiment, Figure 9The following are the comparison results of specific parameters.

[0200] Table 4 shows the comparison between the tree growth parameters obtained by the method of this embodiment and the actual measured values:

[0201]

[0202] Note: (F): actual measurement value; (O): method of this paper.

[0203] Figure 9 The prediction results of DBH and volume parameters are shown respectively. The experimental points of the predicted and measured values ​​of the parameters of the four rubber trees are evenly distributed near the 45° regression line, and the two are in a linear relationship.

[0204] Figure 9 (a) shows the comparison results of the predicted and measured values ​​of the diameter at breast height of four rubber trees obtained by the method of this embodiment. The comparison results of Reken 525 and Reyan 72059 are (R 2 =92.24%, RMSE=1.70cm, MAPE=5.08%) and (R 2 =91.42%, RMSE=1.75cm, MAPE=5.10%). Compared with the first two varieties, the research model of this embodiment predicted the DBH of Reken 628 and PR107 with better results, which were (R 2 =94.31%, RMSE=1.44cm, MAPE=4.87%) and (R 2 =93.87%, RMSE =1.48cm, MAPE =5.03%). This is primarily due to wind damage and tilting of trees in the Reyan 72059 rubber plantation. Adjacent rubber trees obstruct each other, resulting in incomplete point cloud data acquisition, which in turn affects the final parameter prediction results. Reken 525 has a more complex growth morphology, with branches located low and numerous and densely packed. Diameter at breast height (DBH) parameters vary significantly between rubber trees. Reken 628 and PR107, on the other hand, are more wind-resistant and less prone to lodging. They have simpler growth morphologies and fewer branches, resulting in more complete branch data, higher point cloud quality, and better prediction accuracy.

[0205] Figure 9 (b) shows the comparison between the volume prediction and actual measurement of four rubber trees. 2 =91.25%, RMSE=0.050m 3 , MAPE = 6.06%) and Reken 525 (R 2 =90.14%, RMSE=0.052m 3 , MAPE = 8.19%), the RMSE of Reken 628 (R 2=93.88%, RMSE=0.027m 3 , MAPE=5.02%) and PR107 (R 2 =93.73%, RMSE=0.028m 3 , MAPE = 5.33%) was significantly higher. This phenomenon can be explained by the smaller number of branches in Reken 628 and PR107. Under the same forest plot, climate conditions, and species, the differences in volume parameters between different trees are relatively small. Furthermore, the smaller crowns of Reken 628 and PR107 are less affected by shading from neighboring trees of the same species, making it more accurate to obtain tree parameters.

[0206] 4. Conclusion

[0207] By establishing a tree parameter prediction model, the diameter at breast height and volume of a single rubber tree are predicted using some tree parameters automatically obtained from the airborne laser point cloud. This prediction model combines the advantages of fuzzy deep networks, attention mechanisms, and pigeon flock optimization algorithms. It can achieve adaptive model building based on the nonlinear relationship between tree parameters, the weight distribution of abnormal samples, and the model parameter optimization strategy. It is applicable to the establishment of most complex tree relationships and has good universality and robustness. Compared with the three prediction methods of BP, RBF, and GRNN, the prediction model of this embodiment achieved better results in predicting diameter at breast height and volume, with RMSE of 1.59±0.15cm and 0.040±0.013m, respectively. 3 . The experimental results show that the forest parameter prediction model of the present embodiment can achieve good prediction results for a variety of rubber trees. The average MAPE of the four rubber trees predicting the diameter at breast height is 5.02% disturbance, and the average MAPE of the predicted timber volume is 6.15% disturbance. It can effectively obtain the parameters of individual trees in artificial forests, and is superior to the traditional prediction model in the experimental samples. In the prediction study of forest parameters, the differences in the performance of different rubber tree varieties can be found from the influence of the natural environment on the rubber tree plot, the incomplete acquisition of the airborne point cloud caused by the mutual occlusion between trees, and the growth morphological characteristics of different varieties. Based on the combination of fuzzy deep network and multiple artificial intelligence algorithms, the training of the model is automatically completed for different varieties of rubber trees according to the wood parameter prediction model, and it is adapted to the prediction of key parameters of the same variety of rubber trees in the same forest, which is one of the landing applications of artificial intelligence technology in the forestry neighborhood.

Claims

1. A method for accurately predicting tree diameter at breast height volume based on optimized fuzzy deep network, characterized by: include: Step 1: Obtain forest point cloud data using airborne lidar; Step 2: De-noise the point cloud data, filter the point cloud data using the point cloud ground point filtering method, and perform single tree segmentation on the filtered point cloud data; Step 3: Based on the segmented point cloud of individual trees, obtain the forest parameters of individual trees, including east-west crown width, north-south crown width, tree height, point cloud density, and crown volume; Step 4: Use manual mapping methods to obtain the diameter at breast height and volume of the corresponding trees; Step 5: Take the east-west crown width, north-south crown width, tree height, point cloud density, crown volume, diameter at breast height and timber volume of multiple trees as training sample datasets; Step 6: Establish a tree parameter prediction network, which includes a fuzzy deep network and a pigeon flock optimization module. The tree parameter prediction network is trained using a training sample data set. The input of the tree parameter prediction network is the east-west crown width, north-south crown width, tree height, point cloud density and crown volume of the tree, and the output is the diameter at breast height and volume of the tree. After the fuzzy deep network training is completed, the predicted value is output and the predicted value is transmitted to the pigeon flock optimization module. The pigeon flock optimization module updates the parameters of the fuzzy deep network. When the pigeon flock optimization module completes the optimal parameter search of the fuzzy deep network, the fuzzy deep network completes adaptive training based on the optimal parameters to establish the final tree parameter prediction model. The fuzzy deep network includes an adaptive fuzzy layer, a fuzzy reasoning layer, and an attention-based weight update layer from input to output. Step 7: Collect the point cloud data of the forest to be tested and obtain the east-west crown width, north-south crown width, tree height, point cloud density and crown volume of each tree in the forest to be tested according to the methods of steps 2 and 3. Input the east-west crown width, north-south crown width, tree height, point cloud density and crown volume of the trees into the final forest parameter prediction model to obtain the predicted values ​​of the diameter at breast height and volume of the corresponding trees.

2. The method for accurately predicting tree diameter at breast height volume based on optimized fuzzy deep network according to claim 1, characterized in that: The method for obtaining the forest parameters of individual trees in step 3 is: The maximum distance in the east-west direction of the crown point cloud of a single tree is selected as the east-west crown width; the maximum distance in the north-south direction of the crown point cloud of a single tree is selected as the north-south crown width; the vertical distance between the highest point of the crown point cloud of a single tree and the horizontal plane is the tree height; the total number of point clouds of a single tree divided by the projected area of ​​the crown is the point cloud density; the convex hull volume of the crown point cloud of a single tree is calculated, and the convex hull volume is the crown area.

3. The method for accurately predicting tree diameter at breast height volume based on optimized fuzzy deep network according to claim 2, characterized in that: The method for obtaining the diameter at breast height and volume of the tree in step 4 is: At the trunk position of the tree 1.3m above the ground, use a tape measure to obtain the circumference of the trunk, which is the diameter at breast height of the tree; The upper diameter of the trunk of the selected tree is The point is the shape point, D is the tree diameter at breast height, and the length from the shape point to the treetop is H. t The cross-sectional area of ​​the diameter at breast height is S D , the tree height is measured as H, then the volume of the tree is: Where: r is the stem shape index; S vl The volume of trees.

4. The method for accurately predicting tree diameter at breast height volume based on optimized fuzzy deep network according to claim 1, characterized in that: The calculation process of the adaptive fuzzy layer is: Input attribute x u There are k fuzzy subsets in the adaptive fuzzy layer, u∈{1,2,3,4,5}, where x 1 、x 2 、x 3 、x 4 、x 5 The input attributes of the adaptive fuzzy layer are the east-west crown width, north-south crown width, tree height, point cloud density and crown area. Each input attribute has k fuzzy subsets. The adaptive fuzzy layer has a total of 5×k fuzzy subsets. The training samples Input attribute x into the adaptive fuzzy layer u The jth fuzzy subset to which it belongs, outputs the membership degree As shown in formula (2): Where: x u The center of the jth fuzzy subset in the adaptive fuzzy layer is x u The variance of the jth fuzzy subset in the adaptive fuzzy layer is The total number of fuzzy subsets of input attributes is k, and k is updated by the pigeon swarm optimization module, i = 1, ..., n, n is the number of training samples, that is, the total number of trees in the training sample; According to x u Solve for n training samples The local density and distance Local density The calculation formula is: Where: yes The local density, d u is x u The cutoff distance, u∈{1,2,3,4,5}, i=1,2,…,n, m=1,2,…,n; Calculate the local density greater than All training samples and The Euclidean distance of distance As shown in formula (4): Where: yes The local density and for and The Euclidean distance between them, u∈{1,2,3,4,5}; if The local density of is the largest, then And m≠i; calculate As input attribute x u The probability value of the initial center of clustering As shown in formula (5): Where: max(ρ u ) is ρ u The maximum value of n local densities; min(ρ u ) is ρ u The minimum value of n local densities; max(δ u ) is δ u The maximum value of n distances; min(δ u ) is δ u The minimum value of n distances; ρ u Take n training samples as input attributes x u The local density, including n local densities; δ u Take n training samples as input attributes x u The distance, including n distances; Step (a), input attribute x u Among the n training samples according to Sort the values ​​in descending order and take the first k sample points as x u The initial center point of DPKM clustering for n training samples; Step (b), calculate the input attribute x u The Euclidean distance between the n training samples and the k class centers is calculated. According to the minimum distance distribution principle, the training samples are assigned to the class with the nearest center. After the n training samples are assigned once, the mean of each class is calculated, and the mean is used as the new class center and the class center is updated. Repeat the above steps (b) and (c) until the change of the class center is less than the set error. The DPKM clustering is completed. At this time, the center of the k clusters is x u The center of the jth fuzzy subset in the adaptive fuzzy layer When the input attribute x u The center of the jth fuzzy subset in the adaptive fuzzy layer as well as DPKM cluster After establishment, calculate x according to the following adaptive algorithm u The variance of the jth fuzzy subset in the adaptive fuzzy layer First calculate x u The center of the k fuzzy subset centers As shown in formula (6): Then solve the clustering Internal samples and The average Euclidean distance As shown in formula (7): Where: express Internal samples and Euclidean distance; for Internal samples the number of Where: for The width factor of α is the variance scaling factor, which is updated by the pigeon swarm optimization module. For input attribute x u The maximum distance between the centers of the k fuzzy subsets; u∈{1,2,3,4,5}, j=1,2,…,k.

5. The method for accurately predicting tree diameter at breast height volume based on optimized fuzzy deep network according to claim 4 is characterized in that: The calculation process of the fuzzy reasoning layer is: The membership degree is the input of the fuzzy reasoning layer. The product reasoning method is used to establish the fuzzy unit of the fuzzy reasoning layer and calculate the output value of the jth unit of the i-th training sample. As the fuzzy unit output, as shown in formula (9): Where: is the product of the membership degree of the jth fuzzy subset of the i-th training sample under different input attributes; After normalization, u∈{1,2,3,4,5}, i=1,…,n,j=1,2,…,k, n is the total number of trees in the training sample, and the total number of fuzzy units is k.

6. The method for accurately predicting tree diameter at breast height volume based on optimized fuzzy deep network according to claim 5, characterized in that: The calculation process of the attention-based weight update layer is: The loss function is shown in formula (10): Where: y 1 is the predicted value of DBH, y 2 is the predicted value of timber volume; is the measured value of DBH, is the measured value of timber volume; Q i is the attention weight of the i-th training sample, i = 1, 2, ..., n; The output of the fuzzy inference layer The connection weight between the diameter at breast height is The connection weight between and volume is Then for all training samples: Where: represents the jth output of the i-th training sample from the fuzzy inference layer value; represents the predicted value of DBH of the i-th training sample, represents the volume prediction value of the i-th training sample; j = 1, 2, ..., k, i = 1, 2, ..., n, n is the total number of training samples, and the total number of fuzzy units is k; The training samples are normalized and preprocessed, and then the parameter matrix of the i-th tree in the training sample is Contains 7 measured tree parameters, the tree parameter matrix z i From left to right in the middle are the east-west crown width, north-south crown width, tree height, point cloud density, crown volume, timber volume and diameter at breast height. The average east-west crown width of all training samples is avg 1 The average north-south crown width of all training samples is avg 2 , the average tree height of all training samples is avg 3 , the average point cloud density of all training samples is avg 4 , the average crown volume of all training samples is avg 5 , the average volume of all training samples is avg 6 , the average DBH value of all training samples is avg 7 The average value matrix of each tree parameter in the training sample is avg = [avg 1 ,…,avg 7 ], Q i The calculation method of is shown in formula (12): Where: Represents the tree parameter matrix z i The pth tree parameter from left to right, avg p Represents the pth average value of the average matrix avg from left to right, p∈{1,2,3,4,5,6,7}; cosine similarity Sim(z i ,avg) and Euclidean distance Dist(z i ,avg) represent the intrinsic relationship between the forest parameters of the i-th tree and the average value avg; τ is the attention weight scaling coefficient, which is updated by the pigeon flock optimization module; i = 1, 2, ..., n; The attention-based connection weights are continuously updated through back propagation, and the update method is shown in formula (13): Where: η ranges from 0 to 1, indicating learning efficiency; t is the current number of iterations; The initial value of the attention-based connection weight is randomly given and continuously updated through backpropagation. When the number of iterations reaches the maximum, the update is terminated.

7. The method for accurately predicting tree diameter at breast height volume based on optimized fuzzy deep network according to claim 6, characterized in that: The calculation process of the pigeon flock optimization module is: The pigeon flock optimization module is used to optimize the parameters k, α, and τ in the fuzzy deep network. The combination of k, α, and τ is called the model parameter combination δ[k, α, τ]. In the parameter search space, k is 0 to 100 and must be an integer, α is 0 to 30, and τ is 0 to 10. The pigeon flock optimization module is initially composed of L groups of model parameters δ l [k l ,α l ,τ l ](l=1,2,…L), that is, the pigeon group optimization module has L groups of model parameters to optimize and t max iterations; The lth group of model parameter combinations in the tth iteration is Its fitness is shown in formula (14): Where: t is the current iteration number of the pigeon group; l = 1, 2, ..., L; i = 1, 2, ..., n, n is the total number of training samples; is the DBH prediction value of the fuzzy depth network, is the volume prediction value of the fuzzy depth network, is the measured value of DBH, is the measured value of timber volume; The search strategy for model parameter combinations is divided into two stages according to the number of iterations. The initial stage is the first stage. When the number of iterations reaches the maximum number of iterations t max When the search reaches 80%, it enters the second stage. The specific search strategy is as follows: The calculation process of the first stage of the search strategy for model parameter combinations is shown in formula (15): Where: is the cosine iteration weight term; t max is the maximum number of iterations of the pigeon group; ε is a very small constant; rand(0,1) is a random number between [0,1]; array is the lth group of model parameter combinations for the tth iteration For arrays Speed; l=1,2,…L; When the fitness of the optimal parameter combination of L group model parameter combinations does not change for a long time, the L group model parameter combinations are sorted in descending order according to fitness, and the ones with higher fitness are ranked. The group parameters are subjected to population mutation, and the mutation method is shown in formula (16): Where: for The updated value of population mutation, Do rounding; if In the range of 0 to 100, otherwise If the value is within the range of 0 to 30, it will be updated; otherwise, it will not be updated. If the value is within the range of 0 to 10, it will be updated; otherwise, it will not be updated. rand(-1,1) is a random number in the range [-1,1]; The calculation process of the second stage of the model parameter combination search strategy is shown in formula (17): Where: is the tentative value of the lth group of model parameters at time t; if Then update the lth group of model parameters, that is, Otherwise, no update. After each iteration, some model parameter combinations with higher fitness values ​​are discarded and the number of model parameter groups L is updated. When the number of iterations reaches the maximum number of iterations t max Or when only one set of model parameter combinations is left, the iteration ends, and the parameter combination with the lowest fitness among the L sets of model parameter combinations is output, that is, the optimal values ​​of the given parameters k, α, and τ, and the model parameter combination is passed into the fuzzy deep network to complete the training.

Images